agent-nuvira 3.3.13 → 3.3.15

package/dist/learning/behavioural-labels.d.ts

@@ -0,0 +1,86 @@
+/**
+ * Tier-3 behavioural labels — the verdict the harness can INFER from what the
+ * user DID, without the user saying anything.
+ *
+ * WHY A THIRD TIER. An explicit verdict (`nuvira rate`, the dashboard Trace tab)
+ * is the only source of a POSITIVE label, but it only exists when the user
+ * bothers to give one. Between "the user rated it" and "the user said nothing at
+ * all" there is behaviour the harness can OBSERVE: a re-ask, a hand-edit of the
+ * delivered file, an untouched artifact the user then referenced. Those are
+ * inferences, not statements, so they carry `source: 'derived'` and a fit can
+ * weigh them BELOW a human judgement.
+ *
+ * THE RULE THAT KEEPS THIS HONEST. Silence is still NOT acceptance. An untouched
+ * artifact is not a label on its own — it becomes a (weak) acceptance only when
+ * the user's next message REFERENCES that artifact without reporting a regression.
+ * A raw untouched file the user never mentioned stays `null`, exactly like a turn
+ * nobody rated. The three signals, in priority order:
+ *
+ *   1. repeat-ask         — the new ask is near-identical to the delivered ask
+ *                           (Jaccard similarity), so the delivery did not land.  → rejected
+ *   2. hand-edit          — the delivered file changed after the turn made it,
+ *                           so the user had to fix it.                          → rejected
+ *   3. unchanged-referenced — the file is untouched AND the new message names it
+ *                           AND no regression was reported.                      → accepted
+ *
+ * NOTHING ROUTES ON THIS. It labels the deliverable corpus and the turn's trace,
+ * the same two places an explicit verdict writes, and that is all.
+ *
+ * The similarity and reference tests are MEASURED (token overlap, a filename
+ * appearing in the text) — never a phrase list of "sounds like they liked it".
+ */
+/** Why a behavioural label was derived — kept so a reader can audit the inference. */
+export type BehaviouralReason = 'repeat-ask' | 'hand-edit' | 'unchanged-referenced';
+/** A derived verdict. `source` is ALWAYS `'derived'` — that is the point of the tier. */
+export interface BehaviouralLabel {
+    verdict: 'accepted' | 'rejected';
+    source: 'derived';
+    reason: BehaviouralReason;
+}
+/**
+ * Token-set Jaccard at or above which two asks are the SAME request.
+ *
+ * Deliberately high: a re-ask is usually near-verbatim, and a false positive here
+ * would mark a genuinely NEW request as a rejection of the last one. Two different
+ * asks about the same subject share only their common words, which pulls the score
+ * well below this line.
+ */
+export declare const REPEAT_ASK_SIMILARITY = 0.9;
+/** Alphanumeric word tokens, lowercased. No stoplist — overlap is measured, not curated. */
+export declare function normalizeAskTokens(text: string): Set<string>;
+/** Jaccard similarity of two asks' token sets (0 when either is empty). */
+export declare function askSimilarity(a: string, b: string): number;
+/** Does the message NAME the artifact — its filename (with or without extension)? */
+export declare function referencesArtifact(newAsk: string, path: string): boolean;
+/** True when the file exists and its mtime is AFTER the delivery (+ slack). Never throws. */
+export declare function wasModifiedAfter(path: string, deliveredAt: number, epsilonMs?: number): boolean;
+/**
+ * The pure decision, so the inference can be unit-tested without a filesystem.
+ * `null` means "no label is justified" — which is the common, honest answer.
+ */
+export declare function classifyBehaviouralLabel(input: {
+    newAsk: string;
+    previousAsk: string;
+    reportsRegression: boolean;
+    artifactPath: string | null;
+    artifactModified: boolean;
+    referencesArtifact: boolean;
+}): BehaviouralLabel | null;
+/**
+ * Read the newest UNLABELLED delivery, infer a behavioural label, and apply it.
+ *
+ * Returns the label that was applied, or null when nothing was justified. Every
+ * write is best-effort.
+ */
+export declare function deriveAndApplyBehaviouralLabel(input: {
+    newAsk: string;
+    reportsRegression: boolean;
+    now?: number;
+}): BehaviouralLabel | null;
+/**
+ * Write a derived label to the same two places an explicit verdict writes: the
+ * delivery's corpus row and (when the row knows its trace) the turn's trace.
+ * Best-effort — a label write must never break a turn.
+ */
+export declare function applyBehaviouralLabel(traceId: string | undefined, label: BehaviouralLabel, now?: number): void;
+//# sourceMappingURL=behavioural-labels.d.ts.map

package/dist/learning/behavioural-labels.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"behavioural-labels.d.ts","sourceRoot":"","sources":["../../src/learning/behavioural-labels.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAWH,sFAAsF;AACtF,MAAM,MAAM,iBAAiB,GAAG,YAAY,GAAG,WAAW,GAAG,sBAAsB,CAAC;AAEpF,yFAAyF;AACzF,MAAM,WAAW,gBAAgB;IAC/B,OAAO,EAAE,UAAU,GAAG,UAAU,CAAC;IACjC,MAAM,EAAE,SAAS,CAAC;IAClB,MAAM,EAAE,iBAAiB,CAAC;CAC3B;AAED;;;;;;;GAOG;AACH,eAAO,MAAM,qBAAqB,MAAM,CAAC;AAKzC,4FAA4F;AAC5F,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,GAAG,CAAC,MAAM,CAAC,CAG5D;AAED,2EAA2E;AAC3E,wBAAgB,aAAa,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,CAQ1D;AAED,qFAAqF;AACrF,wBAAgB,kBAAkB,CAAC,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,OAAO,CAQxE;AAED,6FAA6F;AAC7F,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,SAAS,SAAwB,GAAG,OAAO,CAQ9G;AAED;;;GAGG;AACH,wBAAgB,wBAAwB,CAAC,KAAK,EAAE;IAC9C,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,iBAAiB,EAAE,OAAO,CAAC;IAC3B,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;IAC5B,gBAAgB,EAAE,OAAO,CAAC;IAC1B,kBAAkB,EAAE,OAAO,CAAC;CAC7B,GAAG,gBAAgB,GAAG,IAAI,CAkB1B;AAED;;;;;GAKG;AACH,wBAAgB,8BAA8B,CAAC,KAAK,EAAE;IACpD,MAAM,EAAE,MAAM,CAAC;IACf,iBAAiB,EAAE,OAAO,CAAC;IAC3B,GAAG,CAAC,EAAE,MAAM,CAAC;CACd,GAAG,gBAAgB,GAAG,IAAI,CAyB1B;AAED;;;;GAIG;AACH,wBAAgB,qBAAqB,CACnC,OAAO,EAAE,MAAM,GAAG,SAAS,EAC3B,KAAK,EAAE,gBAAgB,EACvB,GAAG,GAAE,MAAmB,GACvB,IAAI,CAaN"}

package/dist/learning/behavioural-labels.js

@@ -0,0 +1,166 @@
+/**
+ * Tier-3 behavioural labels — the verdict the harness can INFER from what the
+ * user DID, without the user saying anything.
+ *
+ * WHY A THIRD TIER. An explicit verdict (`nuvira rate`, the dashboard Trace tab)
+ * is the only source of a POSITIVE label, but it only exists when the user
+ * bothers to give one. Between "the user rated it" and "the user said nothing at
+ * all" there is behaviour the harness can OBSERVE: a re-ask, a hand-edit of the
+ * delivered file, an untouched artifact the user then referenced. Those are
+ * inferences, not statements, so they carry `source: 'derived'` and a fit can
+ * weigh them BELOW a human judgement.
+ *
+ * THE RULE THAT KEEPS THIS HONEST. Silence is still NOT acceptance. An untouched
+ * artifact is not a label on its own — it becomes a (weak) acceptance only when
+ * the user's next message REFERENCES that artifact without reporting a regression.
+ * A raw untouched file the user never mentioned stays `null`, exactly like a turn
+ * nobody rated. The three signals, in priority order:
+ *
+ *   1. repeat-ask         — the new ask is near-identical to the delivered ask
+ *                           (Jaccard similarity), so the delivery did not land.  → rejected
+ *   2. hand-edit          — the delivered file changed after the turn made it,
+ *                           so the user had to fix it.                          → rejected
+ *   3. unchanged-referenced — the file is untouched AND the new message names it
+ *                           AND no regression was reported.                      → accepted
+ *
+ * NOTHING ROUTES ON THIS. It labels the deliverable corpus and the turn's trace,
+ * the same two places an explicit verdict writes, and that is all.
+ *
+ * The similarity and reference tests are MEASURED (token overlap, a filename
+ * appearing in the text) — never a phrase list of "sounds like they liked it".
+ */
+import { statSync } from 'node:fs';
+import { basename, extname } from 'node:path';
+import { readDeliverableCandidates, labelNewestUnlabelled, labelDeliverableByTrace, } from './deliverable-corpus.js';
+import { recordTraceVerdict } from './reasoning-trace.js';
+/**
+ * Token-set Jaccard at or above which two asks are the SAME request.
+ *
+ * Deliberately high: a re-ask is usually near-verbatim, and a false positive here
+ * would mark a genuinely NEW request as a rejection of the last one. Two different
+ * asks about the same subject share only their common words, which pulls the score
+ * well below this line.
+ */
+export const REPEAT_ASK_SIMILARITY = 0.9;
+/** Filesystem timestamp slack: a write finishing at the same instant as delivery is not an edit. */
+const FILE_MTIME_EPSILON_MS = 250;
+/** Alphanumeric word tokens, lowercased. No stoplist — overlap is measured, not curated. */
+export function normalizeAskTokens(text) {
+    const tokens = (text || '').toLowerCase().match(/[\p{L}\p{N}]+/gu) ?? [];
+    return new Set(tokens.filter((t) => t.length >= 2));
+}
+/** Jaccard similarity of two asks' token sets (0 when either is empty). */
+export function askSimilarity(a, b) {
+    const sa = normalizeAskTokens(a);
+    const sb = normalizeAskTokens(b);
+    if (sa.size === 0 || sb.size === 0)
+        return 0;
+    let inter = 0;
+    for (const t of sa)
+        if (sb.has(t))
+            inter++;
+    const union = sa.size + sb.size - inter;
+    return union === 0 ? 0 : inter / union;
+}
+/** Does the message NAME the artifact — its filename (with or without extension)? */
+export function referencesArtifact(newAsk, path) {
+    if (!path)
+        return false;
+    const text = (newAsk || '').toLowerCase();
+    const base = basename(path).toLowerCase();
+    const stem = base.slice(0, base.length - extname(base).length);
+    if (stem.length >= 2 && normalizeAskTokens(newAsk).has(stem))
+        return true;
+    // The full filename mentioned verbatim (e.g. "guide.md") also counts.
+    return base.length >= 2 && text.includes(base);
+}
+/** True when the file exists and its mtime is AFTER the delivery (+ slack). Never throws. */
+export function wasModifiedAfter(path, deliveredAt, epsilonMs = FILE_MTIME_EPSILON_MS) {
+    try {
+        const st = statSync(path);
+        return st.mtimeMs > deliveredAt + epsilonMs;
+    }
+    catch {
+        // A path we cannot resolve (different cwd, deleted file) is not evidence of an edit.
+        return false;
+    }
+}
+/**
+ * The pure decision, so the inference can be unit-tested without a filesystem.
+ * `null` means "no label is justified" — which is the common, honest answer.
+ */
+export function classifyBehaviouralLabel(input) {
+    // A reported regression is already labelled by the explicit correction path
+    // (`detectRegressionSignal`). Deriving a second label for it would double-count
+    // one event, so the behavioural tier stands aside.
+    if (input.reportsRegression)
+        return null;
+    if (askSimilarity(input.newAsk, input.previousAsk) >= REPEAT_ASK_SIMILARITY) {
+        return { verdict: 'rejected', source: 'derived', reason: 'repeat-ask' };
+    }
+    // An edit is checked BEFORE any acceptance: a file the user rewrote is the
+    // strongest behavioural negative, whether or not they mentioned it.
+    if (input.artifactModified) {
+        return { verdict: 'rejected', source: 'derived', reason: 'hand-edit' };
+    }
+    if (input.referencesArtifact) {
+        return { verdict: 'accepted', source: 'derived', reason: 'unchanged-referenced' };
+    }
+    return null;
+}
+/**
+ * Read the newest UNLABELLED delivery, infer a behavioural label, and apply it.
+ *
+ * Returns the label that was applied, or null when nothing was justified. Every
+ * write is best-effort.
+ */
+export function deriveAndApplyBehaviouralLabel(input) {
+    if (input.reportsRegression)
+        return null;
+    let row;
+    try {
+        const rows = readDeliverableCandidates();
+        row = [...rows].reverse().find((r) => r.verdict === null);
+    }
+    catch {
+        return null;
+    }
+    if (!row)
+        return null;
+    const artifactModified = row.path ? wasModifiedAfter(row.path, row.ts) : false;
+    const referenced = row.path ? referencesArtifact(input.newAsk, row.path) : false;
+    const label = classifyBehaviouralLabel({
+        newAsk: input.newAsk,
+        previousAsk: row.ask,
+        reportsRegression: false,
+        artifactPath: row.path ?? null,
+        artifactModified,
+        referencesArtifact: referenced,
+    });
+    if (!label)
+        return null;
+    applyBehaviouralLabel(row.traceId, label, input.now ?? Date.now());
+    return label;
+}
+/**
+ * Write a derived label to the same two places an explicit verdict writes: the
+ * delivery's corpus row and (when the row knows its trace) the turn's trace.
+ * Best-effort — a label write must never break a turn.
+ */
+export function applyBehaviouralLabel(traceId, label, now = Date.now()) {
+    try {
+        if (traceId) {
+            labelDeliverableByTrace(traceId, label.verdict, now);
+            recordTraceVerdict(traceId, label.verdict, label.source, now);
+        }
+        else {
+            // No trace to hang it on: label the newest unlabelled row directly. The
+            // cue is real (we are looking at that delivery), so this is not invention.
+            labelNewestUnlabelled(label.verdict, now);
+        }
+    }
+    catch {
+        // Best-effort.
+    }
+}
+//# sourceMappingURL=behavioural-labels.js.map

package/dist/learning/behavioural-labels.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"behavioural-labels.js","sourceRoot":"","sources":["../../src/learning/behavioural-labels.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,OAAO,EAAE,QAAQ,EAAE,MAAM,SAAS,CAAC;AACnC,OAAO,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AAC9C,OAAO,EACL,yBAAyB,EACzB,qBAAqB,EACrB,uBAAuB,GACxB,MAAM,yBAAyB,CAAC;AACjC,OAAO,EAAE,kBAAkB,EAAE,MAAM,sBAAsB,CAAC;AAY1D;;;;;;;GAOG;AACH,MAAM,CAAC,MAAM,qBAAqB,GAAG,GAAG,CAAC;AAEzC,oGAAoG;AACpG,MAAM,qBAAqB,GAAG,GAAG,CAAC;AAElC,4FAA4F;AAC5F,MAAM,UAAU,kBAAkB,CAAC,IAAY;IAC7C,MAAM,MAAM,GAAG,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC,KAAK,CAAC,iBAAiB,CAAC,IAAI,EAAE,CAAC;IACzE,OAAO,IAAI,GAAG,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,CAAC,CAAC;AACtD,CAAC;AAED,2EAA2E;AAC3E,MAAM,UAAU,aAAa,CAAC,CAAS,EAAE,CAAS;IAChD,MAAM,EAAE,GAAG,kBAAkB,CAAC,CAAC,CAAC,CAAC;IACjC,MAAM,EAAE,GAAG,kBAAkB,CAAC,CAAC,CAAC,CAAC;IACjC,IAAI,EAAE,CAAC,IAAI,KAAK,CAAC,IAAI,EAAE,CAAC,IAAI,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAC7C,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,CAAC,IAAI,EAAE;QAAE,IAAI,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC;YAAE,KAAK,EAAE,CAAC;IAC3C,MAAM,KAAK,GAAG,EAAE,CAAC,IAAI,GAAG,EAAE,CAAC,IAAI,GAAG,KAAK,CAAC;IACxC,OAAO,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,GAAG,KAAK,CAAC;AACzC,CAAC;AAED,qFAAqF;AACrF,MAAM,UAAU,kBAAkB,CAAC,MAAc,EAAE,IAAY;IAC7D,IAAI,CAAC,IAAI;QAAE,OAAO,KAAK,CAAC;IACxB,MAAM,IAAI,GAAG,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC;IAC1C,MAAM,IAAI,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,WAAW,EAAE,CAAC;IAC1C,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC,CAAC;IAC/D,IAAI,IAAI,CAAC,MAAM,IAAI,CAAC,IAAI,kBAAkB,CAAC,MAAM,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAC;IAC1E,sEAAsE;IACtE,OAAO,IAAI,CAAC,MAAM,IAAI,CAAC,IAAI,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;AACjD,CAAC;AAED,6FAA6F;AAC7F,MAAM,UAAU,gBAAgB,CAAC,IAAY,EAAE,WAAmB,EAAE,SAAS,GAAG,qBAAqB;IACnG,IAAI,CAAC;QACH,MAAM,EAAE,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC;QAC1B,OAAO,EAAE,CAAC,OAAO,GAAG,WAAW,GAAG,SAAS,CAAC;IAC9C,CAAC;IAAC,MAAM,CAAC;QACP,qFAAqF;QACrF,OAAO,KAAK,CAAC;IACf,CAAC;AACH,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,wBAAwB,CAAC,KAOxC;IACC,4EAA4E;IAC5E,gFAAgF;IAChF,mDAAmD;IACnD,IAAI,KAAK,CAAC,iBAAiB;QAAE,OAAO,IAAI,CAAC;IAEzC,IAAI,aAAa,CAAC,KAAK,CAAC,MAAM,EAAE,KAAK,CAAC,WAAW,CAAC,IAAI,qBAAqB,EAAE,CAAC;QAC5E,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,YAAY,EAAE,CAAC;IAC1E,CAAC;IACD,2EAA2E;IAC3E,oEAAoE;IACpE,IAAI,KAAK,CAAC,gBAAgB,EAAE,CAAC;QAC3B,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,WAAW,EAAE,CAAC;IACzE,CAAC;IACD,IAAI,KAAK,CAAC,kBAAkB,EAAE,CAAC;QAC7B,OAAO,EAAE,OAAO,EAAE,UAAU,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,sBAAsB,EAAE,CAAC;IACpF,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,8BAA8B,CAAC,KAI9C;IACC,IAAI,KAAK,CAAC,iBAAiB;QAAE,OAAO,IAAI,CAAC;IACzC,IAAI,GAAG,CAAC;IACR,IAAI,CAAC;QACH,MAAM,IAAI,GAAG,yBAAyB,EAAE,CAAC;QACzC,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,KAAK,IAAI,CAAC,CAAC;IAC5D,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,IAAI,CAAC;IACd,CAAC;IACD,IAAI,CAAC,GAAG;QAAE,OAAO,IAAI,CAAC;IAEtB,MAAM,gBAAgB,GAAG,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,gBAAgB,CAAC,GAAG,CAAC,IAAI,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC;IAC/E,MAAM,UAAU,GAAG,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,kBAAkB,CAAC,KAAK,CAAC,MAAM,EAAE,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC;IACjF,MAAM,KAAK,GAAG,wBAAwB,CAAC;QACrC,MAAM,EAAE,KAAK,CAAC,MAAM;QACpB,WAAW,EAAE,GAAG,CAAC,GAAG;QACpB,iBAAiB,EAAE,KAAK;QACxB,YAAY,EAAE,GAAG,CAAC,IAAI,IAAI,IAAI;QAC9B,gBAAgB;QAChB,kBAAkB,EAAE,UAAU;KAC/B,CAAC,CAAC;IACH,IAAI,CAAC,KAAK;QAAE,OAAO,IAAI,CAAC;IAExB,qBAAqB,CAAC,GAAG,CAAC,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,IAAI,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC;IACnE,OAAO,KAAK,CAAC;AACf,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,qBAAqB,CACnC,OAA2B,EAC3B,KAAuB,EACvB,MAAc,IAAI,CAAC,GAAG,EAAE;IAExB,IAAI,CAAC;QACH,IAAI,OAAO,EAAE,CAAC;YACZ,uBAAuB,CAAC,OAAO,EAAE,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC;YACrD,kBAAkB,CAAC,OAAO,EAAE,KAAK,CAAC,OAAO,EAAE,KAAK,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;QAChE,CAAC;aAAM,CAAC;YACN,wEAAwE;YACxE,2EAA2E;YAC3E,qBAAqB,CAAC,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC;QAC5C,CAAC;IACH,CAAC;IAAC,MAAM,CAAC;QACP,eAAe;IACjB,CAAC;AACH,CAAC"}

package/dist/learning/capability-evidence.d.ts

@@ -0,0 +1,167 @@
+/**
+ * CAPABILITY BY MEASUREMENT (Bundle 3, B1–B4) — the scorecard.
