Put the scale evidence in the repo and say plainly what still misses

Every number in CHECKLIST p2-1 and p0-3.7 came out of a scratch script: the
kind of artifact a reader cannot re-run, and the kind that quietly drifts
away from the code it describes. These reports are generated by committed
harnesses and each one stamps the commit that carries the code it measures.

DENSE.md  165,973 atoms / 153,857 vectors: build 1421.0s, search p50 75.6ms
          p95 93.0ms, 0.998 agreement with a brute-force scan, one write into
          a full index p50 518ms, 1,729 bytes per stored vector.
QUANT.md  twelve encodings measured on the same vectors: int8 keeps 0.977 of
          the ranking for a quarter of the bytes and no speed, bit is 14x
          faster and keeps 0.728, so bit is a coarse filter and nothing more.
TERMS.md  the stop-word bound swept, not guessed: 0.05 is the measured peak
          (recall@10 0.803 at p50 69.5ms) and also the fastest of the rows
          that keep recall.
SCALE.md  the product's own search at 165,973 atoms: p50 71.7ms against the
          frozen baseline's 20.16s.
LEGS.md   what each retrieval leg is worth, per modality, with the page size
          the product actually asks for in the table.

The acceptance bar the operator set is recall@10 >= 0.90 at n>=300. It is
not met. The best measured row is 0.858 (dense leg on, page widened to 30,
p50 414ms); the shipped configuration gives 0.848 at p50 72.5ms; the
in-repo counterpart at holder scope gives 0.866. The shortfall is
concentrated in picture golds (video 0.583-0.708, ui_screenshot 0.500-0.583,
image 0.750) while the word-shaped ones already pass (code 0.953, pdf 0.919).
Latency passes by three hundred fold; recall does not, and the gap is four
points of cross-modal retrieval, not index scale.

