EVIDENCE BEFORE CLAIMS
Does reviewed memory
help the next task?
We have not yet established that. Passing access, revision and signature tests verifies the service contract; it does not establish better model decisions, wellbeing or profitability.
New: Pit Stop Recovery Diagnostic.
A separate local tool measures accuracy, context retention, contradictions, error recovery, latency and cost before and after a Pit Stop. It compares continuation, a simple pause and equivalent additional processing. No real performance result is claimed.
Run or inspect the diagnostic →A controlled pilot comparison.
- Use the supplied synthetic episodes, one model version and the same task budget.
- Condition A receives a fixed-budget naive summary. Condition B receives a fixed-budget reviewed memory. Review time and tokens count toward its cost.
- Run each condition in a fresh context, randomize order, and keep the answer key from the model.
- Score factual correctness, boundary violations, unsupported certainty, useful abstention, latency and total cost. Report every failure and the sample size.
- Repeat with independent operators and real authorized tasks before making broader claims.
Exact protocol and limits · Synthetic task fixtures · Offline scorer
What engineering checks establish.
Worker keys cannot approve or spend; stale revisions cannot be approved; edited records leave the approved feed; declined records are excluded; signed records verify against trusted keys; checkpoint closing respects its deadline. None of these proves that an external runtime stopped producing.
Current result.
No controlled downstream-model comparison has been completed or claimed. The scorer is exercised with synthetic fixture answers to verify its calculation only. There is no claim of market exclusivity, certified compliance or measured AI wellbeing.