@principles/pd-cli 1.96.0 → 1.97.0

package/dist/services/quality-scorecard/local-evaluator.js

@@ -0,0 +1,112 @@
+/**
+ * PRI-361 — Local Evaluator (I/O layer in pd-cli)
+ *
+ * Calls LM Studio for advisory scoring. Uses core validation
+ * to parse LLM responses — no unsafe casts.
+ */
+import { RUBRIC_LABELS, RUBRIC_PROMPTS, RUBRIC_DIMENSIONS as DIMS, meetsMvpThreshold, sumScores, validateLlmScoreResponse, extractJsonFromLlmResponse, } from '@principles/core/quality-scorecard';
+function buildEvaluationPrompt(episode) {
+    const dimensions = DIMS.map(d => `${d} (${RUBRIC_LABELS[d]}): ${RUBRIC_PROMPTS[d]}`).join('\n');
+    return `You are a quality evaluator for an AI agent's pain-signal -> diagnosis -> principle pipeline.
+
+## Task
+Evaluate this pain episode on a 7-dimension rubric. Each dimension scores 0 (fail), 1 (partial), or 2 (pass).
+
+## Pain Episode
+- ID: ${episode.episodeId}
+- Source: ${episode.source}
+- Pain Score: ${episode.score}
+- Severity: ${episode.severity}
+- Summary: ${episode.summary}
+- Created: ${episode.createdAt}
+- Evolution Task Resolution: ${episode.evolutionTaskResolution ?? 'none'}
+- Linked Principles: ${episode.linkedPrinciples.length > 0 ? episode.linkedPrinciples.join(', ') : 'none'}
+- Gate Blocks: ${episode.gateBlockCount}
+
+## Rubric Dimensions
+${dimensions}
+
+## Additional Checks
+- Is the language consistent (not mixing Chinese and English incoherently)?
+- Is the diagnosis/principle overly abstract (no concrete actionable guidance)?
+- Does it fabricate non-existent evidence, axioms, or references?
+
+## Output Format (STRICT JSON)
+Respond with ONLY a JSON object:
+{
+  "scores": { "G1": 0-2, "G2": 0-2, "G3": 0-2, "G4": 0-2, "G5": 0-2, "G6": 0-2, "G7": 0-2 },
+  "rationales": { "G1": "...", "G2": "...", "G3": "...", "G4": "...", "G5": "...", "G6": "...", "G7": "..." },
+  "flags": ["list of issues found"]
+}
+
+Do NOT output anything other than this JSON object.`;
+}
+export async function evaluateWithLocalModel(episode, config, log) {
+    const prompt = buildEvaluationPrompt(episode);
+    const url = `${config.baseUrl.replace(/\/+$/, '')}/chat/completions`;
+    try {
+        const resp = await fetch(url, {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({
+                model: config.model,
+                messages: [
+                    { role: 'system', content: 'You are a precise JSON-output quality evaluator. Output only valid JSON.' },
+                    { role: 'user', content: prompt },
+                ],
+                temperature: 0.1,
+                max_tokens: 2000,
+            }),
+            signal: AbortSignal.timeout(120_000),
+        });
+        if (!resp.ok) {
+            throw new Error(`LM Studio request failed: ${resp.status}`);
+        }
+        const data = (await resp.json());
+        const content = data.choices?.[0]?.message?.content ?? '';
+        const parsed = extractJsonFromLlmResponse(content);
+        if (parsed === null) {
+            throw new Error(`LM Studio returned non-JSON response`);
+        }
+        const { scores, rationales, flags } = validateLlmScoreResponse(parsed);
+        const totalScore = sumScores(scores);
+        return {
+            model: config.model,
+            dimensionScores: scores,
+            dimensionRationales: rationales,
+            totalScore,
+            maxScore: 14,
+            mvpMet: meetsMvpThreshold(scores),
+            flags: flags,
+        };
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        log(`Evaluation error for ${episode.episodeId}: ${msg}`);
+        const zeroScores = Object.fromEntries(DIMS.map(d => [d, 0]));
+        return {
+            model: config.model,
+            dimensionScores: zeroScores,
+            dimensionRationales: Object.fromEntries(DIMS.map(d => [d, `Evaluation failed: ${msg}`])),