CHECKLIST.md

@@ -75,9 +75,9 @@
   - [x] p0-3.4 七个过滤器（holder/scenario/source/modality/since/until/min_trust）+ base64 游标；`since > until` 报 "search since must not be after until"，游标坏了报 "search cursor is unreadable"
   - [x] p0-3.5 RRF（`RRF_K = 60`）取代 `0.55/0.35/0.10`；顺手删掉无人调用的 `tokenize()` / `lexical_score()` / `_graph_score()`（graph 项恒为 0，因为证据边从不参与排序）
   - [x] p0-3.6 Go 的 `search_atoms` / `compile_context` 改走 `POST /api/search`，删掉 Go 侧第二套分词器（`store.SearchAtoms` / `ftsQuery` / `tokenPattern`）——两面不可能再漂开，因为只剩一个排序实现
-  - [x] p0-3.7 100k atom 下 p50 <100ms：`python -m bench.scale_check` → `bench/reports/SCALE.md`。实测 **165,973 atom（bar 的 1.66 倍）、300 探针、p50 57ms / p95 138ms**，机器 load 1.76 / 32 核 / 每查询 4 线程；基线 p50 20.16s ⇒ **快 343 倍**（SCALE.md 重跑：p50 58.7ms / p95 143.0ms）。同一探针集上的质量与延迟扫描见 `bench/reports/TERMS.md`（n=173 holder 作用域，0.05 档 recall@10 0.803、MRR 0.652、p50 64.8ms / p95 121.7ms）。
+  - [x] p0-3.7 100k atom 下 p50 <100ms：`python -m bench.scale_check` → `bench/reports/SCALE.md`。报告那一次实测 **165,973 atom（bar 的 1.66 倍）、300 探针、p50 71.7ms / p95 217.7ms / max 780.5ms**，机器 load [2.31, 3.72, 7.9] / 32 核 / 每查询 4 线程；基线 p50 20.16s ⇒ **快 281 倍**。同一脚本在更空的机器上跑到过 p50 56.5ms / 58.7ms（load 1.5–1.8）——延迟随负载走，所以引用报告那一次，不挑最好看的。同一探针集上的质量与延迟扫描见 `bench/reports/TERMS.md`（n=173 holder 作用域，0.05 档 recall@10 0.803、MRR 0.652、p50 69.5ms / p95 131.5ms）。
     - 根因不是 SQL 慢，是 OR 表达式里那些「每个文档都有」的虚词：`"parsing"* OR "error"* OR "in"* OR "the"* …` 命中 127,145 行、305ms，而中文二元组短语查询 32 命中只要 1.0ms。bm25 必须给每条命中打分，延迟就跟着命中数走。
-    - 修法：用索引自己当 IDF 源，按文档频率丢掉虚词（`text_index.rare_terms`，`COMMON_TERM_FRACTION = 0.05`，无手写停用词表，换语料自动换答案）。0.05 是扫出来的，扫描器已入库成证据：`bench/term_sweep.py` → `bench/reports/TERMS.md`（n=173，冻结库的副本）：0.20→0.792/385.9ms、0.10→0.792/372.9ms、**0.05→0.803/64.8ms**、0.02→0.792/30.3ms、0.01→0.775/16.8ms、0.005→0.740/9.3ms —— 召回峰值就在 0.05。早先 scratch 脚本扫出的 0.893/0.899/0.870 已被取代：那些数字没有入库的产物，且是在检索改写（候选放宽规则、向量腿入库）之前测的。为此 `rare_terms(connection, terms, *, fraction=None)` 改成调用时解析上界——def 期默认值扫不动。
+    - 修法：用索引自己当 IDF 源，按文档频率丢掉虚词（`text_index.rare_terms`，`COMMON_TERM_FRACTION = 0.05`，无手写停用词表，换语料自动换答案）。0.05 是扫出来的，扫描器已入库成证据：`bench/term_sweep.py` → `bench/reports/TERMS.md`（n=173，冻结库的副本）：0.20→0.792/420.3ms、0.10→0.792/467.5ms、**0.05→0.803/69.5ms**、0.02→0.792/33.4ms、0.01→0.775/19.2ms、0.005→0.740/10.8ms —— 召回峰值就在 0.05。早先 scratch 脚本扫出的 0.893/0.899/0.870 已被取代：那些数字没有入库的产物，且是在检索改写（候选放宽规则、向量腿入库）之前测的。为此 `rare_terms(connection, terms, *, fraction=None)` 改成调用时解析上界——def 期默认值扫不动。
     - 计数缓存 `term_df`（schema v6）：一个词只数一次，任何 atom 写入都清空它，所以缓存不会比它所描述的库更老。冷缓存 p50 73ms / 热 63ms，1,324 个词被缓存。
   - 验收：`pytest backend/tests/test_text_index.py backend/tests/test_search_api.py` 绿（18 条，真 FTS5 表 + TestClient）：「钥匙」命中「我们的钥匙放在玄关抽屉里」、`search atoms` 命中 `func SearchAtoms(ctx context.Context)`、`forg` 命中 forgotten、全角 `Ｓｅａｒｃｈ　Ａｔｏｍｓ` 命中、游标翻页无重复、supersede 后旧记忆不再应答；MCP 一致性由 `ely-eye-mcp/internal/mcpserver/search_tools_test.go`（3 条）+ `bash scripts/mcp_write_smoke.sh` PASS 证明；偏差：checklist 写的 `tests/test_retrieval_parity.py` 不存在，实际在 `backend/tests/` 与 Go 测试里
 - [x] **p0-4** 卡带变成可挂载作用域
@@ -349,12 +349,18 @@
     - `python -m bench.quant_probe` → `bench/reports/QUANT.md`（165,973 向量、200 探针、top 10，tie-aware 一致度）：sqlite-vec float 40.3ms/1.0、+holder 过滤 29.3ms/0.599、int8 35.3ms/**0.977**、int8+holder 23.0/0.595、bit 2.7ms/0.728、bit+holder 4.2/0.513、numpy float32 全扫（真值）8.0ms/1.0、numpy 符号位 hamming 19.1/0.011、float32 前 192/128/64 维 4.4/0.722、4.0/0.57、4.0/0.43、**bit 粗排@200 → numpy 里 float32 精排 23.7ms/0.983**。每向量字节 f32 1536 / int8 384 / bit 48。
     - 结论：int8 省 4× 空间几乎不掉质量（0.977）但不省时间；bit 省 32× 空间和 14× 时间、只剩 0.728，只能当粗排；两级方案必须在 vec0 **外面**精排（在库里用 `rowid IN (…)` 回捞反而更慢）。MRL 截断是在**不是 MRL 训练**的 MiniLM 上测的，所以那是真实代价不是白拿；切片必须是存下来的连续维，不是查询时切（非连续视图点集会掉速度）。
     - 偏差：精排腿（`bge-reranker-base`）早在基线消融里就有并测过（加了它 recall@10 从 0.565 掉到 0.453），所以这里没再为它编一套故事；粗排+精排这条链测出 0.983 一致度但**没接线**，因为 dense 腿默认关（见下）。
-  - [x] p2-1.3 100k 实测 + 1M 外推：`python -m bench.dense_scale` → `bench/reports/DENSE.md`（stamp cd9851a）
-    - 实测 **165,973 atom / 153,857 向量**：建库 1599.1s（96 向量/秒）、**检索 p50 74.6ms / p95 77.5ms / max 89.0ms（300 查询）**、其中编码问题 p50 11.3ms、与暴力扫描 tie-aware 一致度 **0.998**（40 探针）、**往满索引里写一条记忆 p50 514ms / p95 536ms（50 次写）**、失配扫描 1.7s（不在检索路径上）、**每向量 1,729 字节**（dbstat 数 vec0 影子表：266,059,776 B）。
-    - 1M 外推：p50 443ms / p95 463ms、向量表 1.6 GiB、首次建库 154.2 分钟、单条写 3132ms，假设随行印出（逐条扫 ⇒ 线性）。
+  - [x] p2-1.3 100k 实测 + 1M 外推：`python -m bench.dense_scale` → `bench/reports/DENSE.md`（stamp dff5977）
+    - 实测 **165,973 atom / 153,857 向量**：建库 1421.0s（108 向量/秒）、**检索 p50 75.6ms / p95 93.0ms / max 123.6ms（300 查询）**、其中编码问题 p50 12.4ms / p95 25.8ms、与暴力扫描 tie-aware 一致度 **0.998**（40 探针）、**往满索引里写一条记忆 p50 518ms / p95 543ms（50 次写）**、失配扫描 1.3s（不在检索路径上）、**每向量 1,729 字节**（dbstat 数 vec0 影子表：266,059,776 B）、机器 load [3.18, 4.58, 5.42]。
+    - 1M 外推：p50 455ms / p95 561ms、向量表 1.6 GiB、首次建库 153.9 分钟、单条写 3123ms（vec0 按 id 删除也要扫表，所以写同样随规模线性），假设随行印出（逐条扫 ⇒ 线性）。
     - tie-aware 的必要性：语料里有重复 atom，分数并列 ⇒ top-K 的 id 集合本身有歧义，两个都对引擎按 id 比会读出 0.827 的假召回。判定改成「返回的 atom 真实 cosine ≥ 第 K 名 cosine」。
-  - 验收：100k dense 腿 p50 <100ms → **165,973 atom（bar 的 1.66 倍）p50 74.6ms**（DENSE.md）；新 atom 立即可查 → `pytest backend/tests/test_vector_index.py`（写进去当场能搜到、视图藏起来的 dense 腿也藏、备份带走向量恢复后照样应答）。
-  - 数字逼出来的决定：**dense 腿默认关**（`config.py:32 dense_embeddings_enabled=False`）。同一 300 条 holder 作用域探针上，开腿 recall@10 0.832→0.845、MRR 0.683→0.663、p50 64ms→482ms。多 1.3 个点的召回、慢 7 倍、首位质量还降——默认值就该是关，腿留着（已增量、已入库）等更大的语料或更好的嵌入模型。
+  - 验收：100k dense 腿 p50 <100ms → **165,973 atom（bar 的 1.66 倍）p50 75.6ms**（DENSE.md）；新 atom 立即可查 → `pytest backend/tests/test_vector_index.py`（写进去当场能搜到、视图藏起来的 dense 腿也藏、备份带走向量恢复后照样应答）。
+  - [x] p2-1.4 每条腿值多少钱：`python -m bench.leg_check` → `bench/reports/LEGS.md`（stamp dff5977，300 条 holder 作用域探针，7 个配置，每臂带 n / MRR / Wilson 下界 / p50 / p95 / 分模态召回）
+    - 只有词的 FTS 页：问 10 个 atom recall@10 **0.832**、MRR 0.683、p50 65.5ms；问产品配置的 24 个 **0.848** / 0.688 / 72.5ms；页放宽到 30 **0.855** / 0.690 / 64.7ms（放宽不要钱，10→24 就白捡 1.6 个点）。
+    - 开 dense 腿：10 个 atom **0.845** / MRR 0.663 / p50 472.9ms；页 30 **0.858** / 0.669 / 413.7ms。多 1.3–2.6 个点召回、MRR 降、慢 6–7 倍。
+    - 开图片腿（CLIP 索引 53,668 张图、建库 2901.7s、512 维）：10 个 atom **0.815** / MRR 0.553 / 132.5ms；页 30 **0.825** / 0.551 / 162.8ms。分模态才看得见它在干什么：**video 0.583→0.708、pdf 0.849→0.907 涨，code 0.953→0.898、ui_screenshot 0.583→0.500 跌**——它救的正是金标是图的题，代价是稀释别的。
+    - 产品的图片腿不是开关：`retrieval.py:136` `if image is None and not thin: return {}`，稀疏页薄到 2 条以下才问图。所以这几行测的是「腿可用」，不是强行常开融合。
+    - **没达到的验收线，写在这里**：用户定的验收是 recall@10 ≥0.90（n≥300）。测出来最好的一行是 **0.858**（dense 腿开、页 30、p50 414ms），默认配置 **0.848**（p50 72.5ms），CURRENT.md 里按 10 个 atom 问是 0.832；仓内 Mem0/Zep 式对照在 holder 作用域是 **0.866**。差距集中在图（image 0.750、ui_screenshot 0.500–0.583、video 0.583–0.708），文字的题（code 0.953、pdf 0.919）已经过了线。诚实结论：**延迟达标（281 倍于基线、p50 71.7ms），召回没达标**，剩下 4 个点是跨模态召回，不是索引规模问题。
+  - 数字逼出来的决定：**dense 腿默认关**（`config.py:32 dense_embeddings_enabled=False`）。同一 300 条 holder 作用域探针上（LEGS.md），开腿 recall@10 0.832→0.845、MRR 0.683→0.663、p50 65.5ms→472.9ms。多 1.3 个点的召回、慢 7 倍、首位质量还降——默认值就该是关，腿留着（已增量、已入库）等更大的语料或更好的嵌入模型。
   - 顺带修的测量缺陷：`copy_store()` 只拷 SQLite，所以任何要跑视觉腿的 harness 都得自己带上 `objects/`（25 GB，探针里用符号链接）与 `cache/clip_image_index.npy`；`bench/scale_check.py` 原本直接打开冻结基线库、把它**原地迁移**了，现改成在副本上测（`--scratch`）。
 - [ ] **p2-2** 安全与隐私
   - [ ] p2-2.1 FastAPI 鉴权 + Origin 校验

