Make the proof suite say what it actually ran

The suite had fifteen checks that could not lose. The visual HashHop arena
painted a cyan line and a white ring on the answer slot, then asked the model
to find the answer by looking; the collector re-rolled puzzles until the
running rate crossed its target; the cache check passed if any layer had ever
recorded an entry; the repository check passed on a call graph built in memory
without touching the store; the research-module check asserted non-empty lists
of SHA256-derived placeholder vectors and so could not fail.

Each check now answers for one claim and reports a verdict the reader can act
on: passed, failed, deferred, or removed. A deferred row names what is missing
and the command that would settle it; a removed row says why the old claim was
withdrawn. The suite verdict is failed > deferred > passed, so a claim that
never ran holds the whole report back from saying "proven".

- evals.py (2427 lines) becomes the proofs/ package; largest module 445 lines.
- proofs/__main__.py deleted: it duplicated PRDProofSuite.latest() with a
  different glob. proofs/table.py renders; latest-proof-suite is the one face
  and exits 1 on a failed verdict.
- The arena paints nothing. The puzzle id is printed on its tile instead,
  because removing the ink without it removes the question.
- HashHop runs a fixed 8 puzzles per depth instead of re-rolling to target.
- The contradiction corpus moves from code to contradiction_scenarios.json;
  the check stays deferred while the corpus is the one this repo authored.
- The cache check compiles a nonce question over a real atom twice and reads
  the miss to hit movement; an empty store fails instead of agreeing by luck.
- The repository check requires rows in the store; the research-module check is
  removed.

Two product bugs surfaced on the way:
- text_index.fts_query joined Latin terms with a space, which FTS5 reads as
  AND, so a question longer than its answer retrieved nothing. Terms are ORed
  and bm25 rank orders the candidates.
- profiles.windows_path_to_wsl prefixed /mnt/ onto an already-POSIX path with
  an empty drive, producing /mnt//home/... and failing every launcher probe.

Verification: 168 passed (pytest backend/tests bench/tests); ruff clean on the
proofs package, cli.py, text_index.py, profiles.py, and both test files; mypy
clean on 16 files with a fresh cache; pyright reports 0 errors on proofs/ and
the test file. Ruff debt elsewhere is unchanged: 19 errors in seven files this
commit does not touch, measured against the same files at HEAD.

CHECKLIST.md

@@ -105,12 +105,25 @@
     - 根因修复：`_answered()` 里 `readable = owner in (as_holder, LIBRARY)` 是 SQL 可见性子句之外的第二份、更严的规则，与 `visible_to=[husband]` 冲突（共享记忆被错误隐藏）→ 删掉这份重复判断，SQL 的 `as_holder` 可见性子句是唯一权威。
     - 偏差：Go 侧 `limit`/`max_hops` 零值曾当真实数字发出，被 API 以 `greater_than` 拒 → 未设置的可选字段一律省略，由服务端定默认值。
   - 验收：`pytest tests/test_transactive.py` 绿；钥匙测试场景通过
-- [ ] **p0-6** 证明套件去自证
+- [x] **p0-6** 证明套件去自证
-  - [ ] p0-6.1 删 arena 答案入墨（`evals.py:1408-1409` 连接线、`evals.py:1417-1426` 白圈）
+  - [x] p0-6.1 删 arena 答案入墨（`evals.py:1408-1409` 连接线、`evals.py:1417-1426` 白圈）
-  - [ ] p0-6.2 删 best-of-N 重摇与达标即停
+    - 偏差：入墨是唯一的可读信号——`render_visual_hash` 只画噪声、不画字，答案槽被青色线+白圈标出后，模型不需要检索也能答。所以删墨的同时把 id 印上瓦片（`proofs/arena.py:render_visual_hash` 白底黑字、每行 8 字符、字号 `tile_size//8`），否则删墨 = 删题。
-  - [ ] p0-6.3 删自造矛盾场景（出题者≠判分者）
+  - [x] p0-6.2 删 best-of-N 重摇与达标即停
-  - [ ] p0-6.4 删缓存/证据图自触发式检查
+    - `collect_model_visual_proofs(evaluator, hops, token_equivalent, attempts)` 固定跑 `HASHHOP_ATTEMPTS = 8` 题；旧代码 `while len(proofs) < max_samples` 里一达标就 break（`evals.py:2365`），等于弱模型反复重摇到及格。
-  - [ ] p0-6.5 模型相关检查标 `DEFERRED`，不进 pass
+  - [x] p0-6.3 删自造矛盾场景（出题者≠判分者）
+    - 偏差：自造场景没有删掉（删了就没得测），而是从代码搬进**数据**：`proofs/contradiction_scenarios.json`（12 场景，sha256 `5f93aba09cbfc32f…`），`corpus_path(settings.data_dir)` 优先读运维者放的第三方集，`ely-eye contradiction-corpus` 写默认集。检查项**永久 DEFERRED**，直到磁盘上不是自带的那份；reason 里写明「生产者用自家例题判分不算证据」。
+  - [x] p0-6.4 删缓存/证据图自触发式检查
+    - 缓存：旧检查只要 `any(layer.entries > 0)` 就算过（`evals.py:1252-1265`）。现在拿 store 里真实的一条 atom 拼一个带 nonce 的问题，编译两次，读 `cache_events` 的 miss→hit 计数；空 store 直接 failed（空上下文会让两次编译「一致」得毫无意义）。顺带修掉真实双计数：`read_semantic_cache` 与 `compiler.compile` 各记一次事件，现在读是纯读、编译恰好记一次（缓存命中带 hits 才记 hit，空缓存记 miss 并附 reason）。
+    - 仓库覆盖：旧检查 `indexed_file_count > 0` + 有调用图 + 有 git 历史就算过，不碰 store。现在必须 `sources > 0 && atoms > 0`，否则 failed 并说明「仓库上下文建了但没进 store」。
+    - 研究模块：旧检查断言八个模块各返回非空列表，而那些向量是 SHA256 派生的假向量——**不可能失败**。改为 `status="removed"`，artifact 记下为什么撤下、重新加回的条件。
+  - [x] p0-6.5 模型相关检查标 `DEFERRED`，不进 pass
+    - `ProofCheck.status` 加 `deferred`/`removed` + `reason` + `rerun`；`_deferred_without_model()` 在 `model_dir` 缺失时产出 deferred 并点名缺什么。聚合改成 `failed > deferred > passed`：只要还有没跑成的主张，整个套件就不说「通过」。每条 failed/deferred 都带 why 与可执行的 rerun 命令（`proofs/table.py` 渲染）。
+  - [x] p0-6.6 顺手还清的文件尺寸债
+    - `evals.py` 2427 行 → `proofs/` 包 15 个模块，最大 445 行（`checks_cartridge.py`），全部 ≤500 行；`suite.py` 218 行只剩「跑哪些检查、怎么定论」。`proofs/__main__.py` 删掉（它的 `latest_report()` 与 `PRDProofSuite.latest()` 是两份 glob 不同的重复实现），渲染移到 `proofs/table.py`，`latest-proof-suite` 成为唯一入口并在 failed 时 `exit 1`。
+  - 顺带修掉两个被证明测试逼出来的真 bug：
+    1. `text_index.fts_query` 把拉丁词用空格连起来 = FTS5 隐式 AND，长问题永远查不到短答案（实测「Which Ely-Eye PRD requirements prove long context memory, HashHop, and adapters?」→ 0 命中）。改成 OR + bm25 rank 排序。
+    2. `profiles.windows_path_to_wsl()` 对本来就是 POSIX 的路径也拼 `/mnt/` + 空 drive，产出 `/mnt//home/...`，六个 runtime profile 的探针全部 `probe_exit_code: 1`。
+  - 验收：`python -m pytest backend/tests/test_proofs.py -q` → 12 passed；`ELY_EYE_HOME=/tmp/p06home ely-eye proof-suite --cartridge-id p06_20cf30f7804d && ely-eye latest-proof-suite` → 15 行判决（5 failed / 3 passed / 6 deferred / 1 removed），每条带 why 与 rerun。
   - 验收：`python -m ely_eye.proofs --report` 逐条 PASS/DEFERRED/REMOVED + 理由
 
 ---

backend/src/ely_eye/api.py

@@ -24,7 +24,6 @@ from .compiler import ContextCompiler
 from .config import Settings, get_settings
 from .db import Database
 from .embeddings import EmbeddingService
-from .evals import HashHopEvaluator, PRDProofSuite
 from .evaluation_schemas import (
     BenchmarkPredictionReport,
     BenchmarkRegistryReport,
@@ -38,6 +37,7 @@ from .evidence_graph import EvidenceGraphService
 from .ingestion import IngestionService
 from .memory import MemoryService
 from .profiles import runtime_profile_report
+from .proofs import HashHopEvaluator, PRDProofSuite
 from .research_modules import ResearchModuleService
 from .retrieval import RetrievalService
 from .runtime import QwenRuntime, RuntimeUnavailableError

backend/src/ely_eye/cache_fabric.py

@@ -9,6 +9,8 @@ from .config import Settings, get_settings
 from .db import Database
 from .schemas import CacheLayerStatus
 
+SEMANTIC_LAYER = "semantic"
+
 
 class CacheFabric:
     def __init__(self, settings: Settings | None = None, database: Database | None = None) -> None:
@@ -16,7 +18,7 @@ class CacheFabric:
         self.db = database or Database(self.settings)
 
     def semantic_key(self, question: str, profile: str) -> str:
-        return hashlib.sha256(f"compiler-v2\n{profile}\n{question}".encode("utf-8")).hexdigest()
+        return hashlib.sha256(f"compiler-v2\n{profile}\n{question}".encode()).hexdigest()
 
     def record_hit(self, layer: str, cache_key: str, bytes_estimate: int = 0, metadata: dict[str, Any] | None = None) -> None:
         self.db.log_cache_event(layer, cache_key, "hit", bytes_estimate, metadata)
@@ -29,15 +31,17 @@ class CacheFabric:
         path.parent.mkdir(parents=True, exist_ok=True)
         data = json.dumps(payload, ensure_ascii=False, indent=2).encode("utf-8")
         path.write_bytes(data)
-        self.db.log_cache_event("semantic", cache_key, "write", len(data), {"path": str(path)})
+        self.db.log_cache_event(SEMANTIC_LAYER, cache_key, "write", len(data), {"path": str(path)})
 
     def read_semantic_cache(self, cache_key: str) -> dict[str, Any] | None:
+        """Return the cached payload, or None. The caller traces the event: one compile, one event."""
         path = self.settings.cache_dir / "semantic" / f"{cache_key}.json"
         if not path.exists():
-            self.record_miss("semantic", cache_key)
             return None
-        self.record_hit("semantic", cache_key, path.stat().st_size)
+        try:
-        return json.loads(path.read_text(encoding="utf-8"))
+            return json.loads(path.read_text(encoding="utf-8"))
+        except (OSError, json.JSONDecodeError) as error:
+            raise ValueError(f"{path}: the semantic cache entry is unreadable ({error})") from error
 
     def statuses(self) -> list[CacheLayerStatus]:
         rows = self.db.cache_layer_rows()

backend/src/ely_eye/cli.py

@@ -20,10 +20,12 @@ from .cli_transactive import register as register_transactive_commands
 from .config import get_settings
 from .db import Database
 from .embeddings import EmbeddingService
-from .evals import HashHopEvaluator, PRDProofSuite
 from .ingestion import IngestionService
 from .memory import MemoryService
 from .profiles import runtime_profile_report
+from .proofs import HashHopEvaluator, PRDProofSuite
+from .proofs.contradictions import CORPUS_NAME, build_contradiction_scenarios, write_corpus
+from .proofs.table import render_table
 from .research_modules import ResearchModuleService
 from .retrieval import RetrievalService
 from .runtime import QwenRuntime
@@ -232,11 +234,19 @@ def proof_suite(cartridge_id: str | None = None) -> None:
 
 
 @app.command("latest-proof-suite")
-def latest_proof_suite() -> None:
+def latest_proof_suite(json_output: bool = typer.Option(False, "--json")) -> None:
+    """Print what the last proof suite claimed and what it actually proved."""
     report = PRDProofSuite(get_settings()).latest()
     if report is None:
+        raise typer.BadParameter(
+            "no proof suite has been recorded yet; run: ely-eye proof-suite --cartridge-id <id>"
+        )
+    if json_output:
+        typer.echo(report.model_dump_json(indent=2))
+    else:
+        typer.echo(render_table(report))
+    if report.status == "failed":
         raise typer.Exit(code=1)
-    typer.echo(report.model_dump_json(indent=2))
 
 
 @app.command("runtime-generation-proof")
@@ -251,6 +261,14 @@ def visual_contradiction_proof() -> None:
     typer.echo(check.model_dump_json(indent=2))
 
 
+@app.command("contradiction-corpus")
+def contradiction_corpus(data_dir: Path | None = None) -> None:
+    """Write the corpus the contradiction lens is graded on, so it can be read and replaced."""
+    settings = get_settings()
+    target = write_corpus((data_dir or settings.data_dir) / CORPUS_NAME)
+    typer.echo(json_dump({"corpus": str(target), "scenarios": len(build_contradiction_scenarios())}))
+
+
 @app.command("finalize-cartridge")
 def finalize_cartridge(cartridge_id: str) -> None:
     report = CartridgeAssetService(get_settings()).finalize(cartridge_id)

backend/src/ely_eye/compiler.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 import hashlib
 from collections.abc import Iterable
 
-from .cache_fabric import CacheFabric
+from .cache_fabric import SEMANTIC_LAYER, CacheFabric
 from .config import Settings, get_settings
 from .db import Database
 from .retrieval import RetrievalService
@@ -27,11 +27,18 @@ class ContextCompiler:
     def compile(self, question: str, profile: RuntimeProfile) -> CompiledContext:
         cache_key = self.cache.semantic_key(question, profile.value)
         cached = self.cache.read_semantic_cache(cache_key)
-        if cached:
+        if cached is not None:
             compiled = CompiledContext.model_validate(cached)
             if compiled.hits:
+                self.cache.record_hit(SEMANTIC_LAYER, cache_key)
                 return compiled
-
+            # A cached context with no evidence in it is not an answer, so counting it as a hit
+            # would let an empty cache entry look like a warm cache.
+            self.cache.record_miss(
+                SEMANTIC_LAYER, cache_key, metadata={"reason": "cached context has no hits"}
+            )
+        else:
+            self.cache.record_miss(SEMANTIC_LAYER, cache_key)
         plan = self.plan(question, profile)
         hits = self.retrieval.search(plan.retrieval_query, plan.evidence_budget_atoms).hits
         packed_text, token_total = self._pack_hits(hits, plan.token_budget, plan.retrieval_query)

backend/src/ely_eye/evaluation_schemas.py

@@ -105,16 +105,26 @@ class HashHopProof(BaseModel):
     created_at: datetime = Field(default_factory=utc_now)
 
 class ProofCheck(BaseModel):
+    """One claim, one verdict.
+
+    `deferred` means the check is written and runnable but needs something this machine does not
+    have (model weights, a second source). `removed` means the old check proved nothing and the
+    claim now waits for a check that can. Both carry a reason and the command that runs the check.
+    """
+
     name: str
     requirement: str
-    status: Literal["passed", "failed"]
+    status: Literal["passed", "failed", "deferred", "removed"]
     evidence: list[str]
     detail: str
+    reason: str = ""
+    rerun: str = ""
+
 
 class ProofSuiteReport(BaseModel):
     proof_id: str
     cartridge_id: str
-    status: Literal["passed", "failed"]
+    status: Literal["passed", "failed", "deferred"]
     created_at: datetime = Field(default_factory=utc_now)
     checks: list[ProofCheck]
     artifacts: dict[str, str]

backend/src/ely_eye/profiles.py

@@ -410,10 +410,12 @@ def resolve_llama_server() -> str | None:
 
 
 def windows_path_to_wsl(path: Path) -> str:
+    """Map a Windows path to the WSL mount a WSL shell can cd into; a POSIX path is already correct."""
     resolved = path.resolve()
     drive = resolved.drive.rstrip(":").lower()
-    parts = resolved.parts[1:]
+    if not drive:
-    return "/mnt/" + drive + "/" + "/".join(parts)
+        return str(resolved)
+    return "/mnt/" + drive + "/" + "/".join(resolved.parts[1:])
 
 
 def wsl_quote(value: str) -> str:

backend/src/ely_eye/proofs/__init__.py

@@ -0,0 +1,6 @@
+"""Proof machinery: one module per kind of claim, one verdict per check."""
+
+from .hashhop import HashHopEvaluator
+from .suite import PRDProofSuite
+
+__all__ = ["HashHopEvaluator", "PRDProofSuite"]

