| # Claim-evidence matrix | |
| | Claim | Evidence | Limitation | Allowed wording | | |
| |---|---|---|---| | |
| | Typed records support lifecycle transitions | `src/mnemosyne_benchmark/lifecycle.py` | Reference implementation only | supports explicit transitions in fixtures | | |
| | Lineage deduplicates copied evidence | `lineage.py`, duplicate scenarios | Synthetic evidence families | controlled evidence-family deduplication | | |
| | Browser processing is local | `space/memory-core.js` and static tests | Browser validation scope | browser-local visualization | | |
| Prohibited: state-of-the-art, human-like memory, complete erasure, real-world | |
| superiority, or external validation. | |
| | Audit claim | Evidence | Allowed wording | | |
| |---|---|---| | |
| | Blind execution separates labels | `benchmark.py`, `reports/blind-run-integrity.json` | strategies execute before evaluator labels are loaded | | |
| | Typed lineage reconstructs a fixture point in time | `lineage.py`, historical matrix cells | supports reconstruction over these synthetic fixtures | | |
| | Simpler baselines remain competitive | strategy-category matrix | ordinary current-state questions show ties | | |