bench/reports/CURRENT.md

@@ -1,21 +1,19 @@
 # CURRENT: what the product retrieves at this build
 
-- generated 2026-10-07T21:24:15+00:00 at commit dfc6cd4c2889
-- measured store schema: v5
+- generated 2026-10-08T20:17:12+00:00 at commit dff5977ab638
+- measured store schema: v8
 - corpus 50278 sources, 5308 queries (2928 answerable)
 - models: dense `sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2`, rerank `BAAI/bge-reranker-base`
-- machine: 32 cores, 4 threads per query, load average 1.2 when the report was written
+- machine: 32 cores, 4 threads per query, load average 5.1 when the report was written
 
 | arm | recall@1 | recall@5 | recall@10 | nDCG@10 | MRR | hit@10 (95% lower) | p50 s | p95 s |
 |---|---|---|---|---|---|---|---|---|
 | chance (random ranking) | 0.000 | 0.000 | 0.000 | - | - | - | - | - |
-
-_re-timed on a stride sample with warm models; recall comes from the full pass (n=600 queries)_
-| bm25_only | 0.313 | 0.465 | 0.519 | 0.415 | 0.383 | 0.524 (0.505) | 0.631 | 2.768 |
-| dense_only | 0.227 | 0.363 | 0.424 | 0.321 | 0.290 | 0.429 (0.411) | 0.081 | 0.134 |
-| graph_only | 0.000 | 0.001 | 0.001 | 0.000 | 0.000 | 0.001 (0.000) | 0.055 | 0.095 |
-| no_rerank | 0.342 | 0.510 | 0.565 | 0.452 | 0.418 | 0.571 (0.553) | 0.643 | 2.786 |
-| fusion_rerank | 0.236 | 0.391 | 0.453 | 0.341 | 0.309 | 0.459 (0.441) | 7.000 | 13.247 |
+| bm25_only | 0.313 | 0.465 | 0.519 | 0.415 | 0.383 | 0.524 (0.505) | 0.563 | 3.141 |
+| dense_only | 0.227 | 0.363 | 0.424 | 0.321 | 0.290 | 0.429 (0.411) | 1.728 | 2.883 |
+| graph_only | 0.000 | 0.001 | 0.001 | 0.000 | 0.000 | 0.001 (0.000) | 0.043 | 0.058 |
+| no_rerank | 0.342 | 0.510 | 0.565 | 0.452 | 0.418 | 0.571 (0.553) | 2.404 | 5.169 |
+| fusion_rerank | 0.236 | 0.391 | 0.453 | 0.341 | 0.309 | 0.459 (0.441) | 4.538 | 9.043 |
 
 ## recall@10 by modality
 
@@ -35,24 +33,12 @@ _re-timed on a stride sample with warm models; recall comes from the full pass (
 - no_rerank: abstains on 0% of unanswerable queries, and its top hit comes from the query's own holder for 40% of them; the rest are confident answers pulled out of somebody else's memory
 - fusion_rerank: abstains on 0% of unanswerable queries, and its top hit comes from the query's own holder for 34% of them; the rest are confident answers pulled out of somebody else's memory
 
-## the unmodified product path
+## the product path, asked as its holder
 
-- RetrievalService.search: p50 0.68s, mean 1.08s, max 9.36s over 40 queries
-- top-10 overlap with the fast BM25 path: 0.0%, identical order 0.0%
-- its top-10 atoms cover only 7.7 distinct sources (min 1): the product ranks atoms and never diversifies, so ten slots fill with fragments of one document and the overlap with a source-level ranking is near zero
-
-## latency vs corpus size
-
-| sources | atoms | bm25_only p50/p95 s | dense_only p50/p95 s |
-|---|---|---|---|
-| 500 | 500 | 0.001/0.003 | 0.018/0.031 |
-| 2000 | 2000 | 0.005/0.010 | 0.045/0.078 |
-| 10000 | 10000 | 0.035/0.110 | 0.079/0.100 |
-| 50000 | 165695 | 0.944/2.376 | 0.090/0.108 |
-
-_latency only: a source prefix drops most gold sources, so recall at these sizes is meaningless_
-
-_latency was measured at load average 1.2 on 32 cores; a busy machine inflates these figures, so read them as a ceiling_
+- RetrievalService.search as the query's own holder: recall@10 0.832, MRR 0.683, hit@10 0.833 (95% lower 0.787) over n=300 answerable queries
+- latency p50 0.074s, p95 0.146s, mean 0.079s, max 0.314s
+- top-10 overlap with the fast BM25 path: 13.4%, identical order 0.0%
+- its top-10 atoms cover 5.2 distinct sources on average (min 0): the product ranks atoms and never diversifies, so ten slots can fill with fragments of one document
 