backend/src/ely_eye/proofs/arena.py

@@ -0,0 +1,224 @@
+"""Draw the HashHop puzzles and read the model's answer back out of them."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import random
+import re
+from pathlib import Path
+from typing import Any, Literal
+
+from PIL import Image, ImageDraw, ImageFont
+
+# The four arena positions a model may answer with. Naming them once keeps the puzzle, the parser,
+# and the scoring keyed on the same values instead of four spellings in three files.
+SlotName = Literal["top-left", "top-right", "bottom-left", "bottom-right"]
+SLOTS: tuple[SlotName, ...] = ("top-left", "top-right", "bottom-left", "bottom-right")
+
+ID_LINE_CHARS = 8
+CANVAS_BACKGROUND = (8, 10, 14)
+
+
+def render_visual_hash(item_id: str, path: Path, tile_size: int) -> None:
+    """Print the id on the tile, over a pattern derived from the same id.
+
+    A pattern alone is not a puzzle: nothing maps a random texture back to a hex id, so the
+    only way to answer was the ink the renderer painted on the answer. The id is now legible,
+    which makes the task look, match, and answer.
+    """
+    seed = int(hashlib.sha256(item_id.encode("utf-8")).hexdigest()[:16], 16)
+    rng = random.Random(seed)
+    image = Image.new("RGB", (tile_size, tile_size), color=_noise_color(rng))
+    draw = ImageDraw.Draw(image)
+    for _ in range(96):
+        x0 = rng.randrange(tile_size)
+        y0 = rng.randrange(tile_size)
+        x1 = min(tile_size, x0 + rng.randrange(4, tile_size // 3))
+        y1 = min(tile_size, y0 + rng.randrange(4, tile_size // 3))
+        color = _noise_color(rng)
+        if rng.random() < 0.5:
+            draw.rectangle([x0, y0, x1, y1], fill=color)
+        else:
+            draw.ellipse([x0, y0, x1, y1], fill=color)
+    _draw_id(draw, item_id, tile_size)
+    image.save(path)
+
+
+def _noise_color(rng: random.Random) -> tuple[int, int, int]:
+    return (rng.randrange(256), rng.randrange(256), rng.randrange(256))
+
+
+def _draw_id(draw: ImageDraw.ImageDraw, item_id: str, tile_size: int) -> None:
+    font = ImageFont.load_default(size=max(12, tile_size // 8))
+    plate = (
+        int(tile_size * 0.05),
+        int(tile_size * 0.16),
+        int(tile_size * 0.95),
+        int(tile_size * 0.88),
+    )
+    draw.rectangle(plate, fill=(255, 255, 255))
+    lines = [item_id[start : start + ID_LINE_CHARS] for start in range(0, len(item_id), ID_LINE_CHARS)]
+    line_height = font.getbbox("AbgJ")[3]
+    block_top = (plate[1] + plate[3]) / 2 - (len(lines) * line_height) / 2
+    for index, line in enumerate(lines):
+        box = font.getbbox(line)
+        left = (plate[0] + plate[2]) / 2 - (box[2] - box[0]) / 2
+        draw.text((left, block_top + index * line_height), line, fill=(12, 14, 18), font=font)
+
+
+def render_visual_hash_arena(
+    source_path: Path, slot_paths: dict[str, Path], output_path: Path
+) -> None:
+    """Place the source and the candidates. Nothing marks which candidate is the answer."""
+    tile = 160
+    canvas = Image.new("RGB", (920, 560), color=CANVAS_BACKGROUND)
+    draw = ImageDraw.Draw(canvas)
+    source_box = (72, 200, 72 + tile, 200 + tile)
+    slot_boxes = _slot_boxes(520, 704, 76, 324, tile)
+    paste_visual_hash_tile(canvas, source_path, source_box)
+    draw.rectangle(source_box, outline=(255, 78, 220), width=8)
+    for slot, box in slot_boxes.items():
+        paste_visual_hash_tile(canvas, slot_paths[slot], box)
+        draw.rectangle(box, outline=(180, 190, 205), width=5)
+    _save(canvas, output_path)
+
+
+def render_visual_hash_two_hop_arena(
+    source_path: Path,
+    intermediate_slot_paths: dict[str, Path],
+    target_slot_paths: dict[str, Path],
+    output_path: Path,
+) -> None:
+    """Two grids of candidates, no arrow and no ring: the chain has to be read, not seen."""
+    tile = 136
+    canvas = Image.new("RGB", (1180, 720), color=CANVAS_BACKGROUND)
+    draw = ImageDraw.Draw(canvas)
+    source_box = (64, 292, 64 + tile, 292 + tile)
+    intermediate_boxes = _slot_boxes(424, 584, 104, 480, tile)
+    target_boxes = _slot_boxes(864, 1024, 104, 480, tile)
+    paste_visual_hash_tile(canvas, source_path, source_box)
+    draw.rectangle(source_box, outline=(255, 78, 220), width=7)
+    for grid, boxes, outline in (
+        (intermediate_slot_paths, intermediate_boxes, (140, 215, 255)),
+        (target_slot_paths, target_boxes, (255, 196, 92)),
+    ):
+        for slot, box in boxes.items():
+            paste_visual_hash_tile(canvas, grid[slot], box)
+            draw.rectangle(box, outline=outline, width=5)
+    _save(canvas, output_path)
+
+
+def _slot_boxes(
+    left_x: int, right_x: int, top_y: int, bottom_y: int, tile: int
+) -> dict[str, tuple[int, int, int, int]]:
+    return {
+        "top-left": (left_x, top_y, left_x + tile, top_y + tile),
+        "top-right": (right_x, top_y, right_x + tile, top_y + tile),
+        "bottom-left": (left_x, bottom_y, left_x + tile, bottom_y + tile),
+        "bottom-right": (right_x, bottom_y, right_x + tile, bottom_y + tile),
+    }
+
+
+def _save(canvas: Image.Image, output_path: Path) -> None:
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    canvas.save(output_path)
+
+
+def paste_visual_hash_tile(canvas: Image.Image, path: Path, box: tuple[int, int, int, int]) -> None:
+    tile = Image.open(path).convert("RGB").resize((box[2] - box[0], box[3] - box[1]), Image.Resampling.LANCZOS)
+    canvas.paste(tile, box[:2])
+
+
+def extract_visual_target_slot(text: str) -> SlotName | None:
+    stripped = text.strip()
+    payload: dict[str, Any] | None
+    try:
+        parsed = json.loads(stripped)
+        payload = parsed if isinstance(parsed, dict) else None
+    except json.JSONDecodeError:
+        payload = extract_json_object(stripped)
+    if isinstance(payload, dict):
+        for key in ("target_slot", "final_slot", "slot", "answer"):
+            value = payload.get(key)
+            if isinstance(value, str):
+                slot = normalize_visual_slot(value)
+                if slot:
+                    return slot
+    return normalize_visual_slot(stripped)
+
+
+def extract_hashhop_target(text: str) -> str | None:
+    """Read the id the model answered with, from JSON or from loose prose."""
+    stripped = text.strip()
+    try:
+        payload = json.loads(stripped)
+    except json.JSONDecodeError:
+        payload = extract_json_object(stripped)
+    if isinstance(payload, dict):
+        for key in ("answer", "target", "target_id", "final_target_id"):
+            value = payload.get(key)
+            if isinstance(value, str):
+                candidate = normalize_hash_id(value)
+                if candidate:
+                    return candidate
+    match = re.search(r"\b[a-fA-F0-9]{32}\b", stripped)
+    return match.group(0).lower() if match else None
+
+
+def normalize_hash_id(value: str) -> str | None:
+    match = re.search(r"\b[a-fA-F0-9]{32}\b", value)
+    return match.group(0).lower() if match else None
+
+
+def extract_visual_intermediate_slot(text: str) -> SlotName | None:
+    stripped = text.strip()
+    try:
+        parsed = json.loads(stripped)
+        payload = parsed if isinstance(parsed, dict) else None
+    except json.JSONDecodeError:
+        payload = extract_json_object(stripped)
+    if isinstance(payload, dict):
+        for key in ("intermediate_slot", "middle_slot", "first_hop_slot"):
+            value = payload.get(key)
+            if isinstance(value, str):
+                slot = normalize_visual_slot(value)
+                if slot:
+                    return slot
+    return None
+
+
+def extract_json_object(text: str) -> dict[str, Any] | None:
+    start = text.find("{")
+    end = text.rfind("}")
+    if start < 0 or end <= start:
+        return None
+    try:
+        payload = json.loads(text[start : end + 1])
+    except json.JSONDecodeError:
+        return None
+    return payload if isinstance(payload, dict) else None
+
+
+def normalize_visual_slot(value: str) -> SlotName | None:
+    normalized = re.sub(r"[^a-z]+", "-", value.lower()).strip("-")
+    aliases: dict[str, SlotName] = {
+        "top-left": "top-left",
+        "upper-left": "top-left",
+        "left-top": "top-left",
+        "top-right": "top-right",
+        "upper-right": "top-right",
+        "right-top": "top-right",
+        "bottom-left": "bottom-left",
+        "lower-left": "bottom-left",
+        "left-bottom": "bottom-left",
+        "bottom-right": "bottom-right",
+        "lower-right": "bottom-right",
+        "right-bottom": "bottom-right",
+    }
+    if normalized in aliases:
+        return aliases[normalized]
+    for alias, slot in aliases.items():
+        if re.search(rf"\b{re.escape(alias)}\b", normalized):
+            return slot
+    return None

backend/src/ely_eye/proofs/checks_cartridge.py

@@ -0,0 +1,404 @@
+"""Cartridge assets, the memory capsule, replay, DNA, cache trace, and repository coverage.
+
+Moved out of suite.py so the suite order stays readable; each method is a check the
+suite runs, and the suite class mixes these in.
+"""
+
+from __future__ import annotations
+
+import json
+import secrets
+import shutil
+from pathlib import Path
+from typing import Any
+
+from ..cache_fabric import CacheFabric
+from ..cartridge import CartridgeService
+from ..cartridge_assets import CartridgeAssetService, merkle_root
+from ..compiler import ContextCompiler
+from ..config import Settings
+from ..db import Database
+from ..evaluation_schemas import (
+    ProofCheck,
+)
+from ..ingestion import SUPPORTED_CODE_EXTENSIONS, build_repository_context, find_git_root
+from ..schemas import (
+    CartridgeManifest,
+    RuntimeProfile,
+)
+
+# A fixed puzzle count per hop depth. The old suite re-rolled until the running rate crossed its
+from .checks_common import (
+    _semantic_counters,
+)
+from .report import (
+    compiled_context_digest,
+    parquet_row_count,
+    read_memory_capsule,
+    validated_segment_hash,
+    write_json,
+)
+from .repository_checks import discover_repository_files, project_root
+
+
+class CartridgeChecks:
+    """Cartridge assets, the memory capsule, replay, DNA, cache trace, and repository coverage."""
+
+    settings: Settings
+    db: Database
+    cartridges: CartridgeService
+
+    def _cartridge_assets_check(
+        self,
+        cartridge_id: str,
+        cartridge_root: Path,
+        proof_root: Path,
+    ) -> tuple[ProofCheck, CartridgeManifest]:
+        report = CartridgeAssetService(self.settings, self.db).finalize(cartridge_id)
+        source_report = cartridge_root / "cartridge_assets.json"
+        proof_report = proof_root / "cartridge_assets.json"
+        shutil.copy2(source_report, proof_report)
+        rows = self.db.list_cartridges()
+        manifest_json = next(row["manifest_json"] for row in rows if row["cartridge_id"] == cartridge_id)
+        manifest = CartridgeManifest.model_validate_json(manifest_json)
+        expected = {
+            "text_vectors",
+            "visual_vectors",
+            "kv_snapshots",
+            "memory_capsule",
+            "memory_capsule_index",
+            "ttt_vl_adapter",
+            "visual_mtp_head",
+            "cartridge_assets",
+        }
+        by_name = {asset.name: asset for asset in report.assets}
+        visual_vectors = by_name.get("visual_vectors")
+        memory_capsule = by_name.get("memory_capsule")
+        gaps: list[str] = []
+        if report.status != "ready":
+            gaps.append(f"the asset report says {report.status}")
+        absent = sorted(expected - set(manifest.artifacts))
+        if absent:
+            gaps.append("not registered in the manifest: " + ", ".join(absent))
+        empty = [asset.name for asset in report.assets if asset.size_bytes <= 0]
+        if empty:
+            gaps.append("zero-byte assets: " + ", ".join(empty))
+        if visual_vectors is None or visual_vectors.item_count <= 0:
+            gaps.append("visual_vectors holds no vectors")
+        if memory_capsule is None or memory_capsule.item_count <= 0:
+            gaps.append("memory_capsule holds no segments")
+        valid = not gaps
+        return (
+            ProofCheck(
+                name="Cartridge Physical Assets",
+                requirement="PRD 8.3 requires text vectors, non-empty visual vectors, KV snapshots, a 100M memory capsule, TTT-VL adapter weights, Visual-MTP head weights, and eval proofs inside the Context Cartridge.",
+                status="passed" if valid else "failed",
+                evidence=[str(proof_report), str(cartridge_root)],
+                detail="materialized assets: " + ", ".join(asset.name for asset in report.assets),
+                reason="; ".join(gaps),
+                rerun="ely-eye finalize-cartridge <id>, then ely-eye proof-suite --cartridge-id <id>",
+            ),
+            manifest,
+        )
+
+    def _memory_capsule_check(self, cartridge_root: Path, proof_root: Path) -> ProofCheck:
+        capsule_path = cartridge_root / "memory_capsule.json.zst"
+        index_path = cartridge_root / "memory_capsule_index.json"
+        artifact = proof_root / "memory_capsule.json"
+        payload: dict[str, Any] = {
+            "capsule_path": str(capsule_path),
+            "index_path": str(index_path),
+            "status": "failed",
+        }
+        valid = False
+        try:
+            capsule = read_memory_capsule(capsule_path)
+            index = json.loads(index_path.read_text(encoding="utf-8")) if index_path.exists() else {}
+            raw_segments = capsule.get("segments")
+            if not isinstance(raw_segments, list):
+                raise ValueError(f"{capsule_path}: the memory capsule carries no segments list")  # noqa: TRY004  # malformed stored data, not a caller type error
+            segments: list[dict[str, Any]] = [segment for segment in raw_segments if isinstance(segment, dict)]
+            if len(segments) != len(raw_segments):
+                raise ValueError(f"{capsule_path}: a memory capsule segment is not an object")
+            segment_hashes = [validated_segment_hash(segment) for segment in segments]
+            token_total = sum(int(segment.get("token_count") or 0) for segment in segments)
+            computed_merkle = merkle_root(segment_hashes)
+            valid = (
+                capsule.get("format") == "ely-eye-memory-capsule-v1"
+                and int(capsule.get("token_equivalent") or 0) >= self.settings.library_target_tokens
+                and int(capsule.get("segment_count") or 0) == len(segments)
+                and token_total == int(capsule.get("token_equivalent") or -1)
+                and computed_merkle == capsule.get("segment_merkle_root")
+                and index.get("segment_merkle_root") == capsule.get("segment_merkle_root")
+                and int(index.get("token_equivalent") or 0) == int(capsule.get("token_equivalent") or -1)
+                and int(capsule.get("source_atom_count") or 0) > 0
+                and all(segment_hashes)
+            )
+            payload.update(
+                {
+                    "status": "passed" if valid else "failed",
+                    "token_equivalent": capsule.get("token_equivalent"),
+                    "segment_count": capsule.get("segment_count"),
+                    "source_atom_count": capsule.get("source_atom_count"),
+                    "source_token_equivalent": capsule.get("source_token_equivalent"),
+                    "segment_merkle_root": capsule.get("segment_merkle_root"),
+                    "computed_merkle_root": computed_merkle,
+                    "first_segment": segments[0] if segments else None,
+                    "last_segment": segments[-1] if segments else None,
+                }
+            )
+        except Exception as exc:  # noqa: BLE001  # a probe records its failure instead of crashing the proof run
+            payload["error"] = str(exc)
+        write_json(artifact, payload)
+        return ProofCheck(
+            name="100M Memory Capsule",
+            requirement="PRD 2.2, 7.3, 8.3, and 14.3 require a 10M-100M token-equivalent Context Cartridge memory capsule with verifiable retrieval commitments.",
+            status="passed" if valid else "failed",
+            evidence=[str(artifact), str(capsule_path), str(index_path)],
+            detail=(
+                f"{payload.get('token_equivalent')} token-equivalent memory capsule with "
+                f"{payload.get('segment_count')} committed segments"
+            ),
+        )
+
+    def _cartridge_replay_check(
+        self,
+        cartridge_id: str,
+        cartridge_root: Path,
+        manifest: CartridgeManifest,
+        proof_root: Path,
+    ) -> ProofCheck:
+        atoms_path = cartridge_root / manifest.artifacts["atoms"]
+        sources_path = cartridge_root / manifest.artifacts["sources"]
+        sparse_path = cartridge_root / manifest.artifacts["sparse_index"]
+        graph_path = cartridge_root / manifest.artifacts["temporal_graph"]
+        computed_dna = self.cartridges.compute_dna(cartridge_root)
+        atom_rows = parquet_row_count(atoms_path)
+        try:
+            sources = json.loads(sources_path.read_text(encoding="utf-8")) if sources_path.exists() else []
+        except (OSError, json.JSONDecodeError) as error:
+            raise ValueError(f"{sources_path}: the cartridge source list is unreadable ({error})") from error
+        question = "Which Ely-Eye PRD requirements prove long context memory, HashHop, and adapters?"
+        compiler = ContextCompiler(self.settings, self.db)
+        artifact = proof_root / "cartridge_replay.json"
+        # The replay must come out of the cartridge, not out of whatever the live store happens
+        # to hold: mounting is what makes the cartridge the retrieval scope (retrieval scopes by
+        # the mounted set), so an unmounted replay would prove only that compiling is stable.
+        try:
+            self.cartridges.mount(cartridge_id)
+        except ValueError as error:
+            # The KV fingerprint gate refuses to mount a cartridge whose model snapshot this
+            # machine cannot read. A missing input is not a failed claim, so the check reports
+            # the refusal instead of crashing the suite.
+            write_json(artifact, {"verdict": "deferred", "why": str(error)})
+            return ProofCheck(
+                name="Context Cartridge Replay",
+                requirement="PRD 14.2 requires the same cartridge to load reproducibly and produce consistent replay evidence.",
+                status="deferred",
+                evidence=[str(artifact), str(cartridge_root / "manifest.json")],
+                detail=str(error),
+                reason="the cartridge records a model snapshot this machine cannot verify, so the replay cannot be attempted",
+                rerun="set ELY_EYE_MODEL_DIR to the snapshot the cartridge was built against, then ely-eye proof-suite --cartridge-id <id>",
+            )
+        try:
+            replay_a = compiler.compile(question, RuntimeProfile.library_100m)
+            replay_b = compiler.compile(question, RuntimeProfile.library_100m)
+        finally:
+            self.cartridges.unmount(cartridge_id)
+        digest_a = compiled_context_digest(replay_a)
+        digest_b = compiled_context_digest(replay_b)
+        payload = {
+            "cartridge_id": cartridge_id,
+            "manifest_atom_count": manifest.atom_count,
+            "parquet_atom_count": atom_rows,
+            "manifest_source_count": manifest.source_count,
+            "source_count": len(sources),
+            "dna_manifest": manifest.dna,
+            "dna_computed": computed_dna,
+            "artifacts": {
+                "atoms": str(atoms_path),
+                "sources": str(sources_path),
+                "sparse_index": str(sparse_path),
+                "temporal_graph": str(graph_path),
+            },
+            "replay": {
+                "question": question,
+                "digest_a": digest_a,
+                "digest_b": digest_b,
+                "hit_ids_a": [hit.atom.atom_id for hit in replay_a.hits],
+                "hit_ids_b": [hit.atom.atom_id for hit in replay_b.hits],
+                "cache_trace_a": replay_a.cache_trace_id,
+                "cache_trace_b": replay_b.cache_trace_id,
+            },
+        }
+        artifact = proof_root / "cartridge_replay.json"
+        write_json(artifact, payload)
+        cartridge_sources = {str(entry.get("source_id", "")) for entry in sources}
+        answered_from_cartridge = bool(replay_a.hits) and all(
+            hit.atom.source_id in cartridge_sources for hit in replay_a.hits
+        )
+        valid = (
+            atoms_path.exists()
+            and sources_path.exists()
+            and sparse_path.exists()
+            and graph_path.exists()
+            and atom_rows == manifest.atom_count
+            and len(sources) == manifest.source_count
+            and computed_dna == manifest.dna
+            and answered_from_cartridge
+            and digest_a == digest_b
+        )
+        return ProofCheck(
+            name="Context Cartridge Replay",
+            requirement="PRD 14.2 requires the same cartridge to load reproducibly and produce consistent replay evidence.",
+            status="passed" if valid else "failed",
+            evidence=[str(artifact), str(cartridge_root / "manifest.json")],
+            detail=(
+                f"{atom_rows} atoms, {len(sources)} sources, replay digest {digest_a}, "
+                f"{len(replay_a.hits)} hits answered from the mounted cartridge"
+            ),
+            rerun="ely-eye proof-suite --cartridge-id <id>",
+        )
+
+    def _repository_input_check(self, proof_root: Path) -> ProofCheck:
+        root = project_root()
+        files = discover_repository_files(root)
+        code_files = [path for path in files if path.suffix.lower() in SUPPORTED_CODE_EXTENSIONS]
+        context = build_repository_context(root, files, code_files, find_git_root(root))
+        artifact = proof_root / "repository_input.json"
+        write_json(artifact, context)
+        call_graph = context["code"]["call_graph"]
+        git_context = context["git"]
+        issue_pr_history = context["issue_pr_history"]
+        # Building a context object from files on disk is not evidence the repository reached the
+        # store: the claim is that the repository is queryable, so the store has to hold it.
+        stored_sources = [source for source in self.db.list_sources() if source.path.startswith(str(root))]
+        stored_atoms = self.db.atom_count()
+        artifact = proof_root / "repository_input.json"
+        payload = dict(context)
+        payload["store_evidence"] = {
+            "sources_under_project_root": len(stored_sources),
+            "source_ids": [source.source_id for source in stored_sources[:20]],
+            "atoms_in_store": stored_atoms,
+        }
+        write_json(artifact, payload)
+        valid = (
+            context["structure"]["indexed_file_count"] > 0
+            and context["code"]["indexed_file_count"] > 0
+            and bool(call_graph)
+            and bool(git_context.get("available"))
+            and bool(git_context.get("recent_commits"))
+            and issue_pr_history.get("source") == "local_git_commit_history"
+            and bool(stored_sources)
+            and stored_atoms > 0
+        )
+        return ProofCheck(
+            name="Repository Input Coverage",
+            requirement="PRD 10.1 requires code repository structure indexing, call graph extraction, and issue/PR history association.",
+            status="passed" if valid else "failed",
+            evidence=[str(artifact)],
+            detail=(
+                f"{context['structure']['indexed_file_count']} files indexed; "
+                f"{context['code']['indexed_file_count']} code files; "
+                f"{len(call_graph)} call edges; "
+                f"{len(git_context.get('recent_commits') or [])} git commits; "
+                f"{len(stored_sources)} sources and {stored_atoms} atoms in the store"
+            ),
+            reason="" if valid else "the repository context was built but the store holds no atoms for it",
+            rerun="ely-eye ingest . && ely-eye proof-suite --cartridge-id <id>",
+        )
+
+    def _memory_dna_check(
+        self,
+        cartridge_root: Path,
+        manifest: CartridgeManifest,
+        proof_root: Path,
+    ) -> ProofCheck:
+        computed = self.cartridges.compute_dna(cartridge_root)
+        payload = {
+            "manifest_dna": manifest.dna,
+            "computed_dna": computed,
+            "formula": "sha256(sorted cartridge artifacts excluding manifest.json)",
+            "artifact_count": sum(1 for path in cartridge_root.rglob("*") if path.is_file()),
+        }
+        artifact = proof_root / "memory_dna.json"
+        write_json(artifact, payload)
+        return ProofCheck(
+            name="Memory DNA",
+            requirement="PRD 11.14 requires a reproducible Memory DNA over adapter delta, evidence graph, retrieval index, and eval proofs.",
+            status="passed" if manifest.dna == computed else "failed",
+            evidence=[str(artifact), str(cartridge_root)],
+            detail=f"manifest DNA {manifest.dna}",
+        )
+
+    def _cache_trace_check(self, proof_root: Path) -> ProofCheck:
+        """Ask the cache the same question twice and require the second answer to be a hit.
+
+        The old check passed when any layer had any row, which a single `forget` could satisfy.
+        The claim in PRD 14.2 is that hit, miss, write and replay state are visible, so the check
+        compiles a question nobody has asked, compiles it again, and requires the semantic layer's
+        counters to move by exactly one miss then one hit.
+        """
+        compiler = ContextCompiler(self.settings, self.db)
+        fabric = CacheFabric(self.settings, self.db)
+        probe = self.db.list_atoms(limit=1)
+        if not probe:
+            artifact = proof_root / "cache_trace.json"
+            write_json(artifact, {"verdict": "failed", "why": "the store holds no atoms"})
+            return ProofCheck(
+                name="Cache Trace",
+                requirement="PRD 14.2 requires cache and memory proof with visible hit, miss, write, and replay state.",
+                status="failed",
+                evidence=[str(artifact)],
+                detail="the store holds no atoms, so there is nothing for the cache to answer",
+                reason="an empty store answers every question with an empty context, which would make the two compiles agree for the wrong reason",
+                rerun="ely-eye remember or ely-eye ingest first, then re-run this proof",
+            )
+        # The nonce keeps the key unique; the quoted text keeps the question answerable, so a pass
+        # means the cache served a real context rather than two identical empty ones.
+        question = f"{probe[0].text[:60]} cache proof {secrets.token_hex(6)}"
+        before = _semantic_counters(fabric)
+        first = compiler.compile(question, RuntimeProfile.library_100m)
+        middle = _semantic_counters(fabric)
+        second = compiler.compile(question, RuntimeProfile.library_100m)
+        after = _semantic_counters(fabric)
+        digest_first = compiled_context_digest(first)
+        digest_second = compiled_context_digest(second)
+        payload = {
+            "question": question,
+            "counters": {"before": before, "after_first": middle, "after_second": after},
+            "digest_first": digest_first,
+            "digest_second": digest_second,
+            "cache_trace_first": first.cache_trace_id,
+            "cache_trace_second": second.cache_trace_id,
+        }
+        artifact = proof_root / "cache_trace.json"
+        write_json(artifact, payload)
+        miss_recorded = middle["miss"] - before["miss"] == 1 and middle["hit"] - before["hit"] == 0
+        hit_recorded = after["hit"] - middle["hit"] == 1
+        same_answer = digest_first == digest_second and bool(first.hits)
+        valid = miss_recorded and hit_recorded and same_answer
+        return ProofCheck(
+            name="Cache Trace",
+            requirement="PRD 14.2 requires cache and memory proof with visible hit, miss, write, and replay state.",
+            status="passed" if valid else "failed",
+            evidence=[str(artifact)],
+            detail=(
+                f"semantic layer counters miss {before['miss']}→{middle['miss']}→{after['miss']}, "
+                f"hit {before['hit']}→{middle['hit']}→{after['hit']}, replay digest {digest_first}"
+            ),
+            reason="" if valid else "the second compile was not recorded as a cache hit",
+            rerun="ely-eye proof-suite --cartridge-id <id>",
+        )
+
+    def _resolve_cartridge(self, cartridge_id: str | None) -> dict[str, Any]:
+        rows = self.db.list_cartridges()
+        if not rows:
+            raise ValueError("No Context Cartridge is available.")
+        if cartridge_id is None:
+            return rows[0]
+        for row in rows:
+            if row["cartridge_id"] == cartridge_id:
+                return row
+        raise ValueError(f"Unknown cartridge: {cartridge_id}")
+

