Auto backup before evolver run - 2026年 03月 03日星期二 23:12:56 CST

2026-03-03 23:12:56 +08:00
parent 429e5a797d
commit f11d322c3f
99 changed files with 17051 additions and 138 deletions
@@ -0,0 +1,771 @@
+const fs = require('fs');
+const path = require('path');
+const { getMemoryDir } = require('./paths');
+const { normalizePersonalityState, isValidPersonalityState, personalityKey } = require('./personality');
+const { isValidMutation, normalizeMutation } = require('./mutation');
+
+function ensureDir(dir) {
+  try {
+    if (!fs.existsSync(dir)) fs.mkdirSync(dir, { recursive: true });
+  } catch (e) {}
+}
+
+function stableHash(input) {
+  const s = String(input || '');
+  let h = 2166136261;
+  for (let i = 0; i < s.length; i++) {
+    h ^= s.charCodeAt(i);
+    h = Math.imul(h, 16777619);
+  }
+  return (h >>> 0).toString(16).padStart(8, '0');
+}
+
+function nowIso() {
+  return new Date().toISOString();
+}
+
+function normalizeErrorSignature(text) {
+  const s = String(text || '').trim();
+  if (!s) return null;
+  return (
+    s
+      .toLowerCase()
+      // normalize Windows paths
+      .replace(/[a-z]:\\[^ \n\r\t]+/gi, '<path>')
+      // normalize Unix paths
+      .replace(/\/[^ \n\r\t]+/g, '<path>')
+      // normalize hex and numbers
+      .replace(/\b0x[0-9a-f]+\b/gi, '<hex>')
+      .replace(/\b\d+\b/g, '<n>')
+      // normalize whitespace
+      .replace(/\s+/g, ' ')
+      .slice(0, 220)
+  );
+}
+
+function normalizeSignalsForMatching(signals) {
+  const list = Array.isArray(signals) ? signals : [];
+  const out = [];
+  for (const s of list) {
+    const str = String(s || '').trim();
+    if (!str) continue;
+    if (str.startsWith('errsig:')) {
+      const norm = normalizeErrorSignature(str.slice('errsig:'.length));
+      if (norm) out.push(`errsig_norm:${stableHash(norm)}`);
+      continue;
+    }
+    out.push(str);
+  }
+  return out;
+}
+
+function computeSignalKey(signals) {
+  // Key must be stable across runs; normalize noisy signatures (paths, numbers).
+  const list = normalizeSignalsForMatching(signals);
+  const uniq = Array.from(new Set(list.filter(Boolean))).sort();
+  return uniq.join('|') || '(none)';
+}
+
+function extractErrorSignatureFromSignals(signals) {
+  // Convention: signals can include "errsig:<raw>" emitted by signals extractor.
+  const list = Array.isArray(signals) ? signals : [];
+  for (const s of list) {
+    const str = String(s || '');
+    if (str.startsWith('errsig:')) return normalizeErrorSignature(str.slice('errsig:'.length));
+  }
+  return null;
+}
+
+function memoryGraphPath() {
+  const { getEvolutionDir } = require('./paths');
+  const evoDir = getEvolutionDir();
+  return process.env.MEMORY_GRAPH_PATH || path.join(evoDir, 'memory_graph.jsonl');
+}
+
+function memoryGraphStatePath() {
+  const { getEvolutionDir } = require('./paths');
+  return path.join(getEvolutionDir(), 'memory_graph_state.json');
+}
+
+function appendJsonl(filePath, obj) {
+  const dir = path.dirname(filePath);
+  ensureDir(dir);
+  fs.appendFileSync(filePath, JSON.stringify(obj) + '\n', 'utf8');
+}
+
+function readJsonIfExists(filePath, fallback) {
+  try {
+    if (!fs.existsSync(filePath)) return fallback;