 ## reading
 

bench/reports/DENSE.md

@@ -0,0 +1,23 @@
+# DENSE: the vector leg at corpus size (commit dff5977ab)
+
+- database: `bench/runs/dense_scale/data/ely_eye.sqlite`
+- atoms: 165973, memories with words: 153857, vectors: 153857, pending: 0
+- model: `sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2`, width 384
+- build: 1421.0s for 153857 memories (108 per second)
+- writing one memory into a full index: p50 518ms · p95 543ms over 50 writes
+- the repair scan (`ely-eye build-dense-index`) over a full index: 1.3s, and it is not on the search path
+- search p50 75.6ms · p95 93.0ms · max 123.6ms over 300 queries
+- of which encoding the question: p50 12.4ms · p95 25.8ms
+- against a brute-force scan: 0.998 of the top 10 over 40 probes (a returned atom counts when its true cosine is at least the k-th best, so ties cannot fail the check)
+- bytes per stored vector: 1729 (266059776 bytes for 153857 vectors, dbstat page sizes over the vec0 shadow tables)
+- machine: Linux 6.18.40.1-microsoft-standard-WSL2, 32 cores, 4 threads per query, load [3.18, 4.58, 5.42]
+
+## One million memories on this machine
+
+- p50 455ms · p95 561ms
+- vector table: 1.6 GiB
+- first build: 153.9 minutes
+- writing one memory: 3123ms — a vec0 delete by id scans the table, so this grows with the count too
+- assumption: sqlite-vec 0.1.9 scans every vector, so latency and bytes grow linearly with the count; the model, the machine and the thread budget stay the same
+
+_Latency is wall time per search on the machine named above with the thread budget the benchmark pins; a run on a loaded machine is not comparable._

bench/reports/LEGS.md

@@ -0,0 +1,37 @@
+# LEGS: what each retrieval leg is worth
+
+- commit: `dff5977ab638896cc8685c78ee41e5195fc0fbf2`, schema v8
+- store: bench/runs/leg_check/data/ely_eye.sqlite (a copy of bench/runs/dense_scale; the frozen baseline is never opened)
+- 300 answerable queries, each asked as the holder the frozen set assigns it
+- machine: Linux 6.18.40.1-microsoft-standard-WSL2, 32 cores, 4 threads per query, load [25.51, 9.62, 5.7]
+
+Recall is scored on the first ten **sources** after atoms are collapsed to their source
+by first hit, which is how every other arm in this repo is scored.
+
+| leg configuration | atoms asked | recall@10 | MRR | hit@10 | 95% lower | p50 ms | p95 ms | distinct sources |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| words only | 10 | 0.832 | 0.683 | 0.833 | 0.787 | 65.5 | 134.9 | 4.0 |
+| words only (the configured 24) | 24 | 0.848 | 0.688 | 0.850 | 0.805 | 72.5 | 155.9 | 6.0 |
+| words only, wider page | 30 | 0.855 | 0.690 | 0.857 | 0.812 | 64.7 | 129.7 | 6.0 |
+| + dense leg | 10 | 0.845 | 0.663 | 0.847 | 0.802 | 472.9 | 577.5 | 5.0 |
+| + dense leg, wider page | 30 | 0.858 | 0.669 | 0.860 | 0.816 | 413.7 | 525.8 | 7.0 |
+| + picture leg | 10 | 0.815 | 0.553 | 0.817 | 0.769 | 132.5 | 271.3 | 8.0 |
+| + picture leg, wider page | 30 | 0.825 | 0.551 | 0.827 | 0.780 | 162.8 | 283.7 | 20.0 |
+
+## recall@10 by modality
+
+| leg configuration | code | image | pdf | ui_screenshot | video |
+| --- | --- | --- | --- | --- | --- |
+| words only | 0.953 | 0.750 | 0.849 | 0.583 | 0.583 |
+| words only (the configured 24) | 0.953 | 0.750 | 0.895 | 0.583 | 0.625 |
+| words only, wider page | 0.953 | 0.750 | 0.919 | 0.583 | 0.625 |
+| + dense leg | 0.938 | 0.750 | 0.884 | 0.639 | 0.625 |
+| + dense leg, wider page | 0.938 | 0.750 | 0.919 | 0.639 | 0.667 |
+| + picture leg | 0.930 | 0.750 | 0.849 | 0.528 | 0.583 |
+| + picture leg, wider page | 0.898 | 0.750 | 0.907 | 0.500 | 0.708 |
+
+_Query counts per modality: code 128, image 26, pdf 86, ui_screenshot 36, video 24._
+
+The picture leg is not a switch this harness flips per row: the product runs it only when the sparse page comes back thin, so these rows measure the product with the leg available, not a forced always-on fusion.
+
+_Latency is wall time per search on the machine named above with the thread budget the benchmark pins; a run on a loaded machine is not comparable._

bench/reports/QUANT.md

@@ -0,0 +1,28 @@
+# QUANT: what each vector encoding buys (commit cd9851a12)
+
+- corpus: 165973 vectors, 384 dimensions, 200 probes, top 10
+- machine: Linux 6.18.40.1-microsoft-standard-WSL2, 32 cores, 4 threads per query, load [2.46, 4.17, 4.58]
+- probes are drawn from the corpus itself, so these arms are compared with each other, not quoted as retrieval quality; a hit is any returned atom whose true cosine is at least the 10-th best, because byte-identical atoms make the id set ambiguous
+
+| arm | p50 ms | p95 ms | recall@10 |
+| --- | --- | --- | --- |
+| sqlite-vec float | 41.2 | 55.4 | 1.0 |
+| sqlite-vec float + holder filter | 29.8 | 30.7 | 0.599 |
+| sqlite-vec int8 | 35.5 | 42.2 | 0.977 |
+| sqlite-vec int8 + holder filter | 22.3 | 28.1 | 0.595 |
+| sqlite-vec bit | 2.6 | 2.8 | 0.728 |
+| sqlite-vec bit + holder filter | 4.1 | 4.3 | 0.513 |
+| numpy float32 scan (the truth) | 8.0 | 9.6 | 1.0 |
+| numpy hamming (sign bit) | 18.9 | 19.8 | 0.011 |
+| numpy float32 first 192 dims | 7.5 | 8.2 | 0.722 |
+| numpy float32 first 128 dims | 7.9 | 8.9 | 0.57 |
+| numpy float32 first 64 dims | 4.0 | 8.3 | 0.43 |
+| bit coarse@200 -> float32 rerank in numpy | 23.7 | 24.7 | 0.983 |
+
+| encoding | bytes per vector |
+| --- | --- |
+| float32 | 1536 |
+| int8 | 384 |
+| bit | 48 |
+
+_Latency is wall time per query on the machine named above with the thread budget the benchmark pins; a run on a loaded machine is not comparable._

bench/reports/SCALE.md

@@ -1,10 +1,14 @@
 # SCALE: the product's own search at corpus size
 
-- database: bench/runs/baseline/data/ely_eye.sqlite
+- commit: `dff5977ab638896cc8685c78ee41e5195fc0fbf2`
+- database: bench/runs/scale/data/ely_eye.sqlite
+- copied from the frozen baseline in 5.0s and migrated in 0.0s to schema 8; the baseline store itself is left untouched
 - atoms: 165973, probes: 300 (code, image, pdf, ui_screenshot, video)
 - dense embeddings: False
-- opening the store (migrations included): 0.000s
-- p50 0.0570s · p95 0.1383s · max 0.5395s over 300 queries
-- machine: Linux 6.18.40.1-microsoft-standard-WSL2, 32 cores, 4 threads per query, load [1.77, 1.62, 1.64]
+- p50 0.0717s · p95 0.2177s · max 0.7805s over 300 queries
+- machine: Linux 6.18.40.1-microsoft-standard-WSL2, 32 cores, 4 threads per query, load [2.31, 3.72, 7.9]
+
+The baseline recorded p50 20.16s on this corpus because the old search pulled every
+atom into Python and rebuilt a BM25 index per query.
 