backend/src/ely_eye/proofs/checks_common.py

@@ -0,0 +1,65 @@
+"""Shared verdict helpers for the proof checks.
+
+Kept out of suite.py so the check modules can import them without importing the suite.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from ..cache_fabric import SEMANTIC_LAYER, CacheFabric
+from ..config import Settings
+from ..evaluation_schemas import (
+    ProofCheck,
+)
+from .report import (
+    write_json,
+)
+
+# A fixed puzzle count per hop depth. The old suite re-rolled until the running rate crossed its
+
+HASHHOP_ATTEMPTS = 8
+
+GOLD_ADAPTERS = (
+    "HME-Core-LoRA",
+    "HME-Vision-LoRA",
+    "HME-TTT-VL",
+    "HME-Visual-MTP",
+    "HME-Router",
+    "HME-Retrieval",
+)
+
+
+def _deferred_without_model(
+    settings: Settings, artifact: Path, name: str, requirement: str, rerun: str
+) -> ProofCheck | None:
+    """A check that needs weights proves nothing on a machine that has no weights.
+
+    Calling that a failure would blame the machine; calling it a pass would be a lie. The suite
+    records what is missing and the command that runs the check once the weights arrive.
+    """
+    model_dir = settings.model_dir
+    if model_dir is not None and model_dir.exists():
+        return None
+    write_json(
+        artifact,
+        {"status": "deferred", "reason": "no local model weights", "model_dir": str(model_dir)},
+    )
+    return ProofCheck(
+        name=name,
+        requirement=requirement,
+        status="deferred",
+        evidence=[str(artifact)],
+        detail=f"model_dir={model_dir}",
+        reason="the generator or encoder this check measures is not installed on this machine",
+        rerun=rerun,
+    )
+
+
+def _semantic_counters(fabric: CacheFabric) -> dict[str, int]:
+    """Hit and miss counts for the one cache layer that exists; zero when the layer has no rows."""
+    for layer in fabric.statuses():
+        if layer.layer == SEMANTIC_LAYER:
+            return {"hit": layer.hit_count, "miss": layer.miss_count}
+    return {"hit": 0, "miss": 0}
+

backend/src/ely_eye/proofs/checks_runtime.py

@@ -0,0 +1,270 @@
+"""Runtime adapter binding, gold-adapter generation, launch profiles, and the removed module check.
+
+Moved out of suite.py so the suite order stays readable; each method is a check the
+suite runs, and the suite class mixes these in.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+from ..compiler import ContextCompiler
+from ..config import Settings
+from ..db import Database
+from ..evaluation_schemas import (
+    ProofCheck,
+)
+from ..evidence_graph import EvidenceGraphService
+from ..profiles import runtime_profile_report
+from ..runtime import QwenRuntime, resolve_runtime_adapter
+from ..schemas import (
+    RuntimeProfile,
+    TrainingAdapterKind,
+)
+from ..verifier import Verifier
+
+# A fixed puzzle count per hop depth. The old suite re-rolled until the running rate crossed its
+from .checks_common import (
+    GOLD_ADAPTERS,
+    _deferred_without_model,
+)
+from .report import (
+    compact_generation_context,
+    write_json,
+)
+from .training_trace import (
+    count_safetensor_keys,
+    peft_config_loads,
+)
+
+
+class RuntimeChecks:
+    """Runtime adapter binding, gold-adapter generation, launch profiles, and the removed module check."""
+
+    settings: Settings
+    db: Database
+
+    def _runtime_adapter_check(self, proof_root: Path) -> ProofCheck:
+        artifact = proof_root / "runtime_adapter.json"
+        deferred = _deferred_without_model(
+            self.settings,
+            artifact,
+            "Runtime Adapter Binding",
+            "PRD 18 requires the local model package to run with the trained adapter, and PRD 12.1 requires HME-Core-LoRA to be usable by the Qwen runtime.",
+            "ely-eye runtime-generation-proof with ELY_EYE_MODEL_DIR set",
+        )
+        if deferred is not None:
+            return deferred
+        runtime = QwenRuntime(self.settings)
+        status = runtime.status()
+        adapter = resolve_runtime_adapter(self.settings)
+        weight_path = adapter.path / "adapter_model.safetensors" if adapter else None
+        payload = {
+            "runtime_status": status.model_dump(mode="json"),
+            "resolved_adapter": {
+                "path": str(adapter.path) if adapter else None,
+                "kind": adapter.manifest.kind.value if adapter else None,
+                "adapter_id": adapter.manifest.adapter_id if adapter else None,
+                "sha256": adapter.manifest.sha256 if adapter else None,
+                "base_model": adapter.manifest.base_model if adapter else None,
+                "weight_path": str(weight_path) if weight_path else None,
+                "safetensor_keys": count_safetensor_keys(weight_path) if weight_path else 0,
+                "peft_config_ok": peft_config_loads(adapter.path) if adapter else False,
+            },
+        }
+        artifact = proof_root / "runtime_adapter.json"
+        write_json(artifact, payload)
+        valid = (
+            status.available
+            and adapter is not None
+            and adapter.manifest.kind == TrainingAdapterKind.hme_core_lora
+            and adapter.manifest.base_model == self.settings.model_id
+            and status.adapter_path == str(adapter.path)
+            and weight_path is not None
+            and count_safetensor_keys(weight_path) > 0
+            and peft_config_loads(adapter.path)
+        )
+        return ProofCheck(
+            name="Runtime Adapter Binding",
+            requirement="PRD 18 requires the local model package to run with the trained adapter, and PRD 12.1 requires HME-Core-LoRA to be usable by the Qwen runtime.",
+            status="passed" if valid else "failed",
+            evidence=[str(artifact)],
+            detail=(
+                f"{status.backend} runtime resolves {adapter.manifest.kind.value} at {adapter.path}"
+                if adapter
+                else "runtime adapter resolution is empty"
+            ),
+        )
+
+    def _runtime_generation_check(self, proof_root: Path) -> ProofCheck:
+        artifact = proof_root / "runtime_generation.json"
+        deferred = _deferred_without_model(
+            self.settings,
+            artifact,
+            "Runtime Adapter Generation",
+            "PRD 12.1 and PRD 18 require the trained HME-Core-LoRA adapter to be loaded by the local Qwen runtime and used for an evidence-cited generation.",
+            "ely-eye proof-suite --cartridge-id <id> with ELY_EYE_MODEL_DIR set",
+        )
+        if deferred is not None:
+            return deferred
+        retrieval_question = (
+            "Using the Ely-Eye PRD evidence, state the PRD 12.1 local training adapter "
+            "requirements for HME-Core-LoRA, HME-Vision-LoRA, HME-TTT-VL, "
+            "HME-Visual-MTP, HME-Router, and HME-Retrieval. Return JSON with atom id "
+            "citations."
+        )
+        proof_settings = self.settings.model_copy(update={"max_new_tokens": 256, "temperature": 0.0})
+        runtime = QwenRuntime(proof_settings)
+        payload: dict[str, Any] = {
+            "question": retrieval_question,
+            "status": "failed",
+        }
+        try:
+            compiler = ContextCompiler(proof_settings, self.db)
+            context = compiler.compile(retrieval_question, RuntimeProfile.library_100m)
+            context = compact_generation_context(context, "12.1")
+            if not context.hits:
+                raise RuntimeError("No PRD text evidence was available for runtime generation proof.")
+            generation_question = (
+                "According to the retrieved PRD evidence, which local training adapters does "
+                "PRD 12.1 require? Answer with one compact JSON object: a \"citations\" field "
+                "naming the atom ids you actually used, and a \"sentence\" field listing them."
+            )
+            payload["generation_question"] = generation_question
+            status = runtime.status()
+            adapter = resolve_runtime_adapter(proof_settings)
+            generation = runtime.generate(generation_question, context)
+            answer, citations, citation_report = Verifier().verify_answer(generation.text, context)
+            claim_id = EvidenceGraphService(proof_settings, self.db).record_answer(
+                generation_question,
+                answer,
+                citation_report,
+                context,
+            )
+            payload.update(
+                {
+                    "status": "passed",
+                    "runtime_status": status.model_dump(mode="json"),
+                    "adapter": {
+                        "path": generation.adapter_path,
+                        "kind": generation.adapter_kind,
+                        "adapter_id": generation.adapter_id,
+                        "sha256": generation.adapter_sha256,
+                    },
+                    "tokens": {
+                        "input": generation.input_tokens,
+                        "output": generation.output_tokens,
+                    },
+                    "context": {
+                        "profile": context.plan.profile.value,
+                        "hit_ids": [hit.atom.atom_id for hit in context.hits],
+                        "token_equivalent": context.token_equivalent,
+                        "cache_trace_id": context.cache_trace_id,
+                    },
+                    "raw_generation": generation.text,
+                    "verified_answer": answer,
+                    "citations": citations,
+                    "evidence_graph_claim_id": claim_id,
+                    "verifier": citation_report.model_dump(mode="json"),
+                }
+            )
+            named = [name for name in GOLD_ADAPTERS if name in answer]
+            cited_atoms = [
+                hit.atom for hit in context.hits if hit.atom.atom_id in set(citations)
+            ]
+            supported = [atom for atom in cited_atoms if sum(name in (atom.text or "") for name in GOLD_ADAPTERS) >= 3]
+            valid = (
+                status.available
+                and generation.backend == "transformers"
+                and adapter is not None
+                and generation.adapter_path == str(adapter.path)
+                and generation.adapter_kind == TrainingAdapterKind.hme_core_lora.value
+                and generation.adapter_sha256 == adapter.manifest.sha256
+                and bool(context.hits)
+                and bool(generation.text.strip())
+                and (generation.output_tokens or 0) > 0
+                and len(named) >= 5
+                and bool(citations)
+                and citation_report.citation_accuracy == 1.0
+                and bool(supported)
+            )
+            payload["answer_adapters"] = named
+            payload["cited_atoms"] = [atom.atom_id for atom in cited_atoms]
+            payload["supporting_citations"] = [atom.atom_id for atom in supported]
+            if not valid:
+                payload["status"] = "failed"
+            write_json(artifact, payload)
+            return ProofCheck(
+                name="Runtime Adapter Generation",
+                requirement="PRD 12.1 and PRD 18 require the trained HME-Core-LoRA adapter to be loaded by the local Qwen runtime and used for an evidence-cited generation.",
+                status="passed" if valid else "failed",
+                evidence=[str(artifact)],
+                detail=(
+                    f"{generation.backend} generated {generation.output_tokens} tokens with "
+                    f"{generation.adapter_kind} citation accuracy "
+                    f"{citation_report.citation_accuracy:.2f}"
+                ),
+            )
+        except Exception as exc:  # noqa: BLE001  # a probe records its failure instead of crashing the proof run
+            payload["error"] = str(exc)
+            write_json(artifact, payload)
+            return ProofCheck(
+                name="Runtime Adapter Generation",
+                requirement="PRD 12.1 and PRD 18 require the trained HME-Core-LoRA adapter to be loaded by the local Qwen runtime and used for an evidence-cited generation.",
+                status="failed",
+                evidence=[str(artifact)],
+                detail=f"runtime generation failed: {exc}",
+            )
+        finally:
+            runtime.unload()
+
+    def _runtime_profile_check(self, proof_root: Path) -> ProofCheck:
+        report = runtime_profile_report()
+        artifact = proof_root / "runtime_profiles.json"
+        artifact.write_text(report.model_dump_json(indent=2), encoding="utf-8")
+        required_runtimes = {"SGLang", "vLLM", "KTransformers", "llama.cpp"}
+        runtimes = {profile.runtime for profile in report.profiles}
+        gaps = [f"{item.name}: {item.detail}" for item in report.validations if item.status == "missing"]
+        unregistered = sorted(required_runtimes - runtimes)
+        if unregistered:
+            gaps.append("no profile registered for: " + ", ".join(unregistered))
+        if not gaps and report.status != "ready":
+            gaps.append(f"the profile report says {report.status}")
+        valid = not gaps
+        return ProofCheck(
+            name="Runtime Profile Delivery",
+            requirement="PRD 13.1 requires SGLang as the main runtime plus vLLM, KTransformers, and llama.cpp deployment profiles.",
+            status="passed" if valid else "failed",
+            evidence=[str(artifact)],
+            detail="registered runtimes: " + ", ".join(profile.name for profile in report.profiles),
+            reason="; ".join(gaps),
+            rerun="install the launcher in the environment the profile probes (the probe command is in runtime_profiles.json), then ely-eye runtime-profiles",
+        )
+
+    def _research_module_check(self, proof_root: Path) -> ProofCheck:
+        """The old version of this check could not fail, so it proved nothing.
+
+        It compiled a query, asked the research-module service for eight module reports, and
+        passed when every list was non-empty. The module vectors it ranks with are SHA256-derived
+        placeholders, so a non-empty list is a property of the loop, not of the system.
+        """
+        artifact = proof_root / "research_modules.json"
+        write_json(
+            artifact,
+            {
+                "claim": "PRD 11 research modules",
+                "verdict": "removed",
+                "why": "the check asserted non-empty lists built from SHA256-derived placeholder vectors",
+            },
+        )
+        return ProofCheck(
+            name="Research Module Delivery",
+            requirement="PRD 11 requires V-NSA, DVKV, X-Modal KV Sharing, Speculative Context Prefetch, Predictive KV Eviction, Learned Evidence Graph, Differentiable Retrieval Adapter, and Sleep-Time Visual Consolidation.",
+            status="removed",
+            evidence=[str(artifact)],
+            detail="no measurable claim exists per module; the previous check passed on non-empty lists",
+            reason="the check could not fail: eight named modules produced non-empty lists by construction",
+            rerun="write one measurable claim per module, then re-add the check",
+        )
+

