@mossbear/eval-core 0.1.0-alpha.65 → 0.1.0-alpha.67

package/dist/index.d.ts

@@ -7,5 +7,6 @@
 export { type EvalAdapter, type EvalUsage } from './adapter.js';
 export { sanitizeCommand } from './sanitize-command.js';
 export { stripCodeBlocks, type CodeBlockStripResult } from './strip-code-blocks.js';
+export { STEERING_LABELS, STEERING_CLASSIFIER_VERSION, classifySteeringTurn, classifySteeringTurnV1, scoreSteeringBaseline, type SteeringBaselineScore, type SteeringLabel, } from './steering-classification.js';
 export { MAX_PROSE_CHARS, PROSE_REDACTION_CLASSES, redactProse, type ProseRedactionClass, type ProseRedactionReport, } from './redact-prose.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAE,KAAK,WAAW,EAAE,KAAK,SAAS,EAAE,MAAM,cAAc,CAAA;AAE/D,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAA;AAEvD,OAAO,EAAE,eAAe,EAAE,KAAK,oBAAoB,EAAE,MAAM,wBAAwB,CAAA;AAEnF,OAAO,EACL,eAAe,EACf,uBAAuB,EACvB,WAAW,EACX,KAAK,mBAAmB,EACxB,KAAK,oBAAoB,GAC1B,MAAM,mBAAmB,CAAA"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAE,KAAK,WAAW,EAAE,KAAK,SAAS,EAAE,MAAM,cAAc,CAAA;AAE/D,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAA;AAEvD,OAAO,EAAE,eAAe,EAAE,KAAK,oBAAoB,EAAE,MAAM,wBAAwB,CAAA;AAEnF,OAAO,EACL,eAAe,EACf,2BAA2B,EAC3B,oBAAoB,EACpB,sBAAsB,EACtB,qBAAqB,EACrB,KAAK,qBAAqB,EAC1B,KAAK,aAAa,GACnB,MAAM,8BAA8B,CAAA;AAErC,OAAO,EACL,eAAe,EACf,uBAAuB,EACvB,WAAW,EACX,KAAK,mBAAmB,EACxB,KAAK,oBAAoB,GAC1B,MAAM,mBAAmB,CAAA"}

package/dist/index.js

@@ -6,5 +6,6 @@
  */
 export { sanitizeCommand } from './sanitize-command.js';
 export { stripCodeBlocks } from './strip-code-blocks.js';
+export { STEERING_LABELS, STEERING_CLASSIFIER_VERSION, classifySteeringTurn, classifySteeringTurnV1, scoreSteeringBaseline, } from './steering-classification.js';
 export { MAX_PROSE_CHARS, PROSE_REDACTION_CLASSES, redactProse, } from './redact-prose.js';
 //# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAIH,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAA;AAEvD,OAAO,EAAE,eAAe,EAA6B,MAAM,wBAAwB,CAAA;AAEnF,OAAO,EACL,eAAe,EACf,uBAAuB,EACvB,WAAW,GAGZ,MAAM,mBAAmB,CAAA"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAIH,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAA;AAEvD,OAAO,EAAE,eAAe,EAA6B,MAAM,wBAAwB,CAAA;AAEnF,OAAO,EACL,eAAe,EACf,2BAA2B,EAC3B,oBAAoB,EACpB,sBAAsB,EACtB,qBAAqB,GAGtB,MAAM,8BAA8B,CAAA;AAErC,OAAO,EACL,eAAe,EACf,uBAAuB,EACvB,WAAW,GAGZ,MAAM,mBAAmB,CAAA"}

package/dist/steering-classification.d.ts

@@ -0,0 +1,19 @@
+export declare const STEERING_LABELS: readonly ["correction", "steering", "new request", "acknowledgement", "insufficient evidence"];
+export type SteeringLabel = (typeof STEERING_LABELS)[number];
+/** Stored with each prediction so a baseline change is explicit in history. */
+export declare const STEERING_CLASSIFIER_VERSION = "v1";
+/** Conservative, deterministic prediction for one already-redacted user turn. */
+export declare function classifySteeringTurnV1(text: string): SteeringLabel;
+/** The active baseline classifier. Persisted v1 predictions remain immutable. */
+export declare const classifySteeringTurn: typeof classifySteeringTurnV1;
+export type SteeringBaselineScore = {
+    total: number;
+    classified: number;
+    coverage: number;
+    counts: Record<SteeringLabel, number>;
+};
+/** Score label distribution and coverage without writing per-turn predictions. */
+export declare function scoreSteeringBaseline(turns: readonly {
+    text: string;
+}[]): SteeringBaselineScore;
+//# sourceMappingURL=steering-classification.d.ts.map

