mnemosyne-memory-lifecycle-benchmark / reports /benchmark-validity-summary.md
solsticestudioai's picture
Publish Mnemosyne Memory Lifecycle Benchmark v0.2.0
6a700f4 verified
|
Raw
History Blame Contribute Delete
1.59 kB
# Benchmark validity summary
## Findings
1. The benchmark measures distinct capabilities, but discrimination is uneven:
25 of 40 scenarios are tied under the current scored outcomes.
2. Belief revision and historical reconstruction meaningfully separate typed
lineage from simpler baselines. Many straightforward categories remain tied.
3. No expected-answer fields are available during blind strategy execution.
4. The baseline audit found the prior full-Scenario input contract was too
permissive; it is now replaced with `StrategyInput`.
5. The shared v0.1.0 0.975 score was one fixture/evaluator alignment defect,
not evidence of universal strategy equivalence.
6. Several metrics remain `not_applicable` because the current fixtures do not
instantiate their denominators. This is reported, not converted to success.
7. The conflicting-authority fixtures now expect explicit abstention where
authority remains unresolved.
## Version decision
Scenario labels and strategy semantics changed to correct documented validity
defects, so this audit is benchmark v0.2.0. v0.1.0 fixtures and reports remain
available unchanged under the frozen directories.
## Supported claims
Within these synthetic fixtures, typed lineage supports point-in-time
reconstruction and authority-aware supersession. Simpler strategies remain
competitive on ordinary current-state questions.
## Unsupported claims
No claim of real-world memory accuracy, production privacy, complete erasure,
human-like memory, universal belief revision, or superiority over other memory
systems is supported.