backend/src/ely_eye/proofs/checks_training.py

@@ -0,0 +1,232 @@
+"""Adapter training and the learned-memory adapters bound into a cartridge.
+
+Moved out of suite.py so the suite order stays readable; each method is a check the
+suite runs, and the suite class mixes these in.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+from ..config import Settings
+from ..db import Database
+from ..evaluation_schemas import (
+    ProofCheck,
+)
+from ..schemas import (
+    AdapterManifest,
+    TrainingAdapterKind,
+)
+from ..training import (
+    safetensor_weight_stats,
+    sha256_file,
+    sha256_tree,
+)
+
+# A fixed puzzle count per hop depth. The old suite re-rolled until the running rate crossed its
+from .report import (
+    read_json_or_empty,
+    write_json,
+)
+from .training_trace import (
+    count_safetensor_keys,
+    peft_config_loads,
+    validate_training_contract,
+    validate_training_proof,
+    validate_training_trace,
+)
+
+
+class TrainingChecks:
+    """Adapter training and the learned-memory adapters bound into a cartridge."""
+
+    settings: Settings
+    db: Database
+    required_adapters: tuple[TrainingAdapterKind, ...]
+
+    def _adapter_training_check(self, cartridge_id: str, cartridge_root: Path, proof_root: Path) -> ProofCheck:
+        rows: list[dict[str, Any]] = []
+        missing: list[str] = []
+        manifests = self._adapter_manifests(cartridge_id)
+        for kind in self.required_adapters:
+            item = manifests.get(kind)
+            if item is None:
+                missing.append(f"{kind.value}: no adapter manifest")
+                continue
+            manifest, adapter_root = item
+            weight_path = adapter_root / "adapter_model.safetensors"
+            dataset_path = Path(manifest.dataset_path)
+            cartridge_adapter = cartridge_root / "adapters" / kind.value / "adapter_manifest.json"
+            weight_report = safetensor_weight_stats(weight_path)
+            keys = int(weight_report["tensor_count"])
+            summary = read_json_or_empty(adapter_root / "training_summary.json")
+            summary_weight_sha = summary.get("adapter_weight_sha256")
+            summary_weight_sha_ok = (
+                isinstance(summary_weight_sha, str) and summary_weight_sha == weight_report["sha256"]
+            )
+            manifest_weight_sha_ok = (
+                isinstance(manifest.adapter_weight_sha256, str)
+                and manifest.adapter_weight_sha256 == weight_report["sha256"]
+            )
+            manifest_weight_stats_ok = (
+                manifest.adapter_tensor_count == weight_report["tensor_count"]
+                and manifest.adapter_lora_tensor_count == weight_report["lora_tensor_count"]
+                and manifest.adapter_total_elements == weight_report["total_elements"]
+                and manifest.adapter_nonzero_elements == weight_report["nonzero_elements"]
+                and manifest.adapter_nonzero_elements > 0
+                and manifest.adapter_lora_tensor_count > 0
+                and manifest.adapter_weights_finite
+            )
+            peft_ok = peft_config_loads(adapter_root)
+            dataset_sha_ok = dataset_path.exists() and sha256_file(dataset_path) == manifest.dataset_sha256
+            tree_sha_ok = sha256_tree(adapter_root) == manifest.sha256
+            cartridge_bound = cartridge_adapter.exists()
+            trace_report = validate_training_trace(manifest, adapter_root)
+            proof_report = validate_training_proof(manifest, adapter_root)
+            contract_report = validate_training_contract(manifest)
+            # Named conditions, so a failure can say which promise broke instead of listing the
+            # adapters that happened to be present.
+            conditions: list[tuple[str, bool]] = [
+                ("sample_count", manifest.sample_count > 0),
+                ("max_steps", manifest.max_steps > 0),
+                ("trainable_params", manifest.trainable_params > 0),
+                ("final_loss recorded", manifest.final_loss is not None),
+                ("safetensors readable", weight_report["valid"]),
+                ("weight sha256 matches the training summary", summary_weight_sha_ok),
+                ("weight sha256 matches the manifest", manifest_weight_sha_ok),
+                ("weight statistics match the manifest", manifest_weight_stats_ok),
+                ("peft config loads", peft_ok),
+                ("dataset sha256 matches", dataset_sha_ok),
+                ("adapter tree sha256 matches", tree_sha_ok),
+                ("adapter is bound into the cartridge", cartridge_bound),
+                ("training trace verifies", trace_report["valid"]),
+                ("training proof verifies", proof_report["valid"]),
+                ("training contract verifies", contract_report["valid"]),
+            ]
+            valid = all(ok for _, ok in conditions)
+            if not valid:
+                gaps = ", ".join(name for name, ok in conditions if not ok)
+                missing.append(f"{kind.value}: {gaps}")
+            rows.append(
+                {
+                    "kind": kind.value,
+                    "adapter": str(adapter_root),
+                    "dataset": str(dataset_path),
+                    "sample_count": manifest.sample_count,
+                    "max_steps": manifest.max_steps,
+                    "trainable_params": manifest.trainable_params,
+                    "final_loss": manifest.final_loss,
+                    "safetensor_keys": keys,
+                    "adapter_weight_sha256": weight_report["sha256"],
+                    "adapter_weight_sha256_ok": summary_weight_sha_ok and manifest_weight_sha_ok,
+                    "adapter_weight_stats_ok": manifest_weight_stats_ok,
+                    "adapter_weight_stats": weight_report,
+                    "dataset_sha256_ok": dataset_sha_ok,
+                    "adapter_sha256_ok": tree_sha_ok,
+                    "peft_config_ok": peft_ok,
+                    "cartridge_bound": cartridge_bound,
+                    "training_contract": contract_report,
+                    "training_trace": trace_report,
+                    "training_proof": proof_report,
+                }
+            )
+        artifact = proof_root / "adapter_training.json"
+        write_json(artifact, {"adapters": rows})
+        return ProofCheck(
+            name="Adapter Training Proof",
+            requirement="PRD 12.1 requires HME-Core-LoRA, HME-Vision-LoRA, HME-TTT-VL, HME-Visual-MTP, HME-Router, and HME-Retrieval trained adapters.",
+            status="failed" if missing else "passed",
+            evidence=[str(artifact), str(self.settings.adapters_dir), str(cartridge_root / "adapters")],
+            detail="adapters with verified training evidence: " + ", ".join(row["kind"] for row in rows),
+            reason="; ".join(missing),
+            rerun="train the missing adapters (ely-eye train-lora) or attach them (ely-eye attach-adapter), then ely-eye proof-suite --cartridge-id <id>",
+        )
+
+    def _learned_memory_recall_check(
+        self,
+        cartridge_id: str,
+        cartridge_root: Path,
+        proof_root: Path,
+    ) -> ProofCheck:
+        manifests = self._adapter_manifests(cartridge_id)
+        learned_kinds = (
+            TrainingAdapterKind.hme_ttt_vl,
+            TrainingAdapterKind.hme_core_lora,
+            TrainingAdapterKind.hme_retrieval,
+        )
+        rows: list[dict[str, Any]] = []
+        missing: list[str] = []
+        for kind in learned_kinds:
+            item = manifests.get(kind)
+            if item is None:
+                missing.append(f"{kind.value}: no adapter manifest")
+                continue
+            manifest, adapter_root = item
+            weight_path = adapter_root / "adapter_model.safetensors"
+            cartridge_weight = cartridge_root / "adapters" / kind.value / "adapter_model.safetensors"
+            rows.append(
+                {
+                    "kind": kind.value,
+                    "adapter": str(adapter_root),
+                    "cartridge_weight": str(cartridge_weight),
+                    "base_model": manifest.base_model,
+                    "dataset_sha256": manifest.dataset_sha256,
+                    "sample_count": manifest.sample_count,
+                    "trainable_params": manifest.trainable_params,
+                    "safetensor_keys": count_safetensor_keys(weight_path),
+                    "cartridge_weight_present": cartridge_weight.exists(),
+                    "peft_config_ok": peft_config_loads(adapter_root),
+                }
+            )
+            if not cartridge_weight.exists():
+                missing.append(f"{kind.value}: the cartridge holds no adapter weights")
+            elif count_safetensor_keys(weight_path) <= 0:
+                missing.append(f"{kind.value}: the adapter weight file holds no tensors")
+        artifact = proof_root / "learned_memory_recall.json"
+        write_json(
+            artifact,
+            {
+                "cartridge_id": cartridge_id,
+                "learned_memory_adapters": rows,
+                "runtime_binding": "Adapters are stored inside the Context Cartridge and can be loaded without source files.",
+            },
+        )
+        return ProofCheck(
+            name="Learned Memory Recall",
+            requirement="PRD 14.2 and 15.4 require TTT-VL and LoRA capsule memory to be stored as cartridge-loadable adapter weights.",
+            status="failed" if missing else "deferred",
+            evidence=[str(artifact), str(cartridge_root / "adapters" / TrainingAdapterKind.hme_ttt_vl.value)],
+            detail=(
+                "weights load for " + ", ".join(row["kind"] for row in rows)
+                + "; no recall number was measured"
+            ),
+            reason=(
+                "the check loads adapter weights and never asks the model to remember anything, so it "
+                "cannot carry the word recall; the load evidence is real, the recall claim is not yet measured"
+            ),
+            rerun="add a recall probe that answers questions with and without the adapter, then ely-eye proof-suite --cartridge-id <id>",
+        )
+
+    def _adapter_manifests(
+        self,
+        cartridge_id: str,
+    ) -> dict[TrainingAdapterKind, tuple[AdapterManifest, Path]]:
+        candidates: dict[TrainingAdapterKind, list[tuple[AdapterManifest, Path]]] = {}
+        for path in sorted(self.settings.adapters_dir.glob("*/adapter_manifest.json")):
+            manifest = AdapterManifest.model_validate_json(path.read_text(encoding="utf-8"))
+            candidates.setdefault(manifest.kind, []).append((manifest, path.parent))
+        selected: dict[TrainingAdapterKind, tuple[AdapterManifest, Path]] = {}
+        for kind, items in candidates.items():
+            items.sort(
+                key=lambda item: (
+                    item[0].cartridge_id == cartridge_id,
+                    (item[1] / "adapter_model.safetensors").stat().st_mtime
+                    if (item[1] / "adapter_model.safetensors").exists()
+                    else 0.0,
+                ),
+                reverse=True,
+            )
+            selected[kind] = items[0]
+        return selected
+

backend/src/ely_eye/proofs/checks_visual.py

@@ -0,0 +1,315 @@
+"""Visual HashHop retrieval, the benchmark registry, and the visual contradiction lens.
+
+Moved out of suite.py so the suite order stays readable; each method is a check the
+suite runs, and the suite class mixes these in.
+"""
+
+from __future__ import annotations
+
+import math
+from pathlib import Path
+from typing import Any
+
+from ..benchmarks import (
+    benchmark_prediction_report,
+    benchmark_registry_report,
+    benchmark_sample_report,
+    benchmark_source_report,
+)
+from ..config import Settings
+from ..db import Database
+from ..evaluation_schemas import (
+    ProofCheck,
+)
+from ..runtime import QwenRuntime
+from ..verifier import VisualContradictionLens
+
+# A fixed puzzle count per hop depth. The old suite re-rolled until the running rate crossed its
+from .checks_common import (
+    HASHHOP_ATTEMPTS,
+    _deferred_without_model,
+)
+from .contradictions import CORPUS_NAME, corpus_path, load_corpus, scenario_context
+from .hashhop import HashHopEvaluator
+from .hashhop_verify import (
+    collect_model_visual_proofs,
+    copy_proof_artifacts,
+    hashhop_chain_valid,
+    pass_rate,
+)
+from .report import (
+    format_scores,
+    write_json,
+)
+
+
+class VisualChecks:
+    """Visual HashHop retrieval, the benchmark registry, and the visual contradiction lens."""
+
+    settings: Settings
+    db: Database
+
+    def _visual_hashhop_check(self, proof_root: Path) -> ProofCheck:
+        artifact = proof_root / "visual_hashhop_suite.json"
+        deferred = _deferred_without_model(
+            self.settings,
+            artifact,
+            "Visual HashHop Retrieval",
+            "PRD 14.2 requires the model to resolve visual hash hops inside 262K, 1.01M and 100M token contexts.",
+            "ely-eye hashhop visual 1 262144 with ELY_EYE_MODEL_DIR set",
+        )
+        if deferred is not None:
+            return deferred
+        proof_settings = self.settings.model_copy(update={"max_new_tokens": 128, "temperature": 0.0})
+        runtime = QwenRuntime(proof_settings)
+        evaluator = HashHopEvaluator(proof_settings, runtime)
+        try:
+            base_proofs = [
+                evaluator.generate_visual_proof(1, 262_144),
+                evaluator.generate_visual_proof(2, 262_144),
+                evaluator.generate_visual_proof(1, 1_010_000),
+                evaluator.generate_visual_proof(1, 100_000_000),
+                evaluator.generate_text_proof(2, 1_010_000),
+                evaluator.generate_model_text_proof(2, 1_010_000),
+            ]
+            model_visual_1hop = collect_model_visual_proofs(evaluator, 1, 262_144, HASHHOP_ATTEMPTS)
+            model_visual_2hop = collect_model_visual_proofs(evaluator, 2, 262_144, HASHHOP_ATTEMPTS)
+            proofs = base_proofs + model_visual_1hop + model_visual_2hop
+        finally:
+            runtime.unload()
+        payloads: list[dict[str, Any]] = []
+        for proof in proofs:
+            target = proof_root / "hashhop" / proof.proof_id
+            copy_proof_artifacts(proof, target)
+            payloads.append(
+                {
+                    "proof": proof.model_dump(mode="json"),
+                    "chain_valid": hashhop_chain_valid(proof),
+                    "artifact_dir": str(target),
+                }
+            )
+        artifact = proof_root / "visual_hashhop_suite.json"
+        model_solved = [
+            item
+            for item in payloads
+            if item["proof"]["model_target_id"] is not None or item["proof"]["passed"] is not None
+        ]
+        model_text_solved = [
+            proof
+            for proof in base_proofs
+            if proof.kind == "text_hashhop" and proof.model_target_id is not None
+        ]
+        model_solved_passed = [item for item in model_solved if item["proof"]["passed"]]
+        visual_1hop_rate = pass_rate(model_visual_1hop)
+        visual_2hop_rate = pass_rate(model_visual_2hop)
+        metrics = {
+            "visual_hashhop_262k_1hop": {
+                "target": 0.90,
+                "sample_count": len(model_visual_1hop),
+                "passed": sum(1 for proof in model_visual_1hop if proof.passed),
+                "pass_rate": visual_1hop_rate,
+            },
+            "visual_hashhop_262k_2hop": {
+                "target": 0.75,
+                "sample_count": len(model_visual_2hop),
+                "passed": sum(1 for proof in model_visual_2hop if proof.passed),
+                "pass_rate": visual_2hop_rate,
+            },
+        }
+        write_json(artifact, {"proofs": payloads, "metrics": metrics})
+        valid = (
+            all(item["chain_valid"] for item in payloads)
+            and bool(model_text_solved)
+            and all(bool(proof.passed) for proof in model_text_solved)
+            and len(model_visual_1hop) >= 3
+            and len(model_visual_2hop) >= 3
+            and visual_1hop_rate >= 0.90
+            and visual_2hop_rate >= 0.75
+        )
+        return ProofCheck(
+            name="Visual HashHop",
+            requirement="PRD 14.2 and 14.3 require Visual HashHop 262K 1-hop, 262K 2-hop, 1M 1-hop, 100M memory retrieval assets, text HashHop 1M 2-hop, model-solved text HashHop evidence, and model-solved visual 1-hop and 2-hop HashHop evidence.",
+            status="passed" if valid else "failed",
+            evidence=[str(artifact), str(proof_root / "hashhop")],
+            detail=(
+                f"{len(proofs)} HashHop proofs generated; "
+                f"{len(model_solved_passed)}/{len(model_solved)} model-solved proofs passed; "
+                f"262K 1-hop={visual_1hop_rate:.2f}, 2-hop={visual_2hop_rate:.2f}"
+            ),
+        )
+
+    def _standard_benchmark_check(self, proof_root: Path) -> ProofCheck:
+        artifact = proof_root / "standard_benchmarks.json"
+        deferred = _deferred_without_model(
+            self.settings,
+            artifact,
+            "Standard Benchmark Registry",
+            "PRD 14.1 requires MMMU-Pro, OCRBench, MMLongBench-Doc, VideoMME, LongVideoBench, RULER, and LongBench v2 in the evaluation proof chain with real benchmark source metadata, loadable official samples, and local model prediction artifacts.",
+            "ely-eye benchmark-predictions with ELY_EYE_MODEL_DIR set",
+        )
+        if deferred is not None:
+            return deferred
+        registry = benchmark_registry_report()
+        sources = benchmark_source_report()
+        samples = benchmark_sample_report(proof_root / "benchmark_samples")
+        predictions = benchmark_prediction_report(self.settings, proof_root / "benchmark_predictions")
+        artifact = proof_root / "standard_benchmarks.json"
+        source_artifact = proof_root / "standard_benchmark_sources.json"
+        sample_artifact = proof_root / "standard_benchmark_samples.json"
+        prediction_artifact = proof_root / "standard_benchmark_predictions.json"
+        artifact.write_text(registry.model_dump_json(indent=2), encoding="utf-8")
+        source_artifact.write_text(sources.model_dump_json(indent=2), encoding="utf-8")
+        sample_artifact.write_text(samples.model_dump_json(indent=2), encoding="utf-8")
+        prediction_artifact.write_text(predictions.model_dump_json(indent=2), encoding="utf-8")
+        required = {
+            "MMMU-Pro",
+            "OCRBench",
+            "MMLongBench-Doc",
+            "VideoMME",
+            "LongVideoBench",
+            "RULER",
+            "LongBench v2",
+        }
+        registered = {benchmark.name for benchmark in registry.benchmarks}
+        source_names = {source.name for source in sources.sources if source.status == "ready" and source.sha}
+        loaded_sample_names = {sample.name for sample in samples.samples if sample.status == "loaded"}
+        gated_sample_names = {sample.name for sample in samples.samples if sample.status == "gated"}
+        source_proved_sample_names = {
+            sample.name
+            for sample in samples.samples
+            if sample.status == "loaded" and sample.sample_sha256 and sample.source_dataset and sample.source_sha
+        }
+        predicted_sample_names = {
+            prediction.name for prediction in predictions.predictions if prediction.status == "predicted"
+        }
+        prediction_scores = {
+            prediction.name: prediction.score
+            for prediction in predictions.predictions
+            if prediction.status == "predicted"
+        }
+        scored_predictions_pass = {
+            "MMMU-Pro",
+            "OCRBench",
+        } <= {
+            name
+            for name, score in prediction_scores.items()
+            if isinstance(score, (int, float)) and math.isfinite(float(score)) and float(score) >= 0.5
+        }
+        sample_names = loaded_sample_names | gated_sample_names
+        valid = (
+            registry.status == "ready"
+            and sources.status == "ready"
+            and samples.status == "ready"
+            and predictions.status == "ready"
+            and required <= registered
+            and required <= source_names
+            and required <= sample_names
+            and required <= loaded_sample_names
+            and required <= source_proved_sample_names
+            and {"MMMU-Pro", "OCRBench"} <= predicted_sample_names
+            and scored_predictions_pass
+        )
+        return ProofCheck(
+            name="Standard Benchmark Registry",
+            requirement="PRD 14.1 requires MMMU-Pro, OCRBench, MMLongBench-Doc, VideoMME, LongVideoBench, RULER, and LongBench v2 in the evaluation proof chain with real benchmark source metadata, loadable official samples, and local model prediction artifacts.",
+            status="passed" if valid else "failed",
+            evidence=[str(artifact), str(source_artifact), str(sample_artifact), str(prediction_artifact)],
+            detail=(
+                f"verified {len(loaded_sample_names)} loadable official benchmark samples; "
+                f"{len(predicted_sample_names)} local model sample predictions; "
+                f"scores: {format_scores(prediction_scores)}; "
+                f"gated samples: {', '.join(sorted(gated_sample_names)) or 'none'}"
+            ),
+        )
+
+    def _visual_contradiction_check(self, proof_root: Path) -> ProofCheck:
+        lens = VisualContradictionLens()
+        artifact = proof_root / "visual_contradiction_lens.json"
+        path = corpus_path(self.settings.data_dir)
+        if path is None:
+            write_json(
+                artifact,
+                {"verdict": "deferred", "why": f"no {CORPUS_NAME} found in {self.settings.data_dir}"},
+            )
+            return ProofCheck(
+                name="Visual Contradiction Lens",
+                requirement="PRD 14.2 and 14.3 require the Visual Contradiction Lens to localize cross-version visual conflicts.",
+                status="deferred",
+                evidence=[str(artifact)],
+                detail="no contradiction corpus on disk",
+                reason="the check refuses to invent the examples it grades itself on inside the code under test",
+                rerun=f"ely-eye contradiction-corpus, or drop a {CORPUS_NAME} in {self.settings.data_dir}",
+            )
+        scenarios, corpus_sha256 = load_corpus(path)
+        conflicts = [scenario for scenario in scenarios if scenario["expect_drift"]]
+        controls = [scenario for scenario in scenarios if not scenario["expect_drift"]]
+        detected = 0
+        false_positives = 0
+        categories: set[str] = set()
+        rows: list[dict[str, Any]] = []
+        known_categories = (
+            "design-token drift",
+            "layout drift",
+            "typography drift",
+            "copy drift",
+            "visual-code drift",
+            "temporal drift",
+        )
+        for scenario in scenarios:
+            notes = lens.detect(scenario_context(scenario))
+            flagged = bool(notes)
+            if scenario["expect_drift"]:
+                if flagged:
+                    detected += 1
+                    for note in notes:
+                        label = note.split(":", 1)[0]
+                        categories.update(category for category in known_categories if label.endswith(category))
+            elif flagged:
+                false_positives += 1
+            rows.append(
+                {
+                    "scenario": scenario["name"],
+                    "expect_drift": scenario["expect_drift"],
+                    "flagged": flagged,
+                    "notes": notes,
+                }
+            )
+        recall = detected / len(conflicts) if conflicts else 0.0
+        false_positive_rate = false_positives / len(controls) if controls else 0.0
+        valid = recall >= 0.80 and false_positive_rate == 0.0 and len(categories) >= 5
+        write_json(
+            artifact,
+            {
+                "corpus": str(path),
+                "corpus_sha256": corpus_sha256,
+                "scenarios": rows,
+                "metrics": {
+                    "conflict_count": len(conflicts),
+                    "control_count": len(controls),
+                    "detected": detected,
+                    "false_positives": false_positives,
+                    "recall": recall,
+                    "false_positive_rate": false_positive_rate,
+                    "detected_categories": sorted(categories),
+                    "gate_met": valid,
+                },
+            },
+        )
+        return ProofCheck(
+            name="Visual Contradiction Lens",
+            requirement="PRD 14.2 and 14.3 require the Visual Contradiction Lens to localize cross-version visual conflicts with contradiction detection >= 80% on internal conflict samples and without false positives on consistent samples.",
+            status="deferred",
+            evidence=[str(artifact)],
+            detail=(
+                f"measured recall {recall:.2f} on {len(conflicts)} conflict samples, "
+                f"false positives {false_positives}/{len(controls)}, "
+                f"categories {', '.join(sorted(categories))}"
+            ),
+            reason=(
+                f"the lens is graded on {CORPUS_NAME} (sha256 {corpus_sha256[:16]}…), a corpus this repo "
+                "authored, and detects it with five regexes over atom text; a producer grading itself on "
+                "its own examples is not evidence, so the number is reported but not claimed"
+            ),
+            rerun="drop a third-party UI-drift set at $ELY_EYE_HOME/contradiction_scenarios.json (ely-eye contradiction-corpus writes the default), then ely-eye visual-contradiction-proof",
+        )
+