+    const raw = fs.readFileSync(filePath, 'utf8');
+    if (!raw.trim()) return fallback;
+    return JSON.parse(raw);
+  } catch (e) {
+    return fallback;
+  }
+}
+
+function writeJsonAtomic(filePath, obj) {
+  const dir = path.dirname(filePath);
+  ensureDir(dir);
+  const tmp = `${filePath}.tmp`;
+  fs.writeFileSync(tmp, JSON.stringify(obj, null, 2) + '\n', 'utf8');
+  fs.renameSync(tmp, filePath);
+}
+
+function tryReadMemoryGraphEvents(limitLines = 2000) {
+  try {
+    const p = memoryGraphPath();
+    if (!fs.existsSync(p)) return [];
+    const raw = fs.readFileSync(p, 'utf8');
+    const lines = raw
+      .split('\n')
+      .map(l => l.trim())
+      .filter(Boolean);
+    const recent = lines.slice(Math.max(0, lines.length - limitLines));
+    return recent
+      .map(l => {
+        try {
+          return JSON.parse(l);
+        } catch (e) {
+          return null;
+        }
+      })
+      .filter(Boolean);
+  } catch (e) {
+    return [];
+  }
+}
+
+function jaccard(aList, bList) {
+  const aNorm = normalizeSignalsForMatching(aList);
+  const bNorm = normalizeSignalsForMatching(bList);
+  const a = new Set((Array.isArray(aNorm) ? aNorm : []).map(String));
+  const b = new Set((Array.isArray(bNorm) ? bNorm : []).map(String));
+  if (a.size === 0 && b.size === 0) return 1;
+  if (a.size === 0 || b.size === 0) return 0;
+  let inter = 0;
+  for (const x of a) if (b.has(x)) inter++;
+  const union = a.size + b.size - inter;
+  return union === 0 ? 0 : inter / union;
+}
+
+function decayWeight(updatedAtIso, halfLifeDays) {
+  const hl = Number(halfLifeDays);
+  if (!Number.isFinite(hl) || hl <= 0) return 1;
+  const t = Date.parse(updatedAtIso);
+  if (!Number.isFinite(t)) return 1;
+  const ageDays = (Date.now() - t) / (1000 * 60 * 60 * 24);
+  if (!Number.isFinite(ageDays) || ageDays <= 0) return 1;
+  // Exponential half-life decay: weight = 0.5^(age/hl)
+  return Math.pow(0.5, ageDays / hl);
+}
+
+function aggregateEdges(events) {
+  // Aggregate by (signal_key, gene_id) from outcome events.
+  // Laplace smoothing to avoid 0/1 extremes.
+  const map = new Map();
+  for (const ev of events) {
+    if (!ev || ev.type !== 'MemoryGraphEvent') continue;
+    if (ev.kind !== 'outcome') continue;
+    const signalKey = ev.signal && ev.signal.key ? String(ev.signal.key) : '(none)';
+    const geneId = ev.gene && ev.gene.id ? String(ev.gene.id) : null;
+    if (!geneId) continue;
+
+    const k = `${signalKey}::${geneId}`;
+    const cur = map.get(k) || { signalKey, geneId, success: 0, fail: 0, last_ts: null, last_score: null };
+    const status = ev.outcome && ev.outcome.status ? String(ev.outcome.status) : 'unknown';
+    if (status === 'success') cur.success += 1;
+    else if (status === 'failed') cur.fail += 1;
+
+    const ts = ev.ts || ev.created_at || ev.at;
+    if (ts && (!cur.last_ts || Date.parse(ts) > Date.parse(cur.last_ts))) {
+      cur.last_ts = ts;
+      cur.last_score =
+        ev.outcome && Number.isFinite(Number(ev.outcome.score)) ? Number(ev.outcome.score) : cur.last_score;
+    }
+    map.set(k, cur);
+  }
+  return map;
+}
+
+function aggregateGeneOutcomes(events) {
+  // Aggregate by gene_id from outcome events (gene -> outcome success probability).
