New information is not necessarily independent evidence.
ICI Sentinel distinguishes new information, dependent evidence and independent evidence while preserving the last verified state over time.
INTERNAL REGISTERED BENCHMARK · EXTERNAL HELD-OUT VALIDATION PENDING
A registered result, with explicit boundaries
On the registered 64-claim stress benchmark: 100% evidence-novelty classification accuracy; zero continuity violations, missed verified transitions or dependent-evidence inflation violations. Results apply to this internal benchmark, not to every possible operational evidence source.
Four different evidentiary conditions
NEW_INFORMATIONA new message appears without a newly established evidentiary artifact.
NEW_EVIDENCE_DEPENDENTThere is a new artifact, but its provenance shares source roots with evidence already registered for the same claim.
NEW_INDEPENDENT_EVIDENCEThere is a new artifact whose declared ultimate source roots do not overlap the existing roots for the claim.
EVIDENCE_UNRESOLVEDProvenance is insufficient to determine whether the new artifact is independent.
A contradiction does not automatically replace a verified state
Unsupported or source-dependent contradictions leave the previous verified state intact. A first independent contradictory item marks the situation contested. Only sufficient, temporally applicable independent support can establish a verified transition. Late-arriving historical evidence does not roll the current verified state backward.
Previously verified state retained as anchor
New information or dependent/unresolved contradiction
First independent contradictory support
Registered independence threshold satisfied
Why simpler baselines fail
Latest assertion wins
384continuity errors on the registered sequences.
Apparent evidence counter
128false state transitions when provenance dependence is ignored.
Sentinel IAB-16
0false overwrites on the same registered stress set.
Why singleton verification is insufficient
Individual changes can be safe while their combination is unsafe; conversely, two individually unsafe changes may compensate. IAB-16 tested ANY, ALL, TWO-OF-THREE and XOR rules across a finite five-resource Boolean decision domain.
| Method (long lifecycle) | Accuracy | Unsafe false accept | False reject | Oracle calls / decision |
|---|---|---|---|---|
| Direct runtime oracle | 100% | 0% | 0% | 1.00 |
| Singleton certificate | 75% | 37.5% | 12.5% | 0.078 |
| Interaction certificate | 100% | 0% | 0% | 0.50 |
| Adaptive verifier | 100% | 0% | 0% | 0.50 |
With 5 Boolean resources, a full finite-domain certificate takes 2⁵ = 32 oracle queries per certificate identity. Direct verification is cheaper below 32 decisions of reuse; the interaction certificate is cheaper above that threshold under equal query cost and a stable decision identity. The long lifecycle tested 64 decisions per certificate.
What can be claimed
Sentinel does not merely test whether evidence exists. On the registered IAB-16 benchmark it distinguishes informational novelty, dependent evidence and independent evidence; preserves the last verified state through unsupported contradictions; and changes verified state only when sufficient independent, temporally applicable evidence supports the transition.
This is a reproducible, internally registered synthetic benchmark. The two-independent-source threshold is an explicit configuration choice, not a universal rule. The outcome does not establish unique algorithmic novelty, universal superiority, real-world source independence or independent external validation. External evaluator-authored held-out testing remains pending.