package/dist/steering-classification.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"steering-classification.d.ts","sourceRoot":"","sources":["../src/steering-classification.ts"],"names":[],"mappings":"AAAA,eAAO,MAAM,eAAe,gGAMlB,CAAA;AAEV,MAAM,MAAM,aAAa,GAAG,CAAC,OAAO,eAAe,CAAC,CAAC,MAAM,CAAC,CAAA;AAE5D,+EAA+E;AAC/E,eAAO,MAAM,2BAA2B,OAAO,CAAA;AAW/C,iFAAiF;AACjF,wBAAgB,sBAAsB,CAAC,IAAI,EAAE,MAAM,GAAG,aAAa,CAYlE;AAED,iFAAiF;AACjF,eAAO,MAAM,oBAAoB,+BAAyB,CAAA;AAE1D,MAAM,MAAM,qBAAqB,GAAG;IAClC,KAAK,EAAE,MAAM,CAAA;IACb,UAAU,EAAE,MAAM,CAAA;IAClB,QAAQ,EAAE,MAAM,CAAA;IAChB,MAAM,EAAE,MAAM,CAAC,aAAa,EAAE,MAAM,CAAC,CAAA;CACtC,CAAA;AAED,kFAAkF;AAClF,wBAAgB,qBAAqB,CACnC,KAAK,EAAE,SAAS;IAAE,IAAI,EAAE,MAAM,CAAA;CAAE,EAAE,GACjC,qBAAqB,CAkBvB"}

package/dist/steering-classification.js

@@ -0,0 +1,55 @@
+export const STEERING_LABELS = [
+    'correction',
+    'steering',
+    'new request',
+    'acknowledgement',
+    'insufficient evidence',
+];
+/** Stored with each prediction so a baseline change is explicit in history. */
+export const STEERING_CLASSIFIER_VERSION = 'v1';
+const ACKNOWLEDGEMENT = /^(?:ok(?:ay)?|yes|yep|got it|thanks|thank you|great|perfect|looks good|lgtm|sounds good)[.! \t\r\n]*$/i;
+const CORRECTION = /\b(?:that's wrong|that is wrong|not what i (?:asked|meant)|you misunderstood|you missed|you forgot|i meant|i said|i asked for|you did(?:n't| not)|undo that|revert that|wrong file|wrong branch)\b/i;
+const STEERING = /\b(?:instead|also|make sure|don't|do not|remember to|keep it|keep the|focus on|before you|after that|one more thing|please preserve)\b/i;
+const REQUEST = /\b(?:can you|could you|please|implement|add|create|write|fix|update|change|remove|build|explain|show me|summarize|compare|find)\b/i;
+/** Conservative, deterministic prediction for one already-redacted user turn. */
+export function classifySteeringTurnV1(text) {
+    const normalized = text
+        .normalize('NFKC')
+        .replace(/[’‘]/g, "'")
+        .trim()
+        .replace(/\s+/g, ' ');
+    if (normalized.length < 3)
+        return 'insufficient evidence';
+    if (ACKNOWLEDGEMENT.test(normalized))
+        return 'acknowledgement';
+    if (CORRECTION.test(normalized))
+        return 'correction';
+    if (STEERING.test(normalized))
+        return 'steering';
+    if (REQUEST.test(normalized))
+        return 'new request';
+    return 'insufficient evidence';
+}
+/** The active baseline classifier. Persisted v1 predictions remain immutable. */
+export const classifySteeringTurn = classifySteeringTurnV1;
+/** Score label distribution and coverage without writing per-turn predictions. */
+export function scoreSteeringBaseline(turns) {
+    const counts = {
+        correction: 0,
+        steering: 0,
+        'new request': 0,
+        acknowledgement: 0,
+        'insufficient evidence': 0,
+    };
+    for (const turn of turns)
+        counts[classifySteeringTurn(turn.text)] += 1;
+    const total = turns.length;
+    const classified = total - counts['insufficient evidence'];
+    return {
+        total,
+        classified,
+        coverage: total === 0 ? 0 : classified / total,
+        counts,
+    };
+}
+//# sourceMappingURL=steering-classification.js.map