+  const map = new Map();
+  for (const ev of events) {
+    if (!ev || ev.type !== 'MemoryGraphEvent') continue;
+    if (ev.kind !== 'outcome') continue;
+    const geneId = ev.gene && ev.gene.id ? String(ev.gene.id) : null;
+    if (!geneId) continue;
+    const cur = map.get(geneId) || { geneId, success: 0, fail: 0, last_ts: null, last_score: null };
+    const status = ev.outcome && ev.outcome.status ? String(ev.outcome.status) : 'unknown';
+    if (status === 'success') cur.success += 1;
+    else if (status === 'failed') cur.fail += 1;
+    const ts = ev.ts || ev.created_at || ev.at;
+    if (ts && (!cur.last_ts || Date.parse(ts) > Date.parse(cur.last_ts))) {
+      cur.last_ts = ts;
+      cur.last_score =
+        ev.outcome && Number.isFinite(Number(ev.outcome.score)) ? Number(ev.outcome.score) : cur.last_score;
+    }
+    map.set(geneId, cur);
+  }
+  return map;
+}
+
+function edgeExpectedSuccess(edge, opts) {
+  const e = edge || { success: 0, fail: 0, last_ts: null };
+  const succ = Number(e.success) || 0;
+  const fail = Number(e.fail) || 0;
+  const total = succ + fail;
+  const p = (succ + 1) / (total + 2); // Laplace smoothing
+  const halfLifeDays = opts && Number.isFinite(Number(opts.half_life_days)) ? Number(opts.half_life_days) : 30;
+  const w = decayWeight(e.last_ts || '', halfLifeDays);
+  return { p, w, total, value: p * w };
+}
+
+function getMemoryAdvice({ signals, genes, driftEnabled }) {
+  const events = tryReadMemoryGraphEvents(2000);
+  const edges = aggregateEdges(events);
+  const geneOutcomes = aggregateGeneOutcomes(events);
+  const curSignals = Array.isArray(signals) ? signals : [];
+  const curKey = computeSignalKey(curSignals);
+
+  const bannedGeneIds = new Set();
+  const scoredGeneIds = [];
+
+  // Similarity: consider exact key first, then any key with overlap.
+  const seenKeys = new Set();
+  const candidateKeys = [];
+  candidateKeys.push({ key: curKey, sim: 1 });
+  seenKeys.add(curKey);
+
+  for (const ev of events) {
+    if (!ev || ev.type !== 'MemoryGraphEvent') continue;
+    const k = ev.signal && ev.signal.key ? String(ev.signal.key) : '(none)';
+    if (seenKeys.has(k)) continue;
+    const sigs = ev.signal && Array.isArray(ev.signal.signals) ? ev.signal.signals : [];
+    const sim = jaccard(curSignals, sigs);
+    if (sim >= 0.34) {
+      candidateKeys.push({ key: k, sim });
+      seenKeys.add(k);
+    }
+  }
+
+  const byGene = new Map();
+  for (const ck of candidateKeys) {
+    for (const g of Array.isArray(genes) ? genes : []) {
+      if (!g || g.type !== 'Gene' || !g.id) continue;
+      const k = `${ck.key}::${g.id}`;
+      const edge = edges.get(k);
+      const cur = byGene.get(g.id) || { geneId: g.id, best: 0, attempts: 0, prior: 0, prior_attempts: 0 };
+
+      // Signal->Gene edge score (if available)
+      if (edge) {
+        const ex = edgeExpectedSuccess(edge, { half_life_days: 30 });
+        const weighted = ex.value * ck.sim;
+        if (weighted > cur.best) cur.best = weighted;
+        cur.attempts = Math.max(cur.attempts, ex.total);
+      }
+
+      // Gene->Outcome prior (independent of signal): stabilizer when signal edges are sparse.
+      const gEdge = geneOutcomes.get(String(g.id));
+      if (gEdge) {
+        const gx = edgeExpectedSuccess(gEdge, { half_life_days: 45 });
+        cur.prior = Math.max(cur.prior, gx.value);
+        cur.prior_attempts = Math.max(cur.prior_attempts, gx.total);
+      }
+
+      byGene.set(g.id, cur);
+    }
+  }
+
+  for (const [geneId, info] of byGene.entries()) {
+    const combined = info.best > 0 ? info.best + info.prior * 0.12 : info.prior * 0.4;
+    scoredGeneIds.push({ geneId, score: combined, attempts: info.attempts, prior: info.prior });
+    // Low-efficiency path suppression (unless drift is explicit).
