@nexus-cortex/cli 4.129.1 → 4.131.1

package/dist/commands/autoresearch/loop.js

@@ -13,16 +13,49 @@
  *      --missing-provider-key-policy semantics).
  *   3. Mutate each arm: --fixer-cmd (any transformer; sees CORTEX_ARM_INDEX/MODEL/STRATEGY
  *      in its environment), else the LLM `cortex autoresearch fix --model <arm model>`.
- *   4. Commit each candidate (no change → that arm is skipped).
+ *   4. Commit each candidate (no change → that arm is skipped). EVAL GUARD (core EvalGuard.ts):
+ *      the task-sets were sealed into a private copy at loop start and every experiment
+ *      grades against it; in-repo task-sets are hidden from the Fixer's worktree; a
+ *      candidate that touches a protected path (the task-sets, --protect) or strays outside
+ *      --editable is refused unbenched (skipped: 'eval-guard').
  *   5. Run `cortex autoresearch experiment` per arm (base vs candidate, FWER-adjusted with
  *      --n-family = the number of arms launched) and read the verdicts.
  *   6. ACCEPT = mergeEligible when a holdout is given (keep + FWER + holdout-verified); with
  *      no holdout, accept = keep-on-train (logged as UNVERIFIED — provide --holdout-set for
  *      the overfitting guard). The WINNER is the gate-accepted arm with the highest effect;
  *      with --require-judge the judge reviews accepted arms in effect order until one is
- *      approved. On accept: advance base to the winner + anchor the loop branch to it.
+ *      approved. CONFIRM (default on, --no-confirm to skip): the selected winner is
+ *      re-benched base vs candidate on FRESH train runs in a separate store (no pooling with
+ *      the screening records) with a new gate seed — choosing the best of N noisy arms
+ *      inflates its effect (winner's curse) and FWER does not remove that bias. The fresh
+ *      run is judged by core decideConfirmation (--confirm-mode, default 'pooled': the
+ *      screening and confirmation p-values Stouffer-combined at the gate's own bar, plus a
+ *      fresh effect > 0 — protects against lucky winners without demanding significance
+ *      twice). One confirmation per round. A winner that passes screening but fails
+ *      confirmation is a NEAR-MISS: its commit is kept under
+ *      refs/autoresearch/near-miss/<branch>/r<round>-a<arm> and a row goes to
+ *      <store>/.cortex/autoresearch-near-misses.jsonl, so it can be re-tested with more
+ *      runs. On accept: advance base to the winner + anchor the loop branch to it.
  *      Losing/rejected arms are dropped.
+ *   GUARDS (B4, core AutoResearchGuards, --guards block|warn|off, default block): a gain on
+ *      the aggregate must not hide a loss in one task family. A candidate is blocked when a
+ *      family's loss is material AND evident (estimate < -margin and its one-sided upper
+ *      bound < 0, alpha split across families) — per STEP (vs its base, --guard-margin 2)
+ *      and as a running TOTAL from the round-0 base (Σ of merged steps' family effects,
+ *      --guard-total-margin 5). The confirmation run's own family effects are checked too
+ *      and are what the running total records.
+ *   APPROACHES + ATTEMPTS (B6, approachPlan.ts): with --width > 1 each arm's prompt names a
+ *      different approach from a rotating queue; the Fixer may declare known flaws in
+ *      .autoresearch-flaws (read + removed before commit; such an arm is benched, never
+ *      merged); every arm that does not merge is kept under refs/autoresearch/attempts/…
+ *      with a row in <store>/.cortex/autoresearch-attempts.jsonl saying why.
+ *   RUNS: every experiment reports a run plan (core planRuns: the smallest effect its runs
+ *      can detect, and the runs per arm needed for --target-effect). The loop warns when
+ *      underpowered; --auto-runs raises the runs for later rounds (capped by --max-runs).
  *   7. Stop on: success metric met, max rounds, max consecutive stale rounds, or a dry backlog.
+ *      The success metric reads TRAIN scores (the confirmation's when it ran), never the
+ *      holdout: holdout results only gate the merge, so repeated rounds cannot steer the
+ *      loop toward the holdout (adaptive overfitting).
  *
  * GIT SAFETY: all work happens in throwaway worktrees off a dedicated loop branch
  * (`autoresearch/loop-*`). The user's branch and working tree are never touched. Accepted
@@ -30,13 +63,14 @@
  * worktree removal; the operator merges the loop branch when satisfied.
  */
 import { execFileSync, spawn, spawnSync } from 'node:child_process';
-import { mkdtempSync } from 'node:fs';
+import { appendFileSync, existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync } from 'node:fs';
 import { tmpdir } from 'node:os';
 import { join, resolve } from 'node:path';
-import { ResearchBacklog } from '@nexus-cortex/core';
+import { ResearchBacklog, checkCandidateChanges, decideConfirmation, describeViolations, protectGlobFor, sealTaskSet, taskIdsIn, stepGuards, totalGuards, recordDrift, describeGuardBlocks, } from '@nexus-cortex/core';
 import { ThemeManager } from '../../themes/ThemeManager.js';
 import { resolveRepoDir, resolveTaskSet } from './repoResolve.js';
 import { planArms, parseMissingKeyPolicy } from './armPlan.js';
+import { DEFAULT_APPROACHES, FLAWS_FILE, approachesForRound, fixerPrompt, parseApproaches } from './approachPlan.js';
 /** Default judge rubric used when --require-judge is set without an explicit
  *  --judge-rubric / --judge-rubric-file. Tuned for the autoresearch loop: the
  *  statistical gate already proved a measurable, holdout-verified gain, so the
@@ -121,6 +155,74 @@ export async function autoResearchLoop(options) {
     const successMetric = options.successMetric
         ? { taskId: options.successMetric.split(':')[0], threshold: Number(options.successMetric.split(':')[1]) }
         : undefined;
+    const splitList = (s) => s?.split(',').map((x) => x.trim()).filter(Boolean);
+    // ---- Eval guard: seal the task-sets, hide in-repo copies, protect paths ------
+    const sealDir = mkdtempSync(join(tmpdir(), 'arseal-'));
+    let trainSet;
+    let holdoutSet;
+    try {
+        trainSet = sealTaskSet(resolveTaskSet(options.taskSet, repo), sealDir, 'train', repo);
+        holdoutSet = options.holdoutSet ? sealTaskSet(resolveTaskSet(options.holdoutSet, repo), sealDir, 'holdout', repo) : undefined;
+    }
+    catch (e) {
+        console.error(theme.colors.error(`Error: cannot read task-set: ${e.message}`));
+        process.exit(1);
+        return;
+    }
+    const sealedSets = [trainSet, holdoutSet].filter((x) => !!x);
+    const hiddenPaths = sealedSets.map((x) => x.inRepo).filter((x) => !!x);
+    const guard = {
+        protect: [...sealedSets.filter((x) => x.inRepo).map((x) => protectGlobFor(x.inRepo, x.isDir)), ...(splitList(options.protect) ?? [])],
+        editable: splitList(options.editable) ?? [],
+    };
+    emit({
+        event: 'eval_seal',
+        train: { source: trainSet.source, sha256: trainSet.sha256, inRepo: trainSet.inRepo ?? null },
+        ...(holdoutSet ? { holdout: { source: holdoutSet.source, sha256: holdoutSet.sha256, inRepo: holdoutSet.inRepo ?? null } } : {}),
+        protect: guard.protect, editable: guard.editable,
+    });
+    if (holdoutSet?.inRepo) {
+        const msg = `eval-guard: the holdout set is inside the repo (${holdoutSet.inRepo}) — it is hidden from the Fixer's worktree and protected, but still readable from git history; keep it outside the repo for real secrecy`;
+        log(msg);
+        emit({ event: 'warning', message: msg });
+    }
+    if (successMetric && !taskIdsIn(trainSet.sealed).has(successMetric.taskId)) {
+        const msg = `--success-metric task '${successMetric.taskId}' is not in the train set — the success stop reads train scores only (never the holdout), so it will never fire`;
+        log(msg);
+        emit({ event: 'warning', message: msg });
+    }
+    // ---- Confirmation + run planning ------------------------------------------
+    const confirmMode = (options.confirmMode ?? 'pooled');
+    if (!['pooled', 'direction', 'strict'].includes(confirmMode)) {
+        console.error(theme.colors.error(`Error: --confirm-mode must be pooled | direction | strict (got '${options.confirmMode}')`));
+        process.exit(1);
+        return;
+    }
+    const confirmLossMargin = options.confirmLossMargin != null ? Number(options.confirmLossMargin) : undefined;
+    const maxRuns = Number(options.maxRuns ?? '12');
+    let runsPerArm = Number(options.runs ?? '3');
+    const nearMissLog = join(store, '.cortex', 'autoresearch-near-misses.jsonl');
+    const guardMode = options.guards ?? 'block';
+    if (!['block', 'warn', 'off'].includes(guardMode)) {
+        console.error(theme.colors.error(`Error: --guards must be block | warn | off (got '${options.guards}')`));
+        process.exit(1);
+        return;
+    }
+    const guardOpts = {
+        margin: options.guardMargin != null ? Number(options.guardMargin) : undefined,
+        totalMargin: options.guardTotalMargin != null ? Number(options.guardTotalMargin) : undefined,
+    };
+    // B6: distinct approaches per arm (rotating queue); losing attempts are kept.
+    const approaches = options.approaches === false ? []
+        : typeof options.approaches === 'string' && options.approaches.trim() ? parseApproaches(options.approaches)
+            : []; // filled below once the width is known
+    const attemptsLog = join(store, '.cortex', 'autoresearch-attempts.jsonl');
+    const attempts = [];
+    /** Per-family drift from the round-0 base: Σ of merged steps' family effects. */
+    const drift = new Map();
+    /** Step + total guard verdicts for one measurement; [] when guards are off. */
+    const guardCheck = (families) => guardMode === 'off' || !families?.length ? [] : [...stepGuards(families, guardOpts), ...totalGuards(drift, families, guardOpts)];
+    const nearMisses = [];
     // ---- Multi-provider arm plan (width 1 = classic single-arm loop) ------------
     let policy;
     try {
@@ -131,7 +233,6 @@ export async function autoResearchLoop(options) {
         process.exit(1);
         return;
     }
-    const splitList = (s) => s?.split(',').map((x) => x.trim()).filter(Boolean);
     const plan = planArms({
         width: Number(options.width ?? '1'),
         baseModel: options.model,
@@ -141,6 +242,11 @@ export async function autoResearchLoop(options) {
         policy,
     });
     const width = plan.arms.length;
+    if (options.approaches !== false && !approaches.length && width > 1)
+        approaches.push(...DEFAULT_APPROACHES);
+    if (approaches.length && width > approaches.length) {
+        log(`approaches: width ${width} > ${approaches.length} approaches — some arms in a round share one`);
+    }
     if (width > 1 || plan.notes.length) {
         emit({ event: 'arm_plan', width, arms: plan.arms, notes: plan.notes });
         for (const n of plan.notes)
@@ -170,21 +276,161 @@ export async function autoResearchLoop(options) {
     let stop = 'max-rounds';
     if (!json)
         console.log(`\n ${theme.colors.highlight('Auto-research loop')}  repo=${repo}  branch=${branch}  base=${baseRef.slice(0, 8)}  width=${width}  max=${maxRounds}\n`);
+    /** `cortex autoresearch experiment` args for base vs one candidate, graded against the
+     *  SEALED task-sets. `holdout:false` = train only (the confirmation run). */
+    const experimentArgs = (a) => {
+        const args = ['autoresearch', 'experiment',
+            '--experiment-tag', a.tag,
+            '--base-dir', baseDir, '--candidate-dir', a.candDir,
+            '--base-ref', git(['rev-parse', '--short', 'HEAD'], baseDir), '--candidate-ref', a.candRef,
+            '--task-set', trainSet.sealed, '--no-build',
+            '--n-family', String(a.nFamily),
+            '--cortex-dir', a.store, '--runs', String(runsPerArm), '--json'];
+        if (options.targetEffect)
+            args.push('--target-effect', options.targetEffect);
+        if (options.power)
+            args.push('--power', options.power);
+        if (a.holdout && holdoutSet)
+            args.push('--holdout-set', holdoutSet.sealed);
+        if (a.seed != null)
+            args.push('--seed', String(a.seed));
+        if (options.model)
+            args.push('--model', options.model);
+        if (options.temperature)
+            args.push('--temperature', options.temperature);
+        if (a.strategy)
+            args.push('--strategy', a.strategy);
+        if (options.runCmd) {
+            args.push('--run-cmd', options.runCmd, '--accept-exit', options.acceptExit ?? '0');
+            if (options.buildCmd)
+                args.push('--build-cmd', options.buildCmd);
+        }
+        return args;
+    };
+    /** Re-bench the selected winner on FRESH train runs in a private store (so the gate sees
+     *  only the new records) with a new seed; it must KEEP again. Never throws. */
+    const confirmWinner = async (w, r) => {
+        const confirmStore = mkdtempSync(join(tmpdir(), 'arconfirm-'));
+        try {
+            const args = experimentArgs({ tag: `${w.tag}-confirm`, candDir: w.candDir, candRef: w.candRef, store: confirmStore, nFamily: 1, holdout: false, strategy: w.strategy, seed: 7919 * r + w.arm });
