The reviewer's round 2 verified all nine round-1 corrections by re-measuring
them, then failed the commit on the one thing it added that the repo's own bar
forbids: a two-line docstring on _TinyDense. p1-6.4 still lists "comments to one
line" as open work, so a new commit should not widen it. The 470MB rationale
already lives in CHECKLIST.md.
Also cites the CI run for the commit under review (37744547992 @ 20c0416, also
success) instead of stopping at the previous one.
Nine claims in CHECKLIST.md did not reproduce. All nine were prose, not
product, and all nine are now the measured value:
- test files: 19 (18 test_*.py + conftest.py), not 24
- suite size: 253 passed (239 at 922d012), not 252
- build context: 81G on this machine (bench/runs 60.2GB, bench/corpus 20.7GB,
.venv 2.5GB), not 19GB
- profiles.py probes: linux_package_versions catches the 3-tuple because
json.loads needs ValueError; runtime_launcher_probe and gpu_available catch
the 2-tuple. The old sentence claimed one tuple for all three.
- training.py: the CUDA raise is at :1271, not :1272
- cli.py is 490 lines at this commit, not 488
- the bf16 falsifier said "a 4090 that does not support bf16". A 4090 is Ada
(compute 8.9) and does support BF16; the test fabricates the flag. Naming the
card was exactly the error this commit removes, so the sentence now names the
flag.
- the backup digest is a fingerprint of one store snapshot: hashing is
deterministic per state, but the store was rewritten after the commit, so the
current digest differs (498b0cd9...). The record now pins mtime, size and row
counts next to the digest.
- wall times moved between runs (backup 4.0s vs 2.24s, restore 2.8s vs 5.65s),
so they are reported as a range with the reason instead of a single number.
CI was a workflow that could not run its own tests: pytest is not a
dependency of anything, so a fresh clone fails with "No module named
pytest". Test tooling now lives in the backend's dev extra, so CI and a
new contributor install the same tools, and ruff is pinned there instead
of drifting as a bare version in a workflow step.
backup, restore and doctor move the store. Backup uses SQLite's own
online API so a running server is safe, and the manifest records the
schema version, row counts, sha256 and size read back from the copy.
Restore verifies the digest, integrity_check, required tables and schema
age before it touches the live file, and keeps what it replaces as
.before-restore.
Doctor separates broken from degraded. On the 906 MB baseline store
(165,973 atoms) it reports no problems and one true warning: 21,362
picture atoms with no CLIP snapshot configured, so those atoms are
findable only by the text already on them. The first version called that
store healthy because it only complained about paths that were set and
missing, never about a sense the store needs and never had.
AGPL-3.0-only, because PyMuPDF is dual-licensed AGPL-or-commercial and
the repository is public. A permissive license needs the Artifex
commercial license bought first; the README says so rather than hiding it.
Gates: 253 tests pass; ruff clean repo-wide (17 errors at HEAD); mypy
unchanged from HEAD on touched files; pyright 0 errors on the new files;
backup 4.0s and restore 2.8s on real data with counts verified equal.
The reviewer ran its own probes and mutation tests rather than reading the
prose, and confirmed the repo rules. One of its observations, that mount
state is per-connection, I disproved by measurement: mounts live in a table
and a second Database object on the same home reads them back, so the test
asserts an empty mount list only because that test mounts nothing.
Round three of the adapter review found two leftovers from the previous fix.
With three or more live runtimes the reason read "another live model still
holds adapter-a and adapter-b", a singular subject attributing several
adapters to one model, when each runtime attaches one adapter per kind. The
subject now agrees in number with what is held.
LoadRecord still stored its own copy of the manifest sha256, which is the
same second source of truth this work removed for kind, adapter_id and path:
the only constructor copied it, and nothing stopped a caller from passing a
hash that disagreed with the source handed in. It is a property now, so
runtime.py keeps reading record.sha256 and there is one value.
Also recorded, measured by the reviewer and deliberately not treated as
defects: a row for a kind with nothing on disk now reports base_model as null
rather than the runtime's model id, since that row describes no adapter at
all; the same adapter id can appear in both a loaded row and a shadowed row
when it is attached while another copy wins the mount, and both rows are
true; and when the attached adapter is deleted from disk the row says missing
with null identity instead of claiming a gone model.
Three tests carry this: the three-runtime case asserts the plural wording, and
both fixes were mutation-checked (dropping the plural branch fails two tests,
restoring the stored hash fails the generation test).
pytest -q: 228 passed. ruff clean. mypy cold cache on the three touched source
files: 2 errors, identical to the stashed HEAD tree, net zero. pyright: 2
errors, both the known qwen_vl_utils tuple artifact.
