@mnemom/agent-alignment-protocol 0.1.2 → 0.1.3

package/dist/index.d.mts

@@ -678,6 +678,6 @@ declare const CONFLICT_PENALTY_MULTIPLIER = 0.5;
 declare const MIN_WORD_LENGTH = 3;
 /** Maximum features to extract from TF-IDF vectorization */
 declare const MAX_TFIDF_FEATURES = 500;
-declare const ALGORITHM_VERSION = "1.0.0";
+declare const ALGORITHM_VERSION = "1.1.0";
 
 export { ALGORITHM_VERSION, type APTrace, type Action, type ActionCategory, type ActionTarget, type ActionType, type AlignmentCard, type AlignmentCardRequest, type AlignmentCardResponse, type Alternative, type AuditCommitment, type AuditStorage, type AutonomyEnvelope, type AutonomyScope, CONFLICT_PENALTY_MULTIPLIER, type Coherence, type CoherenceResult, type CoherenceResultMessage, DEFAULT_SIMILARITY_THRESHOLD, DEFAULT_SUSTAINED_TURNS_THRESHOLD, type DataSharing, type Decision, type DriftAlert, type DriftAnalysis, type DriftDirection, type DriftIndicator, type Escalation, type EscalationStatus, type EscalationTrigger, type HierarchyType, MAX_TFIDF_FEATURES, MIN_COHERENCE_FOR_PROCEED, MIN_WORD_LENGTH, type MonetaryValue, NEAR_BOUNDARY_THRESHOLD, type Principal, type PrincipalResponse, type PrincipalType, type ProposedCollaboration, type ProposedResolution, type RelationshipType, type RequesterInfo, type Severity, type Signature, type StorageType, type TamperEvidence, type TaskContext, type TraceContext, type TriggerAction, type TriggerCheck, VIOLATION_SEVERITY, type ValueAlignment, type ValueAlignmentDetail, type ValueCoherenceCheck, type ValueCoherenceMessage, type ValueConflict, type ValueConflictResult, type ValueDefinition, type Values, type VerificationMetadata, type VerificationResult, type Violation, type ViolationType, type Warning, checkCoherence, cosineSimilarity, createViolation, detectDrift, extractCardFeatures, extractTraceFeatures, getSelectedAlternative, hadViolations, hasValue, isActionBounded, isActionForbidden, isCardExpired, verifyTrace, wasEscalated };

package/dist/index.d.ts

@@ -678,6 +678,6 @@ declare const CONFLICT_PENALTY_MULTIPLIER = 0.5;
 declare const MIN_WORD_LENGTH = 3;
 /** Maximum features to extract from TF-IDF vectorization */
 declare const MAX_TFIDF_FEATURES = 500;
-declare const ALGORITHM_VERSION = "1.0.0";
+declare const ALGORITHM_VERSION = "1.1.0";
 
 export { ALGORITHM_VERSION, type APTrace, type Action, type ActionCategory, type ActionTarget, type ActionType, type AlignmentCard, type AlignmentCardRequest, type AlignmentCardResponse, type Alternative, type AuditCommitment, type AuditStorage, type AutonomyEnvelope, type AutonomyScope, CONFLICT_PENALTY_MULTIPLIER, type Coherence, type CoherenceResult, type CoherenceResultMessage, DEFAULT_SIMILARITY_THRESHOLD, DEFAULT_SUSTAINED_TURNS_THRESHOLD, type DataSharing, type Decision, type DriftAlert, type DriftAnalysis, type DriftDirection, type DriftIndicator, type Escalation, type EscalationStatus, type EscalationTrigger, type HierarchyType, MAX_TFIDF_FEATURES, MIN_COHERENCE_FOR_PROCEED, MIN_WORD_LENGTH, type MonetaryValue, NEAR_BOUNDARY_THRESHOLD, type Principal, type PrincipalResponse, type PrincipalType, type ProposedCollaboration, type ProposedResolution, type RelationshipType, type RequesterInfo, type Severity, type Signature, type StorageType, type TamperEvidence, type TaskContext, type TraceContext, type TriggerAction, type TriggerCheck, VIOLATION_SEVERITY, type ValueAlignment, type ValueAlignmentDetail, type ValueCoherenceCheck, type ValueCoherenceMessage, type ValueConflict, type ValueConflictResult, type ValueDefinition, type Values, type VerificationMetadata, type VerificationResult, type Violation, type ViolationType, type Warning, checkCoherence, cosineSimilarity, createViolation, detectDrift, extractCardFeatures, extractTraceFeatures, getSelectedAlternative, hadViolations, hasValue, isActionBounded, isActionForbidden, isCardExpired, verifyTrace, wasEscalated };

package/dist/index.js

@@ -54,7 +54,7 @@ var MIN_COHERENCE_FOR_PROCEED = 0.7;
 var CONFLICT_PENALTY_MULTIPLIER = 0.5;
 var MIN_WORD_LENGTH = 3;
 var MAX_TFIDF_FEATURES = 500;
-var ALGORITHM_VERSION = "1.0.0";
+var ALGORITHM_VERSION = "1.1.0";
 