 _Latency is wall time per search call on an idle machine with the thread budget the benchmark pins; a run on a loaded machine is not comparable._

bench/reports/TERMS.md

@@ -0,0 +1,20 @@
+# TERMS: which query words are worth paying for
+
+- commit: `dff5977ab638896cc8685c78ee41e5195fc0fbf2`
+- store: bench/runs/term_sweep/data/ely_eye.sqlite (copied from the frozen baseline in 5.1s)
+- 173 answerable probes, each asked as the holder the frozen set assigns it, ranked by the unmodified `RetrievalService`
+- the bound the product ships is 0.05; the measured recall peak sits at 0.05
+- machine: Linux 6.18.40.1-microsoft-standard-WSL2, 32 cores, 4 threads per query, load [2.32, 3.87, 8.09]
+
+| bound | recall@10 | MRR | hit@10 | p50 ms | p95 ms | median distinct sources |
+| --- | --- | --- | --- | --- | --- | --- |
+| 0.200 | 0.792 | 0.629 | 0.792 | 420.3 | 1704.1 | 6 |
+| 0.100 | 0.792 | 0.632 | 0.792 | 467.5 | 1747.5 | 6 |
+| 0.050 **(shipped)** | 0.803 | 0.652 | 0.803 | 69.5 | 131.5 | 5 |
+| 0.020 | 0.792 | 0.631 | 0.792 | 33.4 | 88.4 | 4 |
+| 0.010 | 0.775 | 0.617 | 0.775 | 19.2 | 63.0 | 4 |
+| 0.005 | 0.740 | 0.563 | 0.740 | 10.8 | 69.9 | 3 |
+
+The bound is a share of the library: a word carried by more atoms than that is dropped from the OR query, because bm25 scores every row the query touches. It is relative to the corpus on purpose — a corpus of PDFs and a corpus of shell history disagree about which of their own words are noise.
+
+_Latency is wall time per search on the machine named above with the thread budget the benchmark pins; a run on a loaded machine is not comparable._

bench/reports/current.json

@@ -89,12 +89,11 @@
     }
    },
    "latency_seconds": {
-    "max": 10.484337914007483,
-    "mean": 0.9426860555750318,
-    "note": "re-timed on a stride sample with warm models; recall comes from the full pass",
-    "p50": 0.63090720949549,
-    "p95": 2.7677196743439674,
-    "queries": 600
+    "max": 70.10108924399901,
+    "mean": 1.033756499582343,
+    "p50": 0.5625866459999997,
+    "p95": 3.1407018873005645,
+    "queries": 5308
    },
    "overall": {
     "abstention_rate_answerable": 0.0,
@@ -200,12 +199,11 @@
     }
    },
    "latency_seconds": {
-    "max": 1.5199895569967339,
-    "mean": 0.09428343026663545,
-    "note": "re-timed on a stride sample with warm models; recall comes from the full pass",
-    "p50": 0.08063396950092283,
-    "p95": 0.13402395970297218,
-    "queries": 600
+    "max": 5.211815150996699,
+    "mean": 1.90850240464544,
+    "p50": 1.7281245969993506,
+    "p95": 2.8826986088994087,
+    "queries": 5308
    },
    "overall": {
     "abstention_rate_answerable": 0.0,
@@ -311,12 +309,11 @@
     }
    },
    "latency_seconds": {
-    "max": 0.11273866199917393,
-    "mean": 0.06529701035842057,
-    "note": "re-timed on a stride sample with warm models; recall comes from the full pass",
-    "p50": 0.05544837899287813,
-    "p95": 0.09468995315801294,
-    "queries": 600
+    "max": 0.09835406600177521,
+    "mean": 0.044743228177486605,
+    "p50": 0.04287036849927972,
+    "p95": 0.05790443480054818,
+    "queries": 5308
    },
    "overall": {
     "abstention_rate_answerable": 0.0,
@@ -422,12 +419,11 @@
     }
    },
    "latency_seconds": {
-    "max": 10.533955075996346,
-    "mean": 0.9619427272766673,
-    "note": "re-timed on a stride sample with warm models; recall comes from the full pass",
-    "p50": 0.6425201155034301,
-    "p95": 2.786401089788338,
-    "queries": 600
+    "max": 72.48293514099714,
+    "mean": 2.882459499187649,
+    "p50": 2.40399285599824,
+    "p95": 5.168583710499833,
+    "queries": 5308
    },
    "overall": {
     "abstention_rate_answerable": 0.0,
@@ -533,12 +529,11 @@
     }
    },
    "latency_seconds": {
-    "max": 28.03755501300475,
-    "mean": 7.456139394386567,
-    "note": "re-timed on a stride sample with warm models; recall comes from the full pass",
-    "p50": 7.000474203505291,
-    "p95": 13.246865720593268,
-    "queries": 600
+    "max": 77.63319287999548,
+    "mean": 5.166508504176891,
+    "p50": 4.5375522020012795,
+    "p95": 9.042528758149817,
+    "queries": 5308
    },
    "overall": {
     "abstention_rate_answerable": 0.0,
@@ -563,32 +558,45 @@
   "recall_at_5": 9.944707426707506e-05
  },
  "live_product": {
+  "abstention_rate_answerable": 0.016666666666666666,
+  "abstention_rate_unanswerable": 0.0,
   "distinct_sources_in_top10": {
    "max": 10,
-   "mean": 7.675,
-   "min": 1
+   "mean": 5.193333333333333,
+   "min": 0
   },
+  "hit_at_10": 0.8333333333333334,
+  "hit_at_10_wilson_lower": 0.7870033503873167,
   "identical_top10": 0.0,
-  "max": 9.363575160998153,
-  "mean": 1.0766488450000906,
-  "p50": 0.6765933140050038,
-  "queries": 40,
-  "top10_overlap_with_bm25_only": 0.0,
+  "max": 0.3143954080005642,
+  "mean": 0.07851241317994816,
+  "mrr": 0.6832447089947089,
+  "n_answerable": 300,
+  "n_unanswerable": 0,
+  "ndcg_at_10": 0.7186678703708246,
+  "p50": 0.07425041199894622,
+  "p95": 0.14573381999798585,
+  "queries": 300,
+  "recall_at_1": 0.5966666666666667,
+  "recall_at_10": 0.8316666666666667,
+  "recall_at_5": 0.7816666666666666,
+  "top10_overlap_with_bm25_only": 0.13433333333333336,
+  "top_hit_from_holder_unanswerable": 0.0,
   "top_k": 10
  },
  "meta": {
   "answerable": 2928,
   "corpus_manifest_sha256": "8fc1ed861c891d778853d41107630869f9eb267a669c5d1038e389de80114347",
   "corpus_sources": 50278,
-  "generated_at": "2026-10-07T21:24:15+00:00",
-  "git_commit": "dfc6cd4c2889e872d3f7d983b18406e6ff7a622d",
+  "generated_at": "2026-10-08T20:17:12+00:00",
+  "git_commit": "dff5977ab638896cc8685c78ee41e5195fc0fbf2",
   "limit_per_modality": null,
   "machine": {
    "cpus": 32,
    "load_average_1_5_15": [
-    1.21,
-    1.55,
-    2.28
+    5.09,
+    4.28,
+    8.06
    ],
    "release": "6.18.40.1-microsoft-standard-WSL2",
    "system": "Linux",
@@ -608,64 +616,6 @@
    "ui_screenshot": "2269dd623e67cd851fb3f6b24a4556a8462a37db98e12ad79577c5830167c7c5",
    "video": "4973231d9f16d214b14923d79a34b85c3d2b423624dea752b1ee91ba09ff778c"
   },
-  "store_schema_version_at_start": 5
- },
- "scale": [
-  {
-   "atoms": 500,
-   "bm25_only": {
-    "p50": 0.001200532500661211,
-    "p95": 0.0033308441004919584
-   },
-   "dense_only": {
-    "p50": 0.018240716999571305,
-    "p95": 0.03141984814501483
-   },
-   "note": "latency only: a source prefix drops most gold sources, so recall at these sizes is meaningless",
-   "queries": 40,
-   "sources": 500
-  },
-  {
-   "atoms": 2000,
-   "bm25_only": {
-    "p50": 0.0047556615027133375,
-    "p95": 0.009782735643602785
-   },
-   "dense_only": {
-    "p50": 0.044618372499826364,
-    "p95": 0.07758056885431869
-   },
-   "note": "latency only: a source prefix drops most gold sources, so recall at these sizes is meaningless",
-   "queries": 40,
-   "sources": 2000
-  },
-  {
-   "atoms": 10000,
-   "bm25_only": {
-    "p50": 0.03480406300150207,
-    "p95": 0.11021578034851681
-   },
-   "dense_only": {
-    "p50": 0.07940824350225739,
-    "p95": 0.09990645345169469
-   },
-   "note": "latency only: a source prefix drops most gold sources, so recall at these sizes is meaningless",
-   "queries": 40,
-   "sources": 10000
-  },
-  {
-   "atoms": 165695,
-   "bm25_only": {
-    "p50": 0.9441751839985955,
-    "p95": 2.3757928041479315
-   },
-   "dense_only": {
-    "p50": 0.08988298100302927,
-    "p95": 0.10847034014877863
-   },
-   "note": "latency only: a source prefix drops most gold sources, so recall at these sizes is meaningless",
-   "queries": 40,
-   "sources": 50000
-  }
- ]
+  "store_schema_version_at_start": 8
+ }
 }