+    if (!driftEnabled && info.attempts >= 2 && info.best < 0.18) {
+      bannedGeneIds.add(geneId);
+    }
+    // Also suppress genes with consistently poor global outcomes when signal edges are sparse.
+    if (!driftEnabled && info.attempts < 2 && info.prior_attempts >= 3 && info.prior < 0.12) {
+      bannedGeneIds.add(geneId);
+    }
+  }
+
+  scoredGeneIds.sort((a, b) => b.score - a.score);
+  const preferredGeneId = scoredGeneIds.length ? scoredGeneIds[0].geneId : null;
+
+  const explanation = [];
+  if (preferredGeneId) explanation.push(`memory_prefer:${preferredGeneId}`);
+  if (bannedGeneIds.size) explanation.push(`memory_ban:${Array.from(bannedGeneIds).slice(0, 6).join(',')}`);
+  if (preferredGeneId) {
+    const top = scoredGeneIds.find(x => x && x.geneId === preferredGeneId);
+    if (top && Number.isFinite(Number(top.prior)) && top.prior > 0) explanation.push(`gene_prior:${top.prior.toFixed(3)}`);
+  }
+  if (driftEnabled) explanation.push('random_drift:enabled');
+
+  return {
+    currentSignalKey: curKey,
+    preferredGeneId,
+    bannedGeneIds,
+    explanation,
+  };
+}
+
+function recordSignalSnapshot({ signals, observations }) {
+  const signalKey = computeSignalKey(signals);
+  const ts = nowIso();
+  const errsig = extractErrorSignatureFromSignals(signals);
+  const ev = {
+    type: 'MemoryGraphEvent',
+    kind: 'signal',
+    id: `mge_${Date.now()}_${stableHash(`${signalKey}|signal|${ts}`)}`,
+    ts,
+    signal: {
+      key: signalKey,
+      signals: Array.isArray(signals) ? signals : [],
+      error_signature: errsig || null,
+    },
+    observed: observations && typeof observations === 'object' ? observations : null,
+  };
+  appendJsonl(memoryGraphPath(), ev);
+  return ev;
+}
+
+function buildHypothesisText({ signalKey, signals, geneId, geneCategory, driftEnabled }) {
+  const sigCount = Array.isArray(signals) ? signals.length : 0;
+  const drift = driftEnabled ? 'drift' : 'directed';
+  const g = geneId ? `${geneId}${geneCategory ? `(${geneCategory})` : ''}` : '(none)';
+  return `Given signal_key=${signalKey} with ${sigCount} signals, selecting gene=${g} under mode=${drift} is expected to reduce repeated errors and improve stability.`;
+}
+
+function recordHypothesis({
+  signals,
+  mutation,
+  personality_state,
+  selectedGene,
+  selector,
+  driftEnabled,
+  selectedBy,
+  capsulesUsed,
+  observations,
+}) {
+  const signalKey = computeSignalKey(signals);
+  const geneId = selectedGene && selectedGene.id ? String(selectedGene.id) : null;
+  const geneCategory = selectedGene && selectedGene.category ? String(selectedGene.category) : null;
+  const ts = nowIso();
+  const errsig = extractErrorSignatureFromSignals(signals);
+  const hypothesisId = `hyp_${Date.now()}_${stableHash(`${signalKey}|${geneId || 'none'}|${ts}`)}`;
+  const personalityState = personality_state || null;
+  const mutNorm = mutation && isValidMutation(mutation) ? normalizeMutation(mutation) : null;
+  const psNorm = personalityState && isValidPersonalityState(personalityState) ? normalizePersonalityState(personalityState) : null;
+  const ev = {
+    type: 'MemoryGraphEvent',
+    kind: 'hypothesis',
+    id: `mge_${Date.now()}_${stableHash(`${hypothesisId}|${ts}`)}`,
+    ts,
+    signal: { key: signalKey, signals: Array.isArray(signals) ? signals : [], error_signature: errsig || null },
+    hypothesis: {
+      id: hypothesisId,
+      text: buildHypothesisText({ signalKey, signals, geneId, geneCategory, driftEnabled }),
+      predicted_outcome: { status: null, score: null },
+    },
+    mutation: mutNorm
+      ? {
+          id: mutNorm.id,
+          category: mutNorm.category,
+          trigger_signals: mutNorm.trigger_signals,
+          target: mutNorm.target,
+          expected_effect: mutNorm.expected_effect,
+          risk_level: mutNorm.risk_level,
+        }
+      : null,
+    personality: psNorm
+      ? {
+          key: personalityKey(psNorm),
+          state: psNorm,
+        }
+      : null,
+    gene: { id: geneId, category: geneCategory },
+    action: {
+      drift: !!driftEnabled,
+      selected_by: selectedBy || 'selector',
+      selector: selector || null,
+    },
+    capsules: {
+      used: Array.isArray(capsulesUsed) ? capsulesUsed.map(String).filter(Boolean) : [],
+    },
+    observed: observations && typeof observations === 'object' ? observations : null,
+  };
+  appendJsonl(memoryGraphPath(), ev);
+  return { hypothesisId, signalKey };
+}
+
+function hasErrorSignal(signals) {
+  const list = Array.isArray(signals) ? signals : [];
+  return list.includes('log_error');
+}
+
+function recordAttempt({
+  signals,
+  mutation,
+  personality_state,
+  selectedGene,
+  selector,
+  driftEnabled,
+  selectedBy,
+  hypothesisId,
+  capsulesUsed,
+  observations,
+}) {
+  const signalKey = computeSignalKey(signals);
+  const geneId = selectedGene && selectedGene.id ? String(selectedGene.id) : null;
+  const geneCategory = selectedGene && selectedGene.category ? String(selectedGene.category) : null;
+  const ts = nowIso();
+  const errsig = extractErrorSignatureFromSignals(signals);
+  const actionId = `act_${Date.now()}_${stableHash(`${signalKey}|${geneId || 'none'}|${ts}`)}`;
+  const personalityState = personality_state || null;
+  const mutNorm = mutation && isValidMutation(mutation) ? normalizeMutation(mutation) : null;
+  const psNorm = personalityState && isValidPersonalityState(personalityState) ? normalizePersonalityState(personalityState) : null;
+  const ev = {
+    type: 'MemoryGraphEvent',
+    kind: 'attempt',
+    id: `mge_${Date.now()}_${stableHash(actionId)}`,
+    ts,
+    signal: { key: signalKey, signals: Array.isArray(signals) ? signals : [], error_signature: errsig || null },
+    mutation: mutNorm
+      ? {
+          id: mutNorm.id,
+          category: mutNorm.category,
+          trigger_signals: mutNorm.trigger_signals,
+          target: mutNorm.target,
+          expected_effect: mutNorm.expected_effect,
+          risk_level: mutNorm.risk_level,
+        }
+      : null,
+    personality: psNorm
+      ? {
+          key: personalityKey(psNorm),
+          state: psNorm,
+        }
+      : null,
+    gene: { id: geneId, category: geneCategory },
+    hypothesis: hypothesisId ? { id: String(hypothesisId) } : null,
+    action: {
+      id: actionId,
+      drift: !!driftEnabled,
+      selected_by: selectedBy || 'selector',
+      selector: selector || null,
+    },
+    capsules: {
+      used: Array.isArray(capsulesUsed) ? capsulesUsed.map(String).filter(Boolean) : [],
+    },
+    observed: observations && typeof observations === 'object' ? observations : null,
+  };
+
+  appendJsonl(memoryGraphPath(), ev);
+
+  // State is mutable; graph is append-only.
+  const statePath = memoryGraphStatePath();
+  const state = readJsonIfExists(statePath, { last_action: null });
+  state.last_action = {
+    action_id: actionId,
+    signal_key: signalKey,
+    signals: Array.isArray(signals) ? signals : [],
+    mutation_id: mutNorm ? mutNorm.id : null,
+    mutation_category: mutNorm ? mutNorm.category : null,
+    mutation_risk_level: mutNorm ? mutNorm.risk_level : null,
+    personality_key: psNorm ? personalityKey(psNorm) : null,
+    personality_state: psNorm || null,
+    gene_id: geneId,
+    gene_category: geneCategory,
+    hypothesis_id: hypothesisId ? String(hypothesisId) : null,
+    capsules_used: Array.isArray(capsulesUsed) ? capsulesUsed.map(String).filter(Boolean) : [],
+    had_error: hasErrorSignal(signals),
+    created_at: ts,
+    outcome_recorded: false,
+    baseline_observed: observations && typeof observations === 'object' ? observations : null,
+  };
+  writeJsonAtomic(statePath, state);
+
+  return { actionId, signalKey };
+}
+
+function inferOutcomeFromSignals({ prevHadError, currentHasError }) {
+  if (prevHadError && !currentHasError) return { status: 'success', score: 0.85, note: 'error_cleared' };
+  if (prevHadError && currentHasError) return { status: 'failed', score: 0.2, note: 'error_persisted' };
+  if (!prevHadError && currentHasError) return { status: 'failed', score: 0.15, note: 'new_error_appeared' };
+  return { status: 'success', score: 0.6, note: 'stable_no_error' };
+}
+
+function clamp01(x) {
+  const n = Number(x);
+  if (!Number.isFinite(n)) return 0;
+  return Math.max(0, Math.min(1, n));
+}
+
+function tryParseLastEvolutionEventOutcome(evidenceText) {
+  // Scan tail text for an EvolutionEvent JSON line and extract its outcome.
+  const s = String(evidenceText || '');
+  if (!s) return null;
+  const lines = s.split('\n').slice(-400);
+  for (let i = lines.length - 1; i >= 0; i--) {
+    const line = lines[i].trim();
+    if (!line) continue;
+    if (!line.includes('"type"') || !line.includes('EvolutionEvent')) continue;
+    try {
+      const obj = JSON.parse(line);
+      if (!obj || obj.type !== 'EvolutionEvent') continue;
+      const o = obj.outcome && typeof obj.outcome === 'object' ? obj.outcome : null;
+      if (!o) continue;
+      const status = o.status === 'success' || o.status === 'failed' ? o.status : null;
+      const score = Number.isFinite(Number(o.score)) ? clamp01(Number(o.score)) : null;
+      if (!status && score == null) continue;
+      return {
+        status: status || (score != null && score >= 0.5 ? 'success' : 'failed'),
+        score: score != null ? score : status === 'success' ? 0.75 : 0.25,
+        note: 'evolutionevent_observed',
+      };
+    } catch (e) {
+      continue;
+    }
+  }
+  return null;
+}
+
+function inferOutcomeEnhanced({ prevHadError, currentHasError, baselineObserved, currentObserved }) {
+  const evidence =
+    currentObserved &&
+    currentObserved.evidence &&
+    (currentObserved.evidence.recent_session_tail || currentObserved.evidence.today_log_tail)
+      ? currentObserved.evidence
+      : null;
+  const combinedEvidence = evidence
+    ? `${String(evidence.recent_session_tail || '')}\n${String(evidence.today_log_tail || '')}`
+    : '';
+  const observed = tryParseLastEvolutionEventOutcome(combinedEvidence);
+  if (observed) return observed;
+
+  const base = inferOutcomeFromSignals({ prevHadError, currentHasError });
+
+  const prevErrCount =
+    baselineObserved && Number.isFinite(Number(baselineObserved.recent_error_count))
+      ? Number(baselineObserved.recent_error_count)
+      : null;
+  const curErrCount =
+    currentObserved && Number.isFinite(Number(currentObserved.recent_error_count))
+      ? Number(currentObserved.recent_error_count)
+      : null;
+
+  let score = base.score;
+  if (prevErrCount != null && curErrCount != null) {
+    const delta = prevErrCount - curErrCount;
+    score += Math.max(-0.12, Math.min(0.12, delta / 50));
+  }
+
+  const prevScan =
+    baselineObserved && Number.isFinite(Number(baselineObserved.scan_ms)) ? Number(baselineObserved.scan_ms) : null;
+  const curScan =
+    currentObserved && Number.isFinite(Number(currentObserved.scan_ms)) ? Number(currentObserved.scan_ms) : null;
+  if (prevScan != null && curScan != null && prevScan > 0) {
+    const ratio = (prevScan - curScan) / prevScan;
+    score += Math.max(-0.06, Math.min(0.06, ratio));
+  }
+
+  return { status: base.status, score: clamp01(score), note: `${base.note}|heuristic_delta` };
+}
+
+function buildConfidenceEdgeEvent({ signalKey, signals, geneId, geneCategory, outcomeEventId, halfLifeDays }) {
+  const events = tryReadMemoryGraphEvents(2000);
+  const edges = aggregateEdges(events);
+  const k = `${signalKey}::${geneId}`;
+  const edge = edges.get(k) || { success: 0, fail: 0, last_ts: null };
+  const ex = edgeExpectedSuccess(edge, { half_life_days: halfLifeDays });
+  const ts = nowIso();
+  return {
+    type: 'MemoryGraphEvent',
+    kind: 'confidence_edge',
+    id: `mge_${Date.now()}_${stableHash(`${signalKey}|${geneId}|confidence|${ts}`)}`,
+    ts,
+    signal: { key: signalKey, signals: Array.isArray(signals) ? signals : [] },
+    gene: { id: geneId, category: geneCategory || null },
+    edge: { signal_key: signalKey, gene_id: geneId },
+    stats: {
+      success: Number(edge.success) || 0,
+      fail: Number(edge.fail) || 0,
+      attempts: Number(ex.total) || 0,
+      p: ex.p,
+      decay_weight: ex.w,
+      value: ex.value,
+      half_life_days: halfLifeDays,
+      updated_at: ts,
+    },
+    derived_from: { outcome_event_id: outcomeEventId || null },
+  };
+}
+
+function buildGeneOutcomeConfidenceEvent({ geneId, geneCategory, outcomeEventId, halfLifeDays }) {
+  const events = tryReadMemoryGraphEvents(2000);
+  const geneOutcomes = aggregateGeneOutcomes(events);
+  const edge = geneOutcomes.get(String(geneId)) || { success: 0, fail: 0, last_ts: null };
+  const ex = edgeExpectedSuccess(edge, { half_life_days: halfLifeDays });
+  const ts = nowIso();
+  return {
+    type: 'MemoryGraphEvent',
+    kind: 'confidence_gene_outcome',
+    id: `mge_${Date.now()}_${stableHash(`${geneId}|gene_outcome|confidence|${ts}`)}`,
+    ts,
+    gene: { id: String(geneId), category: geneCategory || null },
+    edge: { gene_id: String(geneId) },
+    stats: {
+      success: Number(edge.success) || 0,
+      fail: Number(edge.fail) || 0,
+      attempts: Number(ex.total) || 0,
+      p: ex.p,
+      decay_weight: ex.w,
+      value: ex.value,
+      half_life_days: halfLifeDays,
+      updated_at: ts,
+    },
+    derived_from: { outcome_event_id: outcomeEventId || null },
+  };
+}
+
+function recordOutcomeFromState({ signals, observations }) {
+  const statePath = memoryGraphStatePath();
+  const state = readJsonIfExists(statePath, { last_action: null });
+  const last = state && state.last_action ? state.last_action : null;
+  if (!last || !last.action_id) return null;
+  if (last.outcome_recorded) return null;
+
+  const currentHasError = hasErrorSignal(signals);
+  const inferred = inferOutcomeEnhanced({
+    prevHadError: !!last.had_error,
+    currentHasError,
+    baselineObserved: last.baseline_observed || null,
+    currentObserved: observations || null,
+  });
+  const ts = nowIso();
+  const errsig = extractErrorSignatureFromSignals(signals);
+  const ev = {
+    type: 'MemoryGraphEvent',
+    kind: 'outcome',
+    id: `mge_${Date.now()}_${stableHash(`${last.action_id}|outcome|${ts}`)}`,
+    ts,
+    signal: {
+      key: String(last.signal_key || '(none)'),
+      signals: Array.isArray(last.signals) ? last.signals : [],
+      error_signature: errsig || null,
+    },
+    mutation:
+      last.mutation_id || last.mutation_category || last.mutation_risk_level
+        ? {
+            id: last.mutation_id || null,
+            category: last.mutation_category || null,
+            risk_level: last.mutation_risk_level || null,
+          }
+        : null,
+    personality:
+      last.personality_key || last.personality_state
+        ? {
+            key: last.personality_key || null,
+            state: last.personality_state || null,
+          }
+        : null,
+    gene: { id: last.gene_id || null, category: last.gene_category || null },
+    action: { id: String(last.action_id) },
+    hypothesis: last.hypothesis_id ? { id: String(last.hypothesis_id) } : null,
+    outcome: {
+      status: inferred.status,
+      score: inferred.score,
+      note: inferred.note,
+      observed: { current_signals: Array.isArray(signals) ? signals : [] },
+    },
+    confidence: {
+      // This is an interpretable, decayed success estimate derived from outcomes; aggregation is computed at read-time.
+      half_life_days: 30,
+    },
+    observed: observations && typeof observations === 'object' ? observations : null,
+    baseline: last.baseline_observed || null,
+    capsules: {
+      used: Array.isArray(last.capsules_used) ? last.capsules_used : [],
+    },
+  };
+
+  appendJsonl(memoryGraphPath(), ev);
+
+  // Persist explicit confidence snapshots (append-only) for auditability.
+  try {
+    if (last.gene_id) {
+      const edgeEv = buildConfidenceEdgeEvent({
+        signalKey: String(last.signal_key || '(none)'),
+        signals: Array.isArray(last.signals) ? last.signals : [],
+        geneId: String(last.gene_id),
+        geneCategory: last.gene_category || null,
+        outcomeEventId: ev.id,
+        halfLifeDays: 30,
+      });
+      appendJsonl(memoryGraphPath(), edgeEv);
+
+      const geneEv = buildGeneOutcomeConfidenceEvent({
+        geneId: String(last.gene_id),
+        geneCategory: last.gene_category || null,
+        outcomeEventId: ev.id,
+        halfLifeDays: 45,
+      });
+      appendJsonl(memoryGraphPath(), geneEv);
+    }
+  } catch (e) {}
+
+  last.outcome_recorded = true;
+  last.outcome_recorded_at = ts;
+  state.last_action = last;
+  writeJsonAtomic(statePath, state);
+
+  return ev;
+}
+
+function recordExternalCandidate({ asset, source, signals }) {
+  // Append-only annotation: external assets enter as candidates only.
+  // This does not affect outcome aggregation (which only uses kind === 'outcome').
+  const a = asset && typeof asset === 'object' ? asset : null;
+  const type = a && a.type ? String(a.type) : null;
+  const id = a && a.id ? String(a.id) : null;
+  if (!type || !id) return null;
+
+  const ts = nowIso();
+  const signalKey = computeSignalKey(signals);
+  const ev = {
+    type: 'MemoryGraphEvent',
+    kind: 'external_candidate',
+    id: `mge_${Date.now()}_${stableHash(`${type}|${id}|external|${ts}`)}`,
+    ts,
+    signal: { key: signalKey, signals: Array.isArray(signals) ? signals : [] },
+    external: {
+      source: source || 'external',
+      received_at: ts,
+    },
+    asset: { type, id },
+    candidate: {
+      // Minimal hints for later local triggering/validation.
+      trigger: type === 'Capsule' && Array.isArray(a.trigger) ? a.trigger : [],
+      gene: type === 'Capsule' && a.gene ? String(a.gene) : null,
+      confidence: type === 'Capsule' && Number.isFinite(Number(a.confidence)) ? Number(a.confidence) : null,
+    },
+  };
+
+  appendJsonl(memoryGraphPath(), ev);
+  return ev;
+}
+
+module.exports = {
+  memoryGraphPath,
+  computeSignalKey,
+  tryReadMemoryGraphEvents,
+  getMemoryAdvice,
+  recordSignalSnapshot,
+  recordHypothesis,
+  recordAttempt,
+  recordOutcomeFromState,
+  recordExternalCandidate,
+};
+