Skip to content

feat(m2a-wettest): She-Proves Tier A wet test — 8 clips, M2a prosody#2

Merged
shaypal5 merged 7 commits into
mainfrom
feat/initial-debug-run-1
Apr 16, 2026
Merged

feat(m2a-wettest): She-Proves Tier A wet test — 8 clips, M2a prosody#2
shaypal5 merged 7 commits into
mainfrom
feat/initial-debug-run-1

Conversation

@shaypal5

Copy link
Copy Markdown
Member

Summary

  • 8 She-Proves Tier A clips across all 4 violence typologies (SV × 2, IT × 2, NEG × 2, NEU × 2)
  • Generated with M2a SSML prosody defaults from SynthBanshee PR #24
  • Includes TTS utterance cache (assets/speech/) and LLM script cache (assets/scripts/) for full pipeline reproducibility

Wet test results

Check Value Threshold Status
VIC I1 F0 160.1 Hz ≤ 200 Hz
VIC I4 F0 185.0 Hz < 250 Hz
VIC I5 F0 200.1 Hz < 250 Hz
AGG pitch escalation flat / rate-driven no helium artefact ✅ (subjective)
AGG RMS escalation +0.4 dB ≥ 8 dB ⏳ pending M3

AGG RMS escalation is a known Azure normalization limitation — per-turn gain will be applied in M3 (SceneMixer).

Test plan

  • Listen to at least one clip per typology and confirm no obvious synthesis artefacts
  • Confirm VIC sounds subdued/flat under stress (not rising pitch)
  • Confirm AGG sounds faster/more intense at high intensity (not helium)

🤖 Generated with Claude Code

8 clips across all 4 violence typologies (SV, IT, NEG, NEU) for
she_proves Tier A. Generated with M2a SSML prosody defaults:
  - VIC pitch_delta_st capped falling (-4 → -1 st across I1–I5)
  - AGG pitch_delta_st flattened (0/0/+1/+1 across I2–I5)

Includes TTS utterance cache (assets/speech/) and LLM script cache
(assets/scripts/) for full pipeline reproducibility.

Wet test results:
  VIC I1 F0: 160.1 Hz ✓ (≤200 Hz)
  VIC I4 F0: 185.0 Hz ✓ (<250 Hz)
  VIC I5 F0: 200.1 Hz ✓ (<250 Hz)
  AGG pitch: flat/rate-driven ✓ (subjective — no helium artefact)
  AGG RMS escalation: pending M3 (SceneMixer per-turn gain)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
@shaypal5 shaypal5 force-pushed the feat/initial-debug-run-1 branch from 07f969d to 2a20b15 Compare April 16, 2026 06:36
pre-commit-ci Bot and others added 2 commits April 16, 2026 06:40
DELIVERIES.md — master log table (one row per merged PR) linked from
README, with clip counts, duration, typology breakdown, pipeline
milestone, status, and PR link.

deliveries/001-debug-run-1/ — metadata.yaml + notes.md for PR #1
(single debug clip, v1 pipeline, status: superseded).

deliveries/002-m2a-wettest/ — metadata.yaml + notes.md for PR #2
(8 clips, M2a prosody milestone, status: provisional). Includes full
prosody QA table and subjective QA results.

README.md — adds "Delivery history" section linking to DELIVERIES.md.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Adds Delivery 002 (“m2a-wettest”) synthetic Hebrew She-Proves Tier A wet-test batch (8 clips across SV/IT/NEG/NEU) along with delivery logging and pipeline reproducibility caches/metadata.

Changes:

  • Adds delivery notes + structured metadata for deliveries 001 and 002, and introduces a top-level delivery log (DELIVERIES.md) referenced from the README.
  • Adds 8 new Tier A clips (transcripts + strong labels + clip metadata) and a language-level data/he/manifest.csv.
  • Adds cached scene scripts under assets/scripts/ for reproducibility.

Reviewed changes

Copilot reviewed 39 out of 193 changed files in this pull request and generated 17 comments.