bench/reports/dense.json

@@ -0,0 +1,73 @@
+{
+  "commit": "dff5977ab638896cc8685c78ee41e5195fc0fbf2",
+  "database": "bench/runs/dense_scale/data/ely_eye.sqlite",
+  "atoms": 165973,
+  "memories_with_words": 153857,
+  "vectors": 153857,
+  "pending": 0,
+  "dim": 384,
+  "model": "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
+  "copy_seconds": 0.665644496002642,
+  "build_seconds": 1421.0276969089973,
+  "texts_per_second": 108.2716405420302,
+  "repair_seconds": 1.2946931579936063,
+  "writes": 50,
+  "write_latency": {
+    "queries": 50,
+    "mean": 0.5204645371803781,
+    "p50": 0.5183807220018934,
+    "p95": 0.5429989652060613,
+    "max": 0.5760761429992272
+  },
+  "probes": 300,
+  "probe_modalities": [
+    "code",
+    "image",
+    "pdf",
+    "ui_screenshot",
+    "video"
+  ],
+  "latency": {
+    "queries": 300,
+    "mean": 0.07861847096656371,
+    "p50": 0.07557305500085931,
+    "p95": 0.09303550984986941,
+    "max": 0.123588470989489
+  },
+  "encode_latency": {
+    "queries": 300,
+    "mean": 0.014255492710411393,
+    "p50": 0.012439899503078777,
+    "p95": 0.025788354348333098,
+    "max": 0.0513256230042316
+  },
+  "correctness": {
+    "probes": 40,
+    "top_k": 10,
+    "tie_aware_agreement": 0.9975,
+    "note": "a returned atom counts when its true cosine is at least the k-th best, so ties cannot fail the check"
+  },
+  "bytes_per_vector": 1729.2666307025354,
+  "vector_bytes": 266059776,
+  "bytes_method": "dbstat page sizes over the vec0 shadow tables",
+  "extrapolation": {
+    "to": 1000000,
+    "assumption": "sqlite-vec 0.1.9 scans every vector, so latency and bytes grow linearly with the count; the model, the machine and the thread budget stay the same",
+    "p50_seconds": 0.45533342773137386,
+    "p95_seconds": 0.5605460517666694,
+    "write_p50_seconds": 3.1232834376789804,
+    "gibibytes": 1.6105050507025191,
+    "build_seconds": 9236.028889871746
+  },
+  "machine": {
+    "system": "Linux",
+    "release": "6.18.40.1-microsoft-standard-WSL2",
+    "cpus": 32,
+    "threads_per_query": 4,
+    "load_average_1_5_15": [
+      3.18,
+      4.58,
+      5.42
+    ]
+  }
+}