+            emit({ event: 'confirm_start', round: r, arm: w.arm, candidateRef: w.candRef });
+            const exp = await spawnAsync('node', [self, ...args], { stdio: 'capture' });
+            let res;
+            try {
+                res = JSON.parse(exp.stdout);
+            }
+            catch {
+                res = undefined;
+            }
+            if (!res?.verdict) {
+                const reason = String(res?.error ?? ((exp.stderr || '').trim().split('\n').filter(Boolean).slice(-3).join(' ') || `exit ${exp.status}`)).slice(0, 500);
+                return { confirmed: false, mode: confirmMode, decision: 'no-verdict', reason, runs: runsPerArm };
+            }
+            const fresh = res.verdict;
+            const d = decideConfirmation(w.res.verdict, fresh, { mode: confirmMode, lossMargin: confirmLossMargin });
+            // The fresh run's own family effects must not show an evident family loss either.
+            const families = res.families;
+            const blocks = guardCheck(families).filter((g) => g.blocked);
+            const guardFail = guardMode === 'block' && blocks.length > 0;
+            return {
+                confirmed: d.confirmed && !guardFail, mode: d.mode, decision: fresh.decision, effect: fresh.effect, pValue: fresh.pValue,
+                ciLow: fresh.ciLow, combinedP: d.combinedP,
+                reason: guardFail ? `${d.reason}; GUARD on fresh run: ${describeGuardBlocks(blocks)}` : d.reason,
+                runs: runsPerArm, trainSummary: res.benchSummaries?.candidate?.train, families,
+            };
+        }
+        finally {
+            rmSync(confirmStore, { recursive: true, force: true });
+        }
+    };
+    /** A winner that passed screening (and the judge) but not its confirmation: keep its
+     *  commit under a ref and log a row, so it can be re-tested with more runs. */
+    const recordNearMiss = (w, r, c) => {
+        const ref = `refs/autoresearch/near-miss/${branch}/r${r}-a${w.arm}`;
+        try {
+            git(['update-ref', ref, w.candRef]);
+        }
+        catch { /* best effort; the row still records the sha */ }
+        const sha = (() => { try {
+            return git(['rev-parse', w.candRef]);
+        }
+        catch {
+            return w.candRef;
+        } })();
+        const recommended = Math.max(runsPerArm * 2, w.res?.runPlan?.requiredRunsPerArm ?? 0);
+        const row = {
+            ts: new Date().toISOString(), branch, round: r, arm: w.arm, model: w.model ?? null,
+            ref, candidateSha: sha, baseRef,
+            screening: { decision: w.res?.verdict?.decision, effect: w.res?.verdict?.effect, pValue: w.res?.verdict?.pValue, ciLow: w.res?.verdict?.ciLow, ciHigh: w.res?.verdict?.ciHigh, alphaAdjusted: w.res?.verdict?.alphaAdjusted, mergeEligible: !!w.res?.mergeEligible },
+            confirmation: { mode: c.mode, decision: c.decision, effect: c.effect, pValue: c.pValue, ciLow: c.ciLow, combinedP: c.combinedP, reason: c.reason, runs: c.runs },
+            runPlan: w.res?.runPlan ?? null,
+            retest: `cortex autoresearch experiment --base-ref ${baseRef} --candidate-ref ${ref} --runs ${recommended} …  (check out both refs as worktrees; graded against the same task-set)`,
+        };
+        try {
+            mkdirSync(join(store, '.cortex'), { recursive: true });
+            appendFileSync(nearMissLog, JSON.stringify(row) + '\n');
+        }
+        catch (e) {
+            log(`near-miss log write failed: ${e.message}`);
+        }
+        nearMisses.push({ round: r, arm: w.arm, ref, candidateSha: sha, effect: w.res?.verdict?.effect, confirmEffect: c.effect });
+        emit({ event: 'near_miss', round: r, arm: w.arm, ref, candidateSha: sha, log: nearMissLog });
+        log(`r${r}: near-miss kept → ${ref} (row in ${nearMissLog})`);
+    };
+    /** Keep a non-merged arm's commit under refs/autoresearch/attempts/… and log why. */
+    const recordAttempt = (a, r, status) => {
+        const ref = `refs/autoresearch/attempts/${branch}/r${r}-a${a.arm}`;
+        try {
+            git(['update-ref', ref, a.candRef]);
+        }
+        catch { /* best effort */ }
+        let shortstat = '';
+        try {
+            shortstat = git(['diff', '--shortstat', `${a.candRef}~1`, a.candRef]);
+        }
+        catch { /* */ }
+        const row = {
+            ts: new Date().toISOString(), branch, round: r, arm: a.arm, model: a.model ?? null, approach: a.approach ?? null,
+            status, ref, candidateRef: a.candRef, baseRef, diff: shortstat,
+            ...(a.res?.verdict ? { verdict: { decision: a.res.verdict.decision, effect: a.res.verdict.effect, pValue: a.res.verdict.pValue, mergeEligible: !!a.res.mergeEligible } } : {}),
+            ...(a.flaws ? { flaws: a.flaws } : {}),
+            ...(a.violations ? { violations: a.violations } : {}),
+            ...(a.guardBlocks ? { guardBlocks: a.guardBlocks } : {}),
+            ...(a.reason ? { reason: a.reason } : {}),
+        };
+        try {
+            mkdirSync(join(store, '.cortex'), { recursive: true });
+            appendFileSync(attemptsLog, JSON.stringify(row) + '\n');
+        }
+        catch (e) {
+            log(`attempts log write failed: ${e.message}`);
+        }
+        attempts.push({ round: r, arm: a.arm, status, ref });
+        emit({ event: 'attempt_kept', round: r, arm: a.arm, status, ref });
+    };
     /** Run one arm: worktree → fixer → commit → experiment. Never throws. */
-    const runArm = async (assign, r, goal) => {
-        const out = { arm: assign.arm, model: assign.model, strategy: assign.strategy };
+    const runArm = async (assign, r, goal, approach) => {
+        const strategy = [assign.strategy, approach && `approach:${approach.slug}`].filter(Boolean).join('+') || undefined;
+        const out = { arm: assign.arm, model: assign.model, strategy, ...(approach ? { approach: approach.slug } : {}) };
+        const prompt = fixerPrompt(goal, approach, width);
         const candDir = addWorktree(baseRef);
         out.candDir = candDir;
+        // Hide in-repo task-sets from the Fixer (restored to the base version before commit).
+        const hidden = hiddenPaths.filter((p) => existsSync(join(candDir, p)));
+        for (const p of hidden)
+            rmSync(join(candDir, p), { recursive: true, force: true });
         // 3. Mutate (fixer). Single-arm keeps today's interactive stdio; parallel arms are quiet.
         const fixerStdio = json || width > 1 ? 'ignore' : 'inherit';
         emit({ event: 'fix', round: r, arm: assign.arm, model: assign.model, fixer: options.fixerCmd ? 'command' : 'llm' });
         if (options.fixerCmd) {
-            const cmd = options.fixerCmd.replace(/\{prompt\}/g, shQuote(goal));
+            const cmd = options.fixerCmd.replace(/\{prompt\}/g, shQuote(prompt));
             const env = {
                 ...process.env,
                 CORTEX_ARM_INDEX: String(assign.arm),
+                CORTEX_FLAWS_FILE: join(candDir, FLAWS_FILE),
                 ...(assign.model ? { CORTEX_ARM_MODEL: assign.model } : {}),
-                ...(assign.strategy ? { CORTEX_ARM_STRATEGY: assign.strategy } : {}),
+                ...(strategy ? { CORTEX_ARM_STRATEGY: strategy } : {}),
+                ...(approach ? { CORTEX_ARM_APPROACH: approach.slug } : {}),
             };
             const f = width > 1
                 ? await spawnAsync('sh', ['-c', cmd], { cwd: candDir, env, stdio: fixerStdio })
@@ -196,7 +442,7 @@ export async function autoResearchLoop(options) {
             }
         }
         else {
-            const fixArgs = [self, 'autoresearch', 'fix', '--cwd', candDir, '--prompt', goal, '--json'];
+            const fixArgs = [self, 'autoresearch', 'fix', '--cwd', candDir, '--prompt', prompt, '--json'];
             if (assign.model)
                 fixArgs.push('--model', assign.model);
             const f = await spawnAsync('node', fixArgs, { stdio: fixerStdio });
@@ -206,38 +452,52 @@ export async function autoResearchLoop(options) {
                 return out;
             }
         }
-        // 4. Commit candidate (no change → skip this arm).