+            totalScore: 0,
+            maxScore: 14,
+            mvpMet: false,
+            flags: ['evaluation_error'],
+        };
+    }
+}
+export async function checkLmStudioAvailable(baseUrl) {
+    try {
+        const url = `${baseUrl.replace(/\/+$/, '')}/models`;
+        const resp = await fetch(url, { signal: AbortSignal.timeout(5000) });
+        if (!resp.ok)
+            return { available: false, models: [], error: `HTTP ${resp.status}` };
+        const data = (await resp.json());
+        const models = (data.data || []).map((m) => m.id);
+        return { available: true, models };
+    }
+    catch (err) {
+        return { available: false, models: [], error: err instanceof Error ? err.message : String(err) };
+    }
+}
+//# sourceMappingURL=local-evaluator.js.map

package/dist/services/quality-scorecard/local-evaluator.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"local-evaluator.js","sourceRoot":"","sources":["../../../src/services/quality-scorecard/local-evaluator.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAQH,OAAO,EACL,aAAa,EACb,cAAc,EACd,iBAAiB,IAAI,IAAI,EACzB,iBAAiB,EACjB,SAAS,EACT,wBAAwB,EACxB,0BAA0B,GAC3B,MAAM,oCAAoC,CAAC;AAE5C,SAAS,qBAAqB,CAAC,OAAoB;IACjD,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,KAAK,aAAa,CAAC,CAAC,CAAC,MAAM,cAAc,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAEhG,OAAO;;;;;;QAMD,OAAO,CAAC,SAAS;YACb,OAAO,CAAC,MAAM;gBACV,OAAO,CAAC,KAAK;cACf,OAAO,CAAC,QAAQ;aACjB,OAAO,CAAC,OAAO;aACf,OAAO,CAAC,SAAS;+BACC,OAAO,CAAC,uBAAuB,IAAI,MAAM;uBACjD,OAAO,CAAC,gBAAgB,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,gBAAgB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,MAAM;iBACxF,OAAO,CAAC,cAAc;;;EAGrC,UAAU;;;;;;;;;;;;;;;oDAewC,CAAC;AACrD,CAAC;AAOD,MAAM,CAAC,KAAK,UAAU,sBAAsB,CAC1C,OAAoB,EACpB,MAA4B,EAC5B,GAA0B;IAE1B,MAAM,MAAM,GAAG,qBAAqB,CAAC,OAAO,CAAC,CAAC;IAC9C,MAAM,GAAG,GAAG,GAAG,MAAM,CAAC,OAAO,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC,mBAAmB,CAAC;IAErE,IAAI,CAAC;QACH,MAAM,IAAI,GAAG,MAAM,KAAK,CAAC,GAAG,EAAE;YAC5B,MAAM,EAAE,MAAM;YACd,OAAO,EAAE,EAAE,cAAc,EAAE,kBAAkB,EAAE;YAC/C,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC;gBACnB,KAAK,EAAE,MAAM,CAAC,KAAK;gBACnB,QAAQ,EAAE;oBACR,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,0EAA0E,EAAE;oBACvG,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE;iBAClC;gBACD,WAAW,EAAE,GAAG;gBAChB,UAAU,EAAE,IAAI;aACjB,CAAC;YACF,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,OAAO,CAAC;SACrC,CAAC,CAAC;QAEH,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,CAAC;YACb,MAAM,IAAI,KAAK,CAAC,6BAA6B,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;QAC9D,CAAC;QAED,MAAM,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,IAAI,EAAE,CAAoD,CAAC;QACpF,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,OAAO,IAAI,EAAE,CAAC;QAE1D,MAAM,MAAM,GAAG,0BAA0B,CAAC,OAAO,CAAC,CAAC;QACnD,IAAI,MAAM,KAAK,IAAI,EAAE,CAAC;YACpB,MAAM,IAAI,KAAK,CAAC,sCAAsC,CAAC,CAAC;QAC1D,CAAC;QAED,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,KAAK,EAAE,GAAG,wBAAwB,CAAC,MAAM,CAAC,CAAC;QACvE,MAAM,UAAU,GAAG,SAAS,CAAC,MAAM,CAAC,CAAC;QAErC,OAAO;YACL,KAAK,EAAE,MAAM,CAAC,KAAK;YACnB,eAAe,EAAE,MAAM;YACvB,mBAAmB,EAAE,UAAU;YAC/B,UAAU;YACV,QAAQ,EAAE,EAAE;YACZ,MAAM,EAAE,iBAAiB,CAAC,MAAM,CAAC;YACjC,KAAK,EAAE,KAAK;SACb,CAAC;IACJ,CAAC;IAAC,OAAO,GAAY,EAAE,CAAC;QACtB,MAAM,GAAG,GAAG,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAC7D,GAAG,CAAC,wBAAwB,OAAO,CAAC,SAAS,KAAK,GAAG,EAAE,CAAC,CAAC;QACzD,MAAM,UAAU,GAAG,MAAM,CAAC,WAAW,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAyC,CAAC;QACrG,OAAO;YACL,KAAK,EAAE,MAAM,CAAC,KAAK;YACnB,eAAe,EAAE,UAAU;YAC3B,mBAAmB,EAAE,MAAM,CAAC,WAAW,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,sBAAsB,GAAG,EAAE,CAAC,CAAC,CAAoC;YAC3H,UAAU,EAAE,CAAC;YACb,QAAQ,EAAE,EAAE;YACZ,MAAM,EAAE,KAAK;YACb,KAAK,EAAE,CAAC,kBAAkB,CAAC;SAC5B,CAAC;IACJ,CAAC;AACH,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,sBAAsB,CAAC,OAAe;IAC1D,IAAI,CAAC;QACH,MAAM,GAAG,GAAG,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC,SAAS,CAAC;QACpD,MAAM,IAAI,GAAG,MAAM,KAAK,CAAC,GAAG,EAAE,EAAE,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACrE,IAAI,CAAC,IAAI,CAAC,EAAE;YAAE,OAAO,EAAE,SAAS,EAAE,KAAK,EAAE,MAAM,EAAE,EAAE,EAAE,KAAK,EAAE,QAAQ,IAAI,CAAC,MAAM,EAAE,EAAE,CAAC;QACpF,MAAM,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,IAAI,EAAE,CAA+B,CAAC;QAC/D,MAAM,MAAM,GAAG,CAAC,IAAI,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;QAClD,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;IACrC,CAAC;IAAC,OAAO,GAAY,EAAE,CAAC;QACtB,OAAO,EAAE,SAAS,EAAE,KAAK,EAAE,MAAM,EAAE,EAAE,EAAE,KAAK,EAAE,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC;IACnG,CAAC;AACH,CAAC"}

package/dist/services/quality-scorecard/strong-model-gate.d.ts

@@ -0,0 +1,14 @@
+/**
+ * PRI-361 — Strong Model Adjudication Gate (I/O layer in pd-cli)
+ *
+ * Calls cloud model for adjudication. Uses core validation
+ * to parse responses — no unsafe casts.
+ */
+import type { PainEpisode, LocalEvaluation, StrongModelAdjudication, AdjudicationStatus } from '@principles/core/quality-scorecard';
+export declare function adjudicate(episode: PainEpisode, localEval: LocalEvaluation, config: {
+    modelId: string;
+    log: (msg: string) => void;
+}): Promise<StrongModelAdjudication>;
+export declare function skippedAdjudication(reason: string): StrongModelAdjudication;
+export declare function determineFinalLabel(localEval: LocalEvaluation, adjudication: StrongModelAdjudication | null): AdjudicationStatus;
+//# sourceMappingURL=strong-model-gate.d.ts.map

package/dist/services/quality-scorecard/strong-model-gate.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"strong-model-gate.d.ts","sourceRoot":"","sources":["../../../src/services/quality-scorecard/strong-model-gate.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EACV,WAAW,EACX,eAAe,EACf,uBAAuB,EACvB,kBAAkB,EACnB,MAAM,oCAAoC,CAAC;AAkD5C,wBAAsB,UAAU,CAC9B,OAAO,EAAE,WAAW,EACpB,SAAS,EAAE,eAAe,EAC1B,MAAM,EAAE;IAAE,OAAO,EAAE,MAAM,CAAC;IAAC,GAAG,EAAE,CAAC,GAAG,EAAE,MAAM,KAAK,IAAI,CAAA;CAAE,GACtD,OAAO,CAAC,uBAAuB,CAAC,CAsElC;AAED,wBAAgB,mBAAmB,CAAC,MAAM,EAAE,MAAM,GAAG,uBAAuB,CAS3E;AAED,wBAAgB,mBAAmB,CACjC,SAAS,EAAE,eAAe,EAC1B,YAAY,EAAE,uBAAuB,GAAG,IAAI,GAC3C,kBAAkB,CAOpB"}

package/dist/services/quality-scorecard/strong-model-gate.js

@@ -0,0 +1,128 @@
+/**
+ * PRI-361 — Strong Model Adjudication Gate (I/O layer in pd-cli)
+ *
+ * Calls cloud model for adjudication. Uses core validation
+ * to parse responses — no unsafe casts.
+ */
+import { RUBRIC_LABELS, RUBRIC_DIMENSIONS as DIMS, meetsMvpThreshold, validateAdjudicationResponse, extractJsonFromLlmResponse, } from '@principles/core/quality-scorecard';
+function buildAdjudicationPrompt(episode, localEval) {
+    const localScores = DIMS.map(d => `- ${d} (${RUBRIC_LABELS[d]}): ${localEval.dimensionScores[d]}/2 — ${localEval.dimensionRationales[d]}`).join('\n');
+    return `You are a senior quality adjudicator for an AI agent evolution pipeline.
+Your job is to independently re-evaluate a pain episode that was first scored by a local (smaller) model.
+You must provide your own scores — do NOT simply copy the local model's scores.
+
+## Pain Episode
+- ID: ${episode.episodeId}
+- Source: ${episode.source}
+- Pain Score: ${episode.score}
+- Severity: ${episode.severity}
+- Summary: ${episode.summary}
+- Evolution Task Resolution: ${episode.evolutionTaskResolution ?? 'none'}
+- Linked Principles: ${episode.linkedPrinciples.length > 0 ? episode.linkedPrinciples.join(', ') : 'none'}
+
+## Local Model Scores (${localEval.model})
+${localScores}
+Flags: ${localEval.flags.length > 0 ? localEval.flags.join(', ') : 'none'}
+
+## Your Task
+1. Independently score each dimension (0/1/2) based on the evidence.
+2. Check for: language inconsistency, over-abstraction, fabricated evidence.
+3. If your scores differ from the local model by >=2 points on any dimension, explain why.
+4. Give a final verdict: pass, fail, or needs-review.
+
+## Output Format (STRICT JSON)
+{
+  "scores": { "G1": 0-2, "G2": 0-2, "G3": 0-2, "G4": 0-2, "G5": 0-2, "G6": 0-2, "G7": 0-2 },
+  "rationale": "Overall assessment...",
+  "verdict": "pass" | "fail" | "needs-review"
+}
+
+Do NOT output anything other than this JSON object.`;
+}
+export async function adjudicate(episode, localEval, config) {
+    const { modelId: strongModelId, log } = config;
+    const prompt = buildAdjudicationPrompt(episode, localEval);
+    const baseUrl = process.env.OPENAI_BASE_URL || 'https://api.openai.com/v1';
+    const apiKey = process.env.OPENAI_API_KEY;
+    if (!apiKey) {
+        return {
+            model: strongModelId,
+            adjudicationStatus: 'needs-review',
+            confirmedScores: null,
+            confirmedMvpMet: null,
+            rationale: 'OPENAI_API_KEY not set — cannot run strong-model adjudication',
+            nextAction: 'Set OPENAI_API_KEY and re-run with --strong-model',
+        };
+    }
+    try {
+        const resp = await fetch(`${baseUrl}/chat/completions`, {
+            method: 'POST',
+            headers: {
+                'Content-Type': 'application/json',
+                Authorization: `Bearer ${apiKey}`,
+            },
+            body: JSON.stringify({
+                model: strongModelId,
+                messages: [
+                    { role: 'system', content: 'You are a precise JSON-output quality adjudicator. Output only valid JSON.' },
+                    { role: 'user', content: prompt },
+                ],
+                temperature: 0.1,
+                max_tokens: 2000,
+            }),
+            signal: AbortSignal.timeout(120_000),
+        });
+        if (!resp.ok) {
+            throw new Error(`Strong model request failed: ${resp.status}`);
+        }
+        const data = (await resp.json());
+        const content = data.choices?.[0]?.message?.content ?? '';
+        const parsed = extractJsonFromLlmResponse(content);
+        if (parsed === null) {
+            throw new Error('Strong model returned non-JSON');
+        }
+        const validated = validateAdjudicationResponse(parsed);
+        const { scores, verdict } = validated;
+        return {
+            model: strongModelId,
+            adjudicationStatus: verdict,
+            confirmedScores: scores,
+            confirmedMvpMet: meetsMvpThreshold(scores),
+            rationale: validated.rationale,
+            nextAction: null,
+        };
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        log(`Adjudication error: ${msg}`);
+        return {
+            model: strongModelId,
+            adjudicationStatus: 'needs-review',
+            confirmedScores: null,
+            confirmedMvpMet: null,
+            rationale: `Adjudication failed: ${msg}`,
+            nextAction: 'Retry with strong model or manually review',
+        };
+    }
+}
+export function skippedAdjudication(reason) {
+    return {
+        model: 'none',
+        adjudicationStatus: 'skipped',
+        confirmedScores: null,
+        confirmedMvpMet: null,
+        rationale: reason,
+        nextAction: 'Configure and run strong-model adjudication for final quality verdict',
+    };
+}
+export function determineFinalLabel(localEval, adjudication) {
+    if (!adjudication || adjudication.adjudicationStatus === 'skipped') {
+        if (localEval.mvpMet && localEval.totalScore >= 12)
+            return 'local-pass';
+        if (localEval.totalScore <= 6)
+            return 'local-fail';
+        return 'needs-review';
+    }
+    return adjudication.adjudicationStatus;
+}
+//# sourceMappingURL=strong-model-gate.js.map

package/dist/services/quality-scorecard/strong-model-gate.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"strong-model-gate.js","sourceRoot":"","sources":["../../../src/services/quality-scorecard/strong-model-gate.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAQH,OAAO,EACL,aAAa,EACb,iBAAiB,IAAI,IAAI,EACzB,iBAAiB,EACjB,4BAA4B,EAC5B,0BAA0B,GAC3B,MAAM,oCAAoC,CAAC;AAE5C,SAAS,uBAAuB,CAC9B,OAAoB,EACpB,SAA0B;IAE1B,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAC/B,KAAK,CAAC,KAAK,aAAa,CAAC,CAAC,CAAC,MAAM,SAAS,CAAC,eAAe,CAAC,CAAC,CAAC,QAAQ,SAAS,CAAC,mBAAmB,CAAC,CAAC,CAAC,EAAE,CACxG,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAEb,OAAO;;;;;QAKD,OAAO,CAAC,SAAS;YACb,OAAO,CAAC,MAAM;gBACV,OAAO,CAAC,KAAK;cACf,OAAO,CAAC,QAAQ;aACjB,OAAO,CAAC,OAAO;+BACG,OAAO,CAAC,uBAAuB,IAAI,MAAM;uBACjD,OAAO,CAAC,gBAAgB,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,gBAAgB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,MAAM;;yBAEhF,SAAS,CAAC,KAAK;EACtC,WAAW;SACJ,SAAS,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,MAAM;;;;;;;;;;;;;;;oDAerB,CAAC;AACrD,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,UAAU,CAC9B,OAAoB,EACpB,SAA0B,EAC1B,MAAuD;IAEvD,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,GAAG,EAAE,GAAG,MAAM,CAAC;IAC/C,MAAM,MAAM,GAAG,uBAAuB,CAAC,OAAO,EAAE,SAAS,CAAC,CAAC;IAC3D,MAAM,OAAO,GAAG,OAAO,CAAC,GAAG,CAAC,eAAe,IAAI,2BAA2B,CAAC;IAC3E,MAAM,MAAM,GAAG,OAAO,CAAC,GAAG,CAAC,cAAc,CAAC;IAE1C,IAAI,CAAC,MAAM,EAAE,CAAC;QACZ,OAAO;YACL,KAAK,EAAE,aAAa;YACpB,kBAAkB,EAAE,cAAc;YAClC,eAAe,EAAE,IAAI;YACrB,eAAe,EAAE,IAAI;YACrB,SAAS,EAAE,+DAA+D;YAC1E,UAAU,EAAE,mDAAmD;SAChE,CAAC;IACJ,CAAC;IAED,IAAI,CAAC;QACH,MAAM,IAAI,GAAG,MAAM,KAAK,CAAC,GAAG,OAAO,mBAAmB,EAAE;YACtD,MAAM,EAAE,MAAM;YACd,OAAO,EAAE;gBACP,cAAc,EAAE,kBAAkB;gBAClC,aAAa,EAAE,UAAU,MAAM,EAAE;aAClC;YACD,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC;gBACnB,KAAK,EAAE,aAAa;gBACpB,QAAQ,EAAE;oBACR,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,4EAA4E,EAAE;oBACzG,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE;iBAClC;gBACD,WAAW,EAAE,GAAG;gBAChB,UAAU,EAAE,IAAI;aACjB,CAAC;YACF,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,OAAO,CAAC;SACrC,CAAC,CAAC;QAEH,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,CAAC;YACb,MAAM,IAAI,KAAK,CAAC,gCAAgC,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;QACjE,CAAC;QAED,MAAM,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,IAAI,EAAE,CAAoD,CAAC;QACpF,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,OAAO,IAAI,EAAE,CAAC;QAC1D,MAAM,MAAM,GAAG,0BAA0B,CAAC,OAAO,CAAC,CAAC;QACnD,IAAI,MAAM,KAAK,IAAI,EAAE,CAAC;YACpB,MAAM,IAAI,KAAK,CAAC,gCAAgC,CAAC,CAAC;QACpD,CAAC;QAED,MAAM,SAAS,GAAG,4BAA4B,CAAC,MAAM,CAAC,CAAC;QACvD,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,GAAG,SAAS,CAAC;QAEtC,OAAO;YACL,KAAK,EAAE,aAAa;YACpB,kBAAkB,EAAE,OAAO;YAC3B,eAAe,EAAE,MAAM;YACvB,eAAe,EAAE,iBAAiB,CAAC,MAAM,CAAC;YAC1C,SAAS,EAAE,SAAS,CAAC,SAAS;YAC9B,UAAU,EAAE,IAAI;SACjB,CAAC;IACJ,CAAC;IAAC,OAAO,GAAY,EAAE,CAAC;QACtB,MAAM,GAAG,GAAG,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAC7D,GAAG,CAAC,uBAAuB,GAAG,EAAE,CAAC,CAAC;QAClC,OAAO;YACL,KAAK,EAAE,aAAa;YACpB,kBAAkB,EAAE,cAAc;YAClC,eAAe,EAAE,IAAI;YACrB,eAAe,EAAE,IAAI;YACrB,SAAS,EAAE,wBAAwB,GAAG,EAAE;YACxC,UAAU,EAAE,4CAA4C;SACzD,CAAC;IACJ,CAAC;AACH,CAAC;AAED,MAAM,UAAU,mBAAmB,CAAC,MAAc;IAChD,OAAO;QACL,KAAK,EAAE,MAAM;QACb,kBAAkB,EAAE,SAAS;QAC7B,eAAe,EAAE,IAAI;QACrB,eAAe,EAAE,IAAI;QACrB,SAAS,EAAE,MAAM;QACjB,UAAU,EAAE,uEAAuE;KACpF,CAAC;AACJ,CAAC;AAED,MAAM,UAAU,mBAAmB,CACjC,SAA0B,EAC1B,YAA4C;IAE5C,IAAI,CAAC,YAAY,IAAI,YAAY,CAAC,kBAAkB,KAAK,SAAS,EAAE,CAAC;QACnE,IAAI,SAAS,CAAC,MAAM,IAAI,SAAS,CAAC,UAAU,IAAI,EAAE;YAAE,OAAO,YAAY,CAAC;QACxE,IAAI,SAAS,CAAC,UAAU,IAAI,CAAC;YAAE,OAAO,YAAY,CAAC;QACnD,OAAO,cAAc,CAAC;IACxB,CAAC;IACD,OAAO,YAAY,CAAC,kBAAkB,CAAC;AACzC,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@principles/pd-cli",
-  "version": "1.96.0",
+  "version": "1.97.0",
   "description": "PD CLI — Pain recording, sample management, and evolution tasks",
   "type": "module",
   "bin": {

package/src/commands/quality-scorecard.ts

@@ -0,0 +1,272 @@
+/**
+ * pd quality scorecard — CLI command (PRI-361)
+ *
+ * JSON contract: --json mode outputs EXACTLY one JSON object to stdout.
+ * All progress/diagnostic output goes to stderr.
+ * Errors produce structured JSON: { ok: false, error, nextAction }.
+ */
+
+import { mkdirSync, writeFileSync } from 'fs';
+import { dirname } from 'path';
+import type {
+  EpisodeEvaluation,
+  QualityScorecardReport,
+  StrongModelAdjudication,
+} from '@principles/core/quality-scorecard';
+import {
+  validateCliOptions,
+  needsAdjudication,
+  generateMarkdownReport,
+  generateHtmlReport,
+  generateJsonReport,
+} from '@principles/core/quality-scorecard';
+import { extractEpisodes, extractLogStats } from '../services/quality-scorecard/data-extractor.js';
+import { evaluateWithLocalModel, checkLmStudioAvailable } from '../services/quality-scorecard/local-evaluator.js';
+import { adjudicate, skippedAdjudication, determineFinalLabel } from '../services/quality-scorecard/strong-model-gate.js';
+
+// ── Logging: stderr only, silent in JSON mode ──────────────────────
+
+let jsonMode = false;
+
+function log(msg: string): void {
+  if (!jsonMode) {
+    process.stderr.write(msg + '\n');
+  }
+}
+
+// ── Structured JSON output helpers ─────────────────────────────────
+
+function writeJsonOutput(data: unknown): void {
+  process.stdout.write(JSON.stringify(data, null, 2) + '\n');
+}
+
+function writeJsonError(error: string, nextAction: string): void {
+  writeJsonOutput({ ok: false, error, nextAction });
+}
+
+// ── Summary computation ────────────────────────────────────────────
+
+function computeSummary(evaluations: EpisodeEvaluation[]) {
+  const totalEpisodes = evaluations.length;
+  // localPassCount/localFailCount: based strictly on local model's own conclusion
+  // (localEval.mvpMet + totalScore), NOT finalLabel which may incorporate strong-model adjudication.
+  const localPassCount = evaluations.filter(e => e.localEvaluation.mvpMet && e.localEvaluation.totalScore >= 12).length;
+  const localFailCount = evaluations.filter(e => e.localEvaluation.totalScore <= 6).length;
+  const strongModelReviewedCount = evaluations.filter(e =>
+    e.strongModelAdjudication && e.strongModelAdjudication.adjudicationStatus !== 'skipped'
+  ).length;
+  const finalPassCount = evaluations.filter(e => e.finalLabel === 'pass').length;
+  const finalFailCount = evaluations.filter(e => e.finalLabel === 'fail').length;
+  const needsReviewCount = evaluations.filter(e => e.finalLabel === 'needs-review').length;
+  const localOnlyCount = evaluations.filter(e => e.finalLabel === 'local-pass' || e.finalLabel === 'local-fail').length;
+  const averageLocalScore = totalEpisodes > 0
+    ? evaluations.reduce((s, e) => s + e.localEvaluation.totalScore, 0) / totalEpisodes
+    : 0;
+  const mvpThresholdMetCount = evaluations.filter(e => e.localEvaluation.mvpMet).length;
+
+  return {
+    totalEpisodes, localPassCount, localFailCount, strongModelReviewedCount,
+    finalPassCount, finalFailCount, needsReviewCount, localOnlyCount,
+    averageLocalScore, mvpThresholdMetCount,
+  };
+}
+
+// ── Main handler ───────────────────────────────────────────────────
+
+export async function handleQualityScorecard(opts: Record<string, unknown>): Promise<void> {
+  const isJson = Boolean(opts.json);
+  jsonMode = isJson;
+
+  // Resolve workspace paths
+  const { resolveWorkspaceDir } = await import('../resolve-workspace.js');
+  const { join } = await import('path');
+  const { existsSync } = await import('fs');
+  const workspace = resolveWorkspaceDir(opts.workspace as string | undefined);
+  const dbPath = join(workspace, '.state', 'trajectory.db');
+  const logsDir = join(workspace, '.state', 'logs');
+
+  // 1. Validate CLI options
+  const { options, errors } = validateCliOptions({
+    dbPath,
+    logsDir,
+    localModelBaseUrl: opts.localUrl ?? 'http://localhost:12341/v1',
+    localModelId: opts.localModel ?? 'qwen3.6-27b-mtp',
+    strongModelId: opts.strongModel ?? null,
+    limit: opts.limit ?? '0',
+    format: isJson ? 'json' : (opts.format ?? 'markdown'),
+    output: opts.output,
+    minPainScore: opts.minScore ?? '50',
+    skipStrongModel: opts.skipStrongModel ?? false,
+  });
+
+  if (errors.length > 0) {
+    const msg = errors.map(e => `${e.field}: ${e.message}`).join('; ');
+    if (isJson) {
+      writeJsonError(msg, 'Fix the invalid options and retry');
+    } else {
+      process.stderr.write(`❌ Invalid options:\n${errors.map(e => `  - ${e.field}: ${e.message}`).join('\n')}\n`);
+    }
+    process.exitCode = 1;
+    return;
+  }
+
+  // 2. Check files exist
+  if (!existsSync(options.dbPath)) {
+    const msg = `trajectory.db not found at: ${options.dbPath}`;
+    if (isJson) {
+      writeJsonError(msg, 'Ensure the workspace has PD data (run PD first to generate trajectory.db)');
+    } else {
+      process.stderr.write(`❌ ${msg}\n`);
+    }
+    process.exitCode = 1;
+    return;
+  }
+
+  // 3. Ensure output directory exists
+  const outputDir = dirname(options.output);
+  if (outputDir && !existsSync(outputDir)) {
+    try {
+      mkdirSync(outputDir, { recursive: true });
+      log(`Created output directory: ${outputDir}`);
+    } catch (err: unknown) {
+      const msg = err instanceof Error ? err.message : String(err);
+      if (isJson) {
+        writeJsonError(`Cannot create output directory: ${msg}`, 'Ensure the output path is writable');
+      } else {
+        process.stderr.write(`❌ Cannot create output directory: ${msg}\n`);
+      }
+      process.exitCode = 1;
+      return;
+    }
+  }
+
+  // 4. Check LM Studio
+  log('🔍 PD Quality Scorecard — Starting...');
+  log(`   DB: ${options.dbPath}`);
+  log(`   Local Model: ${options.localModelId} @ ${options.localModelBaseUrl}`);
+  log(`   Strong Model: ${options.strongModelId ?? 'skipped'}`);
+
+  const lmStatus = await checkLmStudioAvailable(options.localModelBaseUrl);
+  if (!lmStatus.available) {
+    if (isJson) {
+      writeJsonError(`LM Studio not available: ${lmStatus.error}`, 'Start LM Studio or check --local-url');
+    } else {
+      process.stderr.write(`❌ LM Studio not available at ${options.localModelBaseUrl}: ${lmStatus.error}\n`);
+    }
+    process.exitCode = 1;
+    return;
+  }
+
+  if (!lmStatus.models.includes(options.localModelId)) {
+    if (isJson) {
+      writeJsonError(`Model "${options.localModelId}" not found. Available: ${lmStatus.models.join(', ')}`, 'Use --local-model with an available model');
+    } else {
+      process.stderr.write(`❌ Model "${options.localModelId}" not found. Available: ${lmStatus.models.join(', ')}\n`);
+    }
+    process.exitCode = 1;
+    return;
+  }
+
+  // 5. Extract data
+  log('\n📊 Extracting dogfood data...');
+  const { episodes, stats: extractStats } = await extractEpisodes(options.dbPath, {
+    minScore: options.minPainScore,
+    limit: options.limit,
+  });
+  log(`   Found ${episodes.length} unique episodes (total pain events: ${extractStats.total})`);
+
+  const logStats = extractLogStats(options.logsDir);
+  log(`   Event logs: ${logStats.totalEvents} events (${logStats.painSignalCount} pain signals)`);
+
+  // 6. Evaluate each episode
+  log('\n🤖 Running local model evaluation...');
+  const evaluations: EpisodeEvaluation[] = [];
+
+  for (let i = 0; i < episodes.length; i++) {
+    const ep = episodes[i];
+    if (!ep) continue;
+    log(`   [${i + 1}/${episodes.length}] ${ep.episodeId} (score=${ep.score})...`);
+
+    const localEval = await evaluateWithLocalModel(ep, {
+      baseUrl: options.localModelBaseUrl,
+      model: options.localModelId,
+    }, (msg: string) => log(`     ${msg}`));
+    log(`     Local: ${localEval.totalScore}/14 MVP=${localEval.mvpMet} flags=[${localEval.flags.join(',')}]`);
+
+    // 7. Strong model adjudication
+    let adjudication: StrongModelAdjudication;
+    if (options.skipStrongModel || !options.strongModelId) {
+      adjudication = skippedAdjudication(
+        options.skipStrongModel
+          ? 'Strong model skipped by --skip-strong-model flag'
+          : 'No strong model configured'
+      );
+    } else {
+      const decision = needsAdjudication(ep, localEval);