+ *
+ * THE DEFECT THIS REPLACES. `AutoModelRouter.getModelCapabilities()` judges a
+ * model from its NAME: `mini`/`tiny`/`flash`/`lite` each SUBTRACT 0.1 from
+ * reasoning (so `gemini-3.1-flash-lite` loses 0.2 for two words), `pro`/`large`/
+ * `opus` add, a `70b` in the id adds 0.45, and a frontier-family keyword adds
+ * 0.2. Nothing in the system had ever measured whether a model is any good, so
+ * max mode's reasoning floor was a NAME threshold, and this programme's own
+ * investigation fell into the trap it was reporting: it called DeepSeek V4.1
+ * Flash weak *because of the word "flash"* while that model served an 82-step
+ * build.
+ *
+ * THE REPLACEMENT. One record per provider × model holding five named
+ * parameters, each 0–100 with its own sample count, fed from what the harness
+ * OBSERVES while doing real work:
+ *
+ *   accuracy     did the work actually verify          (TurnReport.verification)
+ *   performance  how fast it answered, per call        (registry latency EMA)
+ *   cost         what the work cost                    (measured tokens × price)
+ *   robustness   does it hold up across many calls     (errorRate / partialRate)
+ *   ecosystem    what the pair can do besides answer   (tool-calling, window)
+ *
+ * THE PRIOR RULE (§3.2), and the one thing to read before changing anything:
+ * **0 samples returns the declared prior, byte-for-byte.** Measurement replaces
+ * the prior as samples accumulate (`PRIOR_FULL_SAMPLES`) — a wrong prior can
+ * outrank evidence for a while, but never after the record has real samples.
+ * The id-substring terms are REMOVED, not capped: a capped name hint is still a
+ * name judgement, and "the name contributes nothing" is the whole acceptance.
+ *
+ * WHAT THIS MODULE DOES NOT DO. It is pure — no registry, no I/O — so the
+ * arithmetic is testable in isolation and the storage layer decides when to call
+ * it. It never invents a sample: a turn that verified nothing contributes
+ * NOTHING (not a neutral 50), because a run that never checks its work has no
+ * accuracy evidence, and a fabricated neutral sample would dilute the real ones.
+ */
+/** The five named parameters, in the order `model explain` prints them. */
+export type CapabilityParameter = 'accuracy' | 'performance' | 'cost' | 'robustness' | 'ecosystem';
+/** How many observations a parameter needs before it is reported as MEASURED. */
+export declare const MIN_SAMPLES_FOR_EVIDENCE = 5;
+/**
+ * Samples at which the prior's weight reaches zero. Between 0 and here the prior
+ * decays LINEARLY (`priorWeight = 1 - samples / PRIOR_FULL_SAMPLES`), so a single
+ * lucky or unlucky turn cannot swing a ranking, and a model with a real track
+ * record is judged by it. Stated here rather than inline because §7 of the design
+ * calls the decay rate a calibration decision, not a constant to bury.
+ */
+export declare const PRIOR_FULL_SAMPLES: number;
+/** EMA weight for one observation. Matches the registry's latency/token EMAs. */
+export declare const EVIDENCE_ALPHA = 0.3;
+/** Where a reported value came from. Printed beside every number. */
+export type EvidenceSource = 'measured' | 'prior';
+/** One parameter: the measured value (0–100), how many samples, and its source. */
+export interface ParameterScore {
+    /** Measured value 0–100. Meaningless on its own — always pair it with `samples`. */
+    measured: number;
+    /** Observations folded in. 0 = nothing measured yet. */
+    samples: number;
+}
+/**
+ * The scorecard for one provider × model. Every parameter is optional: absent
+ * means "nothing measured", which reports as the prior rather than as a zero.
+ */
+export interface CapabilityRecord {
+    accuracy?: ParameterScore;
+    performance?: ParameterScore;
+    cost?: ParameterScore;
+    robustness?: ParameterScore;
+    ecosystem?: ParameterScore;
+    /** Epoch ms of the last fold. */
+    updatedAt: number;
+}
+/**
+ * Accuracy weights per turn verdict. `verified` is the strongest evidence the
+ * harness can produce (a change was made AND observed). `unverified` is real but
+ * weak evidence: the model answered, and its work could not be confirmed — so it
+ * pulls the score down without being scored as an outright failure, which would
+ * double-count the same event (`applyReward`'s `verificationPassed: false`
+ * already carries that penalty).
+ *
+ * A `blocked` turn is NOT here on purpose: a wall the run hit is not a verdict
+ * about the model.
+ */
+export declare const ACCURACY_BY_VERDICT: {
+    readonly verified: 100;
+    readonly unverified: 40;
+};
+/** A fresh record: nothing measured. */
+export declare function emptyCapabilityRecord(now?: number): CapabilityRecord;
+/**
+ * Fold a turn's verification verdict into `accuracy`. `blocked`/
+ * `not-applicable`/`delivered-and-read-back`/absent contribute nothing at all (see
+ * the header). `delivered-and-read-back` is NEUTRAL on purpose: it proves an authored
+ * document exists and was read back, not that it is correct or complete, so folding
+ * it either way would invent a quality signal the evidence does not carry.
+ */
+export declare function foldVerification(record: CapabilityRecord, verification: 'verified' | 'unverified' | 'blocked' | 'not-applicable' | 'delivered-and-read-back' | undefined | null, now?: number): CapabilityRecord;
+/** Fold one measured call's outcome into `robustness` (0–100 per call). */
+export declare function foldCallOutcome(record: CapabilityRecord, ok: boolean, now?: number): CapabilityRecord;
+/** Fold a measured latency into `performance` (a call under 800 ms scores 100). */
+export declare function foldLatency(record: CapabilityRecord, latencyMs: number | undefined, now?: number): CapabilityRecord;
+/**
+ * The value to USE for a parameter, and whether it is measured or borrowed.
+ *
+ * THE ONE RULE THAT DECIDES EVERYTHING: with no samples the answer is exactly the
+ * prior passed in. That is what makes this change safe to land before the record
+ * has any data — behaviour on a cold start is unchanged — and it is what makes
+ * the acceptance test "two ids differing only by a tier word score identically"
+ * true, since the prior is provider-level and cannot see the id.
+ */
+export declare function effectiveParameter(record: CapabilityRecord | undefined, parameter: CapabilityParameter, priorFraction: number): {
+    value: number;
+    samples: number;
+    source: EvidenceSource;
+};
+/** Capability tier, DERIVED from the parameters — never parsed from an id. */
+export type CapabilityTier = 'Frontier' | 'Balanced' | 'Utility';
+/**
+ * Derive the tier. Deliberately coarse and stated in one place:
+ *   - `Utility`  — accuracy is poor, or the pair is unreliable;
+ *   - `Frontier` — strong accuracy AND a wide ecosystem;
+ *   - `Balanced` — everything else, which is most things.
+ * `ecosystem` participates because a strong answerer that cannot call tools is
+ * not a frontier AGENT model, and this harness only cares about agent models.
+ */
+export declare function deriveTier(view: {
+    accuracy: {
+        value: number;
+        samples: number;
+    };
+    robustness: {
+        value: number;
+        samples: number;
+    };
+    ecosystem: {
+        value: number;
+        samples: number;
+    };
+}): CapabilityTier;
+/**
+ * Declared priors for the parameters a cold start cannot supply.
+ *
+ * `accuracy` and `performance` are deliberately ABSENT: their priors are the
+ * provider's own declared baseline, which only the router knows. `cost` is
+ * absent too — pricing belongs to a provider ACCOUNT, so with no measurement the
+ * honest answer is "no prior" rather than a number that reads as free:
+ *
+ *   - `robustness: 0.7` — most pairs answer most of the time; a prior, replaced
+ *     by measurement as soon as calls are recorded;
+ *   - `ecosystem: 0.5` — nothing is known about tool-calling until observed, and
+ *     0.5 keeps a pair OUT of the `Frontier` gate (which needs ≥ 0.7). A pair we
+ *     have never watched use a tool is not a frontier agent model.
+ */
+export declare const DEFAULT_PRIORS: Record<CapabilityParameter, number | undefined>;
+/**
+ * The whole scorecard as printable rows: `accuracy 78 (measured, n=14)`.
+ *
+ * The sample count is NOT optional — a score without one is exactly the
+ * plausible-looking number this programme exists to remove — and a parameter with
+ * neither a prior nor a sample prints `n/a` rather than a `0` that would read as
+ * "free" or "broken".
+ */
+export declare function capabilityLines(record: CapabilityRecord | undefined, priors?: Partial<Record<CapabilityParameter, number>>, priorLabels?: Partial<Record<CapabilityParameter, {
+    source: string;
+    fetchedAt: number;
+}>>): string[];
+//# sourceMappingURL=capability-evidence.d.ts.map

package/dist/learning/capability-evidence.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"capability-evidence.d.ts","sourceRoot":"","sources":["../../src/learning/capability-evidence.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAmCG;AAEH,2EAA2E;AAC3E,MAAM,MAAM,mBAAmB,GAAG,UAAU,GAAG,aAAa,GAAG,MAAM,GAAG,YAAY,GAAG,WAAW,CAAC;AAEnG,iFAAiF;AACjF,eAAO,MAAM,wBAAwB,IAAI,CAAC;AAE1C;;;;;;GAMG;AACH,eAAO,MAAM,kBAAkB,QAA+B,CAAC;AAE/D,iFAAiF;AACjF,eAAO,MAAM,cAAc,MAAM,CAAC;AAElC,qEAAqE;AACrE,MAAM,MAAM,cAAc,GAAG,UAAU,GAAG,OAAO,CAAC;AAElD,mFAAmF;AACnF,MAAM,WAAW,cAAc;IAC7B,oFAAoF;IACpF,QAAQ,EAAE,MAAM,CAAC;IACjB,wDAAwD;IACxD,OAAO,EAAE,MAAM,CAAC;CACjB;AAED;;;GAGG;AACH,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,EAAE,cAAc,CAAC;IAC1B,WAAW,CAAC,EAAE,cAAc,CAAC;IAC7B,IAAI,CAAC,EAAE,cAAc,CAAC;IACtB,UAAU,CAAC,EAAE,cAAc,CAAC;IAC5B,SAAS,CAAC,EAAE,cAAc,CAAC;IAC3B,iCAAiC;IACjC,SAAS,EAAE,MAAM,CAAC;CACnB;AAED;;;;;;;;;;GAUG;AACH,eAAO,MAAM,mBAAmB;;;CAGtB,CAAC;AAEX,wCAAwC;AACxC,wBAAgB,qBAAqB,CAAC,GAAG,GAAE,MAAmB,GAAG,gBAAgB,CAEhF;AAYD;;;;;;GAMG;AACH,wBAAgB,gBAAgB,CAC9B,MAAM,EAAE,gBAAgB,EACxB,YAAY,EACR,UAAU,GACV,YAAY,GACZ,SAAS,GACT,gBAAgB,GAChB,yBAAyB,GACzB,SAAS,GACT,IAAI,EACR,GAAG,GAAE,MAAmB,GACvB,gBAAgB,CAOlB;AAED,2EAA2E;AAC3E,wBAAgB,eAAe,CAC7B,MAAM,EAAE,gBAAgB,EACxB,EAAE,EAAE,OAAO,EACX,GAAG,GAAE,MAAmB,GACvB,gBAAgB,CAMlB;AAED,mFAAmF;AACnF,wBAAgB,WAAW,CACzB,MAAM,EAAE,gBAAgB,EACxB,SAAS,EAAE,MAAM,GAAG,SAAS,EAC7B,GAAG,GAAE,MAAmB,GACvB,gBAAgB,CAUlB;AAED;;;;;;;;GAQG;AACH,wBAAgB,kBAAkB,CAChC,MAAM,EAAE,gBAAgB,GAAG,SAAS,EACpC,SAAS,EAAE,mBAAmB,EAC9B,aAAa,EAAE,MAAM,GACpB;IAAE,KAAK,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,cAAc,CAAA;CAAE,CAc5D;AAED,8EAA8E;AAC9E,MAAM,MAAM,cAAc,GAAG,UAAU,GAAG,UAAU,GAAG,SAAS,CAAC;AAEjE;;;;;;;GAOG;AACH,wBAAgB,UAAU,CAAC,IAAI,EAAE;IAC/B,QAAQ,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC;IAC7C,UAAU,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC;IAC/C,SAAS,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC;CAC/C,GAAG,cAAc,CAIjB;AAED;;;;;;;;;;;;;GAaG;AACH,eAAO,MAAM,cAAc,EAAE,MAAM,CAAC,mBAAmB,EAAE,MAAM,GAAG,SAAS,CAM1E,CAAC;AAEF;;;;;;;GAOG;AACH,wBAAgB,eAAe,CAC7B,MAAM,EAAE,gBAAgB,GAAG,SAAS,EACpC,MAAM,GAAE,OAAO,CAAC,MAAM,CAAC,mBAAmB,EAAE,MAAM,CAAC,CAAM,EACzD,WAAW,GAAE,OAAO,CAAC,MAAM,CAAC,mBAAmB,EAAE;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,SAAS,EAAE,MAAM,CAAA;CAAE,CAAC,CAAM,GAC5F,MAAM,EAAE,CAmBV"}

package/dist/learning/capability-evidence.js

@@ -0,0 +1,210 @@
+/**
+ * CAPABILITY BY MEASUREMENT (Bundle 3, B1–B4) — the scorecard.
+ *
+ * THE DEFECT THIS REPLACES. `AutoModelRouter.getModelCapabilities()` judges a
+ * model from its NAME: `mini`/`tiny`/`flash`/`lite` each SUBTRACT 0.1 from
+ * reasoning (so `gemini-3.1-flash-lite` loses 0.2 for two words), `pro`/`large`/
+ * `opus` add, a `70b` in the id adds 0.45, and a frontier-family keyword adds
+ * 0.2. Nothing in the system had ever measured whether a model is any good, so
+ * max mode's reasoning floor was a NAME threshold, and this programme's own
+ * investigation fell into the trap it was reporting: it called DeepSeek V4.1
+ * Flash weak *because of the word "flash"* while that model served an 82-step
+ * build.
+ *
+ * THE REPLACEMENT. One record per provider × model holding five named
+ * parameters, each 0–100 with its own sample count, fed from what the harness
+ * OBSERVES while doing real work:
+ *
+ *   accuracy     did the work actually verify          (TurnReport.verification)
+ *   performance  how fast it answered, per call        (registry latency EMA)
+ *   cost         what the work cost                    (measured tokens × price)
+ *   robustness   does it hold up across many calls     (errorRate / partialRate)
+ *   ecosystem    what the pair can do besides answer   (tool-calling, window)
+ *
+ * THE PRIOR RULE (§3.2), and the one thing to read before changing anything:
+ * **0 samples returns the declared prior, byte-for-byte.** Measurement replaces
+ * the prior as samples accumulate (`PRIOR_FULL_SAMPLES`) — a wrong prior can
+ * outrank evidence for a while, but never after the record has real samples.
+ * The id-substring terms are REMOVED, not capped: a capped name hint is still a
+ * name judgement, and "the name contributes nothing" is the whole acceptance.
+ *
+ * WHAT THIS MODULE DOES NOT DO. It is pure — no registry, no I/O — so the
+ * arithmetic is testable in isolation and the storage layer decides when to call
+ * it. It never invents a sample: a turn that verified nothing contributes
+ * NOTHING (not a neutral 50), because a run that never checks its work has no
+ * accuracy evidence, and a fabricated neutral sample would dilute the real ones.
+ */
+/** How many observations a parameter needs before it is reported as MEASURED. */
+export const MIN_SAMPLES_FOR_EVIDENCE = 5;
+/**
+ * Samples at which the prior's weight reaches zero. Between 0 and here the prior
+ * decays LINEARLY (`priorWeight = 1 - samples / PRIOR_FULL_SAMPLES`), so a single
+ * lucky or unlucky turn cannot swing a ranking, and a model with a real track
+ * record is judged by it. Stated here rather than inline because §7 of the design
+ * calls the decay rate a calibration decision, not a constant to bury.
+ */
+export const PRIOR_FULL_SAMPLES = 2 * MIN_SAMPLES_FOR_EVIDENCE;
+/** EMA weight for one observation. Matches the registry's latency/token EMAs. */
+export const EVIDENCE_ALPHA = 0.3;
+/**
+ * Accuracy weights per turn verdict. `verified` is the strongest evidence the
+ * harness can produce (a change was made AND observed). `unverified` is real but
+ * weak evidence: the model answered, and its work could not be confirmed — so it
+ * pulls the score down without being scored as an outright failure, which would
+ * double-count the same event (`applyReward`'s `verificationPassed: false`
+ * already carries that penalty).
+ *
+ * A `blocked` turn is NOT here on purpose: a wall the run hit is not a verdict
+ * about the model.
+ */
+export const ACCURACY_BY_VERDICT = {
+    verified: 100,
+    unverified: 40,
+};
+/** A fresh record: nothing measured. */
+export function emptyCapabilityRecord(now = Date.now()) {
+    return { updatedAt: now };
+}
+/** Fold one 0–100 observation into a parameter's EMA (α = `EVIDENCE_ALPHA`). */
+function foldEMA(score, observation) {
+    const obs = Math.max(0, Math.min(100, observation));
+    if (!score || score.samples <= 0)
+        return { measured: obs, samples: 1 };
+    return {
+        measured: EVIDENCE_ALPHA * obs + (1 - EVIDENCE_ALPHA) * score.measured,
+        samples: score.samples + 1,
+    };
+}
+/**
+ * Fold a turn's verification verdict into `accuracy`. `blocked`/
+ * `not-applicable`/`delivered-and-read-back`/absent contribute nothing at all (see
+ * the header). `delivered-and-read-back` is NEUTRAL on purpose: it proves an authored
+ * document exists and was read back, not that it is correct or complete, so folding
+ * it either way would invent a quality signal the evidence does not carry.
+ */
+export function foldVerification(record, verification, now = Date.now()) {
+    if (verification !== 'verified' && verification !== 'unverified')
+        return record;
+    return {
+        ...record,
+        accuracy: foldEMA(record.accuracy, ACCURACY_BY_VERDICT[verification]),
+        updatedAt: now,
+    };
+}
+/** Fold one measured call's outcome into `robustness` (0–100 per call). */
+export function foldCallOutcome(record, ok, now = Date.now()) {
+    return {
+        ...record,
+        robustness: foldEMA(record.robustness, ok ? 100 : 0),
+        updatedAt: now,
+    };
+}
+/** Fold a measured latency into `performance` (a call under 800 ms scores 100). */
+export function foldLatency(record, latencyMs, now = Date.now()) {
+    if (!latencyMs || latencyMs <= 0)
+        return record;
+    // 800 ms or faster = 100; 10 s or slower = 0; linear between. The scale is
+    // stated, not inferred, so `model explain` can explain a number it prints.
+    const scaled = Math.max(0, Math.min(100, 100 * (1 - (latencyMs - 800) / (10_000 - 800))));
+    return {
+        ...record,
+        performance: foldEMA(record.performance, scaled),
+        updatedAt: now,
+    };
+}
+/**
+ * The value to USE for a parameter, and whether it is measured or borrowed.
+ *
+ * THE ONE RULE THAT DECIDES EVERYTHING: with no samples the answer is exactly the
+ * prior passed in. That is what makes this change safe to land before the record
+ * has any data — behaviour on a cold start is unchanged — and it is what makes
+ * the acceptance test "two ids differing only by a tier word score identically"
+ * true, since the prior is provider-level and cannot see the id.
+ */
+export function effectiveParameter(record, parameter, priorFraction) {
+    const prior = Math.max(0, Math.min(1, priorFraction));
+    const score = record?.[parameter];
+    const samples = score?.samples ?? 0;
+    if (!score || samples <= 0)
+        return { value: prior, samples: 0, source: 'prior' };
+    const priorWeight = Math.max(0, 1 - samples / PRIOR_FULL_SAMPLES);
+    // Stored 0–100 (see `foldEMA`), reported 0–1 like every other capability scalar.
+    const measured = Math.max(0, Math.min(100, score.measured)) / 100;
+    const value = priorWeight * prior + (1 - priorWeight) * measured;
+    return {
+        value: Math.max(0, Math.min(1, value)),
+        samples,
+        source: samples >= MIN_SAMPLES_FOR_EVIDENCE ? 'measured' : 'prior',
+    };
+}
+/**
+ * Derive the tier. Deliberately coarse and stated in one place:
+ *   - `Utility`  — accuracy is poor, or the pair is unreliable;
+ *   - `Frontier` — strong accuracy AND a wide ecosystem;
+ *   - `Balanced` — everything else, which is most things.
+ * `ecosystem` participates because a strong answerer that cannot call tools is
+ * not a frontier AGENT model, and this harness only cares about agent models.
+ */
+export function deriveTier(view) {
+    if (view.accuracy.value < 0.5 || view.robustness.value < 0.5)
+        return 'Utility';
+    if (view.accuracy.value >= 0.8 && view.ecosystem.value >= 0.7)
+        return 'Frontier';
+    return 'Balanced';
+}
+/**
+ * Declared priors for the parameters a cold start cannot supply.
+ *
+ * `accuracy` and `performance` are deliberately ABSENT: their priors are the
+ * provider's own declared baseline, which only the router knows. `cost` is
+ * absent too — pricing belongs to a provider ACCOUNT, so with no measurement the
+ * honest answer is "no prior" rather than a number that reads as free:
+ *
+ *   - `robustness: 0.7` — most pairs answer most of the time; a prior, replaced
+ *     by measurement as soon as calls are recorded;
+ *   - `ecosystem: 0.5` — nothing is known about tool-calling until observed, and
+ *     0.5 keeps a pair OUT of the `Frontier` gate (which needs ≥ 0.7). A pair we
+ *     have never watched use a tool is not a frontier agent model.
+ */
+export const DEFAULT_PRIORS = {
+    accuracy: undefined,
+    performance: undefined,
+    cost: undefined,
+    robustness: 0.7,
+    ecosystem: 0.5,
+};
+/**
+ * The whole scorecard as printable rows: `accuracy 78 (measured, n=14)`.
+ *
+ * The sample count is NOT optional — a score without one is exactly the
+ * plausible-looking number this programme exists to remove — and a parameter with
+ * neither a prior nor a sample prints `n/a` rather than a `0` that would read as
+ * "free" or "broken".
+ */
+export function capabilityLines(record, priors = {}, priorLabels = {}) {
+    return ['accuracy', 'performance', 'cost', 'robustness', 'ecosystem'].map((p) => {
+        const prior = priors[p] ?? DEFAULT_PRIORS[p];
+        if (prior === undefined) {
+            const samples = record?.[p]?.samples ?? 0;
+            return samples > 0
+                ? `${p} ${Math.round((record?.[p]?.measured ?? 0))} (measured, n=${samples})`
+                : `${p} n/a (no prior declared, nothing measured)`;
+        }
+        const { value, samples, source } = effectiveParameter(record, p, prior);
+        const pct = Math.round(value * 100);
+        // §6.2 rule 4: a borrowed number must say WHERE it was borrowed from, so a reader can tell an
+        // observation from a prior. An unlabelled prior still prints the bare `prior` it always did —
+        // the label is added, never substituted.
+        const basis = source === 'measured' ? 'measured' : describePrior(p, priorLabels);
+        return `${p} ${pct} (${basis}${samples > 0 ? `, n=${samples}` : ''})`;
+    });
+}
+/** `prior` alone, or `prior: <source>, <date>` when the prior came from somewhere nameable. */
+function describePrior(p, labels) {
+    const label = labels[p];
+    if (!label)
+        return 'prior';
+    const day = new Date(label.fetchedAt).toISOString().slice(0, 10);
+    return `prior: ${label.source}, ${day}`;
+}
+//# sourceMappingURL=capability-evidence.js.map