The report row contradicted itself. The reason came from the attachment
record while path, adapter_id, base_model, origin and cartridge_id came from
whatever resolution preferred right now. Reproduced without any stubbing: a
live model holds adapter X, training writes a newer adapter Y, and the
endpoint then answers state=loaded adapter_id=adapter-Y while explaining
"attached to the live model from .../x-attached". A JSON consumer reads the
identity columns as the loaded adapter and gets the wrong one.
The same root cause had a second symptom. Attachments were keyed by kind with
last-writer-wins, so with two live runtimes one unload made the row claim
"that model is gone" while the other runtime's model was still alive holding
its adapter.
LoadRecord now carries the AdapterSource it came from, so a row built from a
record has exactly one identity, and the registry is a list of attachments
each holding its own weak reference: loaded means some live model still holds
an adapter, and when more than one does the reason names the other.
clear_loads had no production caller once unload stopped erasing the process
record, so it is gone rather than kept as a test-only entry point.
Two tests pin both symptoms, and both were mutation-checked: restoring the
identity to the resolved source fails the first, and letting
live_attachments ignore liveness fails the second and the unload test.
pytest -q: 227 passed. ruff clean. mypy cold cache on the three touched
source files: 2 errors, byte-identical to the same command on a stashed HEAD
tree, net zero. pyright: 2 errors, both the known qwen_vl_utils tuple
artifact, unchanged from HEAD.
Independent review of the adapter resolution work found six defects, all
real and all fixed here.
A kind whose only candidates were rejected had no winning path, so its row
carried path=None and the rejected candidate fell into the shadow loop: the
report said "another adapter of this kind wins for the current mount" while
nothing won, and the real deferred reason was buried. A shadow row now
requires a winner and names it.
"loaded" was backed by a process-global record, so a fresh runtime with no
model still claimed a live adapter. The record is now held by the instance
that attached it, and the report keeps a weak reference: while the model
lives the row says loaded, once it is gone the row says available and
explains that the model is gone.
The sort key had lost the manifest's cartridge tag, which changed which
adapter loads on any machine holding a tagged library copy, and neither the
commit message nor the checklist said so. The tag is back as the middle key.
Proof checks resolved through the library while generation resolved through
the mount, so one proof payload could name two different copies. Both now
resolve the way generation does.
loaded_at was written and never read, so it is gone. The checklist claim that
peft is lazily imported in one place was true only for generation and is now
scoped that way.
The four generation paths and status() had no test coverage, so "no function
was lost" rested on reading. One test now walks the transformers prompt path
end to end with stub weights and asserts the generation answer carries the
adapter it attached.
pytest -q: 225 passed. ruff clean. mypy on the three touched source files:
2 errors, the identical 2 on a stashed HEAD tree, net zero. pyright: 2
errors, both the known qwen_vl_utils tuple artifact, unchanged from HEAD.
runtime.py resolved adapters at 7 call sites and only ever looked in the
adapter library, so an adapter attached to a cartridge (cartridge.py:230
copies it into <cartridge>/adapters/{kind}/) was never loaded by anything.
Resolution now lives in adapters.py: it reads the mounted cartridge set,
prefers a mounted cartridge's copy over the library, and answers every one
of the six kinds with a state the facts support — loaded, available,
deferred (base_model mismatch or missing weights, both models named in the
reason), missing — plus a shadowed row for each adapter that lost the
choice. A mounted cartridge is the scenario selection; neither manifest
carries a scenario field, so none was invented.
Two bugs the new tests caught: the preference sort lost reverse=True, so a
library adapter beat a mounted cartridge's copy, and the load branch had
never been executed by a test. Stub weights now prove peft is asked to
attach, that the returned model is put in eval mode, and that the endpoint
reports loaded until unload clears it.
Faces: GET /api/runtime/adapters, `ely-eye adapters`, and a log line naming
the reason whenever a generation runs without an adapter.
Gates: 222 passed; ruff clean; mypy cold cache on the 7 touched files gives
the same 2 errors as HEAD's runtime.py, and pyright gives the same 2 as
HEAD — both are qwen_vl_utils typing artifacts (process_vision_info returns
2 or 3 values depending on return_video_kwargs; the product takes 2).