package/dist/steering-classification.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"steering-classification.js","sourceRoot":"","sources":["../src/steering-classification.ts"],"names":[],"mappings":"AAAA,MAAM,CAAC,MAAM,eAAe,GAAG;IAC7B,YAAY;IACZ,UAAU;IACV,aAAa;IACb,iBAAiB;IACjB,uBAAuB;CACf,CAAA;AAIV,+EAA+E;AAC/E,MAAM,CAAC,MAAM,2BAA2B,GAAG,IAAI,CAAA;AAE/C,MAAM,eAAe,GACnB,wGAAwG,CAAA;AAC1G,MAAM,UAAU,GACd,qMAAqM,CAAA;AACvM,MAAM,QAAQ,GACZ,yIAAyI,CAAA;AAC3I,MAAM,OAAO,GACX,oIAAoI,CAAA;AAEtI,iFAAiF;AACjF,MAAM,UAAU,sBAAsB,CAAC,IAAY;IACjD,MAAM,UAAU,GAAG,IAAI;SACpB,SAAS,CAAC,MAAM,CAAC;SACjB,OAAO,CAAC,OAAO,EAAE,GAAG,CAAC;SACrB,IAAI,EAAE;SACN,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAA;IACvB,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,uBAAuB,CAAA;IACzD,IAAI,eAAe,CAAC,IAAI,CAAC,UAAU,CAAC;QAAE,OAAO,iBAAiB,CAAA;IAC9D,IAAI,UAAU,CAAC,IAAI,CAAC,UAAU,CAAC;QAAE,OAAO,YAAY,CAAA;IACpD,IAAI,QAAQ,CAAC,IAAI,CAAC,UAAU,CAAC;QAAE,OAAO,UAAU,CAAA;IAChD,IAAI,OAAO,CAAC,IAAI,CAAC,UAAU,CAAC;QAAE,OAAO,aAAa,CAAA;IAClD,OAAO,uBAAuB,CAAA;AAChC,CAAC;AAED,iFAAiF;AACjF,MAAM,CAAC,MAAM,oBAAoB,GAAG,sBAAsB,CAAA;AAS1D,kFAAkF;AAClF,MAAM,UAAU,qBAAqB,CACnC,KAAkC;IAElC,MAAM,MAAM,GAAkC;QAC5C,UAAU,EAAE,CAAC;QACb,QAAQ,EAAE,CAAC;QACX,aAAa,EAAE,CAAC;QAChB,eAAe,EAAE,CAAC;QAClB,uBAAuB,EAAE,CAAC;KAC3B,CAAA;IACD,KAAK,MAAM,IAAI,IAAI,KAAK;QAAE,MAAM,CAAC,oBAAoB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,CAAA;IAEtE,MAAM,KAAK,GAAG,KAAK,CAAC,MAAM,CAAA;IAC1B,MAAM,UAAU,GAAG,KAAK,GAAG,MAAM,CAAC,uBAAuB,CAAC,CAAA;IAC1D,OAAO;QACL,KAAK;QACL,UAAU;QACV,QAAQ,EAAE,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,UAAU,GAAG,KAAK;QAC9C,MAAM;KACP,CAAA;AACH,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mossbear/eval-core",
-  "version": "0.1.0-alpha.65",
+  "version": "0.1.0-alpha.67",
   "private": false,
   "description": "Portable pieces of Mossbear's evaluation path — command redaction, the LLM adapter seam, and guide normalization.",
   "license": "MIT",
@@ -38,7 +38,7 @@
     "test": "vitest run"
   },
   "dependencies": {
-    "@mossbear/protocol": "0.1.0-alpha.65"
+    "@mossbear/protocol": "0.1.0-alpha.67"
   },
   "devDependencies": {
     "@types/node": "20.19.43",