 // src/verification/features.ts
 var STOPWORDS = /* @__PURE__ */ new Set([
@@ -148,15 +148,7 @@ function extractTraceFeatures(trace) {
       features[`escalation:${trace.escalation.escalation_status}`] = 1;
     }
   }
-  const reasoningTokens = tokenize(trace.decision.selection_reasoning);
-  for (const token of reasoningTokens) {
-    features[`content:${token}`] = (features[`content:${token}`] ?? 0) + 0.5;
-  }
   for (const alt of trace.decision.alternatives_considered) {
-    const altTokens = tokenize(alt.description);
-    for (const token of altTokens) {
-      features[`content:${token}`] = (features[`content:${token}`] ?? 0) + 0.25;
-    }
     if (alt.flags) {
       for (const flag of alt.flags) {
         features[`flag:${flag}`] = 1;
@@ -429,7 +421,7 @@ function detectDrift(card, traces, similarityThreshold = DEFAULT_SIMILARITY_THRE
     } else {
       lowSimilarityStreak = [];
     }
-    if (lowSimilarityStreak.length >= sustainedThreshold) {
+    if (lowSimilarityStreak.length === sustainedThreshold) {
       const latest = lowSimilarityStreak[lowSimilarityStreak.length - 1];
       const direction = inferDriftDirection(
         lowSimilarityStreak,

package/dist/index.mjs

@@ -6,7 +6,7 @@ var MIN_COHERENCE_FOR_PROCEED = 0.7;
 var CONFLICT_PENALTY_MULTIPLIER = 0.5;
 var MIN_WORD_LENGTH = 3;
 var MAX_TFIDF_FEATURES = 500;
-var ALGORITHM_VERSION = "1.0.0";
+var ALGORITHM_VERSION = "1.1.0";
 
 // src/verification/features.ts
 var STOPWORDS = /* @__PURE__ */ new Set([
@@ -100,15 +100,7 @@ function extractTraceFeatures(trace) {
       features[`escalation:${trace.escalation.escalation_status}`] = 1;
     }
   }
-  const reasoningTokens = tokenize(trace.decision.selection_reasoning);
-  for (const token of reasoningTokens) {
-    features[`content:${token}`] = (features[`content:${token}`] ?? 0) + 0.5;
-  }
   for (const alt of trace.decision.alternatives_considered) {
-    const altTokens = tokenize(alt.description);
-    for (const token of altTokens) {
-      features[`content:${token}`] = (features[`content:${token}`] ?? 0) + 0.25;
-    }
     if (alt.flags) {
       for (const flag of alt.flags) {
         features[`flag:${flag}`] = 1;
@@ -381,7 +373,7 @@ function detectDrift(card, traces, similarityThreshold = DEFAULT_SIMILARITY_THRE
     } else {
       lowSimilarityStreak = [];
     }
-    if (lowSimilarityStreak.length >= sustainedThreshold) {
+    if (lowSimilarityStreak.length === sustainedThreshold) {
       const latest = lowSimilarityStreak[lowSimilarityStreak.length - 1];
       const direction = inferDriftDirection(
         lowSimilarityStreak,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mnemom/agent-alignment-protocol",
-  "version": "0.1.2",
+  "version": "0.1.3",
   "description": "Agent Alignment Protocol (AAP) - Verification and drift detection for AI agents",
   "main": "dist/index.js",
   "module": "dist/index.mjs",

package/src/constants.ts

@@ -41,4 +41,4 @@ export const MAX_TFIDF_FEATURES = 500;
 
 // Version
 // -------
-export const ALGORITHM_VERSION = "1.0.0";
+export const ALGORITHM_VERSION = "1.1.0";

package/src/verification/api.ts

@@ -370,8 +370,8 @@ export function detectDrift(
       lowSimilarityStreak = [];
     }
 
-    // Check if we've hit the threshold for alerting
-    if (lowSimilarityStreak.length >= sustainedThreshold) {
+    // Check if we've hit the threshold for alerting (== not >= to fire once)
+    if (lowSimilarityStreak.length === sustainedThreshold) {
       const latest = lowSimilarityStreak[lowSimilarityStreak.length - 1];
 
       // Infer drift direction

package/src/verification/features.ts

@@ -98,18 +98,12 @@ export function extractTraceFeatures(trace: APTrace): FeatureVector {
     }
   }
 
-  // Content features from reasoning
-  const reasoningTokens = tokenize(trace.decision.selection_reasoning);
-  for (const token of reasoningTokens) {
-    features[`content:${token}`] = (features[`content:${token}`] ?? 0) + 0.5;
-  }
+  // Note: Content features from reasoning/alternatives are deliberately excluded.
+  // Card features are purely structural, so content tokens dilute cosine
+  // similarity without adding alignment signal. See CALIBRATION.md Section 3.5.
 
-  // Alternative features
+  // Flag features from alternatives (structural, not content)
   for (const alt of trace.decision.alternatives_considered) {
-    const altTokens = tokenize(alt.description);
-    for (const token of altTokens) {
-      features[`content:${token}`] = (features[`content:${token}`] ?? 0) + 0.25;
-    }
     if (alt.flags) {
       for (const flag of alt.flags) {
         features[`flag:${flag}`] = 1.0;