The library

Rendered from docs/specs/2026-09-26_channel-stopband.md; the markdown file is the record.

Run spec: does the double low-pass stopband move any detector? (Sat Sep 26, 2026, 14:26)

Go: Nathan, in the channel-robustness chat ("go stop band"), after oversight scoped it as measurement only. Follows the open shortcut-ledger entry "Double low-pass stopband" (docs/architecture.md §11) and docs/reports/2026-09-26_channel-robustness.md (A). Short path (sub-hour, measurement only): this spec, one script, hermetic tests, one Claude critique. No TSV, no rule change, no refit.

Question

Every detector applies hearsay.handcrafted.band_limit once inside its own path: M1b in prepare_segment, M3 in prepare_input, M5 in deploy_transform, handcrafted in _crop. NSA test files arrive already low-passed at ~7.2 kHz, so at test time they are filtered twice, while training clips are filtered once. Scoring band_limit(x) instead of x for the same clip approximates the test condition. It matches the floor depth but not the wall's transition-band shape (see Review). The difference is the detectors' response to the stopband depth alone.

Design (fixed before any result)

• Cohort: the 500 outer-holdout clips of action E (scripts/m3_probes.holdout_rows(500), same seeds and test-length crops): 250 real, 250 spoof. These are out of sample for M1b, M5 and handcrafted v5.
• Conditions per clip: once = the raw segment; twice = band_limit(segment). Both are scored through the runner's own paths (hearsay.pipeline.Models, CPU) and fused with models/fusion_v2/constants.json, as in E. The fused probability is recorded as well.
• Readouts:
  • Per detector (M1b, M5, M3, handcrafted v5): Δlogit = twice − once; mean, median, 95th percentile of |Δ|, and signed mean by class.
  • The same for the fused score and the fused probability.
  • The M3 step's firing rate under each condition.
  • Verdict flips at P = 0.5, and at the brief-cost argmin threshold of the once fused scores (descriptive, computed on this cohort).
  • Holdout minDCF under each condition, per detector and fused.
• Reading rule (pre-declared):
  • A detector responds if its 95th-percentile |Δlogit| exceeds 10% of the IQR of its inner-OOF logits (from its export in outputs/detector_scores/).
  • The fused rule responds if its holdout minDCF moves by more than 0.037 (one holdout false alarm).
  • Inconclusive if fewer than 500 complete pairs, any readout input is non-finite, or an IQR is zero.
  • If nothing responds, the ledger entry closes as "measured, no response".
  • If something responds, the report states the refit cost (handcrafted ~30 min; M1b ~1–1.2 h; M5 not feasible today). Any refit goes to Nathan and through scripts/fuse_sweep_m5.py.

Failure modes → tests (tests/test_channel_stopband.py)

Documentation

• Create: this spec; a section in docs/reports/2026-09-26_channel-robustness.md.
• Update: the §11 ledger entry in docs/architecture.md (its status), and the STATUS channel-robustness row.

Results (14:41)

Measured, no response on all four detectors and the fused rule (outputs/channel/stopband.json): 500 of 500 clips paired, 250 real and 250 spoof.

• Fused holdout minDCF: 0.012 → 0.016 (limit 0.037).
• Verdict flips: 0 at P 0.5; 2 at the once argmin threshold.
• M3 step: fires on 2.0% of clips under both conditions.
• Ledger: the entry in architecture.md §11 is closed; no refit.
• Tests: tests/test_channel_stopband.py, 61 passing after the review (full suite 645).

Review

Claude critique (14:55, on f268f35): Overall Acceptable.

• Verified:
  • one band-limit pass in once for every detector path, and two in twice
  • once equals E's clean scores exactly
  • the monkeypatched scorer ran, and E's cache is untouched
  • the reading rule matches this spec
  • stopband.json rebuilds identically from the cache
• Findings, all fixed in the follow-up commit:
  1. The ledger row called the 95th percentile a maximum. Reworded, and the single-clip maxima are now stated.
  2. "Reproduces the test condition" was an overclaim. The extra pass matches floor depth but not the transition band (about 9 dB) or the 7.5–8 kHz stopband (about 11 dB). The conclusion is now scoped to stopband depth, and the shape is recorded as an unprobed residual in the ledger.
  3. Test gaps: a sweep over every column with non-finite values, single-responder ledgers (one detector, or fused only), zero or NaN IQR through readout, and exact boundaries. Tests added.
  4. The fused row mixed probability and logit units. Footnoted.
  5. The header time was wrong (the file was created 14:26, before the run started at 14:27). Corrected, and this critique is recorded here.
• Not fixed (out of scope): the STATUS "Tests" row predates this change.