The reviewer ran the harness and the gates itself rather than trusting the
prose: round one failed on 11 defects, round two on 4, round three passed
after re-running bench.asr_check in a fresh process, re-checking the stamp
chain and matching 18 checklist assertions.
The report is regenerated after 10a39c5 so its stamp names the code it
describes. Speed is quoted as a range in the checklist because transcribe
seconds move with machine load (3.26s then 3.12s on this box) while WER,
denominators and word counts do not move at all.
Numbers: 258.18s of audio, WER 0.242 over 533 reference words, 80 audio atoms
and 563 word timestamps read back from the store, 0 out of order and 0 before
their own segment measured on the stored words.
A second review found three numbers left behind when the report was
regenerated with pinned decoding:
- the transcriber docstring claimed 1.7 s for 88.9 s of speech; the report
says 1.39 s, and 1.7 came from the unpinned run. The docstring now names no
number, so it cannot go stale, and it stops calling container seconds speech
with voice activity detection off.
- the checklist still quoted 3.55 s and 61.6x / 73.0x / 97.7x; the frozen
report says 3.26 s and 63.9x / 82.4x / 112.2x.
- the checklist still said 31 speech segments for one mp4; with greedy decoding
that video yields 38, which is where the store's 80 atoms come from
(38 + 39 + 3).
The report also credited the store with timing invariants that were measured on
the model's answer. The harness now recomputes them from the words as stored
(out_of_order_in_store, before_their_segment_in_store) and says so.
Verification: 212 passed, ruff clean, mypy clean on both touched files with a
cold cache, pyright 0 errors.
A review of 5289cb4 found the ASR report was not honest arithmetic:
- "audio s" was the last segment's end, not the media length. The real
container durations are 88.93 / 109.18 / 60.07 s, so the report understated
the audio by 48 s and every speed cell was wrong.
- Decoding was unpinned, so a rerun was a new draw: WER moved 0.242 to 0.262
and one video moved 0.278 to 0.583. Greedy single-temperature decoding with
no carry-over now answers the same way twice, and the harness runs the pass
twice to show it.
- The report claimed its WER was an upper bound. It is not a bound in either
direction: a caption that drops words also hides words the model dropped.
The denominators (201 / 296 / 36 reference words) are printed instead.
- "66 speech atoms" was bench-side arithmetic. The store now says 80 audio
atoms carrying 563 word timestamps, read back after a real ingest.
Same silent-skip bug in a second place: a text file over max_text_file_bytes
returned an empty batch with no reason. It raises FileTooLarge with its size,
and a walk records that refusal like any other.
Also fixed from the same review: the Go search tool never published the audio
modality, cli_senses carried a second json_dump, the no-audio-track branch had
a test that fed it markdown, _seconds was an empty wrapper, and the docstring
quoted a speed no report measured.
Gate root cause: bench/tests imports the bench package by name with no root
conftest, so bare pytest could not collect it at all. pythonpath in pyproject
fixes both launch styles; 212 tests pass either way.
Verification: 212 passed (bare pytest and python -m pytest), ruff clean,
mypy clean on the six touched files with a cold cache, repo-wide mypy 24/6
unchanged from HEAD, pyright 0 errors on the touched files, go build/vet/test
and gofmt clean.
An audio file used to produce no memory at all: _ingest_file returned an empty
batch for any suffix it did not know, with no log and no reason. It now raises
UnsupportedFile when you name the file and records a SkippedFile when it was a
directory walk, so a refusal is visible either way.
Speech is measured, not promised. Over 210.5 seconds of real Video-MME audio,
faster-whisper-tiny on CPU int8 transcribed at 51x realtime and reached word
error rate 0.242 against community subtitles nobody here wrote (a caption track
drops words a transcript keeps, so that is an upper bound). 66 speech atoms now
carry 550 word-level timestamps, 0 words out of order and 0 starting before
their own segment. The snapshot is 78.2 MB, under the 500 MB cap, and the report
is bench/reports/ASR.md.
Two library facts were measured rather than assumed. faster-whisper's own audio
loader is broken against PyAV 19 (av.open got an unexpected keyword
'metadata_errors') and it declares av with no version bound, so asr.decode_audio
decodes with PyAV directly and hands the model a float32 array instead of
pinning an old av with no cp314 wheel. And av.Error does not exist on PyAV 19 -
the real base is av.error.FFmpegError - so that is the only thing caught and my
own ValueErrors keep their messages.
Files over the 500-line rule were split rather than excused: ingestion.py (748)
into ingestion_common / ingestion_media / ingestion_repository / ingestion, and
cli.py (544) into cli_senses plus cli_common, following the cli_transactive
register pattern. MediaIngest declares _source_record abstract, so a half-wired
host fails at construction instead of returning None.
Verification: 207 tests pass; ruff clean on the 13 touched files; mypy clean on
them, and repo-wide mypy fell from 29 errors to 24 (5 fixed, none added);
pyright reports 0 errors on the touched files.
A screenshot with no OCR-able words used to be a memory nothing could
reach. It now has a vector, and a question can be asked with it.
- picture space: Xenova/chinese-clip-vit-base-patch16 fp16 ONNX, 361MB,
sha256 0fad978faa89b85f…, CPU-runnable on the onnxruntime the repo
already ships. Text 76ms, image 29ms/image at batch 8. The ONNX contract
was measured, not assumed: it is the fused contrastive graph, so the
unused tower is fed a placeholder row, and the tokenizer must be
BertTokenizer — a CLIP tokenizer emits ids past the embedding table.
- visual_index.py keeps picture vectors in their own family beside the
text ones, keyed by atom id, and refuses to answer when the map and the
array disagree. retrieval.py stays under 500 lines.
- a search can be asked with a picture, and when the words come up thin
the picture ranking joins the fusion — but only then, so a text query
never pays for the encoder.
- visual tokens are counted at 32x32 patches (migration v7): 21,325 image
atoms went from 5,331 to 34,618 tokens, and the library from 29.9M to
84.2M. The 100M capsule fits this 50,278-file corpus and nothing much
larger — that is what the claim actually means.
- the vision-language describer is written and gated: with no snapshot an
atom records "deferred: no vision-language model snapshot"; a broken
model becomes a note on the atom instead of a dead ingest.
- forget now voids the picture index and the copy this process holds, so
a forgotten picture stops answering from RAM.
Measured, six arms, ScreenSpot-v2 (554 icon crops, 408 whole screenshots),
bench/reports/VISUAL.md:
icon pool picture 0.208 words 0.175 chance 0.018
screenshot pool words 0.384 picture 0.215
screenshot to icon 0.108 — CLIP is a whole-image model, not a grounding
model: it cannot point at one small icon inside a screen.
within one screen MRR 0.908 (recall@10 is 1.000 there only because
chance is 1.000 too, so MRR is the only honest number in that row).
ScreenSpot-v2 instructions are all English, so this run does not test the
bilingual claim and the report says so.
Gates: 199 tests, ruff clean on 17 files, mypy clean on 16 (fresh cache),
the repo's own pyright 0 errors on the 12 files touched.
The product's own search answered in p50 20.16s on 165,973 atoms. The SQL
rewrite brought that to 0.68s; this brings it to 57ms and says why the rest was
there.
The cost was not SQL. An OR MATCH scores every row it touches, and a question
carries words the whole library carries: `"parsing"* OR "error"* OR "in"* OR
"the"* OR "pdf"* OR "converter"*` matched 127,145 atoms and took 305ms, while a
CJK bigram phrase query with 32 matches took 1.0ms. Latency tracked match count.
- text_index.rare_terms asks the index which of its own words are noise: count
each term, drop the ones over 5% of the library, keep the rarest few when all
of them are common. No stop-word list; a corpus of PDFs and a corpus of shell
history get different answers. 0.05 is swept, not chosen — on n=173
recall@10 peaks there (0.899 / p50 58.6ms) against 0.893 / 112.9ms at 0.20
and 0.844 / 13.4ms at 0.01.
- The bound is a cost, not a claim about rarity: `cap = max(int(total*fraction)+1, 2)`,
so a word matching one atom in eleven is kept — scoring one candidate is free.
A capped count cannot rank words that all saturate, so those few are counted
exactly.
- Schema v6 adds `term_df`: a word is counted once and the next question pays
nothing. Every atom write clears it, so a stored count cannot outlive the
library it described. Cold p50 73ms, warm 63ms, 1,324 terms cached.
- The denominator is `count(*) FROM atoms` (0.01ms) rather than
`count(*) FROM atom_fts` (37ms cold): FTS rows are written and deleted with
their atoms, so the two counts move together and only one of them is free.
SCALE.md, idle machine (load 1.76, 32 cores, 4 threads per query):
165,973 atoms — 1.66x the contract's 100k — 300 probes, p50 57ms, p95 138ms.
Against the frozen baseline of 20.16s that is 354x. Holder-scoped (n=173):
recall@10 0.870, MRR 0.740, p50 62ms.
Verification: pytest backend/tests bench/tests -q -> 173 passed (5 new
text_index tests: the common word is dropped and the rare one still answers, an
all-common query keeps its rarest words in order, a one-word query is never
filtered, an empty library keeps everything, and a write voids the counts);
ruff clean; mypy fresh cache clean; pyright 0 errors on the touched files.
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.
A memory system that only answers from its own store is a filing cabinet,
not a partner. The transactive rule the project is named for — not knowing
is fine when you know who does — needed a holder registry, a directory
query, and a referral that never leaks the text it may not read.
- schema v5 adds `holders` (person/agent/service, declared topics, contact).
A holder with memories exists whether or not anybody registered them; the
registry adds reachability, it does not grant existence.
- `who_knows` ranks holders by their own memories and is deliberately NOT
filtered by visibility: "she knows" is the answer. The answer is filtered.
- `ask` searches the asker's visible scope first (hop 0), then walks the
directory with a visited set and a hop budget. Three terminal states:
answered, referred, no_holder_knows. Nobody knowing is a result, not an error.
- Confidence is ordinal, not calibrated: relative strength, coverage, and
declared topics. Every hit carries the raw numbers and the basis lines that
produced the score, so a ranking can be argued with.
- Faces: HTTP (/api/holders, /api/who_knows, /api/ask), CLI (holders,
holder-set, holder-rm, who-knows, ask), MCP (who_knows, ask, list_holders).
Four bugs found by running the thing, not by reading it:
- FTS5 refuses bm25() inside MAX(); the aggregate uses FTS5's own `rank` column.
- A CJK question is longer than the sentence that answers it, so the whole-run
phrase matched nothing. The query now ORs adjacent bigram phrases; adjacency
is kept, equal length is not. Latin path untouched, baseline unaffected.
- The hop budget counted the asker as a hop, so max_hops=1 asked nobody.
- A second, stricter readability test in _answered() disagreed with the SQL
visibility clause and hid a memory the asker was allowed to read. The SQL
clause is the only authority now.
A method named `list` shadowed the builtin inside its own class and made
mypy reject the return annotation; renamed rather than cast around it.
Proof: 155 pytest, 12 Go tests, ruff and mypy clean on touched files, pyright
ledger unchanged at 62, and scripts/mcp_write_smoke.sh drives two principals
on one SQLite file — the husband asks through MCP, gets `referred` with his
wife's contact channel, and her private sentence appears nowhere in the reply.
The product path pulled every atom into Python and rebuilt a BM25Okapi
index per query: p50 20.16s over 165,973 atoms. It now runs one SQLite
statement with LIMIT/OFFSET.
- text_index.py: one normalizer for both sides of the index. CJK runs are
split per character (unicode61 cannot segment them), camelCase and
separators are split so "search atoms" finds SearchAtoms, latin tokens
get a prefix match. Rows written before this existed hold raw text in
atom_fts, so schema v3 rebuilds the index on upgrade.
- retrieval.py: filters (holder / scenario / source / modality / since /
until / min_trust), a base64 cursor, and reciprocal rank fusion in place
of the 0.55/0.35/0.10 blend whose graph term was always zero.
- An empty result is {status: "no_match"}, not a bare array.
- Go: search_atoms and compile_context rank through POST /api/search. The
second Go tokenizer is deleted; two implementations could drift, which
is what the MCP/HTTP parity requirement forbids.
- db.py passed 500 lines, so the row codec moved to row_codec.py.
- Tooling: installed the repo's own pyright. pyrightconfig.json had no
exclude and swept the downloaded corpus (1171 errors); with excludes the
count is 62, all pre-existing (evals 48, profiles 4, training 4,
ingestion 3, runtime 2, cartridge_assets 1). backend/pyproject.toml
carried a second [tool.pyright] that shadowed the root config; deleted.
- bench/scale_check.py measures the product search at corpus size. The
100k-atom p50 number is not written until it is measured on an idle
machine.
Verification: 112 pytest passed; go build/vet/test clean, gofmt clean;
ruff clean on every touched file; mypy clean on the 14 touched Python
files; repo pyright 62 errors, unchanged per-file ledger.
An EvidenceAtom could say what was remembered and nothing else: not whose memory it
is, not where it came from, not when it was true, not in what situation it may be
used. The baseline measured the cost of that: the same ranker scores recall@10 0.412
over the whole corpus and 0.866 inside one holder's memory, so the missing holder
column is worth more than any ranking change in the ablation.
