Concepts

Evidence, Validity, And Reporting

Evidence nodes, evidence edges, validity labels, validity scoring, and bounded RCA wording.

isplay reports are designed to be bounded: they should state what was observed, what was changed under replay, what evidence supports the conclusion, and what caveats remain.

Validity Labels

LabelMeaning
confirmed_by_replayComparable replay evidence exists without a stronger detected caveat.
sensitive_to_fixtureThe conclusion depends on fixture output.
model_nondeterministicRepeated attempts varied under the same policy.
diverged_but_comparableThe traces diverged but downstream comparison is still useful.
non_comparableBranch drift broke meaningful comparison.
unsupportedEvidence is missing, incomplete, or not replay-backed.

Evidence Graph

Analysis outputs can include evidence nodes and edges. This lets a report cite not just a final answer but the evidence chain behind it.

Evidence nodeExamples
Prompt clauseA policy instruction that changed behavior when masked.
RetrievalA source chunk that introduced wrong facts.
MemoryPersistent state that affected the decision.
StateCheckpoint field changed by an intervention.
Tool fixture/schemaTool output or tool contract that affected replay.
Model configTemperature or provider setting.
Metric deltaDifference in replay metrics.
InterventionThe controlled change being tested.

Evidence edge relations include supports, contradicts, depends_on, changed_by, first_diverged_at, and sensitive_to.

Evidence Strength Ordering

  1. Captured base trace facts are strongest for what happened.
  2. Recorded-only replay is strongest for what changed under a controlled transformation.
  3. Recorded fixtures are stronger than analyst, simulator, AI, or live fixtures.
  4. Repeated comparable trials are stronger than one replay.
  5. Fixture-sensitive or non-comparable findings must stay conditional.
  6. Observability-only managed runtime traces are useful, but do not imply deterministic replay.

Report Wording

Use precise verbs:

SayWhen
observedBase run fact.
under replayControlled branch result.
under fixtureResult depends on supplied fixture.
suggestsEvidence is useful but incomplete.
unsupportedNo replay-backed evidence supports the claim.
non-comparableBranch drift prevents a valid comparison.

Avoid words like "proved" or "guaranteed" unless an external formal proof actually exists.

On this page