# Context Compaction Audit

Audit conversational context compaction, summarization, and message pruning for named entity retention, critical fact loss, and semantic distortion.

## Workflow

1. Ingest original conversation history:
   - Original token count ($T_{\text{orig}}$)
   - Declared critical entities and constraints (e.g. user IDs, file paths, requirements, numerical bounds)
2. Ingest compacted / summarized context:
   - Compacted token count ($T_{\text{comp}}$)
   - Compression ratio: $R = 1 - (T_{\text{comp}} / T_{\text{orig}})$
3. Audit entity retention:
   - Verify every declared critical entity and constraint is preserved in the compacted summary without factual alteration or hallucination.
   - Entity retention score: $S_{\text{ret}} = N_{\text{preserved}} / N_{\text{required}}$
4. Assert zero loss on strict key-value parameters and constraint bounds.

## Synthetic Input Format

JSON object with:
- `schemaVersion`: 1
- `compactionMethod`: `"recursive_summary"` | `"sliding_window_prune"` | `"semantic_dedup"`
- `metrics`:
  - `originalTokens`: positive integer
  - `compactedTokens`: positive integer
- `requiredEntities`: list of strings (must be present in compacted text)
- `compactedText`: string
- `assertions`:
  - `minCompressionRatio`: optional decimal string (e.g. "0.5000")
  - `minEntityRetention`: optional decimal string (e.g. "1.0000")

## Exit Codes

- `0`: Valid compaction with all critical entities retained.
- `1`: Entity loss or insufficient compression.
- `2`: Malformed input or safety violation.

## Verification

```sh
python3 scripts/check.py examples/synthetic.json
```
