Confirmed
Bisect evidence, an explicit root-cause statement, a direct revert, or a validated test demonstrates that the change introduced the failure.
Evidence, confidence, and completeness
A polished answer should never hide missing evidence.
Every fact and inference should expose where it came from, how it was derived, and how certain the system is.
ExactDirectly represented by an authoritative source
DerivedCalculated deterministically from exact evidence
InferredReasoned association with visible uncertainty
Evidence classes
Facts, derivations, and inferences remain different things.
A Git parent SHA, compiler reference, explicit issue link, or human-confirmed decision can be exact evidence. A call-graph edge, merge base, co-change score, or uniquely matched anchor can be derived. A possible regression relationship or likely task invalidation is inferred.
KodeAtlas should never flatten those into one confident narrative.
Regression memory · Planned
Temporal proximity is not causality.
Strongly associated
Corrective work explicitly references the earlier change or the same behavior and affected symbols with compelling supporting evidence.
Possibly associated
A later fix touches the same symbol or dependency shortly after the change, but direct causality is not established.
Temporally related
Events are close in time or share broad scope; the relationship is useful for investigation but must not be framed as causal.
Anchor resolution · In development
Ambiguity is a state, not a reason to fabricate lineage.
The resolver should preserve the original location, report the current candidate, include confidence and rationale, and expose competing candidates when a symbol was split, merged, heavily rewritten, deleted, or otherwise ambiguous.
Manual corrections become auditable evidence with actor, timestamp, rationale, and affected derived outputs. They do not erase the system’s earlier result.
Completeness
Context quality is bounded by the evidence available.
A packet generated without a private fork, missing transcript, unavailable CI history, deleted branch, unresolved symbol mapping, or unarchived incident must say so. The current archive completeness model remains part of every later intelligence surface.
Take the next step
Make uncertainty operational instead of hiding it.
Trust comes from visible evidence, visible limits, and the ability to inspect or correct the path from source record to answer.