Specs
Implementation-ready (or implementation-planned) specifications for changes that require code modifications.
Proposed (Prediction Modes)
| Spec | Title | Description |
|---|---|---|
| 061 | Total PHQ-8 Score Prediction | Predict total score (0-24) instead of item-level |
| 062 | Binary Depression Classification | Binary classification (PHQ-8 >= 10) |
| 063 | Severity Inference Prompt Policy | Allow inference from temporal/intensity markers |
These specs address the task validity problem: PHQ-8 item-level frequency scoring is often underdetermined from DAIC-WOZ transcripts (see docs/clinical/task-validity.md).
Deferred
Archived (Implemented)
Implemented specs are distilled into canonical (non-archive) documentation under docs/:
Pipeline Robustness (Specs 053-057) - PR #92, 2026-01-03
| Spec | Title | Canonical Doc Location |
|---|---|---|
| 053 | Evidence Hallucination Detection | Evidence Extraction, Features |
| 054 | Strict Evidence Schema Validation | Evidence Extraction, Exceptions |
| 055 | Embedding NaN Detection | Artifact Generation, Debugging |
| 056 | Failure Pattern Observability | Error Handling, Debugging |
| 057 | Embedding Dimension Strict Mode | Artifact Generation, Configuration |
JSON Reliability (Specs 058-060) - 2026-01-04
| Spec | Title | Canonical Doc Location |
|---|---|---|
| 058 | Increase PydanticAI Default Retries | Configuration, JSON audit |
| 059 | json-repair Fallback | Evidence Extraction, JSON audit |
| 060 | Retry Telemetry Metrics | Error Handling, Debugging |
Other Implemented Specs
- Quantitative severity bounds (BUG-045): spec → PHQ-8 docs
- Retrieval audit redaction (Spec 064): spec → RAG debugging
- Feature index + defaults: features.md
- RAG runtime features (prompt format, CRAG, batch embedding): runtime-features.md
- RAG debugging workflow: debugging.md
- RAG artifact generation (embeddings + tags): artifact-generation.md
- Chunk scoring setup + schema: chunk-scoring.md
- Error handling philosophy: error-handling.md
- Exception taxonomy: exceptions.md
- Metrics definitions + output schema: metrics-and-evaluation.md
- Selective prediction confidence signals (Spec 046): spec → metrics docs
- Verbalized confidence (Spec 048): spec → metrics docs
- Supervised confidence calibrator (Spec 049): spec → metrics docs
- Consistency-based confidence (Spec 050): spec → metrics docs
- Advanced CSFs from fd-shifts (Spec 051): spec → metrics docs
- Excess AURC/AUGRC metrics (Spec 052): spec → metrics docs
- Remove keyword backfill (Spec 047): spec → configuration philosophy
- Configuration philosophy: configuration-philosophy.md
- DAIC-WOZ transcript preprocessing + variants: spec → user guide
Historical spec texts remain in docs/_archive/specs/ for provenance, but the active documentation
should not require them.