- schemas.py: MemoryOrigin, TimePrecision, Visibility; EvidenceAtom gains holder_id,
origin, scenario, event_time / event_time_end / event_time_precision /
event_time_source, valid_from / valid_to, supersedes, visibility, retention_until.
`time` is renamed `ingested_at` — ingestion time gets one name, not two.
A validator refuses valid_to before valid_from, event_time_end before event_time,
and an atom that supersedes itself.
- migrations.py: schema v2 with PRAGMA user_version. A first-version database is
upgraded in place (ALTER + RENAME COLUMN), rows intact, and a database newer than
this build refuses to open instead of being silently downgraded.
- v_atoms_current hides memories that are closed, lapsed or not yet begun; unknown
bounds read as visible. Time travel is db.atoms_as_of(moment, holder) because a
SQLite view cannot take a parameter.
- evaluation_schemas.py: the proof and benchmark report shapes move out of schemas.py
to keep it under the file-size limit; importers follow.
- backend/tests/test_memory_schema.py: 8 tests, including the v1 -> v2 upgrade and the
validity timeline.
Verification of record: pytest backend/tests bench/tests -q -> 70 passed; ruff check on
every touched file -> all checks passed; mypy (fresh cache) on db.py, migrations.py,
schemas.py, evaluation_schemas.py and backend/tests -> no issues. evals.py carries 40
pre-existing mypy errors at HEAD and 40 after this change (measured on both copies);
its pyright findings are the same debt, and p0-6 rewrites that file. `peft` and
`pyarrow.parquet` "could not be resolved" is the documented pyright artifact: pi-lens
runs it with 0 config files, and both import at runtime.
The frozen baseline showed the unmodified RetrievalService agreeing with an equivalent
BM25 index on 0% of its top ten results. Reproduced by hand: its top ten atoms came
from one document's pages, so a source-level ranking cannot overlap it.
- bench/systems.py: live_product_rankings also counts distinct sources behind the
product's top-k atoms.
- bench/run.py: scale probes no longer filtered by gold (a source prefix keeps almost
no gold, so the 500/2,000/10,000 rows were silently skipped); scale_arm_names reads
only latency dicts, which is what crashed write_markdown on the note field; the
unanswerable share is computed instead of hard-coded.
- bench/tests/test_joint.py: 60 tests total, including the writer regression.
Verification: pytest bench/tests -q -> 60 passed; ruff check bench/ -> all checks
passed; mypy (fresh cache) -> no issues in 25 source files.
- bench/ingest_baseline.py drives the unmodified product IngestionService over
all 50,789 corpus files in sharded SQLite DBs, then merges them. Thread and
memory ceilings are set explicitly: ONNX Runtime defaulted to every core
(95 threads/worker, load 490) and its CPU arena kept freed pages mapped,
which OOM-killed workers.
- bench/metrics.py: recall@k, MRR, nDCG@10, Wilson lower bound, chance
baseline. No model, no product import, so the scorer cannot grade itself.
- bench/systems.py + bench/run.py: the five ablation arms over the merged DB,
latency percentiles, scale curve, and a top-10 equivalence check against the
live RetrievalService.
- bench/counterpart.py + bench/run_external.py: in-repo Mem0-style flat store
and Zep-style temporal store as the external reference point.
- bench/audit.py: enforces single-attempt scoring, frozen n, a chance row
beside every score, and producer != verifier. It already flags the product's
retry-until-target loop at backend/src/ely_eye/evals.py:2351.
- backend/pyproject.toml: rapidocr-onnxruntime>=1.4.4 was unsatisfiable
(PyPI maxes at 1.2.3); the successor is the renamed rapidocr package.
Ground truth comes only from third-party annotations (MMLongBench-Doc
evidence_page_ids, VisRAG qrels, ScreenSpot-v2 bboxes, Video-MME QA,
LongMemEval answer_session_ids, CodeSearchNet func_path+func_name) and
from document structure; no model judgement is used anywhere.
- bench/corpus/manifest.jsonl: 28 sources with url, retrieved_at, sha256,
annotation_source, license, files[]; verified by bench.manifest_check
- bench/sets/{pdf,image,ui_screenshot,video,code,conversation}.{jsonl} plus
corpus indexes and a manifest that freezes every file's sha256
- 2,928 answerable + 2,380 unanswerable queries over 854 holders; every
query carries scope {holder, scenario, as_of} for transactive routing
- bench/validate_sets.py fails on missing gold, holder mismatch between
gold and query, gold observed after as_of, and manifest digest drift
ZacharyZhang-NY