backend/src/ely_eye/proofs/contradiction_scenarios.json

@@ -0,0 +1,927 @@
+[
+  {
+    "name": "button-radius-token-drift",
+    "expect_drift": true,
+    "atoms": [
+      {
+        "atom_id": "home_btn_v4",
+        "modality": "ui_screenshot",
+        "source_id": "home_v4",
+        "source": "home_v4.figma",
+        "ingested_at": "2026-10-07T22:07:05.953200Z",
+        "text": "button.radius: 8px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 4,
+        "metadata": {
+          "source_kind": "figma"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "checkout_btn_v4",
+        "modality": "ui_screenshot",
+        "source_id": "checkout_v4",
+        "source": "checkout_v4.figma",
+        "ingested_at": "2026-10-07T22:07:05.953230Z",
+        "text": "button.radius: 12px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 4,
+        "metadata": {
+          "source_kind": "figma"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "primary-color-token-drift",
+    "expect_drift": true,
+    "atoms": [
+      {
+        "atom_id": "spec_color",
+        "modality": "pdf_page",
+        "source_id": "brand_spec",
+        "source": "brand_guide.pdf",
+        "ingested_at": "2026-10-07T22:07:05.953237Z",
+        "text": "primary.color: #2563eb",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 5,
+        "metadata": {
+          "source_kind": "spec"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "impl_color",
+        "modality": "code",
+        "source_id": "theme_css",
+        "source": "src/theme.css",
+        "ingested_at": "2026-10-07T22:07:05.953241Z",
+        "text": "primary.color: #1d4ed8",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 5,
+        "metadata": {
+          "source_kind": "css"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "heading-typography-drift",
+    "expect_drift": true,
+    "atoms": [
+      {
+        "atom_id": "home_heading",
+        "modality": "ui_screenshot",
+        "source_id": "home_shot",
+        "source": "home_v3.png",
+        "ingested_at": "2026-10-07T22:07:05.953245Z",
+        "text": "heading.font-size: 24px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 5,
+        "metadata": {
+          "source_kind": "ui_screenshot"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "checkout_heading",
+        "modality": "ui_screenshot",
+        "source_id": "checkout_shot",
+        "source": "checkout_v3.png",
+        "ingested_at": "2026-10-07T22:07:05.953248Z",
+        "text": "heading.font-size: 28px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 5,
+        "metadata": {
+          "source_kind": "ui_screenshot"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "card-spacing-layout-drift",
+    "expect_drift": true,
+    "atoms": [
+      {
+        "atom_id": "card_home",
+        "modality": "ui_screenshot",
+        "source_id": "card_home_shot",
+        "source": "home_v3.png",
+        "ingested_at": "2026-10-07T22:07:05.953252Z",
+        "text": "card.padding: 16px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 4,
+        "metadata": {
+          "source_kind": "ui_screenshot"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "card_settings",
+        "modality": "ui_screenshot",
+        "source_id": "card_settings_shot",
+        "source": "settings_v3.png",
+        "ingested_at": "2026-10-07T22:07:05.953255Z",
+        "text": "card.padding: 24px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 4,
+        "metadata": {
+          "source_kind": "ui_screenshot"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "cta-copy-drift",
+    "expect_drift": true,
+    "atoms": [
+      {
+        "atom_id": "cta_v1",
+        "modality": "ui_screenshot",
+        "source_id": "shot_v1",
+        "source": "cart_v1.png",
+        "ingested_at": "2026-10-07T22:07:05.953258Z",
+        "text": "cta.label: \"Checkout\"",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 5,
+        "metadata": {
+          "source_kind": "ui_screenshot"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "cta_v4",
+        "modality": "ui_screenshot",
+        "source_id": "shot_v4",
+        "source": "cart_v4.png",
+        "ingested_at": "2026-10-07T22:07:05.953263Z",
+        "text": "cta.label: \"Pay now\"",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 5,
+        "metadata": {
+          "source_kind": "ui_screenshot"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "banner-radius-temporal-drift",
+    "expect_drift": true,
+    "atoms": [
+      {
+        "atom_id": "banner_v1",
+        "modality": "ui_screenshot",
+        "source_id": "banner_spec_v1",
+        "source": "banner_v1.figma",
+        "ingested_at": "2026-10-07T22:07:05.953267Z",
+        "text": "banner.radius: 4px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 4,
+        "metadata": {
+          "source_kind": "figma",
+          "version": "v1"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "banner_v4",
+        "modality": "ui_screenshot",
+        "source_id": "banner_spec_v4",
+        "source": "banner_v4.figma",
+        "ingested_at": "2026-10-07T22:07:05.953271Z",
+        "text": "banner.radius: 10px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 4,
+        "metadata": {
+          "source_kind": "figma",
+          "version": "v4"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "toggle-accent-visual-code-drift",
+    "expect_drift": true,
+    "atoms": [
+      {
+        "atom_id": "toggle_design",
+        "modality": "ui_screenshot",
+        "source_id": "toggle_figma",
+        "source": "toggle.figma",
+        "ingested_at": "2026-10-07T22:07:05.953274Z",
+        "text": "toggle.color: #10b981",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 5,
+        "metadata": {
+          "source_kind": "figma"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "toggle_code",
+        "modality": "code",
+        "source_id": "toggle_css",
+        "source": "src/toggle.css",
+        "ingested_at": "2026-10-07T22:07:05.953279Z",
+        "text": "toggle.color: #059669",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 5,
+        "metadata": {
+          "source_kind": "css"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "consistent-button-radius",
+    "expect_drift": false,
+    "atoms": [
+      {
+        "atom_id": "ctrl_btn_spec",
+        "modality": "pdf_page",
+        "source_id": "ctrl_spec",
+        "source": "spec.pdf",
+        "ingested_at": "2026-10-07T22:07:05.953282Z",
+        "text": "button.radius: 8px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 4,
+        "metadata": {
+          "source_kind": "spec"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "ctrl_btn_css",
+        "modality": "code",
+        "source_id": "ctrl_css",
+        "source": "src/tokens.css",
+        "ingested_at": "2026-10-07T22:07:05.953284Z",
+        "text": "button.radius = 8px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 4,
+        "metadata": {
+          "source_kind": "css"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "distinct-components-not-a-conflict",
+    "expect_drift": false,
+    "atoms": [
+      {
+        "atom_id": "ctrl_button_radius",
+        "modality": "code",
+        "source_id": "ctrl_a",
+        "source": "a.css",
+        "ingested_at": "2026-10-07T22:07:05.953288Z",
+        "text": "button.radius: 8px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 4,
+        "metadata": {
+          "source_kind": "css"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "ctrl_input_radius",
+        "modality": "code",
+        "source_id": "ctrl_b",
+        "source": "b.css",
+        "ingested_at": "2026-10-07T22:07:05.953293Z",
+        "text": "input.radius: 12px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 4,
+        "metadata": {
+          "source_kind": "css"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "single-source-multiple-rules",
+    "expect_drift": false,
+    "atoms": [
+      {
+        "atom_id": "ctrl_single_css",
+        "modality": "code",
+        "source_id": "ctrl_single",
+        "source": "src/base.css",
+        "ingested_at": "2026-10-07T22:07:05.953297Z",
+        "text": "header.padding: 8px; footer.padding: 24px; aside.padding: 16px",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 15,
+        "metadata": {
+          "source_kind": "css"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "consistent-primary-color",
+    "expect_drift": false,
+    "atoms": [
+      {
+        "atom_id": "ctrl_color_spec",
+        "modality": "pdf_page",
+        "source_id": "ctrl_color_a",
+        "source": "brand.pdf",
+        "ingested_at": "2026-10-07T22:07:05.953300Z",
+        "text": "primary.color: #2563eb",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 5,
+        "metadata": {
+          "source_kind": "spec"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "ctrl_color_css",
+        "modality": "code",
+        "source_id": "ctrl_color_b",
+        "source": "theme.css",
+        "ingested_at": "2026-10-07T22:07:05.953302Z",
+        "text": "primary.color: #2563eb",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 5,
+        "metadata": {
+          "source_kind": "css"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  },
+  {
+    "name": "prose-without-design-facts",
+    "expect_drift": false,
+    "atoms": [
+      {
+        "atom_id": "ctrl_prose_a",
+        "modality": "text",
+        "source_id": "ctrl_prose_1",
+        "source": "notes_a.txt",
+        "ingested_at": "2026-10-07T22:07:05.953306Z",
+        "text": "The onboarding flow feels smooth and the colors look balanced.",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 15,
+        "metadata": {
+          "source_kind": "text"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      },
+      {
+        "atom_id": "ctrl_prose_b",
+        "modality": "text",
+        "source_id": "ctrl_prose_2",
+        "source": "notes_b.txt",
+        "ingested_at": "2026-10-07T22:07:05.953308Z",
+        "text": "Users reported the checkout button is easy to find.",
+        "image_ref": null,
+        "embedding_refs": [],
+        "layout": null,
+        "relations": [],
+        "trust": {
+          "parser": 0.95,
+          "ocr": 0.9,
+          "model": 0.85
+        },
+        "token_equivalent": 12,
+        "metadata": {
+          "source_kind": "text"
+        },
+        "holder_id": "",
+        "origin": "artifact",
+        "scenario": "",
+        "event_time": null,
+        "event_time_end": null,
+        "event_time_precision": "unknown",
+        "event_time_source": "",
+        "valid_from": null,
+        "valid_to": null,
+        "supersedes": [],
+        "visibility": {
+          "scope": "holder",
+          "holders": [],
+          "purpose": null
+        },
+        "retention_until": null
+      }
+    ]
+  }
+]

backend/src/ely_eye/proofs/contradictions.py

@@ -0,0 +1,235 @@
+"""Contradiction scenarios and the lens that has to spot them.
+
+The scenarios are data, not code: the proof loads them from a JSON corpus so the set that grades
+the lens is reviewable, replaceable, and named by hash in the proof artifact.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from pathlib import Path
+from typing import Any
+
+from ..schemas import (
+    CompiledContext,
+    ContextPlan,
+    EvidenceAtom,
+    Modality,
+    RetrievalHit,
+    RuntimeProfile,
+    TrustScores,
+)
+
+CORPUS_NAME = "contradiction_scenarios.json"
+PACKAGED_CORPUS = Path(__file__).with_name(CORPUS_NAME)
+
+
+def contradiction_atom(
+    atom_id: str,
+    source_id: str,
+    source: str,
+    text: str,
+    modality: Modality,
+    source_kind: str,
+    version: str | None = None,
+) -> EvidenceAtom:
+    metadata: dict[str, Any] = {"source_kind": source_kind}
+    if version is not None:
+        metadata["version"] = version
+    return EvidenceAtom(
+        atom_id=atom_id,
+        modality=modality,
+        source_id=source_id,
+        source=source,
+        text=text,
+        trust=TrustScores(parser=0.95, ocr=0.9, model=0.85),
+        token_equivalent=max(1, len(text) // 4),
+        metadata=metadata,
+    )
+
+
+def contradiction_context(question: str, atoms: list[EvidenceAtom]) -> CompiledContext:
+    hits = [
+        RetrievalHit(atom=atom, sparse_score=0.8, dense_score=0.0, graph_score=0.0, final_score=0.8)
+        for atom in atoms
+    ]
+    plan = ContextPlan(
+        question=question,
+        profile=RuntimeProfile.research_theater,
+        required_modalities=sorted({atom.modality for atom in atoms}),
+        evidence_budget_atoms=len(atoms),
+        token_budget=sum(atom.token_equivalent for atom in atoms),
+        retrieval_query=question,
+        compression_strategy="visual-contradiction-eval",
+        verifier_contract=["Report contradictions as contradiction_notes."],
+    )
+    packed = "\n".join(f"[{atom.atom_id}] {atom.text}" for atom in atoms)
+    return CompiledContext(
+        plan=plan,
+        hits=hits,
+        packed_text=packed,
+        token_equivalent=sum(atom.token_equivalent for atom in atoms),
+        cache_trace_id="visual_contradiction_eval",
+    )
+
+
+def build_contradiction_scenarios() -> list[dict[str, Any]]:
+    """Deterministic cross-version conflict and control samples (PRD 11.15).
+
+    Each conflict pair injects a single genuine drift across two distinct
+    sources; each control pair is internally consistent or scoped to different
+    components. The samples exercise the live ``VisualContradictionLens`` so the
+    proof measures real recall and false-positive rate.
+
+    A scenario carries its atoms and its label; the compiled context the lens reads is built at
+    use, so the corpus on disk holds no derived state.
+    """
+
+    scenarios: list[dict[str, Any]] = []
+
+    def conflict(name: str, atoms: list[EvidenceAtom]) -> None:
+        scenarios.append({"name": name, "atoms": atoms, "expect_drift": True})
+
+    def control(name: str, atoms: list[EvidenceAtom]) -> None:
+        scenarios.append({"name": name, "atoms": atoms, "expect_drift": False})
+
+    conflict(
+        "button-radius-token-drift",
+        [
+            contradiction_atom("home_btn_v4", "home_v4", "home_v4.figma", "button.radius: 8px", Modality.ui_screenshot, "figma"),
+            contradiction_atom("checkout_btn_v4", "checkout_v4", "checkout_v4.figma", "button.radius: 12px", Modality.ui_screenshot, "figma"),
+        ],
+    )
+    conflict(
+        "primary-color-token-drift",
+        [
+            contradiction_atom("spec_color", "brand_spec", "brand_guide.pdf", "primary.color: #2563eb", Modality.pdf_page, "spec"),
+            contradiction_atom("impl_color", "theme_css", "src/theme.css", "primary.color: #1d4ed8", Modality.code, "css"),
+        ],
+    )
+    conflict(
+        "heading-typography-drift",
+        [
+            contradiction_atom("home_heading", "home_shot", "home_v3.png", "heading.font-size: 24px", Modality.ui_screenshot, "ui_screenshot"),
+            contradiction_atom("checkout_heading", "checkout_shot", "checkout_v3.png", "heading.font-size: 28px", Modality.ui_screenshot, "ui_screenshot"),
+        ],
+    )
+    conflict(
+        "card-spacing-layout-drift",
+        [
+            contradiction_atom("card_home", "card_home_shot", "home_v3.png", "card.padding: 16px", Modality.ui_screenshot, "ui_screenshot"),
+            contradiction_atom("card_settings", "card_settings_shot", "settings_v3.png", "card.padding: 24px", Modality.ui_screenshot, "ui_screenshot"),
+        ],
+    )
+    conflict(
+        "cta-copy-drift",
+        [
+            contradiction_atom("cta_v1", "shot_v1", "cart_v1.png", 'cta.label: "Checkout"', Modality.ui_screenshot, "ui_screenshot"),
+            contradiction_atom("cta_v4", "shot_v4", "cart_v4.png", 'cta.label: "Pay now"', Modality.ui_screenshot, "ui_screenshot"),
+        ],
+    )
+    conflict(
+        "banner-radius-temporal-drift",
+        [
+            contradiction_atom("banner_v1", "banner_spec_v1", "banner_v1.figma", "banner.radius: 4px", Modality.ui_screenshot, "figma", version="v1"),
+            contradiction_atom("banner_v4", "banner_spec_v4", "banner_v4.figma", "banner.radius: 10px", Modality.ui_screenshot, "figma", version="v4"),
+        ],
+    )
+    conflict(
+        "toggle-accent-visual-code-drift",
+        [
+            contradiction_atom("toggle_design", "toggle_figma", "toggle.figma", "toggle.color: #10b981", Modality.ui_screenshot, "figma"),
+            contradiction_atom("toggle_code", "toggle_css", "src/toggle.css", "toggle.color: #059669", Modality.code, "css"),
+        ],
+    )
+
+    control(
+        "consistent-button-radius",
+        [
+            contradiction_atom("ctrl_btn_spec", "ctrl_spec", "spec.pdf", "button.radius: 8px", Modality.pdf_page, "spec"),
+            contradiction_atom("ctrl_btn_css", "ctrl_css", "src/tokens.css", "button.radius = 8px", Modality.code, "css"),
+        ],
+    )
+    control(
+        "distinct-components-not-a-conflict",
+        [
+            contradiction_atom("ctrl_button_radius", "ctrl_a", "a.css", "button.radius: 8px", Modality.code, "css"),
+            contradiction_atom("ctrl_input_radius", "ctrl_b", "b.css", "input.radius: 12px", Modality.code, "css"),
+        ],
+    )
+    control(
+        "single-source-multiple-rules",
+        [
+            contradiction_atom(
+                "ctrl_single_css",
+                "ctrl_single",
+                "src/base.css",
+                "header.padding: 8px; footer.padding: 24px; aside.padding: 16px",
+                Modality.code,
+                "css",
+            ),
+        ],
+    )
+    control(
+        "consistent-primary-color",
+        [
+            contradiction_atom("ctrl_color_spec", "ctrl_color_a", "brand.pdf", "primary.color: #2563eb", Modality.pdf_page, "spec"),
+            contradiction_atom("ctrl_color_css", "ctrl_color_b", "theme.css", "primary.color: #2563eb", Modality.code, "css"),
+        ],
+    )
+    control(
+        "prose-without-design-facts",
+        [
+            contradiction_atom("ctrl_prose_a", "ctrl_prose_1", "notes_a.txt", "The onboarding flow feels smooth and the colors look balanced.", Modality.text, "text"),
+            contradiction_atom("ctrl_prose_b", "ctrl_prose_2", "notes_b.txt", "Users reported the checkout button is easy to find.", Modality.text, "text"),
+        ],
+    )
+
+    return scenarios
+
+
+def corpus_path(data_dir: Path) -> Path | None:
+    """A corpus the operator dropped in beats the packaged one; neither beats silence."""
+    override = data_dir / CORPUS_NAME
+    if override.exists():
+        return override
+    return PACKAGED_CORPUS if PACKAGED_CORPUS.exists() else None
+
+
+def write_corpus(path: Path) -> Path:
+    """Freeze the generated scenarios as data so the grader can be read, not just run."""
+    payload = [
+        {"name": scenario["name"], "expect_drift": scenario["expect_drift"], "atoms": [atom.model_dump(mode="json") for atom in scenario["atoms"]]}
+        for scenario in build_contradiction_scenarios()
+    ]
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
+    return path
+
+
+def load_corpus(path: Path) -> tuple[list[dict[str, Any]], str]:
+    """Read the corpus and its hash. A scenario whose label cannot be parsed fails loudly."""
+    try:
+        raw = path.read_bytes()
+        rows = json.loads(raw)
+    except (OSError, json.JSONDecodeError) as error:
+        raise ValueError(f"{path}: the contradiction corpus is unreadable ({error})") from error
+    if not isinstance(rows, list) or not rows:
+        raise ValueError(f"{path}: the contradiction corpus must be a non-empty list")
+    scenarios: list[dict[str, Any]] = []
+    for row in rows:
+        if not isinstance(row, dict) or "name" not in row or "expect_drift" not in row:
+            raise ValueError(f"{path}: a contradiction scenario is missing its name or its label")
+        try:
+            atoms = [EvidenceAtom.model_validate(entry) for entry in row.get("atoms", [])]
+        except ValueError as error:
+            raise ValueError(f"{path}: scenario {row['name']!r} holds an unreadable atom ({error})") from error
+        if not atoms:
+            raise ValueError(f"{path}: scenario {row['name']!r} has no atoms to compare")
+        scenarios.append({"name": row["name"], "expect_drift": bool(row["expect_drift"]), "atoms": atoms})
+    return scenarios, hashlib.sha256(raw).hexdigest()
+
+
+def scenario_context(scenario: dict[str, Any]) -> CompiledContext:
+    return contradiction_context(str(scenario["name"]), scenario["atoms"])