Show a summary per file
File Description
deliveries/002-m2a-wettest/notes.md Delivery 002 narrative notes, QA summary, limitations
deliveries/002-m2a-wettest/metadata.yaml Delivery 002 structured metadata (counts, QA metrics, speakers, engine)
deliveries/001-debug-run-1/notes.md Delivery 001 narrative notes (superseded)
deliveries/001-debug-run-1/metadata.yaml Delivery 001 structured metadata (superseded)
DELIVERIES.md New delivery log table + status definitions
README.md Adds “Delivery history” section pointing to DELIVERIES.md
data/he/manifest.csv Adds dataset manifest for the 8 clips
data/he/agg_m_30-45_001/sp_sv_a_0001_00.txt Transcript for SV clip 0001
data/he/agg_m_30-45_001/sp_sv_a_0001_00.jsonl Strong labels for SV clip 0001
data/he/agg_m_30-45_001/sp_sv_a_0001_00.json Clip metadata for SV clip 0001
data/he/agg_m_30-45_001/sp_sv_a_0002_00.txt Transcript for SV clip 0002
data/he/agg_m_30-45_001/sp_sv_a_0002_00.jsonl Strong labels for SV clip 0002
data/he/agg_m_30-45_001/sp_sv_a_0002_00.json Clip metadata for SV clip 0002
data/he/agg_m_30-45_001/sp_it_a_0001_00.txt Transcript for IT clip 0001 (updated content/timing)
data/he/agg_m_30-45_001/sp_it_a_0001_00.jsonl Strong labels for IT clip 0001 (updated)
data/he/agg_m_30-45_001/sp_it_a_0001_00.json Clip metadata for IT clip 0001 (updated)
data/he/agg_m_30-45_001/sp_it_a_0002_00.txt Transcript for IT clip 0002
data/he/agg_m_30-45_001/sp_it_a_0002_00.jsonl Strong labels for IT clip 0002
data/he/agg_m_30-45_001/sp_it_a_0002_00.json Clip metadata for IT clip 0002
data/he/agg_m_30-45_001/sp_neg_a_0001_00.txt Transcript for NEG clip 0001
data/he/agg_m_30-45_001/sp_neg_a_0001_00.jsonl Strong labels for NEG clip 0001
data/he/agg_m_30-45_001/sp_neg_a_0001_00.json Clip metadata for NEG clip 0001
data/he/agg_m_30-45_001/sp_neg_a_0002_00.txt Transcript for NEG clip 0002
data/he/agg_m_30-45_001/sp_neg_a_0002_00.jsonl Strong labels for NEG clip 0002
data/he/agg_m_30-45_001/sp_neg_a_0002_00.json Clip metadata for NEG clip 0002
data/he/agg_m_30-45_001/sp_neu_a_0001_00.txt Transcript for NEU clip 0001
data/he/agg_m_30-45_001/sp_neu_a_0001_00.jsonl Strong labels for NEU clip 0001
data/he/agg_m_30-45_001/sp_neu_a_0001_00.json Clip metadata for NEU clip 0001
data/he/agg_m_30-45_001/sp_neu_a_0002_00.txt Transcript for NEU clip 0002
data/he/agg_m_30-45_001/sp_neu_a_0002_00.jsonl Strong labels for NEU clip 0002
data/he/agg_m_30-45_001/sp_neu_a_0002_00.json Clip metadata for NEU clip 0002
assets/scripts/ca24e3d5f040452f2df8df3b2cd7d8d34491c2f4c896f9e08f397fdd61ce2b69.json Cached generated script (scene turns)
assets/scripts/c507234e77bfe487531218f506cdf223e3ae4d1dfcd356cb09c0c3ecbf154663.json Cached generated script (scene turns)
assets/scripts/a58e4534b8cb00c83b01c358afb9039f17568297dad3c8d257a84194fe28885f.json Cached generated script (scene turns)
assets/scripts/9e184d6024e8ca23d5afd44d01a00d89ae647b03f14faab76cabe66e00ba0965.json Cached generated script (scene turns)
assets/scripts/804926fb260415c74f62fa3332dc185e50180551ba2e5c24b61bf18dd6545fbb.json Cached generated script (scene turns)
assets/scripts/69b29ad8e90d20476c22282fc43528eae4d1a2dda1b7dd1c0fe6628287417903.json Cached generated script (scene turns)
assets/scripts/498b07ded13fdae1ae18e3075afb85373bccfc863d9acefeb9f0ac6b4f3c0cba.json Cached generated script (scene turns)
assets/scripts/0743595033eb507e3f61aaf1a11f150a700182efe045fce0b43bc8bbf13fc881.json Cached generated script (scene turns)
Comments suppressed due to low confidence (1)