+      if (decision.shouldAdjudicate) {
+        log(`     Adjudicating (${decision.priority}: ${decision.reason})...`);
+        adjudication = await adjudicate(ep, localEval, { modelId: options.strongModelId, log: (msg: string) => log(`     ${msg}`) });
+        log(`     Adjudication: ${adjudication.adjudicationStatus}`);
+      } else {
+        adjudication = skippedAdjudication(decision.reason);
+        log(`     Adjudication skipped: ${decision.reason}`);
+      }
+    }
+
+    const finalLabel = determineFinalLabel(localEval, adjudication);
+    evaluations.push({ episode: ep, localEvaluation: localEval, strongModelAdjudication: adjudication, finalLabel });
+  }
+
+  // 8. Build and write report
+  log('\n📝 Generating report...');
+  const summary = computeSummary(evaluations);
+  const report: QualityScorecardReport = {
+    generatedAt: new Date().toISOString(),
+    dataSource: {
+      painEventCount: extractStats.total,
+      evolutionTaskCount: 0,
+      principleEventCount: 0,
+      gateBlockCount: 0,
+      dateRange: extractStats.dateRange,
+    },
+    localEvaluatorConfig: {
+      model: options.localModelId,
+      baseUrl: options.localModelBaseUrl.replace(/\/v\d+$/, '/...'),
+      apiKeyStatus: 'not-required',
+    },
+    strongModelConfig: {
+      model: options.strongModelId,
+      status: options.skipStrongModel || !options.strongModelId ? 'skipped' : 'configured',
+    },
+    evaluations,
+    summary,
+    knownLimitations: [
+      'Local model scores are advisory only — not final quality conclusions.',
+      'Without strong-model adjudication, samples are marked local-pass/local-fail/needs-review.',
+      'Deduplication is based on reason text similarity — may miss distinct episodes.',
+      'Local model output is non-deterministic despite temperature=0.1.',
+    ],
+  };
+
+  let content: string;
+  switch (options.format) {
+    case 'html': content = generateHtmlReport(report); break;
+    case 'json': content = generateJsonReport(report); break;
+    case 'markdown':
+    default: content = generateMarkdownReport(report); break;
+  }
+
+  writeFileSync(options.output, content, 'utf-8');
+
+  log(`\n✅ Report written to: ${options.output}`);
+  log(`   Format: ${options.format}`);
+  log(`   Episodes: ${summary.totalEpisodes}`);
+  log(`   Local Pass: ${summary.localPassCount} | Local Fail: ${summary.localFailCount}`);
+  log(`   Strong Model Reviewed: ${summary.strongModelReviewedCount}`);
+  log(`   Final Pass: ${summary.finalPassCount} | Final Fail: ${summary.finalFailCount} | Needs Review: ${summary.needsReviewCount}`);
+
+  // JSON mode: output exactly one JSON object to stdout
+  if (isJson) {
+    writeJsonOutput({ ok: true, report });
+  }
+}

package/src/index.ts

@@ -927,4 +927,28 @@ consoleCmd.action(async (opts) => {
   });
 });
 
+// ─── Quality Scorecard (PRI-361) ──────────────────────────────────
+
+const qualityCmd = program
+  .command('quality')
+  .description('Quality scoring and evaluation');
+
+qualityCmd
+  .command('scorecard')
+  .description('Generate quality scorecard report for PD pain→diagnosis→principle chain')
+  .option('-w, --workspace <path>', 'Workspace directory')
+  .option('--local-model <id>', 'LM Studio model ID', 'qwen3.6-27b-mtp')
+  .option('--local-url <url>', 'LM Studio base URL', 'http://localhost:12341/v1')
+  .option('--strong-model <id>', 'Strong model for adjudication (provider/model)')
+  .option('--skip-strong-model', 'Skip strong model adjudication', false)
+  .option('--min-score <n>', 'Minimum pain score to evaluate', '50')
+  .option('--limit <n>', 'Max episodes to evaluate (0=all)', '0')
+  .option('--format <fmt>', 'Output format: json, markdown, html', 'markdown')
+  .option('--output <path>', 'Output file path')
+  .option('--json', 'Output as JSON', false)
+  .action(async (opts) => {
+    const { handleQualityScorecard } = await import('./commands/quality-scorecard.js');
+    await handleQualityScorecard(opts);
+  });
+
 program.parse();