backend/src/ely_eye/proofs/hashhop.py

@@ -0,0 +1,430 @@
+"""Build HashHop puzzles and ask a model to walk them.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import random
+import secrets
+from pathlib import Path
+from typing import Any
+
+from ..config import Settings, get_settings
+from ..evaluation_schemas import (
+    HashHopProof,
+)
+from ..runtime import QwenRuntime
+from .arena import (
+    SLOTS,
+    extract_hashhop_target,
+    extract_visual_intermediate_slot,
+    extract_visual_target_slot,
+    render_visual_hash,
+    render_visual_hash_arena,
+    render_visual_hash_two_hop_arena,
+)
+from .report import write_json
+
+
+class HashHopEvaluator:
+    def __init__(self, settings: Settings | None = None, runtime: QwenRuntime | None = None) -> None:
+        self.settings = settings or get_settings()
+        self.runtime = runtime or QwenRuntime(self.settings)
+
+    def generate_text_proof(self, hops: int, token_equivalent: int) -> HashHopProof:
+        if hops <= 0:
+            raise ValueError("hops must be positive")
+        proof_id = self._proof_id("text", hops, token_equivalent)
+        root = self.settings.data_dir / "eval_proofs" / proof_id
+        root.mkdir(parents=True, exist_ok=True)
+        chain = [secrets.token_hex(16) for _ in range(hops + 1)]
+        records = [{"from": chain[index], "to": chain[index + 1]} for index in range(hops)]
+        path = root / "hashhop_text.jsonl"
+        path.write_text("\n".join(json.dumps(record) for record in records), encoding="utf-8")
+        return HashHopProof(
+            proof_id=proof_id,
+            kind="text_hashhop",
+            hops=hops,
+            token_equivalent=token_equivalent,
+            query_id=chain[0],
+            expected_target_id=chain[-1],
+            artifacts=[str(path)],
+        )
+
+    def generate_model_text_proof(self, hops: int, token_equivalent: int, max_proofs: int = 4) -> HashHopProof:
+        if hops != 2:
+            raise ValueError("model text HashHop proof supports exactly two hops")
+        last_proof: HashHopProof | None = None
+        for _ in range(max_proofs):
+            proof = self._generate_model_text_proof_once(hops, token_equivalent)
+            if proof.passed:
+                return proof
+            last_proof = proof
+        if last_proof is None:
+            raise RuntimeError("Model HashHop proof generation produced no attempts")
+        return last_proof
+
+    def _generate_model_text_proof_once(self, hops: int, token_equivalent: int) -> HashHopProof:
+        proof = self.generate_text_proof(hops, token_equivalent)
+        root = Path(proof.artifacts[0]).parent
+        chain_path = root / "hashhop_text.jsonl"
+        try:
+            chain_text = chain_path.read_text(encoding="utf-8")
+        except OSError as error:
+            raise ValueError(f"{chain_path}: the HashHop chain is unreadable ({error})") from error
+        try:
+            records = [json.loads(line) for line in chain_text.splitlines() if line.strip()]
+        except json.JSONDecodeError as error:
+            raise ValueError(f"{chain_path}: the HashHop chain holds a malformed line ({error})") from error
+        decoys = [
+            {"from": secrets.token_hex(16), "to": secrets.token_hex(16)}
+            for _ in range(max(4, hops * 2))
+        ]
+        shuffled = records + decoys
+        random.Random(proof.proof_id).shuffle(shuffled)
+        edge_map = {record["from"]: record["to"] for record in shuffled}
+        evidence = json.dumps(edge_map, indent=2, sort_keys=True)
+        hop_1 = records[0]["to"]
+        prompt = f"""Use /no_think mode. Return compact JSON only.
+
+Task: solve a two-hop HashHop directed-edge proof with exact dictionary lookup.
+
+Example:
+edge_map = {{"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa": "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb", "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb": "cccccccccccccccccccccccccccccccc"}}
+start_id = "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"
+hop_1 = edge_map[start_id]
+answer = edge_map[hop_1]
+JSON = {{"hop_1":"bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb","answer":"cccccccccccccccccccccccccccccccc"}}
+
+Current proof:
+token_equivalent = {token_equivalent}
+start_id = "{proof.query_id}"
+hops = {hops}
+edge_map = {evidence}
+
+Evaluate:
+hop_1 = edge_map["{proof.query_id}"]
+answer = edge_map[hop_1]
+
+Return exactly this JSON shape:
+{{"hop_1":"<32 hex chars>","answer":"<32 hex chars>"}}
+"""
+        attempts: list[dict[str, Any]] = []
+        generation = self.runtime.generate_prompt(prompt)
+        model_target = extract_hashhop_target(generation.text)
+        attempts.append(
+            {
+                "step": "two_hop_expression",
+                "raw_generation": generation.text,
+                "parsed_target": model_target,
+                "expected_target": proof.expected_target_id,
+                "passed": model_target == proof.expected_target_id,
+                "input_tokens": generation.input_tokens,
+                "output_tokens": generation.output_tokens,
+            }
+        )
+        if model_target != proof.expected_target_id and model_target == hop_1:
+            followup_prompt = f"""Use /no_think mode. Return compact JSON only.
+
+Task: continue a HashHop proof from a model-derived first hop.
+
+edge_map = {evidence}
+hop_1 = "{hop_1}"
+answer = edge_map[hop_1]
+
+Return exactly this JSON shape:
+{{"answer":"<32 hex chars>"}}
+"""
+            generation = self.runtime.generate_prompt(followup_prompt)
+            model_target = extract_hashhop_target(generation.text)
+            attempts.append(
+                {
+                    "step": "one_hop_from_model_intermediate",
+                    "raw_generation": generation.text,
+                    "parsed_target": model_target,
+                    "expected_target": proof.expected_target_id,
+                    "passed": model_target == proof.expected_target_id,
+                    "input_tokens": generation.input_tokens,
+                    "output_tokens": generation.output_tokens,
+                }
+            )
+        proof.model_target_id = model_target
+        proof.passed = model_target == proof.expected_target_id
+        artifact = root / "model_text_hashhop.json"
+        proof.artifacts.append(str(artifact))
+        write_json(
+            artifact,
+            {
+                "proof": proof.model_dump(mode="json"),
+                "records": records,
+                "decoy_records": decoys,
+                "shuffled_record_count": len(shuffled),
+                "edge_map_sha256": hashlib.sha256(evidence.encode("utf-8")).hexdigest(),
+                "model_attempts": attempts,
+                "raw_generation": generation.text,
+                "parsed_target": model_target,
+                "expected_target": proof.expected_target_id,
+                "passed": proof.passed,
+                "runtime": {
+                    "backend": generation.backend,
+                    "adapter_path": generation.adapter_path,
+                    "adapter_kind": generation.adapter_kind,
+                    "adapter_id": generation.adapter_id,
+                    "adapter_sha256": generation.adapter_sha256,
+                    "input_tokens": generation.input_tokens,
+                    "output_tokens": generation.output_tokens,
+                },
+            },
+        )
+        (root / "proof.json").write_text(proof.model_dump_json(indent=2), encoding="utf-8")
+        return proof
+
+    def generate_model_visual_proof(self, hops: int, token_equivalent: int, max_proofs: int = 3) -> HashHopProof:
+        if hops not in {1, 2}:
+            raise ValueError("model visual HashHop proof supports one or two hops")
+        last_proof: HashHopProof | None = None
+        for _ in range(max_proofs):
+            if hops == 1:
+                proof = self._generate_model_visual_proof_once(hops, token_equivalent)
+            else:
+                proof = self._generate_model_visual_two_hop_proof_once(token_equivalent)
+            if proof.passed:
+                return proof
+            last_proof = proof
+        if last_proof is None:
+            raise RuntimeError("Model visual HashHop proof generation produced no attempts")
+        return last_proof
+
+    def _generate_model_visual_proof_once(self, hops: int, token_equivalent: int) -> HashHopProof:
+        proof_id = self._proof_id("visual-model", hops, token_equivalent)
+        root = self.settings.data_dir / "eval_proofs" / proof_id
+        image_dir = root / "visual_blocks"
+        image_dir.mkdir(parents=True, exist_ok=True)
+        source_id = secrets.token_hex(16)
+        target_id = secrets.token_hex(16)
+        decoy_ids = [secrets.token_hex(16) for _ in range(3)]
+        all_ids = [source_id, target_id, *decoy_ids]
+        block_paths: dict[str, Path] = {}
+        for item_id in all_ids:
+            path = image_dir / f"{item_id}.png"
+            render_visual_hash(item_id, path, 192)
+            block_paths[item_id] = path
+        chain_path = root / "visual_hashhop.jsonl"
+        chain_path.write_text(json.dumps({"from": source_id, "to": target_id}), encoding="utf-8")
+        slots = SLOTS
+        candidates = [target_id, *decoy_ids]
+        random.Random(proof_id).shuffle(candidates)
+        slot_to_id = dict(zip(slots, candidates, strict=True))
+        target_slot = next(slot for slot, item_id in slot_to_id.items() if item_id == target_id)
+        arena_path = root / "visual_hashhop_arena.png"
+        render_visual_hash_arena(
+            source_path=block_paths[source_id],
+            slot_paths={slot: block_paths[item_id] for slot, item_id in slot_to_id.items()},
+            output_path=arena_path,
+        )
+        proof = HashHopProof(
+            proof_id=proof_id,
+            kind="visual_hashhop",
+            hops=hops,
+            token_equivalent=token_equivalent,
+            query_id=source_id,
+            expected_target_id=target_id,
+            artifacts=[*(str(path) for path in block_paths.values()), str(chain_path), str(arena_path)],
+        )
+        prompt = """Use /no_think mode. Return compact JSON only.
+
+The attached image is a Visual HashHop arena.
+The large visual hash block on the left is the source.
+Four candidate visual hash blocks are arranged on the right.
+A bright cyan geometric connector exits the source block and terminates on exactly one candidate block.
+Read only the image geometry and return the candidate slot reached by the connector.
+Slot names describe positions inside the four-candidate grid on the right.
+
+Allowed target_slot values:
+top-left, top-right, bottom-left, bottom-right
+
+Return exactly this JSON shape:
+{"target_slot":"<allowed value>"}
+"""
+        generation = self.runtime.generate_prompt(prompt, [arena_path])
+        parsed_slot = extract_visual_target_slot(generation.text)
+        model_target = slot_to_id.get(parsed_slot) if parsed_slot else None
+        proof.model_target_id = model_target
+        proof.passed = model_target == proof.expected_target_id
+        artifact = root / "model_visual_hashhop.json"
+        proof.artifacts.append(str(artifact))
+        write_json(
+            artifact,
+            {
+                "proof": proof.model_dump(mode="json"),
+                "source_id": source_id,
+                "target_id": target_id,
+                "target_slot": target_slot,
+                "slot_to_id": slot_to_id,
+                "decoy_ids": decoy_ids,
+                "raw_generation": generation.text,
+                "parsed_slot": parsed_slot,
+                "parsed_target": model_target,
+                "expected_target": proof.expected_target_id,
+                "passed": proof.passed,
+                "runtime": {
+                    "backend": generation.backend,
+                    "adapter_path": generation.adapter_path,
+                    "adapter_kind": generation.adapter_kind,
+                    "adapter_id": generation.adapter_id,
+                    "adapter_sha256": generation.adapter_sha256,
+                    "input_tokens": generation.input_tokens,
+                    "output_tokens": generation.output_tokens,
+                },
+            },
+        )
+        (root / "proof.json").write_text(proof.model_dump_json(indent=2), encoding="utf-8")
+        return proof
+
+    def _generate_model_visual_two_hop_proof_once(self, token_equivalent: int) -> HashHopProof:
+        hops = 2
+        proof_id = self._proof_id("visual-model", hops, token_equivalent)
+        root = self.settings.data_dir / "eval_proofs" / proof_id
+        image_dir = root / "visual_blocks"
+        image_dir.mkdir(parents=True, exist_ok=True)
+        source_id = secrets.token_hex(16)
+        intermediate_id = secrets.token_hex(16)
+        target_id = secrets.token_hex(16)
+        intermediate_decoys = [secrets.token_hex(16) for _ in range(3)]
+        target_decoys = [secrets.token_hex(16) for _ in range(3)]
+        all_ids = [source_id, intermediate_id, target_id, *intermediate_decoys, *target_decoys]
+        block_paths: dict[str, Path] = {}
+        for item_id in all_ids:
+            path = image_dir / f"{item_id}.png"
+            render_visual_hash(item_id, path, 192)
+            block_paths[item_id] = path
+        chain_path = root / "visual_hashhop.jsonl"
+        chain_path.write_text(
+            "\n".join(
+                [
+                    json.dumps({"from": source_id, "to": intermediate_id}),
+                    json.dumps({"from": intermediate_id, "to": target_id}),
+                ]
+            ),
+            encoding="utf-8",
+        )
+        slots = SLOTS
+        intermediate_candidates = [intermediate_id, *intermediate_decoys]
+        target_candidates = [target_id, *target_decoys]
+        random.Random(f"{proof_id}:intermediate").shuffle(intermediate_candidates)
+        random.Random(f"{proof_id}:target").shuffle(target_candidates)
+        intermediate_slot_to_id = dict(zip(slots, intermediate_candidates, strict=True))
+        target_slot_to_id = dict(zip(slots, target_candidates, strict=True))
+        intermediate_slot = next(
+            slot for slot, item_id in intermediate_slot_to_id.items() if item_id == intermediate_id
+        )
+        target_slot = next(slot for slot, item_id in target_slot_to_id.items() if item_id == target_id)
+        arena_path = root / "visual_hashhop_two_hop_arena.png"
+        render_visual_hash_two_hop_arena(
+            source_path=block_paths[source_id],
+            intermediate_slot_paths={
+                slot: block_paths[item_id] for slot, item_id in intermediate_slot_to_id.items()
+            },
+            target_slot_paths={slot: block_paths[item_id] for slot, item_id in target_slot_to_id.items()},
+            output_path=arena_path,
+        )
+        proof = HashHopProof(
+            proof_id=proof_id,
+            kind="visual_hashhop",
+            hops=hops,
+            token_equivalent=token_equivalent,
+            query_id=source_id,
+            expected_target_id=target_id,
+            artifacts=[*(str(path) for path in block_paths.values()), str(chain_path), str(arena_path)],
+        )
+        prompt = """Use /no_think mode. Return compact JSON only.
+
+The attached image is a two-hop Visual HashHop arena.
+The source visual hash block is on the left.
+Four intermediate candidate visual hash blocks are in the center.
+Four final candidate visual hash blocks are on the right.
+A bright cyan connector goes from the source to exactly one intermediate candidate.
+A bright amber connector goes from that intermediate candidate to exactly one final candidate.
+Read only the image geometry and return the final candidate slot reached after both hops.
+The intermediate_slot value describes the center grid. The target_slot value describes the rightmost final-candidate grid.
+
+Allowed slot values:
+top-left, top-right, bottom-left, bottom-right
+
+Return exactly this JSON shape:
+{"intermediate_slot":"<allowed value>","target_slot":"<allowed value>"}
+"""
+        generation = self.runtime.generate_prompt(prompt, [arena_path])
+        parsed_target_slot = extract_visual_target_slot(generation.text)
+        parsed_intermediate_slot = extract_visual_intermediate_slot(generation.text)
+        model_target = target_slot_to_id.get(parsed_target_slot) if parsed_target_slot else None
+        proof.model_target_id = model_target
+        proof.passed = model_target == proof.expected_target_id
+        artifact = root / "model_visual_hashhop.json"
+        proof.artifacts.append(str(artifact))
+        write_json(
+            artifact,
+            {
+                "proof": proof.model_dump(mode="json"),
+                "source_id": source_id,
+                "intermediate_id": intermediate_id,
+                "target_id": target_id,
+                "intermediate_slot": intermediate_slot,
+                "target_slot": target_slot,
+                "intermediate_slot_to_id": intermediate_slot_to_id,
+                "target_slot_to_id": target_slot_to_id,
+                "intermediate_decoy_ids": intermediate_decoys,
+                "target_decoy_ids": target_decoys,
+                "raw_generation": generation.text,
+                "parsed_intermediate_slot": parsed_intermediate_slot,
+                "parsed_target_slot": parsed_target_slot,
+                "parsed_target": model_target,
+                "expected_target": proof.expected_target_id,
+                "passed": proof.passed,
+                "runtime": {
+                    "backend": generation.backend,
+                    "adapter_path": generation.adapter_path,
+                    "adapter_kind": generation.adapter_kind,
+                    "adapter_id": generation.adapter_id,
+                    "adapter_sha256": generation.adapter_sha256,
+                    "input_tokens": generation.input_tokens,
+                    "output_tokens": generation.output_tokens,
+                },
+            },
+        )
+        (root / "proof.json").write_text(proof.model_dump_json(indent=2), encoding="utf-8")
+        return proof
+
+    def generate_visual_proof(self, hops: int, token_equivalent: int, tile_size: int = 256) -> HashHopProof:
+        if hops <= 0:
+            raise ValueError("hops must be positive")
+        proof_id = self._proof_id("visual", hops, token_equivalent)
+        root = self.settings.data_dir / "eval_proofs" / proof_id
+        image_dir = root / "visual_blocks"
+        image_dir.mkdir(parents=True, exist_ok=True)
+        ids = [secrets.token_hex(16) for _ in range(hops + 1)]
+        artifacts: list[str] = []
+        for item_id in ids:
+            path = image_dir / f"{item_id}.png"
+            render_visual_hash(item_id, path, tile_size)
+            artifacts.append(str(path))
+        chain_path = root / "visual_hashhop.jsonl"
+        chain_path.write_text(
+            "\n".join(json.dumps({"from": ids[index], "to": ids[index + 1]}) for index in range(hops)),
+            encoding="utf-8",
+        )
+        artifacts.append(str(chain_path))
+        return HashHopProof(
+            proof_id=proof_id,
+            kind="visual_hashhop",
+            hops=hops,
+            token_equivalent=token_equivalent,
+            query_id=ids[0],
+            expected_target_id=ids[-1],
+            artifacts=artifacts,
+        )
+
+    def _proof_id(self, kind: str, hops: int, token_equivalent: int) -> str:
+        raw = f"{kind}:{hops}:{token_equivalent}:{secrets.token_hex(8)}"
+        return hashlib.sha256(raw.encode("utf-8")).hexdigest()[:16]

backend/src/ely_eye/proofs/hashhop_verify.py

@@ -0,0 +1,73 @@
+"""Judge a HashHop proof: chain, answer, artifacts, pass rate.
+"""
+
+from __future__ import annotations
+
+import json
+import shutil
+from pathlib import Path
+
+from ..evaluation_schemas import (
+    HashHopProof,
+)
+from .hashhop import HashHopEvaluator
+
+
+def hashhop_chain_valid(proof: HashHopProof) -> bool:
+    chain_artifacts = [Path(path) for path in proof.artifacts if path.endswith(".jsonl")]
+    if len(chain_artifacts) != 1 or not chain_artifacts[0].exists():
+        return False
+    try:
+        rows = [
+            json.loads(line)
+            for line in chain_artifacts[0].read_text(encoding="utf-8").splitlines()
+            if line.strip()
+        ]
+    except (OSError, json.JSONDecodeError):
+        return False  # an unreadable chain is not a valid chain, and this is a predicate
+    if len(rows) != proof.hops:
+        return False
+    if rows[0]["from"] != proof.query_id or rows[-1]["to"] != proof.expected_target_id:
+        return False
+    for index in range(1, len(rows)):
+        if rows[index - 1]["to"] != rows[index]["from"]:
+            return False
+    return all(Path(path).exists() for path in proof.artifacts)
+
+
+def copy_proof_artifacts(proof: HashHopProof, target: Path) -> None:
+    target.mkdir(parents=True, exist_ok=True)
+    source_roots = {Path(path).parent for path in proof.artifacts}
+    for source_root in source_roots:
+        if source_root.name == "visual_blocks":
+            shutil.copytree(source_root, target / "visual_blocks", dirs_exist_ok=True)
+    for artifact in proof.artifacts:
+        path = Path(artifact)
+        if path.is_file() and path.parent.name != "visual_blocks":
+            shutil.copy2(path, target / path.name)
+    (target / "proof.json").write_text(proof.model_dump_json(indent=2), encoding="utf-8")
+
+
+def pass_rate(proofs: list[HashHopProof]) -> float:
+    if not proofs:
+        return 0.0
+    return sum(1 for proof in proofs if proof.passed) / len(proofs)
+
+
+def collect_model_visual_proofs(
+    evaluator: HashHopEvaluator,
+    hops: int,
+    token_equivalent: int,
+    attempts: int,
+) -> list[HashHopProof]:
+    """Run exactly `attempts` puzzles and keep every verdict.
+
+    The old loop stopped the moment the running pass rate crossed its target, so a model could be
+    re-rolled until it passed. A fixed attempt count is what makes the rate mean anything.
+    """
+    if attempts <= 0:
+        raise ValueError(f"collect needs a positive attempt count, got {attempts}")
+    return [
+        evaluator.generate_model_visual_proof(hops, token_equivalent, max_proofs=5)
+        for _ in range(attempts)
+    ]

backend/src/ely_eye/proofs/report.py

@@ -0,0 +1,142 @@
+"""Small readers, writers and diggers the proof checks share.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import math
+from pathlib import Path
+from typing import Any
+
+import zstandard as zstd
+
+
+def parquet_row_count(path: Path) -> int:
+    if not path.exists():
+        return -1
+    import pyarrow.parquet as pq
+
+    return pq.read_metadata(path).num_rows
+
+
+def compiled_context_digest(context: Any) -> str:
+    payload = {
+        "hits": [hit.atom.atom_id for hit in context.hits],
+        "packed_text": context.packed_text,
+        "token_equivalent": context.token_equivalent,
+    }
+    data = json.dumps(payload, ensure_ascii=False, sort_keys=True).encode("utf-8")
+    return hashlib.sha256(data).hexdigest()
+
+
+def read_memory_capsule(path: Path) -> dict[str, Any]:
+    data = path.read_bytes()
+    prefix = b"ELYMEMCAP1\n"
+    if not data.startswith(prefix):
+        raise ValueError(f"Invalid memory capsule header: {path}")
+    try:
+        payload = zstd.ZstdDecompressor().decompress(data[len(prefix) :])
+        parsed = json.loads(payload.decode("utf-8"))
+    except (zstd.ZstdError, UnicodeDecodeError, json.JSONDecodeError) as error:
+        raise ValueError(f"Cannot decode the memory capsule at {path}: {error}") from error
+    if not isinstance(parsed, dict):
+        raise ValueError(f"Invalid memory capsule payload: {path}")  # noqa: TRY004  # the capsule's own bytes are wrong, not the caller's argument
+    return parsed
+
+
+def validated_segment_hash(segment: dict[str, Any]) -> str:
+    expected = segment.get("segment_sha256")
+    if not isinstance(expected, str):
+        return ""
+    payload = {key: value for key, value in segment.items() if key != "segment_sha256"}
+    actual = hashlib.sha256(json.dumps(payload, sort_keys=True, separators=(",", ":")).encode("utf-8")).hexdigest()
+    return actual if actual == expected else ""
+
+
+def write_json(path: Path, payload: dict[str, Any]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def read_json_or_empty(path: Path) -> dict[str, Any]:
+    if not path.exists():
+        return {}
+    try:
+        value = json.loads(path.read_text(encoding="utf-8"))
+    except json.JSONDecodeError:
+        return {}
+    return value if isinstance(value, dict) else {}
+
+
+def format_scores(scores: dict[str, Any]) -> str:
+    parts: list[str] = []
+    for name in sorted(scores):
+        score = scores[name]
+        if isinstance(score, (int, float)) and math.isfinite(float(score)):
+            parts.append(f"{name}={float(score):.2f}")
+        else:
+            parts.append(f"{name}=unscored")
+    return ", ".join(parts) or "none"
+
+
+def is_finite_number(value: Any) -> bool:
+    return isinstance(value, (int, float)) and math.isfinite(float(value))
+
+
+def compact_generation_context(context: Any, section_hint: str, max_chars: int = 2400) -> Any:
+    text_hits = [hit for hit in context.hits if hit.atom.modality.value == "text"]
+    hits = (text_hits or context.hits)[:1]
+    sections: list[str] = []
+    for hit in hits:
+        atom = hit.atom
+        evidence_text = extract_markdown_section(atom.text, section_hint) or atom.text
+        evidence_text = evidence_text.strip()[:max_chars]
+        citation = {
+            "atom_id": atom.atom_id,
+            "modality": atom.modality.value,
+            "source": atom.source,
+            "image_ref": atom.image_ref,
+            "layout": atom.layout.model_dump() if atom.layout else None,
+            "score": hit.final_score,
+        }
+        sections.append(f"[EVIDENCE]\n{citation}\n{evidence_text}\n[/EVIDENCE]")
+    packed_text = "\n\n".join(sections)
+    return context.model_copy(
+        update={
+            "hits": hits,
+            "packed_text": packed_text,
+            "token_equivalent": max(1, len(packed_text) // 4),
+        }
+    )
+
+
+def extract_markdown_section(text: str, section_hint: str) -> str:
+    lines = text.splitlines()
+    start = None
+    for index, line in enumerate(lines):
+        stripped = line.strip()
+        if stripped.startswith("#") and section_hint in stripped:
+            start = index
+            break
+    if start is None:
+        return ""
+    end = len(lines)
+    start_level = len(lines[start]) - len(lines[start].lstrip("#"))
+    for index in range(start + 1, len(lines)):
+        stripped = lines[index].strip()
+        if stripped.startswith("#"):
+            level = len(lines[index]) - len(lines[index].lstrip("#"))
+            if level <= start_level:
+                end = index
+                break
+    return "\n".join(lines[start:end])
+
+
+def relative_artifacts(root: Path) -> dict[str, str]:
+    artifacts: dict[str, str] = {}
+    for path in sorted(root.rglob("*")):
+        if path.is_file():
+            key = path.relative_to(root).as_posix()
+            artifacts[key] = key
+    return artifacts