data/he/agg_m_30-45_001/sp_it_a_0001_00.json:58

  • transcript_path is an absolute local path and includes a workstation username, which breaks portability and leaks environment-specific information. Also, weak_label.has_violence introduces a binary violence flag that conflicts with the repo spec in README.md (no binary violence/non-violence labels). Prefer repo-relative transcript_path (or omit it) and remove/rename has_violence unless the spec is updated.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Comment thread data/he/manifest.csv Outdated
Comment thread data/he/agg_m_30-45_001/sp_it_a_0002_00.json Outdated
Comment thread data/he/agg_m_30-45_001/sp_neg_a_0002_00.json Outdated
Comment thread data/he/agg_m_30-45_001/sp_it_a_0002_00.json
Comment thread data/he/agg_m_30-45_001/sp_neg_a_0001_00.json Outdated
Comment thread data/he/agg_m_30-45_001/sp_sv_a_0001_00.json Outdated
Comment thread data/he/agg_m_30-45_001/sp_sv_a_0002_00.json Outdated
Comment thread data/he/agg_m_30-45_001/sp_neg_a_0001_00.json
Comment thread data/he/agg_m_30-45_001/sp_neu_a_0001_00.json
Comment thread data/he/agg_m_30-45_001/sp_neu_a_0002_00.json
shaypal5 and others added 2 commits April 16, 2026 17:21
…hreads

Adds two GitHub Actions workflows using shaypal5/pr-agent-context@v4:
- ci.yml: initial run on pull_request events
- pr-agent-context-refresh.yml: refresh flow triggered by review events and
  check_run completions, using publish_mode=append and
  include_outdated_review_threads=true so stale diff threads remain visible

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…sh on push

- Remove `secrets: github_token:` from both workflows; the reusable workflow
  uses github.token implicitly and defines no secrets inputs
- Add `pull_request: synchronize` trigger to refresh workflow so it runs on
  every new commit pushed to the PR
- Fall back to github.run_id in concurrency group for check_run events where
  pull_requests array may be empty

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
@github-actions

This comment has been minimized.

@github-actions

This comment has been minimized.

- manifest.csv: make wav_path/strong_labels_path repo-relative; lowercase
  speaker_ids to match on-disk directory naming convention
- All clip .json: make transcript_path repo-relative
- 7 clip .json: set dirty_file_path null where dirty WAV is absent from repo
  (sp_it_a_0001_00 retains its path — dirty file is present)
- README: replace prohibition on binary labels with accurate derived-field
  policy — has_violence is a legitimate convenience field computed from
  violence_typology/violence_categories/max_intensity; taxonomy is ground truth

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
@github-actions

This comment has been minimized.

CLAUDE.md — full agent/contributor policy file covering:
  - Cache integrity rules (assets/speech, assets/scripts)
  - Clip file integrity and ASCII filename conventions
  - Delivery log requirements (DELIVERIES.md + deliveries/{slug}/)
  - Label policy: has_violence is a derived convenience field, not to be
    removed; taxonomy is ground truth; replacing taxonomy with a binary
    flag is prohibited
  - Audio format spec
  - How clips get here (SynthBanshee env vars)
  - Validation commands
  - What NOT to do (consolidated list)

README.md — add agent callout box at the top pointing to CLAUDE.md;
add "Agent and contributor guidelines" section at bottom.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown

pr-agent-context report:

No unresolved review comments, failing checks, or actionable patch coverage gaps were found on PR
#2. Treat this PR as all clear unless new signals appear.

Run metadata:

Tool ref: v4
Tool version: 4.0.18
Trigger: commit pushed
Workflow run: 24520427940 attempt 1
Comment timestamp: 2026-04-16T15:59:27.377987+00:00
PR head commit: e191abcd1597d26b3b31c581effbddf69dfd42b8

@shaypal5 shaypal5 merged commit 839555b into main Apr 16, 2026
3 checks passed
@shaypal5 shaypal5 deleted the feat/initial-debug-run-1 branch April 16, 2026 16:04
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants