@hmharness/evolution 0.12.1 → 0.13.3

package/dist/bench.js

@@ -101,6 +101,22 @@ export async function runBench(home, run) {
         }
     }
     const passed = results.filter((r) => r.pass).length;
+    // persist a history record: readiness's "eval-regression-stable" condition
+    // reads evolution/benches/*.json - without these records the RL gate's
+    // stability condition can never be met (day-20 audit gap)
+    try {
+        const dir = join(home, 'evolution', 'benches');
+        await mkdir(dir, { recursive: true });
+        const record = {
+            time: new Date().toISOString(),
+            total: cases.length,
+            passed,
+            passRate: cases.length === 0 ? 1 : passed / cases.length,
+            results: results.map((r) => ({ name: r.name, pass: r.pass, detail: r.detail.slice(0, 200) })),
+        };
+        await writeFile(join(dir, `${record.time.replace(/[:.]/g, '-')}.json`), JSON.stringify(record, null, 2) + '\n', 'utf8');
+    }
+    catch { /* history is best-effort; the run result still returns */ }
     return { results, passRate: cases.length === 0 ? 1 : passed / cases.length };
 }
 /** Starter cases so the evolve gate has a signal from day one. */

package/dist/evolve.js

@@ -178,6 +178,19 @@ export async function runEvolution(opts) {
         }
     }
     report.proposals = proposals;
+    // 4b. Promotion quality floor (day-16 meta-audit): "no regression" alone
+    // can promote a candidate that is still terrible when the BASELINE itself
+    // is weak (baseline 20% -> candidate 25% passes the old gate). An absolute
+    // minimum train pass rate closes that hole. Configurable via
+    // config.json evolution.minPassRate (default 0.6).
+    let minPassRate = 0.6;
+    try {
+        const cfg = JSON.parse(await readFile(join(home, 'config.json'), 'utf8'));
+        if (typeof cfg.evolution?.minPassRate === 'number' && cfg.evolution.minPassRate >= 0 && cfg.evolution.minPassRate <= 1) {
+            minPassRate = cfg.evolution.minPassRate;
+        }
+    }
+    catch { /* default floor */ }
     // 5. A/B gate each proposal on the train set.
     for (const p of proposals.slice(0, maxProposals)) {
         say(`candidate "${p.name}": drafting + candidate bench`);
@@ -236,6 +249,20 @@ export async function runEvolution(opts) {
                 say(`  rejected (${regression ? 'regression' : 'lower pass rate'})`);
                 continue;
             }
+            if (candRate < minPassRate) {
+                await deleteDraft(home, p.name);
+                await recordParetoEntry(home, { name: p.name, parentInsights: insightIds, rejectedReason: `below quality floor (${candRate.toFixed(2)} < ${minPassRate})`, scores: { train: candRate }, metaModel: provider.model, at: new Date().toISOString() });
+                report.outcomes.push({
+                    name: p.name,
+                    action: 'rejected',
+                    reason: `below quality floor: pass rate ${candRate} < min ${minPassRate} (non-regression vs a weak baseline ${baseRate} is not good enough)`,
+                    baseline: { passRate: baseRate, cases: summary(baseResults) },
+                    candidate: { passRate: candRate, cases: summary(candResults) },
+                    lineage: { parentInsights: insightIds, scores: { train: candRate }, metaModel: provider.model, decidedAt: new Date().toISOString() },
+                });
+                say(`  rejected (below quality floor ${minPassRate})`);
+                continue;
+            }
             if (costRegressions.length > 0) {
                 await deleteDraft(home, p.name);
                 await recordParetoEntry(home, { name: p.name, parentInsights: insightIds, rejectedReason: `cost regression on ${costRegressions.join(', ')}`, scores: { train: candRate }, metaModel: provider.model, at: new Date().toISOString() });
@@ -303,7 +330,7 @@ export async function runEvolution(opts) {
             report.outcomes.push({
                 name: p.name,
                 action: 'promoted',
-                reason: `no regression (train ${(candRate * 100).toFixed(0)}% vs ${(baseRate * 100).toFixed(0)}%${holdout.length ? `, holdout ${(holdoutRate * 100).toFixed(0)}%` : ''}) - promoted to CANARY (20% sessions, impact-gated full promotion)${archivedPrevious ? '; previous version archived' : ''}${weakGate ? ' [WEAK GATE: no holdout cases defined]' : ''}`,
+                reason: `no regression (train ${(candRate * 100).toFixed(0)}% vs ${(baseRate * 100).toFixed(0)}%${holdout.length ? `, holdout ${(holdoutRate * 100).toFixed(0)}%` : ''}; floor ≥${(minPassRate * 100).toFixed(0)}%) - promoted to CANARY (20% sessions, impact-gated full promotion)${archivedPrevious ? '; previous version archived' : ''}${weakGate ? ' [WEAK GATE: no holdout cases defined]' : ''}`,
                 baseline: { passRate: baseRate, cases: summary(baseResults) },
                 candidate: { passRate: candRate, cases: summary(candResults) },
                 ...(holdout.length ? { holdout: { baselineRate: holdoutBaseRate, candidateRate: holdoutRate } } : {}),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hmharness/evolution",
-  "version": "0.12.1",
+  "version": "0.13.3",
   "description": "hmharness evolution subsystem: persistent memory, insight capture, skill library, and the bench that gives evolution its fitness signal. First-class kernel citizen, not a plugin.",
   "type": "module",
   "main": "dist/index.js",