Independent Evidence Reinforcement
Can a correlation-aware evidence model assign more confidence to genuinely independent observations than to larger volumes of repeated, shared, or stale observations?
Claim boundaryThis lab tests evidence weighting and temporal confidence using fictional topology and deterministic scenarios. It does not test remediation authority, production incidents, or self-healing safety.
Evidence summary
What the current experiment observed
Method
The evidence path
Findings and revision
The evidence changed the work
Phase A established the basic contrast
Ten thousand correlated repetitions, 50 independent observations, and 5,000 observations hidden behind shared topology test whether raw volume is mistaken for independent support.
Phase B broke score v1
The first scoring candidate passed monotonicity and repetition saturation but failed concentration sensitivity and adversarial ranking. The negative result remains frozen in the repository.
Score v2 repaired the known structural counterexample
Inverse-Simpson effective source counts replaced raw topology cardinality. V2 passed the unchanged Phase B suite while preserving Phase A, establishing regression-repair evidence rather than independent production validity.
Phase C2 broke temporal v1
Temporal v1 passed all seven original Phase C checks but only three of seven adversarial checks covering recovery identity, mixed history, delayed recovery, recurrence, clock skew, and timestamp contradiction. Temporal v2 preserves those counterexamples and is evaluated against the frozen suites.
Limitations
What this lab does not prove
- Synthetic topology is only a proxy for real causal independence.
- The experimental 30-minute half-life is not a production recommendation.
- Repairing known counterexamples does not provide independent holdout validation.
- Incident ground truth, precision, recall, and remediation authority remain outside the current experiment.
Inspect and reproduce