Project Status

Last updated: 6 Oct 2026. Legend: ✅ done · 🟡 in progress · ⬜ not started · ⚠️ done with caveats

Workstream	Item	Status	Notes
Pipeline	Audio validation and decoding	✅	Typed error codes
Pipeline	Alignment	⚠️	Deterministic fallback done; WhisperX untested
Pipeline	Features (F0, energy, pauses), speaker normalization	✅
Scoring	Four rubric rules, scoring, templates	✅	rubric-1.0
Scoring	Insufficient-evidence rules	⚠️	SNR estimate too optimistic (see Stress Testing)
Evaluation	IoU/boundary/severity-order harness	✅	Dev + held-out split
Evaluation	Stress harness	✅	Noise robustness is a known gap
Dataset	Synthetic dev corpus (63 items)	✅	Not human speech
Dataset	Real public-domain baselines	⬜	Rights check needed per recording
Dataset	Human re-recordings and human label review	⬜
API	Endpoints and schema	✅
Dashboard	Upload, playback, overlays, explanation cards	🟡	Built against docs/schema.md
Submission	README + wiki	✅
Submission	Six-page technical document	⬜
Submission	3–10 min demo video with subtitles	⬜
Submission	Devpost page, all teammates added	⬜

Risks

• Real recordings and human review are the biggest remaining effort and carry the largest rubric weight (data engineering, 30%).
• Forced alignment on real audio is unproven.
• Noise robustness may expose false positives in demo recordings; record the demo in a quiet room.

Open decisions

1. Which real recordings to use, and their reuse terms.
2. Whether to prioritise noise robustness or WhisperX validation first.
3. Whether to expand beyond three speeches (only after the minimum demo is solid).