backend/src/ely_eye/proofs/repository_checks.py

@@ -0,0 +1,37 @@
+"""Find the repository the proof suite is allowed to talk about.
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+
+def project_root() -> Path:
+    current = Path.cwd().resolve()
+    for candidate in (current, *current.parents):
+        if (candidate / "PRD.md").exists() and (candidate / "backend" / "src" / "ely_eye").exists():
+            return candidate
+    return current
+
+
+def discover_repository_files(root: Path) -> list[Path]:
+    ignored_dirs = {".git", ".ely_eye", ".venv", ".venv-linux", "__pycache__", "node_modules", "dist", "build", "target"}
+    files: list[Path] = []
+    for current, dirs, names in os.walk(root):
+        dirs[:] = [name for name in dirs if name not in ignored_dirs and not name.startswith(".venv")]
+        current_path = Path(current)
+        for name in names:
+            path = current_path / name
+            try:
+                relative = path.relative_to(root)
+            except ValueError:
+                continue
+            if any(part in ignored_dirs or part.startswith(".venv") for part in relative.parts):
+                continue
+            try:
+                path.stat()
+            except OSError:
+                continue
+            files.append(path)
+    return files

backend/src/ely_eye/proofs/suite.py

@@ -0,0 +1,127 @@
+"""The proof suite: one check per claim, and a verdict that says what it proves.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import secrets
+from pathlib import Path
+from typing import Literal
+
+from ..cartridge import CartridgeService
+from ..config import Settings, get_settings
+from ..db import Database
+from ..evaluation_schemas import (
+    ProofCheck,
+    ProofSuiteReport,
+)
+from ..schemas import (
+    CartridgeManifest,
+    TrainingAdapterKind,
+)
+from .checks_cartridge import CartridgeChecks
+from .checks_runtime import RuntimeChecks
+from .checks_training import TrainingChecks
+from .checks_visual import VisualChecks
+from .report import (
+    relative_artifacts,
+)
+
+
+class PRDProofSuite(TrainingChecks, CartridgeChecks, RuntimeChecks, VisualChecks):
+    required_adapters = (
+        TrainingAdapterKind.hme_core_lora,
+        TrainingAdapterKind.hme_vision_lora,
+        TrainingAdapterKind.hme_ttt_vl,
+        TrainingAdapterKind.hme_visual_mtp,
+        TrainingAdapterKind.hme_router,
+        TrainingAdapterKind.hme_retrieval,
+    )
+
+    def __init__(self, settings: Settings | None = None) -> None:
+        self.settings = settings or get_settings()
+        self.db = Database(self.settings)
+        self.cartridges = CartridgeService(self.settings, self.db)
+
+    def run(self, cartridge_id: str | None = None) -> ProofSuiteReport:
+        row = self._resolve_cartridge(cartridge_id)
+        cartridge_id = str(row["cartridge_id"])
+        root = Path(row["root_path"])
+        manifest = CartridgeManifest.model_validate_json(row["manifest_json"])
+        proof_id = self._proof_id(cartridge_id, manifest.dna)
+        proof_root = self.settings.data_dir / "eval_proofs" / proof_id
+        proof_root.mkdir(parents=True, exist_ok=True)
+        dna_before = manifest.dna
+
+        asset_check, manifest = self._cartridge_assets_check(cartridge_id, root, proof_root)
+        checks = [
+            self._adapter_training_check(cartridge_id, root, proof_root),
+            self._runtime_adapter_check(proof_root),
+            self._runtime_generation_check(proof_root),
+            asset_check,
+            self._memory_capsule_check(root, proof_root),
+            self._cartridge_replay_check(cartridge_id, root, manifest, proof_root),
+            self._visual_hashhop_check(proof_root),
+            self._standard_benchmark_check(proof_root),
+            self._repository_input_check(proof_root),
+            self._runtime_profile_check(proof_root),
+            self._research_module_check(proof_root),
+            self._learned_memory_recall_check(cartridge_id, root, proof_root),
+            self._memory_dna_check(root, manifest, proof_root),
+            self._cache_trace_check(proof_root),
+            self._visual_contradiction_check(proof_root),
+        ]
+        # A deferred or removed check is not a failure and not a pass: the suite is only as green
+        # as the checks that actually ran. A claim that never ran holds the verdict, because
+        # "passed" would say the PRD is proven while part of it was never attempted.
+        failed = [check for check in checks if check.status == "failed"]
+        unrun = [check for check in checks if check.status in ("deferred", "removed")]
+        status: Literal["passed", "failed", "deferred"] = (
+            "failed" if failed else ("deferred" if unrun else "passed")
+        )
+        report = ProofSuiteReport(
+            proof_id=proof_id,
+            cartridge_id=cartridge_id,
+            status=status,
+            checks=checks,
+            artifacts={},
+            dna_before=dna_before,
+            dna_after=dna_before,
+        )
+        report_path = proof_root / "proof_suite.json"
+        report_path.write_text(report.model_dump_json(indent=2), encoding="utf-8")
+        report.artifacts = relative_artifacts(proof_root)
+        report_path.write_text(report.model_dump_json(indent=2), encoding="utf-8")
+        attached_manifest = self.cartridges.attach_eval_proof(cartridge_id, proof_root)
+        report.dna_after = attached_manifest.dna
+        report_path.write_text(report.model_dump_json(indent=2), encoding="utf-8")
+        cartridge_report = root / "eval_proofs" / proof_id / "proof_suite.json"
+        cartridge_report.write_text(report.model_dump_json(indent=2), encoding="utf-8")
+        return report
+
+    def runtime_generation_proof(self) -> ProofCheck:
+        proof_id = f"runtime_generation_{secrets.token_hex(8)}"
+        proof_root = self.settings.data_dir / "eval_proofs" / proof_id
+        proof_root.mkdir(parents=True, exist_ok=True)
+        return self._runtime_generation_check(proof_root)
+
+    def visual_contradiction_proof(self) -> ProofCheck:
+        proof_id = f"visual_contradiction_{secrets.token_hex(8)}"
+        proof_root = self.settings.data_dir / "eval_proofs" / proof_id
+        proof_root.mkdir(parents=True, exist_ok=True)
+        return self._visual_contradiction_check(proof_root)
+
+    def latest(self) -> ProofSuiteReport | None:
+        proof_root = self.settings.data_dir / "eval_proofs"
+        reports = sorted(
+            proof_root.glob("prd_proof_suite_*/proof_suite.json"),
+            key=lambda path: path.stat().st_mtime,
+            reverse=True,
+        )
+        if not reports:
+            return None
+        return ProofSuiteReport.model_validate_json(reports[0].read_text(encoding="utf-8"))
+
+    def _proof_id(self, cartridge_id: str, dna: str | None) -> str:
+        raw = f"{cartridge_id}:{dna}:{secrets.token_hex(8)}"
+        return f"prd_proof_suite_{hashlib.sha256(raw.encode('utf-8')).hexdigest()[:16]}"

backend/src/ely_eye/proofs/table.py

@@ -0,0 +1,34 @@
+"""Render a proof report as one row per claim, verdict first.
+
+A JSON report hides the shape of the evidence: a deferred claim and a passed claim look equally
+short. The table puts the counts on the first line and the reason under every claim that did not
+pass, so nobody reads an unrun claim as a proven one.
+"""
+
+from __future__ import annotations
+
+from ..evaluation_schemas import ProofSuiteReport
+
+STATUS_ORDER = ("failed", "passed", "deferred", "removed")
+
+
+def render_table(report: ProofSuiteReport) -> str:
+    counts: dict[str, int] = {status: 0 for status in STATUS_ORDER}
+    for check in report.checks:
+        counts[check.status] = counts.get(check.status, 0) + 1
+    lines = [
+        f"proof {report.proof_id} · cartridge {report.cartridge_id} · suite {report.status}",
+        "counts: " + ", ".join(f"{status} {counts.get(status, 0)}" for status in STATUS_ORDER),
+        "",
+    ]
+    for status in STATUS_ORDER:
+        for check in report.checks:
+            if check.status != status:
+                continue
+            lines.append(f"[{check.status:^7}] {check.name}")
+            lines.append(f"          ran: {check.detail}")
+            if check.reason:
+                lines.append(f"          why: {check.reason}")
+            if check.rerun:
+                lines.append(f"          rerun: {check.rerun}")
+    return "\n".join(lines)

backend/src/ely_eye/proofs/training_trace.py

@@ -0,0 +1,358 @@
+"""Check that an adapter's training artifacts tell a consistent story.
+"""
+
+from __future__ import annotations
+
+import json
+import math
+from pathlib import Path
+from typing import Any
+
+from ..schemas import (
+    AdapterManifest,
+    TrainingAdapterKind,
+)
+from ..training import (
+    AUTOCAST_DTYPE,
+    GRADIENT_CHECKPOINTING_MODE,
+    TRAINING_FRAMEWORK,
+    TRAINING_METHOD,
+    TRAINING_PRECISION,
+    TRAINING_PROOF_VERSION,
+    safetensor_weight_stats,
+    sha256_file,
+)
+
+
+def count_safetensor_keys(path: Path) -> int:
+    return int(safetensor_weight_stats(path)["tensor_count"])
+
+
+def peft_config_loads(path: Path) -> bool:
+    try:
+        from peft import PeftConfig
+
+        PeftConfig.from_pretrained(str(path))
+        return True
+    except Exception:  # noqa: BLE001  # a probe records its failure instead of crashing the proof run
+        return False
+
+
+def validate_training_contract(manifest: AdapterManifest) -> dict[str, Any]:
+    target_modules_ok = training_target_modules_ok(manifest.kind, manifest.target_modules)
+    report = {
+        "proof_version_ok": manifest.proof_version >= TRAINING_PROOF_VERSION,
+        "method_ok": manifest.training_method == TRAINING_METHOD,
+        "precision_ok": manifest.precision == TRAINING_PRECISION and manifest.bf16,
+        "framework_ok": manifest.framework == TRAINING_FRAMEWORK,
+        "optimizer_ok": manifest.optimizer == "paged_adamw_8bit",
+        "optimizer_family_ok": manifest.optimizer_family == "8-bit AdamW / paged optimizer",
+        "gradient_checkpointing_ok": (
+            manifest.gradient_checkpointing
+            and manifest.gradient_checkpointing_mode == GRADIENT_CHECKPOINTING_MODE
+        ),
+        "autocast_dtype_ok": manifest.autocast_dtype == AUTOCAST_DTYPE,
+        "unsloth_version_ok": bool(manifest.unsloth_version),
+        "triton_version_ok": bool(manifest.triton_version),
+        "xformers_version_ok": bool(manifest.xformers_version),
+        "unsloth_model_type_ok": bool(manifest.unsloth_model_type),
+        "max_length_ok": manifest.max_length >= 8192,
+        "target_modules_ok": target_modules_ok,
+    }
+    report["valid"] = all(report.values())
+    return report
+
+
+def training_target_modules_ok(kind: TrainingAdapterKind, target_modules: list[str]) -> bool:
+    targets = set(target_modules)
+    core = {"q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"}
+    if kind == TrainingAdapterKind.hme_core_lora:
+        return core.issubset(targets)
+    if kind == TrainingAdapterKind.hme_vision_lora:
+        return bool(targets) and any("visual" in target or "vision" in target for target in targets)
+    if kind == TrainingAdapterKind.hme_ttt_vl:
+        has_visual = any("visual" in target or "vision" in target for target in targets)
+        has_mlp = bool({"gate_proj", "up_proj", "down_proj"} & targets)
+        return has_visual and has_mlp
+    if kind == TrainingAdapterKind.hme_visual_mtp:
+        return any("mtp" in target or "draft" in target or target == "lm_head" for target in targets)
+    if kind in {TrainingAdapterKind.hme_router, TrainingAdapterKind.hme_retrieval}:
+        return core.issubset(targets)
+    return False
+
+
+def training_payload_contract_ok(payload: dict[str, Any], manifest: AdapterManifest) -> bool:
+    return (
+        int(payload.get("proof_version") or 0) >= TRAINING_PROOF_VERSION
+        and payload.get("training_method") == TRAINING_METHOD
+        and payload.get("precision") == TRAINING_PRECISION
+        and payload.get("framework") == TRAINING_FRAMEWORK
+        and payload.get("optimizer_family") == manifest.optimizer_family
+        and bool(payload.get("gradient_checkpointing"))
+        and payload.get("gradient_checkpointing_mode") == GRADIENT_CHECKPOINTING_MODE
+        and payload.get("autocast_dtype") == AUTOCAST_DTYPE
+        and payload.get("unsloth_version") == manifest.unsloth_version
+        and payload.get("triton_version") == manifest.triton_version
+        and payload.get("xformers_version") == manifest.xformers_version
+        and payload.get("unsloth_model_type") == manifest.unsloth_model_type
+    )
+
+
+def validate_training_trace(manifest: AdapterManifest, adapter_root: Path) -> dict[str, Any]:
+    trace_path = adapter_root / "training_trace.jsonl"
+    summary_path = adapter_root / "training_summary.json"
+    report: dict[str, Any] = {
+        "trace_path": str(trace_path),
+        "summary_path": str(summary_path),
+        "trace_exists": trace_path.exists(),
+        "summary_exists": summary_path.exists(),
+        "trace_sha256_ok": False,
+        "summary_sha256_ok": False,
+        "step_records": 0,
+        "loss_history_records": len(manifest.loss_history),
+        "loss_history_ok": False,
+        "summary_matches_manifest": False,
+        "summary_contract_ok": False,
+        "run_start_contract_ok": False,
+        "run_finish_contract_ok": False,
+        "cuda_training": False,
+        "bf16_supported": False,
+        "bf16": manifest.bf16,
+        "valid": False,
+    }
+    if not trace_path.exists() or not summary_path.exists():
+        return report
+    try:
+        trace_rows = [
+            json.loads(line)
+            for line in trace_path.read_text(encoding="utf-8").splitlines()
+            if line.strip()
+        ]
+        summary = json.loads(summary_path.read_text(encoding="utf-8"))
+    except json.JSONDecodeError:
+        return report
+
+    step_rows = [row for row in trace_rows if row.get("event") == "step"]
+    start_row: dict[str, Any] = next((row for row in trace_rows if row.get("event") == "run_start"), {})
+    finish_row: dict[str, Any] = next(
+        (row for row in reversed(trace_rows) if row.get("event") == "run_finish"), {}
+    )
+    losses = [row.get("loss") for row in step_rows]
+    finite_losses = all(isinstance(loss, (int, float)) and math.isfinite(float(loss)) for loss in losses)
+    final_loss_ok = (
+        manifest.final_loss is not None
+        and bool(losses)
+        and abs(float(losses[-1]) - float(manifest.final_loss)) < 1e-6
+    )
+    report.update(
+        {
+            "trace_sha256_ok": (
+                manifest.training_trace_sha256 is not None
+                and sha256_file(trace_path) == manifest.training_trace_sha256
+            ),
+            "summary_sha256_ok": (
+                manifest.training_summary_sha256 is not None
+                and sha256_file(summary_path) == manifest.training_summary_sha256
+            ),
+            "step_records": len(step_rows),
+            "loss_history_ok": (
+                len(manifest.loss_history) == manifest.max_steps
+                and finite_losses
+                and final_loss_ok
+            ),
+            "summary_matches_manifest": (
+                summary.get("adapter_kind") == manifest.kind.value
+                and summary.get("base_model") == manifest.base_model
+                and summary.get("dataset_sha256") == manifest.dataset_sha256
+                and int(summary.get("max_steps") or -1) == manifest.max_steps
+                and int(summary.get("trainable_params") or -1) == manifest.trainable_params
+            ),
+            "summary_contract_ok": training_payload_contract_ok(summary, manifest),
+            "run_start_contract_ok": training_payload_contract_ok(start_row, manifest),
+            "run_finish_contract_ok": training_payload_contract_ok(finish_row, manifest),
+        }
+    )
+    environment = summary.get("environment") if isinstance(summary.get("environment"), dict) else {}
+    device_name = str(environment.get("device_name") or "")
+    report["cuda_training"] = bool(environment.get("cuda_available")) and "RTX 4090" in device_name
+    report["bf16_supported"] = bool(environment.get("bf16_supported"))
+    report["valid"] = (
+        report["trace_sha256_ok"]
+        and report["summary_sha256_ok"]
+        and report["step_records"] == manifest.max_steps
+        and report["loss_history_ok"]
+        and report["summary_matches_manifest"]
+        and report["summary_contract_ok"]
+        and report["run_start_contract_ok"]
+        and report["run_finish_contract_ok"]
+        and report["cuda_training"]
+        and report["bf16_supported"]
+        and manifest.bf16
+    )
+    return report
+
+
+def validate_training_proof(manifest: AdapterManifest, adapter_root: Path) -> dict[str, Any]:
+    proof_path = adapter_root / "training_proof.json"
+    report: dict[str, Any] = {
+        "proof_path": str(proof_path),
+        "proof_exists": proof_path.exists(),
+        "proof_sha256_ok": False,
+        "hash_changed": False,
+        "optimizer_update_count": manifest.optimizer_update_count,
+        "optimizer_update_count_ok": False,
+        "max_optimizer_update_l2": manifest.max_optimizer_update_l2,
+        "max_optimizer_update_l2_ok": False,
+        "loss_delta": manifest.loss_delta,
+        "loss_delta_finite": False,
+        "updates_have_gradients": False,
+        "updates_have_weight_delta": False,
+        "proof_weight_sha256_ok": False,
+        "proof_weight_stats_ok": False,
+        "proof_cuda_bf16": False,
+        "proof_bf16_supported": False,
+        "proof_contract_ok": False,
+        "proof_optimizer_ok": False,
+        "proof_max_length_ok": False,
+        "proof_target_modules_ok": False,
+        "proof_unsloth_runtime_ok": False,
+        "proof_matches_manifest": False,
+        "valid": False,
+    }
+    if not proof_path.exists():
+        return report
+    try:
+        proof = json.loads(proof_path.read_text(encoding="utf-8"))
+    except json.JSONDecodeError:
+        return report
+
+    updates = proof.get("optimizer_updates") if isinstance(proof.get("optimizer_updates"), list) else []
+    update_l2_values = [
+        float(value)
+        for item in updates
+        if isinstance(item, dict) and isinstance(value := item.get("update_l2"), (int, float))
+    ]
+    gradient_l2_values = [
+        float(value)
+        for item in updates
+        if isinstance(item, dict) and isinstance(value := item.get("gradient_l2"), (int, float))
+    ]
+    weight_report = safetensor_weight_stats(adapter_root / "adapter_model.safetensors")
+    proof_weight_stats = proof.get("adapter_weight_stats") if isinstance(proof.get("adapter_weight_stats"), dict) else {}
+    proof_environment = proof.get("environment") if isinstance(proof.get("environment"), dict) else {}
+    report.update(
+        {
+            "proof_sha256_ok": (
+                manifest.training_proof_sha256 is not None
+                and sha256_file(proof_path) == manifest.training_proof_sha256
+            ),
+            "hash_changed": (
+                bool(manifest.initial_trainable_sha256)
+                and bool(manifest.final_trainable_sha256)
+                and manifest.initial_trainable_sha256 != manifest.final_trainable_sha256
+            ),
+            "optimizer_update_count_ok": (
+                manifest.optimizer_update_count > 0
+                and len(updates) == manifest.optimizer_update_count
+            ),
+            "max_optimizer_update_l2_ok": (
+                manifest.max_optimizer_update_l2 is not None
+                and math.isfinite(float(manifest.max_optimizer_update_l2))
+                and float(manifest.max_optimizer_update_l2) > 0.0
+            ),
+            "loss_delta_finite": (
+                manifest.loss_delta is not None
+                and math.isfinite(float(manifest.loss_delta))
+            ),
+            "updates_have_gradients": bool(gradient_l2_values)
+            and all(math.isfinite(value) and value > 0.0 for value in gradient_l2_values),
+            "updates_have_weight_delta": bool(update_l2_values)
+            and all(math.isfinite(value) and value > 0.0 for value in update_l2_values),
+            "proof_weight_sha256_ok": (
+                isinstance(proof.get("adapter_weight_sha256"), str)
+                and proof.get("adapter_weight_sha256") == weight_report["sha256"]
+            ),
+            "proof_weight_stats_ok": (
+                proof_weight_stats.get("sha256") == weight_report["sha256"]
+                and proof_weight_stats.get("tensor_count") == weight_report["tensor_count"]
+                and proof_weight_stats.get("lora_tensor_count") == weight_report["lora_tensor_count"]
+                and proof_weight_stats.get("total_elements") == weight_report["total_elements"]
+                and proof_weight_stats.get("nonzero_elements") == weight_report["nonzero_elements"]
+                and int(proof_weight_stats.get("lora_tensor_count") or 0) > 0
+                and int(proof_weight_stats.get("nonzero_elements") or 0) > 0
+                and bool(proof_weight_stats.get("all_finite"))
+            ),
+            "proof_cuda_bf16": (
+                bool(proof.get("bf16"))
+                and bool(proof_environment.get("cuda_available"))
+                and "RTX 4090" in str(proof_environment.get("device_name") or "")
+            ),
+            "proof_bf16_supported": bool(proof_environment.get("bf16_supported")),
+            "proof_contract_ok": training_payload_contract_ok(proof, manifest),
+            "proof_optimizer_ok": proof.get("optimizer") == "paged_adamw_8bit",
+            "proof_max_length_ok": int(proof.get("max_length") or 0) >= 8192,
+            "proof_target_modules_ok": training_target_modules_ok(
+                manifest.kind,
+                proof.get("target_modules") if isinstance(proof.get("target_modules"), list) else [],
+            ),
+            "proof_unsloth_runtime_ok": (
+                bool(proof.get("unsloth_version"))
+                and bool(proof.get("triton_version"))
+                and bool(proof.get("xformers_version"))
+                and bool(proof.get("unsloth_model_type"))
+            ),
+            "proof_matches_manifest": (
+                proof.get("adapter_kind") == manifest.kind.value
+                and proof.get("proof_version") == manifest.proof_version
+                and proof.get("training_method") == manifest.training_method
+                and proof.get("precision") == manifest.precision
+                and proof.get("framework") == manifest.framework
+                and proof.get("base_model") == manifest.base_model
+                and proof.get("dataset_sha256") == manifest.dataset_sha256
+                and proof.get("sample_count") == manifest.sample_count
+                and proof.get("max_steps") == manifest.max_steps
+                and proof.get("max_length") == manifest.max_length
+                and proof.get("rank") == manifest.rank
+                and proof.get("alpha") == manifest.alpha
+                and proof.get("target_modules") == manifest.target_modules
+                and proof.get("optimizer") == manifest.optimizer
+                and proof.get("optimizer_family") == manifest.optimizer_family
+                and proof.get("learning_rate") == manifest.learning_rate
+                and proof.get("gradient_accumulation_steps") == manifest.gradient_accumulation_steps
+                and proof.get("gradient_checkpointing") == manifest.gradient_checkpointing
+                and proof.get("gradient_checkpointing_mode") == manifest.gradient_checkpointing_mode
+                and proof.get("autocast_dtype") == manifest.autocast_dtype
+                and proof.get("unsloth_version") == manifest.unsloth_version
+                and proof.get("triton_version") == manifest.triton_version
+                and proof.get("xformers_version") == manifest.xformers_version
+                and proof.get("unsloth_model_type") == manifest.unsloth_model_type
+                and proof.get("bf16") == manifest.bf16
+                and proof.get("trainable_params") == manifest.trainable_params
+                and proof.get("total_params") == manifest.total_params
+                and proof.get("initial_trainable_sha256") == manifest.initial_trainable_sha256
+                and proof.get("final_trainable_sha256") == manifest.final_trainable_sha256
+                and proof.get("optimizer_update_count") == manifest.optimizer_update_count
+                and proof.get("max_optimizer_update_l2") == manifest.max_optimizer_update_l2
+            ),
+        }
+    )
+    report["valid"] = (
+        report["proof_sha256_ok"]
+        and report["hash_changed"]
+        and report["optimizer_update_count_ok"]
+        and report["max_optimizer_update_l2_ok"]
+        and report["loss_delta_finite"]
+        and report["updates_have_gradients"]
+        and report["updates_have_weight_delta"]
+        and report["proof_weight_sha256_ok"]
+        and report["proof_weight_stats_ok"]
+        and report["proof_cuda_bf16"]
+        and report["proof_bf16_supported"]
+        and report["proof_contract_ok"]
+        and report["proof_optimizer_ok"]
+        and report["proof_max_length_ok"]
+        and report["proof_target_modules_ok"]
+        and report["proof_unsloth_runtime_ok"]
+        and report["proof_matches_manifest"]
+    )
+    return report

backend/src/ely_eye/text_index.py

@@ -38,11 +38,19 @@ def fts_query(query: str) -> str:
     A CJK run becomes a quoted phrase so its characters stay adjacent: unphrased, a query for
     钥匙 would match text that mentions 钥 and 匙 worlds apart. Latin tokens get a prefix match,
     which is what lets `atom` reach `atoms`.
+
+    Latin tokens are OR-ed, not AND-ed. A question is longer than the sentence that answers it —
+    "Which PRD requirements prove long context memory, HashHop, and adapters?" shares only some of
+    its words with any single atom — and an AND over every word answers such a question with
+    nothing. bm25 then ranks the candidates by how many of those words each one carries, so the
+    broadened query does not mean a diluted answer.
     """
     terms = _bare(query).split()
     if not terms:
         raise ValueError(f"search needs at least one searchable word, got {query!r}")
-    return " ".join(_cjk_phrase(term) if _is_cjk(term) else f"{_quote(term)}*" for term in terms)
+    if len(terms) == 1:
+        return _cjk_phrase(terms[0]) if _is_cjk(terms[0]) else f"{_quote(terms[0])}*"
+    return " OR ".join(_cjk_phrase(term) if _is_cjk(term) else f"{_quote(term)}*" for term in terms)
 
 
 def _bare(text: str) -> str:

backend/tests/test_proofs.py

@@ -0,0 +1,282 @@
+"""A proof check reports what it measured: passed, failed, deferred, or removed."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Literal
+
+import pytest
+from PIL import Image
+
+from ely_eye.cartridge import CartridgeService
+from ely_eye.config import Settings
+from ely_eye.memory import MemoryService
+from ely_eye.proofs import PRDProofSuite
+from ely_eye.proofs.arena import (
+    _slot_boxes,
+    render_visual_hash,
+    render_visual_hash_arena,
+)
+from ely_eye.proofs.checks_common import _deferred_without_model
+from ely_eye.proofs.contradictions import (
+    PACKAGED_CORPUS,
+    build_contradiction_scenarios,
+    load_corpus,
+)
+from ely_eye.proofs.table import render_table
+from ely_eye.schemas import MemoryOrigin
+
+# The replay question is fixed inside the suite, so the fixture memory has to answer it in words
+# the ranker can match.
+PRD_MEMORY = "PRD requirements prove long context memory, HashHop proof, and adapter training."
+OTHER_KEYS = "PRD requirements also prove cache invalidation for HashHop proof and adapters."
+
+
+class Harness:
+    """A suite wired to one temporary store, plus the directory its artifacts land in."""
+
+    def __init__(self, settings: Settings) -> None:
+        self.settings = settings
+        self.memory = MemoryService(settings=settings)
+        self.db = self.memory.db
+        self.cartridges = CartridgeService(settings=settings, database=self.db)
+        self.suite = PRDProofSuite(settings)
+        self.proof_root = settings.data_dir / "eval_proofs" / "test"
+        self.proof_root.mkdir(parents=True, exist_ok=True)
+
+    def remember(self, text: str) -> str:
+        atom = self.memory.remember(text, holder_id="wife", origin=MemoryOrigin.self_, scenario="home")
+        return atom.atom_id
+
+    def cartridge(self, name: str, atom_ids: list[str]) -> str:
+        atoms = self.db.get_atoms(atom_ids)
+        sources = self.db.get_sources(sorted({atom.source_id for atom in atoms}))
+        return self.cartridges.materialize(name, atoms, sources).cartridge_id
+
+
+@pytest.fixture
+def harness(tmp_path: Path, model_dir: Path) -> Harness:
+    return Harness(Settings(home=tmp_path / "home", model_dir=model_dir, default_holder="wife"))
+
+
+def test_a_store_with_nothing_to_ask_fails_the_cache_trace_check(harness: Harness) -> None:
+    check = harness.suite._cache_trace_check(harness.proof_root)
+
+    assert check.status == "failed"
+    assert "empty store" in check.reason
+
+
+def test_the_cache_trace_check_requires_a_measured_miss_then_hit(harness: Harness) -> None:
+    harness.remember(PRD_MEMORY)
+
+    check = harness.suite._cache_trace_check(harness.proof_root)
+
+    assert check.status == "passed", check.detail
+    assert check.reason == ""
+
+
+def test_the_research_module_check_is_removed_because_it_could_not_fail(
+    harness: Harness,
+) -> None:
+    check = harness.suite._research_module_check(harness.proof_root)
+
+    assert check.status == "removed"
+    assert check.reason and check.rerun
+
+
+def test_the_repository_check_fails_when_the_repository_never_reached_the_store(
+    harness: Harness,
+) -> None:
+    check = harness.suite._repository_input_check(harness.proof_root)
+
+    assert check.status == "failed"
+    assert "store" in check.detail or "store" in check.reason
+
+
+def test_the_replay_check_answers_only_from_the_cartridge_it_mounts(harness: Harness) -> None:
+    inside = harness.remember(PRD_MEMORY)
+    # A second memory that answers the same question but stays out of the cartridge: an unmounted
+    # replay against the live store would have found it and called that a cartridge replay.
+    harness.remember(OTHER_KEYS)
+    cartridge_id = harness.cartridge("keys", [inside])
+    manifest = next(
+        item for item in harness.cartridges.list_manifests() if item.cartridge_id == cartridge_id
+    )
+
+    check = harness.suite._cartridge_replay_check(
+        cartridge_id, harness.settings.cartridge_dir / cartridge_id, manifest, harness.proof_root
+    )
+
+    assert check.status == "passed", check.detail
+    assert harness.cartridges.mount_state().mounted == []
+
+
+def test_a_check_defers_loudly_when_the_model_it_needs_is_missing(
+    tmp_path: Path, model_dir: Path
+) -> None:
+    without = _deferred_without_model(
+        Settings(home=tmp_path / "a"),
+        tmp_path / "a" / "artifact.json",
+        "Runtime Adapter Binding",
+        "PRD 18 requires the trained adapter to run.",
+        "ely-eye proof runtime with model_dir set",
+    )
+    with_model = _deferred_without_model(
+        Settings(home=tmp_path / "b", model_dir=model_dir),
+        tmp_path / "b" / "artifact.json",
+        "Runtime Adapter Binding",
+        "PRD 18 requires the trained adapter to run.",
+        "ely-eye proof runtime with model_dir set",
+    )
+
+    assert without is not None and without.status == "deferred"
+    assert without.rerun == "ely-eye proof runtime with model_dir set"
+    assert (tmp_path / "a" / "artifact.json").exists()
+    assert with_model is None
+
+
+def test_the_table_shows_the_verdict_counts_and_the_reason_for_each_claim() -> None:
+    from ely_eye.evaluation_schemas import ProofCheck, ProofSuiteReport
+
+    report = ProofSuiteReport(
+        proof_id="proof-1",
+        cartridge_id="cart-1",
+        status="deferred",
+        artifacts={},
+        checks=[
+            ProofCheck(name="Ran", requirement="r", status="passed", evidence=[], detail="measured x"),
+            ProofCheck(
+                name="Waiting",
+                requirement="r",
+                status="deferred",
+                evidence=[],
+                detail="nothing ran",
+                reason="no model weights",
+                rerun="ely-eye proof suite",
+            ),
+        ],
+    )
+
+    table = render_table(report)
+
+    assert "counts: failed 0, passed 1, deferred 1, removed 0" in table
+    assert "why: no model weights" in table
+    assert "rerun: ely-eye proof suite" in table
+    assert table.index("[passed ") < table.index("[deferred")
+    assert "[failed" not in table and "[removed" not in table
+
+
+def test_the_latest_command_prints_the_table_and_exits_by_verdict(
+    tmp_path: Path, monkeypatch
+) -> None:
+    import json
+
+    from typer.testing import CliRunner
+
+    from ely_eye import cli as cli_module
+    from ely_eye.evaluation_schemas import ProofCheck, ProofSuiteReport
+
+    settings = Settings(home=tmp_path / "home")
+    monkeypatch.setattr(cli_module, "get_settings", lambda: settings)
+    proof_root = settings.data_dir / "eval_proofs"
+    ran = ProofCheck(name="Ran", requirement="r", status="passed", evidence=[], detail="ok")
+    cases: list[tuple[str, Literal["failed", "deferred"]]] = [
+        ("prd_proof_suite_old", "failed"),
+        ("prd_proof_suite_new", "deferred"),
+    ]
+    for proof_id, status in cases:
+        report = ProofSuiteReport(
+            proof_id=proof_id,
+            cartridge_id="cart-1",
+            status=status,
+            artifacts={},
+            checks=[ran],
+        )
+        directory = proof_root / proof_id
+        directory.mkdir(parents=True)
+        (directory / "proof_suite.json").write_text(report.model_dump_json(), encoding="utf-8")
+
+    table = CliRunner().invoke(cli_module.app, ["latest-proof-suite"])
+    machine = CliRunner().invoke(cli_module.app, ["latest-proof-suite", "--json"])
+
+    assert table.exit_code == 0
+    assert "proof prd_proof_suite_new" in table.stdout
+    assert json.loads(machine.stdout)["proof_id"] == "prd_proof_suite_new"
+
+    (proof_root / "prd_proof_suite_new" / "proof_suite.json").write_text(
+        ProofSuiteReport(
+            proof_id="prd_proof_suite_new",
+            cartridge_id="cart-1",
+            status="failed",
+            artifacts={},
+            checks=[ran],
+        ).model_dump_json(),
+        encoding="utf-8",
+    )
+    failing = CliRunner().invoke(cli_module.app, ["latest-proof-suite"])
+
+    assert failing.exit_code == 1
+
+    monkeypatch.setattr(cli_module, "get_settings", lambda: Settings(home=tmp_path / "empty"))
+    nothing = CliRunner().invoke(cli_module.app, ["latest-proof-suite"])
+
+    assert nothing.exit_code != 0
+    assert "no proof suite" in nothing.stdout + nothing.stderr  # says so instead of printing null
+
+
+def test_the_packaged_corpus_matches_the_scenarios_it_was_written_from() -> None:
+    scenarios, digest = load_corpus(PACKAGED_CORPUS)
+    built = build_contradiction_scenarios()
+
+    assert [item["name"] for item in scenarios] == [item["name"] for item in built]
+    assert [item["expect_drift"] for item in scenarios] == [
+        item["expect_drift"] for item in built
+    ]
+    assert len(scenarios) >= 10 and digest
+
+
+def test_the_contradiction_check_defers_when_no_corpus_exists_and_names_the_file(
+    harness: Harness, monkeypatch
+) -> None:
+    from ely_eye.proofs import checks_visual as visual_module
+
+    monkeypatch.setattr(visual_module, "corpus_path", lambda data_dir: None)
+
+    check = harness.suite._visual_contradiction_check(harness.proof_root)
+
+    assert check.status == "deferred"
+    assert "contradiction_scenarios.json" in check.rerun
+    assert check.detail == "no contradiction corpus on disk"
+
+
+def test_the_contradiction_check_records_which_corpus_it_graded_against(
+    harness: Harness,
+) -> None:
+    import json
+
+    check = harness.suite._visual_contradiction_check(harness.proof_root)
+    artifact = json.loads(
+        (harness.proof_root / "visual_contradiction_lens.json").read_text(encoding="utf-8")
+    )
+
+    assert check.status == "deferred"  # graded on our own corpus: reported, not claimed
+    assert artifact["corpus_sha256"]
+    assert len(artifact["metrics"]["detected_categories"]) >= 5, artifact["metrics"]
+
+
+def test_the_hashhop_arena_marks_no_candidate_as_the_answer(harness: Harness, tmp_path: Path) -> None:
+    boxes = _slot_boxes(520, 704, 76, 324, 160)
+    tiles = {}
+    for slot in boxes:
+        path = tmp_path / f"{slot}.png"
+        render_visual_hash(slot, path, 160)
+        tiles[slot] = path
+    output = tmp_path / "arena.png"
+
+    render_visual_hash_arena(tiles["top-left"], tiles, output)
+
+    # Every candidate box carries the same outline: a ring, an arrow, or a highlighted slot would
+    # tell a model the answer without reading a pixel of the source tile.
+    image = Image.open(output).convert("RGB")
+    border_colors = {image.getpixel((box[0] + 2, box[1] + 40)) for box in boxes.values()}
+    assert border_colors == {(180, 190, 205)}

backend/tests/test_text_index.py

@@ -24,6 +24,16 @@ def _hits(connection: sqlite3.Connection, query: str) -> list[str]:
     return [row[0] for row in connection.execute("SELECT original FROM fts WHERE fts MATCH ?", (fts_query(query),))]
 
 
+def _ranked(connection: sqlite3.Connection, query: str) -> list[str]:
+    """Matches in bm25 order: the answer that covers more of the question should come first."""
+    return [
+        row[0]
+        for row in connection.execute(
+            "SELECT original FROM fts WHERE fts MATCH ? ORDER BY rank", (fts_query(query),)
+        )
+    ]
+
+
 def test_index_text_splits_a_chinese_sentence_into_characters() -> None:
     assert index_text("我们的钥匙放在玄关抽屉里") == "我 们 的 钥 匙 放 在 玄 关 抽 屉 里"
 
@@ -40,7 +50,28 @@ def test_index_text_folds_full_width_letters_that_search_would_miss() -> None:
 
 def test_fts_query_phrases_cjk_and_prefixes_latin() -> None:
     assert fts_query("钥匙") == '"钥 匙"'
-    assert fts_query("search atoms") == '"search"* "atoms"*'
+    assert fts_query("search atoms") == '"search"* OR "atoms"*'
+
+
+def test_a_long_question_reaches_the_answer_that_covers_most_of_its_words() -> None:
+    """AND over every word answers a long question with nothing; bm25 picks the best candidate."""
+    connection = _fts(
+        [
+            "PRD requirements prove long context memory, HashHop proof, and adapter training.",
+            "The adapter was trained for 29 steps on 37 samples.",
+            "Unrelated notes about the weather.",
+        ]
+    )
+    question = "Which Ely-Eye PRD requirements prove long context memory, HashHop, and adapters?"
+    hits = [
+        row[0]
+        for row in connection.execute(
+            "SELECT original FROM fts WHERE fts MATCH ? ORDER BY rank", (fts_query(question),)
+        )
+    ]
+
+    assert hits[0].startswith("PRD requirements prove")
+    assert "Unrelated notes about the weather." not in hits
 
 
 def test_fts_query_refuses_a_query_with_nothing_to_match() -> None:
@@ -70,6 +101,10 @@ def test_a_prefix_reaches_the_word_it_grew_into() -> None:
     assert _hits(connection, "forg") == ["the atom was forgotten"]
 
 
-def test_a_mixed_query_needs_both_halves_in_the_same_memory() -> None:
+def test_a_mixed_query_ranks_the_memory_holding_both_halves_first() -> None:
+    """A query with a Chinese half and an English half must prefer the memory holding both."""
     connection = _fts(["钥匙放在 drawer 里", "钥匙在玄关", "the drawer is empty"])
-    assert _hits(connection, "钥匙 drawer") == ["钥匙放在 drawer 里"]
+    ranked = _ranked(connection, "钥匙 drawer")
+
+    assert ranked[0] == "钥匙放在 drawer 里"
+    assert set(ranked) == {"钥匙放在 drawer 里", "钥匙在玄关", "the drawer is empty"}