## [speech_context_classifier] - 2025-07-30
### Added
- Finalized `SpeechContextClassifierInput` model with partner speech and vocabulary fields
- Finalized `SpeechContextClassifierOutput` model aligned with A3CPMessage structure
- Defined nested `ClassifierOutput` and `ClassifierRankingItem` classes for result structure
### Generated
- `speech_context_classifier.schema.json` from validated Pydantic models
- `input.example.json` and `output.example.json` for documentation and test use

## [speech_context_classifier] - 2025-07-30
### Updated
- Removed undocumented `classifier_output.flags` field to align with SCHEMA_REFERENCE.md
- Deleted `relevance_scores` from Outputs section (not part of schema; replaced by `ranking`)
- Clarified that `classifier_output.intent` may be `null` and `ranking` empty when no match is found
- Explicitly listed `schema_version` and `record_id` as required fields in Outputs

## 2025-07-30  memory_integrator/README.md
- Rewrote module doc to align with updated schema and coordination flow
- Clarified that input comes from memory_interface and confidence_evaluator
- Specified that it outputs modifiers, not standalone A3CPMessages
- Added JSON output block with memory fields
- Added schema compliance summary
## [2025-07-30] Clarification of Routing and CARE Integration: confidence_evaluator
- Corrected downstream flow in `CARE Integration` section:
  - Now correctly routes through `clarification_planner` before reaching `output_expander`
- Clarified that the module mutates existing `A3CPMessage` records with updated scoring fields
- Confirmed compliance with `SCHEMA_REFERENCE.md` for `classifier_output`, `final_decision`, and `context.flags`

## [2025-07-30] Schema Pass-Through and Fusion Logic: input_broker
- Declared `input_broker` as a non-emitting, schema-pass-through module
  - Receives validated `A3CPMessage` inputs from gesture, sound, speech, and visual classifiers
  - Does not generate or mutate schema fields
  - Optionally annotates message groups with `stream_segment_id`
- Added Module Usage Matrix showing read-only access to all input fields
- Clarified CARE Integration role as first multimodal fusion point
- Documented internal output bundle format (not schema-validated, but traceable)
Maintains strict compliance with `SCHEMA_ARCHITECTURE.md` one-message-per-input rule.

## [2025-07-30] Schema & Model Update: visual_environment_classifier
- Declared `visual_environment_classifier` as schema-compliant (partial `A3CPMessage`)
  - Emits `context_location`, `context.flags`, and standard metadata fields
  - Sets `modality = "image"`, `source = "communicator"`
  - Output consumed by `input_broker` for context fusion
- Added Module Usage Matrix with downstream field mapping
- Included structured JSON output example with location and flags
- Clarified model design:
  - Uses a shared, static environment classifier (e.g., CNN)
  - Does not access or depend on `model_registry`

## [2025-07-30] Schema Alignment & CARE Integration Fixes
- `gesture_classifier`:
  - Declared compliant output via `classifier_output` in `A3CPMessage`
  - Added `ranking` field example for top-N predictions
  - Corrected CARE Integration section to reflect actual graph:
    - Receives input from `landmark_extractor`
    - Outputs only to `input_broker`
    - Removed incorrect links to `CARE Engine` or `clarification_planner`
- `sound_classifier`:
  - Revised CARE Integration section:
    - Clarified output to `input_broker` for fusion
    - Removed incorrect implication of direct CARE Engine interaction
    - Documented optional API exposure (`/api/sound/infer/`)

## [2025-07-30] Schema Compliance: speech_context_classifier
- Updated `speech_context_classifier` documentation to align with A3CP schema rules:
  - Declared production of partial `A3CPMessage` with `classifier_output`
  - Added support for `ranking` field: top-N intent-confidence pairs
  - Ensured `intent` and `confidence` fields are always present for compatibility
  - Included structured example with multiple ranked predictions
  - Added Module Usage Matrix showing downstream field consumption
Clarifies one-message-per-utterance design while supporting soft intent ranking.
## [2025-07-30] Schema Compliance Updates — A3CPMessage Integration
- Updated `speech_transcriber` module documentation:
  - Declared output of finalized transcript segments as `A3CPMessage`
  - Added schema compliance summary and example payload
  - Inserted Module Usage Matrix for compatibility with downstream classifiers
- Updated `landmark_extractor` module documentation:
  - Confirmed schema-compliant output via `raw_features_ref`
  - Declared required metadata fields (modality, source, vector_version, etc.)
  - Added Module Usage Matrix covering `gesture_classifier` and `schema_recorder`
  - Included example A3CPMessage with external vector reference