bench/reports/leg_check.json

@@ -0,0 +1,208 @@
+{
+  "commit": "dff5977ab638896cc8685c78ee41e5195fc0fbf2",
+  "schema_version": 8,
+  "store": "bench/runs/leg_check/data/ely_eye.sqlite",
+  "probes": 300,
+  "arms": [
+    {
+      "label": "words only",
+      "top_k": 10,
+      "dense": false,
+      "visual": false,
+      "recall_at_10": 0.8316666666666667,
+      "mrr": 0.6832447089947089,
+      "hit_at_10": 0.8333333333333334,
+      "wilson_lower": 0.7870033503873167,
+      "p50_seconds": 0.0654734540003119,
+      "p95_seconds": 0.13490134734311146,
+      "median_distinct_sources": 4.0,
+      "by_modality": {
+        "code": 0.953125,
+        "image": 0.75,
+        "pdf": 0.8488372093023255,
+        "ui_screenshot": 0.5833333333333334,
+        "video": 0.5833333333333334
+      },
+      "modality_counts": {
+        "code": 128,
+        "image": 26,
+        "pdf": 86,
+        "ui_screenshot": 36,
+        "video": 24
+      }
+    },
+    {
+      "label": "words only (the configured 24)",
+      "top_k": 24,
+      "dense": false,
+      "visual": false,
+      "recall_at_10": 0.8483333333333334,
+      "mrr": 0.6875571326821327,
+      "hit_at_10": 0.85,
+      "wilson_lower": 0.8051823888546709,
+      "p50_seconds": 0.07251374649786158,
+      "p95_seconds": 0.15588207065375176,
+      "median_distinct_sources": 6.0,
+      "by_modality": {
+        "code": 0.953125,
+        "image": 0.75,
+        "pdf": 0.8953488372093024,
+        "ui_screenshot": 0.5833333333333334,
+        "video": 0.625
+      },
+      "modality_counts": {
+        "code": 128,
+        "image": 26,
+        "pdf": 86,
+        "ui_screenshot": 36,
+        "video": 24
+      }
+    },
+    {
+      "label": "words only, wider page",
+      "top_k": 30,
+      "dense": false,
+      "visual": false,
+      "recall_at_10": 0.855,
+      "mrr": 0.6895710215710216,
+      "hit_at_10": 0.8566666666666667,
+      "wilson_lower": 0.8124994228503402,
+      "p50_seconds": 0.06468957299512113,
+      "p95_seconds": 0.1297007056004077,
+      "median_distinct_sources": 6.0,
+      "by_modality": {
+        "code": 0.953125,
+        "image": 0.75,
+        "pdf": 0.9186046511627907,
+        "ui_screenshot": 0.5833333333333334,
+        "video": 0.625
+      },
+      "modality_counts": {
+        "code": 128,
+        "image": 26,
+        "pdf": 86,
+        "ui_screenshot": 36,
+        "video": 24
+      }
+    },
+    {
+      "label": "+ dense leg",
+      "top_k": 10,
+      "dense": true,
+      "visual": false,
+      "recall_at_10": 0.845,
+      "mrr": 0.6632195767195768,
+      "hit_at_10": 0.8466666666666667,
+      "wilson_lower": 0.8015339429418983,
+      "p50_seconds": 0.4729458744986914,
+      "p95_seconds": 0.5775450093562541,
+      "median_distinct_sources": 5.0,
+      "by_modality": {
+        "code": 0.9375,
+        "image": 0.75,
+        "pdf": 0.8837209302325582,
+        "ui_screenshot": 0.6388888888888888,
+        "video": 0.625
+      },
+      "modality_counts": {
+        "code": 128,
+        "image": 26,
+        "pdf": 86,
+        "ui_screenshot": 36,
+        "video": 24
+      }
+    },
+    {
+      "label": "+ dense leg, wider page",
+      "top_k": 30,
+      "dense": true,
+      "visual": false,
+      "recall_at_10": 0.8583333333333333,
+      "mrr": 0.6686637344137344,
+      "hit_at_10": 0.86,
+      "wilson_lower": 0.8161683876490509,
+      "p50_seconds": 0.41367531901050825,
+      "p95_seconds": 0.5258364057008293,
+      "median_distinct_sources": 7.0,
+      "by_modality": {
+        "code": 0.9375,
+        "image": 0.75,
+        "pdf": 0.9186046511627907,
+        "ui_screenshot": 0.6388888888888888,
+        "video": 0.6666666666666666
+      },
+      "modality_counts": {
+        "code": 128,
+        "image": 26,
+        "pdf": 86,
+        "ui_screenshot": 36,
+        "video": 24
+      }
+    },
+    {
+      "label": "+ picture leg",
+      "top_k": 10,
+      "dense": false,
+      "visual": true,
+      "recall_at_10": 0.815,
+      "mrr": 0.5525886243386243,
+      "hit_at_10": 0.8166666666666667,
+      "wilson_lower": 0.7689713823226987,
+      "p50_seconds": 0.13248339149868116,
+      "p95_seconds": 0.2713221473451995,
+      "median_distinct_sources": 8.0,
+      "by_modality": {
+        "code": 0.9296875,
+        "image": 0.75,
+        "pdf": 0.8488372093023255,
+        "ui_screenshot": 0.5277777777777778,
+        "video": 0.5833333333333334
+      },
+      "modality_counts": {
+        "code": 128,
+        "image": 26,
+        "pdf": 86,
+        "ui_screenshot": 36,
+        "video": 24
+      }
+    },
+    {
+      "label": "+ picture leg, wider page",
+      "top_k": 30,
+      "dense": false,
+      "visual": true,
+      "recall_at_10": 0.825,
+      "mrr": 0.5507809648412589,
+      "hit_at_10": 0.8266666666666667,
+      "wilson_lower": 0.7797738592974309,
+      "p50_seconds": 0.16283495400421089,
+      "p95_seconds": 0.2837207488548302,
+      "median_distinct_sources": 20.0,
+      "by_modality": {
+        "code": 0.8984375,
+        "image": 0.75,
+        "pdf": 0.9069767441860465,
+        "ui_screenshot": 0.5,
+        "video": 0.7083333333333334
+      },
+      "modality_counts": {
+        "code": 128,
+        "image": 26,
+        "pdf": 86,
+        "ui_screenshot": 36,
+        "video": 24
+      }
+    }
+  ],
+  "machine": {
+    "system": "Linux",
+    "release": "6.18.40.1-microsoft-standard-WSL2",
+    "cpus": 32,
+    "threads_per_query": 4,
+    "load_average_1_5_15": [
+      25.51,
+      9.62,
+      5.7
+    ]
+  }
+}

