Report midpoint signal in the baseline status and publish the sweep #7

Merged

peterswimm merged 1 commit from baseline-probe-sweep into main

2026-10-20 23:32:49 +00:00

peterswimm commented

2026-10-20 23:29:35 +00:00

Owner

Scope

Follow-ups named in #6, one commit:

The baseline report counted states only. completed means both captures exist, so it could not tell a silent comparison from a real one, which is how the pre-#6 baseline reported 106 completed modules while comparing silence. The baseline worker now retains each side's midpoint level, and the report adds midpoint_signal_counts (both-signal, both-silent, reference-silent, candidate-silent, not-compared).
docs/G1_BASELINE_STATUS.json is regenerated from a fresh four-shard baseline on plan v21: 92 of 109 midpoints carry signal on both engines, 5 are silent on both, 9 on one side, 3 are not compared. State counts are unchanged (106 completed, 3 unobservable).
tools/probe_sweep.py joins the public export allowlist.
docs/G1_BASELINE_PIPELINE.md describes the new field and the sweep.

Testing

81 tests pass; a new test covers each signal class, including not-compared for results written before levels were retained.
tools/export_public.py into an empty directory verifies the public tree, now including tools/probe_sweep.py.
The status file was produced by module-baseline-run (4 shards) → module-baseline-merge → module-baseline-report against the private G1-Emu adapter.

Impact Statement

Existing features: baseline worker results gain reference_rms and candidate_rms in the midpoint comparison.
Public APIs / contracts: the baseline report gains one field, midpoint_signal_counts; existing fields are unchanged.
Schemas: report format name unchanged (nardwahl-module-baseline-report-v1); the new field is additive.
Data / artifacts: docs/G1_BASELINE_STATUS.json regenerated.
Integrations: no expected impact.
Mappings: no expected impact.
Backward compatibility: older merged results still report, with their midpoints counted as not-compared.

Docs

docs/G1_BASELINE_PIPELINE.md updated as above.

## Scope Follow-ups named in #6, one commit: - The baseline report counted states only. `completed` means both captures exist, so it could not tell a silent comparison from a real one, which is how the pre-#6 baseline reported 106 completed modules while comparing silence. The baseline worker now retains each side's midpoint level, and the report adds `midpoint_signal_counts` (`both-signal`, `both-silent`, `reference-silent`, `candidate-silent`, `not-compared`). - `docs/G1_BASELINE_STATUS.json` is regenerated from a fresh four-shard baseline on plan v21: 92 of 109 midpoints carry signal on both engines, 5 are silent on both, 9 on one side, 3 are not compared. State counts are unchanged (106 completed, 3 unobservable). - `tools/probe_sweep.py` joins the public export allowlist. - `docs/G1_BASELINE_PIPELINE.md` describes the new field and the sweep. ## Testing - 81 tests pass; a new test covers each signal class, including `not-compared` for results written before levels were retained. - `tools/export_public.py` into an empty directory verifies the public tree, now including `tools/probe_sweep.py`. - The status file was produced by `module-baseline-run` (4 shards) → `module-baseline-merge` → `module-baseline-report` against the private G1-Emu adapter. ## Impact Statement - Existing features: baseline worker results gain `reference_rms` and `candidate_rms` in the midpoint comparison. - Public APIs / contracts: the baseline report gains one field, `midpoint_signal_counts`; existing fields are unchanged. - Schemas: report format name unchanged (`nardwahl-module-baseline-report-v1`); the new field is additive. - Data / artifacts: `docs/G1_BASELINE_STATUS.json` regenerated. - Integrations: no expected impact. - Mappings: no expected impact. - Backward compatibility: older merged results still report, with their midpoints counted as `not-compared`. ## Docs `docs/G1_BASELINE_PIPELINE.md` updated as above.

peterswimm added 1 commit

2026-10-20 23:29:35 +00:00

Report midpoint signal in the baseline status and publish the sweep

PR policy check / policy-gate (pull_request) Successful in 0s

Details

fa7863e80f

The baseline report counted states only, and "completed" means both
captures exist, so it could not tell a silent comparison from a real
one. The baseline worker now retains each side's midpoint level, and the
report adds midpoint_signal_counts (both-signal, both-silent,
reference-silent, candidate-silent, not-compared).

Regenerated docs/G1_BASELINE_STATUS.json from a fresh four-shard v21
baseline: 92 of 109 midpoints carry signal on both engines, 5 are silent
on both, 9 on one side, 3 are not compared.

tools/probe_sweep.py joins the public export allowlist, and the pipeline
doc describes the new field and the sweep.

peterswimm merged commit f2a75d6df5 into main

2026-10-20 23:32:49 +00:00

peterswimm referenced this pull request from a commit

2026-10-20 23:32:49 +00:00

Merge pull request 'Report midpoint signal in the baseline status and publish the sweep' (#7) from baseline-probe-sweep into main

peterswimm deleted branch baseline-probe-sweep

2026-10-20 23:32:49 +00:00

peterswimm referenced this pull request

2026-10-21 03:04:00 +00:00

Give silent baseline probes stimuli and fix the model bugs they exposed #8