package/dist/learning/capability-evidence.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"capability-evidence.js","sourceRoot":"","sources":["../../src/learning/capability-evidence.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAmCG;AAKH,iFAAiF;AACjF,MAAM,CAAC,MAAM,wBAAwB,GAAG,CAAC,CAAC;AAE1C;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAG,CAAC,GAAG,wBAAwB,CAAC;AAE/D,iFAAiF;AACjF,MAAM,CAAC,MAAM,cAAc,GAAG,GAAG,CAAC;AA2BlC;;;;;;;;;;GAUG;AACH,MAAM,CAAC,MAAM,mBAAmB,GAAG;IACjC,QAAQ,EAAE,GAAG;IACb,UAAU,EAAE,EAAE;CACN,CAAC;AAEX,wCAAwC;AACxC,MAAM,UAAU,qBAAqB,CAAC,MAAc,IAAI,CAAC,GAAG,EAAE;IAC5D,OAAO,EAAE,SAAS,EAAE,GAAG,EAAE,CAAC;AAC5B,CAAC;AAED,gFAAgF;AAChF,SAAS,OAAO,CAAC,KAAiC,EAAE,WAAmB;IACrE,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,WAAW,CAAC,CAAC,CAAC;IACpD,IAAI,CAAC,KAAK,IAAI,KAAK,CAAC,OAAO,IAAI,CAAC;QAAE,OAAO,EAAE,QAAQ,EAAE,GAAG,EAAE,OAAO,EAAE,CAAC,EAAE,CAAC;IACvE,OAAO;QACL,QAAQ,EAAE,cAAc,GAAG,GAAG,GAAG,CAAC,CAAC,GAAG,cAAc,CAAC,GAAG,KAAK,CAAC,QAAQ;QACtE,OAAO,EAAE,KAAK,CAAC,OAAO,GAAG,CAAC;KAC3B,CAAC;AACJ,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,gBAAgB,CAC9B,MAAwB,EACxB,YAOQ,EACR,MAAc,IAAI,CAAC,GAAG,EAAE;IAExB,IAAI,YAAY,KAAK,UAAU,IAAI,YAAY,KAAK,YAAY;QAAE,OAAO,MAAM,CAAC;IAChF,OAAO;QACL,GAAG,MAAM;QACT,QAAQ,EAAE,OAAO,CAAC,MAAM,CAAC,QAAQ,EAAE,mBAAmB,CAAC,YAAY,CAAC,CAAC;QACrE,SAAS,EAAE,GAAG;KACf,CAAC;AACJ,CAAC;AAED,2EAA2E;AAC3E,MAAM,UAAU,eAAe,CAC7B,MAAwB,EACxB,EAAW,EACX,MAAc,IAAI,CAAC,GAAG,EAAE;IAExB,OAAO;QACL,GAAG,MAAM;QACT,UAAU,EAAE,OAAO,CAAC,MAAM,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;QACpD,SAAS,EAAE,GAAG;KACf,CAAC;AACJ,CAAC;AAED,mFAAmF;AACnF,MAAM,UAAU,WAAW,CACzB,MAAwB,EACxB,SAA6B,EAC7B,MAAc,IAAI,CAAC,GAAG,EAAE;IAExB,IAAI,CAAC,SAAS,IAAI,SAAS,IAAI,CAAC;QAAE,OAAO,MAAM,CAAC;IAChD,2EAA2E;IAC3E,2EAA2E;IAC3E,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC,SAAS,GAAG,GAAG,CAAC,GAAG,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;IAC1F,OAAO;QACL,GAAG,MAAM;QACT,WAAW,EAAE,OAAO,CAAC,MAAM,CAAC,WAAW,EAAE,MAAM,CAAC;QAChD,SAAS,EAAE,GAAG;KACf,CAAC;AACJ,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,kBAAkB,CAChC,MAAoC,EACpC,SAA8B,EAC9B,aAAqB;IAErB,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,aAAa,CAAC,CAAC,CAAC;IACtD,MAAM,KAAK,GAAG,MAAM,EAAE,CAAC,SAAS,CAAC,CAAC;IAClC,MAAM,OAAO,GAAG,KAAK,EAAE,OAAO,IAAI,CAAC,CAAC;IACpC,IAAI,CAAC,KAAK,IAAI,OAAO,IAAI,CAAC;QAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,CAAC,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC;IACjF,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,GAAG,OAAO,GAAG,kBAAkB,CAAC,CAAC;IAClE,iFAAiF;IACjF,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC,GAAG,GAAG,CAAC;IAClE,MAAM,KAAK,GAAG,WAAW,GAAG,KAAK,GAAG,CAAC,CAAC,GAAG,WAAW,CAAC,GAAG,QAAQ,CAAC;IACjE,OAAO;QACL,KAAK,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC;QACtC,OAAO;QACP,MAAM,EAAE,OAAO,IAAI,wBAAwB,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,OAAO;KACnE,CAAC;AACJ,CAAC;AAKD;;;;;;;GAOG;AACH,MAAM,UAAU,UAAU,CAAC,IAI1B;IACC,IAAI,IAAI,CAAC,QAAQ,CAAC,KAAK,GAAG,GAAG,IAAI,IAAI,CAAC,UAAU,CAAC,KAAK,GAAG,GAAG;QAAE,OAAO,SAAS,CAAC;IAC/E,IAAI,IAAI,CAAC,QAAQ,CAAC,KAAK,IAAI,GAAG,IAAI,IAAI,CAAC,SAAS,CAAC,KAAK,IAAI,GAAG;QAAE,OAAO,UAAU,CAAC;IACjF,OAAO,UAAU,CAAC;AACpB,CAAC;AAED;;;;;;;;;;;;;GAaG;AACH,MAAM,CAAC,MAAM,cAAc,GAAoD;IAC7E,QAAQ,EAAE,SAAS;IACnB,WAAW,EAAE,SAAS;IACtB,IAAI,EAAE,SAAS;IACf,UAAU,EAAE,GAAG;IACf,SAAS,EAAE,GAAG;CACf,CAAC;AAEF;;;;;;;GAOG;AACH,MAAM,UAAU,eAAe,CAC7B,MAAoC,EACpC,SAAuD,EAAE,EACzD,cAA2F,EAAE;IAE7F,OAAQ,CAAC,UAAU,EAAE,aAAa,EAAE,MAAM,EAAE,YAAY,EAAE,WAAW,CAA2B,CAAC,GAAG,CAClG,CAAC,CAAC,EAAE,EAAE;QACJ,MAAM,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,IAAI,cAAc,CAAC,CAAC,CAAC,CAAC;QAC7C,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;YACxB,MAAM,OAAO,GAAG,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,IAAI,CAAC,CAAC;YAC1C,OAAO,OAAO,GAAG,CAAC;gBAChB,CAAC,CAAC,GAAG,CAAC,IAAI,IAAI,CAAC,KAAK,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,QAAQ,IAAI,CAAC,CAAC,CAAC,iBAAiB,OAAO,GAAG;gBAC7E,CAAC,CAAC,GAAG,CAAC,4CAA4C,CAAC;QACvD,CAAC;QACD,MAAM,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,EAAE,GAAG,kBAAkB,CAAC,MAAM,EAAE,CAAC,EAAE,KAAK,CAAC,CAAC;QACxE,MAAM,GAAG,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,GAAG,GAAG,CAAC,CAAC;QACpC,8FAA8F;QAC9F,8FAA8F;QAC9F,yCAAyC;QACzC,MAAM,KAAK,GAAG,MAAM,KAAK,UAAU,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,EAAE,WAAW,CAAC,CAAC;QACjF,OAAO,GAAG,CAAC,IAAI,GAAG,KAAK,KAAK,GAAG,OAAO,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC;IACxE,CAAC,CACF,CAAC;AACJ,CAAC;AAED,+FAA+F;AAC/F,SAAS,aAAa,CACpB,CAAsB,EACtB,MAAmF;IAEnF,MAAM,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;IACxB,IAAI,CAAC,KAAK;QAAE,OAAO,OAAO,CAAC;IAC3B,MAAM,GAAG,GAAG,IAAI,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,WAAW,EAAE,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;IACjE,OAAO,UAAU,KAAK,CAAC,MAAM,KAAK,GAAG,EAAE,CAAC;AAC1C,CAAC"}