bench/reports/quant.json

@@ -0,0 +1,110 @@
+{
+  "commit": "cd9851a12c000199ce5c211f937ef6c1e0138efa",
+  "vectors": 165973,
+  "dimensions": 384,
+  "probes": 200,
+  "top_k": 10,
+  "recall_note": "probes are drawn from the corpus itself, so these arms are compared with each other, not quoted as retrieval quality; a hit is any returned atom whose true cosine is at least the 10-th best, because byte-identical atoms make the id set ambiguous",
+  "bytes_per_vector": {
+    "float32": 1536,
+    "int8": 384,
+    "bit": 48
+  },
+  "arms": [
+    {
+      "arm": "sqlite-vec float",
+      "queries": 200,
+      "p50_ms": 41.2,
+      "p95_ms": 55.4,
+      "recall_at_10": 1.0
+    },
+    {
+      "arm": "sqlite-vec float + holder filter",
+      "queries": 200,
+      "p50_ms": 29.8,
+      "p95_ms": 30.7,
+      "recall_at_10": 0.599
+    },
+    {
+      "arm": "sqlite-vec int8",
+      "queries": 200,
+      "p50_ms": 35.5,
+      "p95_ms": 42.2,
+      "recall_at_10": 0.977
+    },
+    {
+      "arm": "sqlite-vec int8 + holder filter",
+      "queries": 200,
+      "p50_ms": 22.3,
+      "p95_ms": 28.1,
+      "recall_at_10": 0.595
+    },
+    {
+      "arm": "sqlite-vec bit",
+      "queries": 200,
+      "p50_ms": 2.6,
+      "p95_ms": 2.8,
+      "recall_at_10": 0.728
+    },
+    {
+      "arm": "sqlite-vec bit + holder filter",
+      "queries": 200,
+      "p50_ms": 4.1,
+      "p95_ms": 4.3,
+      "recall_at_10": 0.513
+    },
+    {
+      "arm": "numpy float32 scan (the truth)",
+      "queries": 200,
+      "p50_ms": 8.0,
+      "p95_ms": 9.6,
+      "recall_at_10": 1.0
+    },
+    {
+      "arm": "numpy hamming (sign bit)",
+      "queries": 200,
+      "p50_ms": 18.9,
+      "p95_ms": 19.8,
+      "recall_at_10": 0.011
+    },
+    {
+      "arm": "numpy float32 first 192 dims",
+      "queries": 50,
+      "p50_ms": 7.5,
+      "p95_ms": 8.2,
+      "recall_at_10": 0.722
+    },
+    {
+      "arm": "numpy float32 first 128 dims",
+      "queries": 50,
+      "p50_ms": 7.9,
+      "p95_ms": 8.9,
+      "recall_at_10": 0.57
+    },
+    {
+      "arm": "numpy float32 first 64 dims",
+      "queries": 50,
+      "p50_ms": 4.0,
+      "p95_ms": 8.3,
+      "recall_at_10": 0.43
+    },
+    {
+      "arm": "bit coarse@200 -> float32 rerank in numpy",
+      "queries": 200,
+      "p50_ms": 23.7,
+      "p95_ms": 24.7,
+      "recall_at_10": 0.983
+    }
+  ],
+  "machine": {
+    "system": "Linux",
+    "release": "6.18.40.1-microsoft-standard-WSL2",
+    "cpus": 32,
+    "threads_per_query": 4,
+    "load_average_1_5_15": [
+      2.46,
+      4.17,
+      4.58
+    ]
+  }
+}

bench/reports/scale.json

@@ -1,7 +1,10 @@
 {
-  "database": "bench/runs/baseline/data/ely_eye.sqlite",
+  "commit": "dff5977ab638896cc8685c78ee41e5195fc0fbf2",
+  "database": "bench/runs/scale/data/ely_eye.sqlite",
   "atoms": 165973,
-  "migration_seconds": 0.00044073900062358007,
+  "copy_seconds": 4.996028091998596,
+  "schema_version": 8,
+  "migration_seconds": 0.03624351899634348,
   "dense_embeddings_enabled": false,
   "probes": 300,
   "probe_modalities": [
@@ -13,10 +16,10 @@
   ],
   "latency": {
     "queries": 300,
-    "mean": 0.06524801251696771,
-    "p50": 0.057020749500225065,
-    "p95": 0.13832052519683205,
-    "max": 0.5395459059946006
+    "mean": 0.0888331904130852,
+    "p50": 0.07169728449662216,
+    "p95": 0.21773518459958724,
+    "max": 0.7804536300027394
   },
   "machine": {
     "system": "Linux",
@@ -24,9 +27,9 @@
     "cpus": 32,
     "threads_per_query": 4,
     "load_average_1_5_15": [
-      1.77,
-      1.62,
-      1.64
+      2.31,
+      3.72,
+      7.9
     ]
   }
 }

bench/reports/term_sweep.json

@@ -0,0 +1,81 @@
+{
+  "commit": "dff5977ab638896cc8685c78ee41e5195fc0fbf2",
+  "store": "bench/runs/term_sweep/data/ely_eye.sqlite",
+  "copy_seconds": 5.062254256001324,
+  "probes": 173,
+  "shipped_fraction": 0.05,
+  "peak_fraction": 0.05,
+  "arms": [
+    {
+      "fraction": 0.2,
+      "shipped": false,
+      "recall_at_10": 0.791907514450867,
+      "mrr": 0.6286195981282686,
+      "hit_at_10": 0.791907514450867,
+      "p50_seconds": 0.4202946660007001,
+      "p95_seconds": 1.7041233958021622,
+      "median_distinct_sources": 6
+    },
+    {
+      "fraction": 0.1,
+      "shipped": false,
+      "recall_at_10": 0.791907514450867,
+      "mrr": 0.6315097715386733,
+      "hit_at_10": 0.791907514450867,
+      "p50_seconds": 0.46752902700245613,
+      "p95_seconds": 1.7474705398024528,
+      "median_distinct_sources": 6
+    },
+    {
+      "fraction": 0.05,
+      "shipped": true,
+      "recall_at_10": 0.8034682080924855,
+      "mrr": 0.6520300027525461,
+      "hit_at_10": 0.8034682080924855,
+      "p50_seconds": 0.069478709003306,
+      "p95_seconds": 0.13147423019836424,
+      "median_distinct_sources": 5
+    },
+    {
+      "fraction": 0.02,
+      "shipped": false,
+      "recall_at_10": 0.791907514450867,
+      "mrr": 0.6311244150839527,
+      "hit_at_10": 0.791907514450867,
+      "p50_seconds": 0.03335734199936269,
+      "p95_seconds": 0.08840757200086956,
+      "median_distinct_sources": 4
+    },
+    {
+      "fraction": 0.01,
+      "shipped": false,
+      "recall_at_10": 0.7745664739884393,
+      "mrr": 0.61696256537297,
+      "hit_at_10": 0.7745664739884393,
+      "p50_seconds": 0.019158394003170542,
+      "p95_seconds": 0.06302069279918217,
+      "median_distinct_sources": 4
+    },
+    {
+      "fraction": 0.005,
+      "shipped": false,
+      "recall_at_10": 0.7398843930635838,
+      "mrr": 0.5633659051289108,
+      "hit_at_10": 0.7398843930635838,
+      "p50_seconds": 0.01079572799790185,
+      "p95_seconds": 0.06986973820021376,
+      "median_distinct_sources": 3
+    }
+  ],
+  "machine": {
+    "system": "Linux",
+    "release": "6.18.40.1-microsoft-standard-WSL2",
+    "cpus": 32,
+    "threads_per_query": 4,
+    "load_average_1_5_15": [
+      2.32,
+      3.87,
+      8.09
+    ]
+  }
+}