The held-out benchmark keeps the 40-commit floor and the union of twelve-commit stratified draws.
A training freeze passes floor=20 (never below the ratio) and the record carries ordinary_floor.
When the union of twelve-commit draws cannot reach the count (small cells stay unallocated on
every draw), one stratified draw of the full count is used; with count equal to the eligible
commits that is every one of them.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Sourced apart from the 12 pilot repositories: 212 candidates (pool-fresh, pool-retry, OSV for PyPI,
npm, Maven and Packagist), 49 dropped because the benchmarks already name their repository, 413
catalog locations, 99 passing all three verifier checks in 28 repositories. Short of the 40-positive
target. catalog.py now reads .java and .php. Freeze waits on the maintainer's spot-check
(CANDIDATES.md, SOURCING.md).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
0 of 12 positives; 12.6 false alarms per 100 ordinary pushes (2.1 on changed
lines only). The misses come from what the rules and facts express (access
control lives in obligations, which do not convert; sources cover request
objects only), not from the engine. Comparison arm only: a Semgrep miss is
never evidence against a hunter finding.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Opt-in filter on the classifier verdict for hunter-sourced defects; the default
keeps everything. Each combined record notes the kept verdicts and the counts.
Live check on the pilot: (a)+(b) filtered to 'vulnerable' gives recall 5/12 and
54.4 false alarms per 100 ordinary pushes, matching the scratch numbers.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Live run: every classification failed with ValueError because Caller.ask
reads client.usage as a dict (the learn CLI always hands it a MeteredClient)
while a bare DeepSeekChatClient keeps usage as a list of rows. build_caller
now wraps a client that carries a usage list in MeteredClient; scripted and
already-metered clients pass through unchanged.
Regression test runs classify_family through the real DeepSeekChatClient
wrapper with a stub transport (the module fixture that stubs classify is
restored for it). The error field stays the exception type only, since a
message can carry a model key.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
The tool hunter's family tag is the model's free text (a live run returned
family:sql-parser and family:readonly-enforcement), so every arm (b) defect
scored family "unknown" and score.catches() could never credit the hunter
even when it landed on the labelled file.
hunter_arm now classifies each finding into the closed taxonomy vocabulary
with the decision-engine classifier (learn.program.classify with
families=FAMILY_IDS minus unknown): a numbered, clamped excerpt around the
finding sets valid_lines, the canonical ANSWER_FORMAT is reused, an off-list
family parses to "none" and lands as "unknown". The benchmark label is never
a fallback. The defect keeps the raw hunter family, title and rationale plus
the classifier's verdict, confidence, rationale, cited lines and parse
failures for audit; the record counts classifier calls and families.
--no-classify keeps the old heuristic path.
14 new offline tests, 67 green with the benchmark tests.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
benchmarks/valid/hunter_arm.py runs the shipped deep-mode tool hunter
(tool_hunter.run_tool_hunter via resolve_hunter_client) over each frozen
change off-cluster: detached worktree at head with git hooks disabled,
hotspots scoped to the changed files, instrument rule on every hotspot
(bytes read printed, empty or unreadable fails the change loudly), model id
from the config resolver, family from the taxonomy, never from the label.
Records land in the exact shape score.py reads (push.admission.defects with
locations), resumable per change.
`combine` joins arm (a) replay records with arm (b) hunter records per change
id, dedupes defects by (path, line, family), keeps (a)'s instrument, records
which arms were present, and writes a file score.py loads directly.
9 offline tests (stub hunter, real scorer), 53 green with the benchmark
tests. No network, no key handling, no live run.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
First live run of the valid detection benchmark. Labels come from the fix
site (fix^1), never SZZ.
- 32 held-out Python/JS advisories sourced from the public pool-p1 set,
CWE + text from OSV. 133 candidate locations -> 125 after dropping
tests/minified.
- 125 DeepSeek verification calls: 31 locations passed all three checks
(vulnerable pre-fix, fix addresses it, reachable in-repo) across 12 repos.
- Maintainer spot-check accepted all 12.
- 20 ordinary negatives per positive change from the same repos, 240 total.
- Held-out guard: 0 dropped against the training-only used set.
- Frozen manifest at benchmarks/valid/pilot/ (252 changes), provenance and
NOTES under benchmarks/measurements/2026-10-12-valid-benchmark-pilot/.
Three fixes found by the live run:
- catalog.py: skip a site the fix renamed or removed instead of aborting.
- assemble.py: negative ratio counts positive changes, not accepted
functions (malla would have demanded 260 negatives). Test updated.
- held-out guard: run against training sources only; build_used() also
scans measurements/populations, which flagged the 12 repos via my own
SZZ-pool analysis outputs (circular).
Replay and metrics not yet run; see NOTES.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Builds the benchmark from .kiro/specs/valid-detection-benchmark/brief.md:
sound labels from the fix (no SZZ), app-shaped verified positives,
ordinary negatives from the same repos, whole tool on unseen repos,
standard metrics.
Spec: requirements.md, design.md, tasks.md. Reuses the existing sourcing,
fix-site labelling, negative drawing, replay, and Wilson scoring; drops
the SZZ layer.
New offline modules under benchmarks/valid/:
- catalog.py positives labelled at the fix site (never introducing/last_touch)
- verify.py one LLM call per case (vulnerable? fix addresses it? reachable
in-repo?) + human spot-check; this is the app-shaped filter.
Honors the instrument rule: bytes read per function, reject on
an empty read.
- assemble.py >=20x ordinary negatives + a frozen, deterministic manifest with
the held-out guard
- metrics.py precision (PPV), false alarms per 100 pushes, exploit-verified
rate, and the plain headline, layered on score.py
36 new offline tests, 495 regression tests green. No network, no harvest,
no model calls in tests. Group 5 (the live pilot) is not started.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
