fix: preserve learning evidence integrity

This commit is contained in:
m
2026-08-22 02:43:06 +03:00
parent 01f4fa27c2
commit 8a3b085855
8 changed files with 144 additions and 14 deletions
+4 -4
View File
@@ -42,13 +42,13 @@ func ObserveSkillOutcome(profile capability.SkillProfile, observation capability
}
next.LastValidatedUnixMillis = observation.ObservedAtUnixMillis
next.ConfidenceMillis = clampConfidence(next.ConfidenceMillis + 25)
case capability.UsageFailed, capability.UsageUnknown:
case capability.UsageFailed:
next.KnownFailureExperienceRefs = appendEvidenceRef(next.KnownFailureExperienceRefs, observation.ExperienceRef)
next.ObservedFailureCount++
next.ConfidenceMillis = clampConfidence(next.ConfidenceMillis - 150)
case capability.UsageSkipped:
next.KnownFailureExperienceRefs = appendEvidenceRef(next.KnownFailureExperienceRefs, observation.ExperienceRef)
next.ConfidenceMillis = clampConfidence(next.ConfidenceMillis - 25)
case capability.UsageUnknown, capability.UsageSkipped:
// UNKNOWN and SKIPPED are evidence of incomplete observation or a safe
// non-application, not evidence that the skill itself failed.
default:
return capability.SkillProfile{}, ErrInvalidSkillProfile
}
+14
View File
@@ -44,6 +44,20 @@ func TestValidatedSkillCanApplyFastPathWhenFreshAndConfident(t *testing.T) {
}
}
func TestUnknownAndSkippedOutcomesDoNotBecomeSkillFailures(t *testing.T) {
profile := skillProfile()
for _, outcome := range []capability.UsageOutcomeClass{capability.UsageUnknown, capability.UsageSkipped} {
next, err := ObserveSkillOutcome(profile, capability.SkillOutcomeObservation{ExperienceRef: capability.SkillEvidenceRef("experience:" + string(outcome)), Outcome: outcome, ObservedAtUnixMillis: profile.RecordedAtUnixMillis + 1})
if err != nil {
t.Fatal(err)
}
if next.ObservedFailureCount != profile.ObservedFailureCount || len(next.KnownFailureExperienceRefs) != len(profile.KnownFailureExperienceRefs) || next.ConfidenceMillis != profile.ConfidenceMillis {
t.Fatalf("outcome %q was incorrectly recorded as failure: %#v", outcome, next)
}
profile = next
}
}
func skillProfile() capability.SkillProfile {
return capability.SkillProfile{
ID: "skill-obvious-reminder-response-v1",
+13 -2
View File
@@ -54,9 +54,11 @@ func NewExperienceRecord(value epistemics.ExperienceRecord) (epistemics.Experien
}
func ReflectionForExperience(value epistemics.ExperienceRecord) (epistemics.ReflectionCandidate, bool, error) {
if _, err := NewExperienceRecord(value); err != nil {
normalized, err := NewExperienceRecord(value)
if err != nil {
return epistemics.ReflectionCandidate{}, false, err
}
value = normalized
trigger, ok := reflectionTrigger(value)
if !ok {
return epistemics.ReflectionCandidate{}, false, nil
@@ -74,7 +76,7 @@ func ReflectionForExperience(value epistemics.ExperienceRecord) (epistemics.Refl
SourceUtteranceID: value.SourceUtteranceID,
Summary: reflectionSummary(trigger, value),
EvidenceRefs: reflectionEvidence(value),
RequiresCognition: trigger == epistemics.ReflectionContradiction || trigger == epistemics.ReflectionRepeatedDegradation,
RequiresCognition: reflectionNeedsCognition(trigger),
RecordedAtUnixMillis: value.RecordedAtUnixMillis,
Revision: 1,
}
@@ -84,6 +86,15 @@ func ReflectionForExperience(value epistemics.ExperienceRecord) (epistemics.Refl
return candidate, true, nil
}
func reflectionNeedsCognition(trigger epistemics.ReflectionTriggerKind) bool {
switch trigger {
case epistemics.ReflectionUserCorrection, epistemics.ReflectionUnexpectedOutcome, epistemics.ReflectionFailure, epistemics.ReflectionLowConfidence, epistemics.ReflectionContradiction, epistemics.ReflectionRepeatedDegradation:
return true
default:
return false
}
}
func ValidateReflectionCandidate(value epistemics.ReflectionCandidate) error {
if value.ID == "" || value.Trigger == "" || value.State == "" || value.SourceExperienceID == "" ||
value.OriginalIntentID == "" || value.SourceUtteranceID == "" || strings.TrimSpace(value.Summary) == "" ||
+18 -1
View File
@@ -50,11 +50,28 @@ func TestUserCorrectionCreatesReflectionCandidate(t *testing.T) {
if err != nil || !ok {
t.Fatalf("expected reflection candidate, ok=%v err=%v", ok, err)
}
if candidate.Trigger != epistemics.ReflectionUserCorrection || candidate.State != epistemics.ReflectionCandidateOpen || len(candidate.EvidenceRefs) == 0 {
if candidate.Trigger != epistemics.ReflectionUserCorrection || candidate.State != epistemics.ReflectionCandidateOpen || !candidate.RequiresCognition || len(candidate.EvidenceRefs) == 0 {
t.Fatalf("unexpected reflection candidate: %#v", candidate)
}
}
func TestReflectionNormalizesExperienceIDBeforeReferencingIt(t *testing.T) {
candidate, ok, err := ReflectionForExperience(epistemics.ExperienceRecord{
SourceKind: epistemics.ExperienceSourceInteraction,
OriginalIntentID: "intent-1",
SourceUtteranceID: "utterance-3",
HelpMethod: epistemics.HelpMethodClarification,
ExpectedOutcome: epistemics.OutcomeSuccess,
ActualOutcome: epistemics.OutcomeFailed,
OutcomeDetail: "deterministic route failed",
ConfidenceMillis: 800,
RecordedAtUnixMillis: 1200,
})
if err != nil || !ok || candidate.SourceExperienceID == "" || !candidate.RequiresCognition {
t.Fatalf("reflection did not preserve normalized source experience: candidate=%#v ok=%v err=%v", candidate, ok, err)
}
}
func TestFastPathDeclineReturnsToThinkingInsteadOfGuessing(t *testing.T) {
experience, err := FastPathDeclineExperience("intent-1", "utterance-1", "skill:obvious-reminder-response-v1", string(epistemics.OutcomeOutOfDistribution), 1200)
if err != nil {