+        // 4. Restore the hidden task-sets, then commit the candidate (no change → skip this arm)
+        // unless the eval guard refuses it.
+        // A hidden task-set that exists again after the Fixer ran was WRITTEN by it (it could not
+        // see the original) — tampering, refused below like any protected-path change.
+        const recreated = hiddenPaths
+            .filter((p) => existsSync(join(candDir, p)))
+            .map((path) => ({ path, rule: 'protected' }));
+        for (const p of hiddenPaths)
+            rmSync(join(candDir, p), { recursive: true, force: true });
+        if (hidden.length)
+            git(['checkout', 'HEAD', '--', ...hidden], candDir);
+        // Declared flaws (B6): read and remove the Fixer's flaws file before commit.
+        const flawsPath = join(candDir, FLAWS_FILE);
+        if (existsSync(flawsPath)) {
+            try {
+                const lines = readFileSync(flawsPath, 'utf8').split('\n').map((l) => l.replace(/^[-*\s]+/, '').trim()).filter(Boolean);
+                if (lines.length)
+                    out.flaws = lines.slice(0, 20).map((l) => l.slice(0, 300));
+            }
+            catch { /* unreadable → treated as no declaration */ }
+            rmSync(flawsPath, { force: true });
+        }
         git(['add', '-A'], candDir);
         const dirty = spawnSync('git', ['-C', candDir, 'diff', '--cached', '--quiet']).status !== 0;
-        if (!dirty) {
+        if (!dirty && !recreated.length) {
             out.skipped = 'no-change';
             return out;
         }
-        git(['commit', '-q', '-m', `autoresearch loop r${r}${width > 1 ? `a${assign.arm}` : ''}: ${goal.slice(0, 60)}`], candDir);
+        const changed = git(['diff', '--cached', '--name-only', '--no-renames'], candDir).split('\n').filter(Boolean);
+        // Commit before the guard so a refused candidate is kept as evidence (attempts ref).
+        git(['commit', '-q', '--allow-empty', '-m', `autoresearch loop r${r}${width > 1 ? `a${assign.arm}` : ''}: ${goal.slice(0, 60)}`], candDir);
         out.candRef = git(['rev-parse', '--short', 'HEAD'], candDir);
+        const checked = checkCandidateChanges(changed, guard);
+        const verdict = { ok: checked.ok && !recreated.length, violations: [...recreated, ...checked.violations] };
+        if (!verdict.ok) {
+            out.skipped = 'eval-guard';
+            out.violations = verdict.violations;
+            out.reason = `refused unbenched: ${describeViolations(verdict.violations)}`;
+            emit({ event: 'eval_guard', round: r, arm: assign.arm, model: assign.model, violations: verdict.violations });
+            return out;
+        }
         // 5. Experiment base vs candidate (subprocess → JSON verdict). FWER-adjusted with the
         // real family width so N parallel arms don't inflate the false-keep rate.
         const tag = width > 1 ? `loop-r${r}a${assign.arm}` : `loop-r${r}`;
-        const expArgs = ['autoresearch', 'experiment',
-            '--experiment-tag', tag,
-            '--base-dir', baseDir, '--candidate-dir', candDir,
-            '--base-ref', git(['rev-parse', '--short', 'HEAD'], baseDir), '--candidate-ref', out.candRef,
-            '--task-set', resolveTaskSet(options.taskSet, repo), '--no-build',
-            '--n-family', String(width),
-            '--cortex-dir', store, '--runs', options.runs ?? '3', '--json'];
-        if (options.holdoutSet)
-            expArgs.push('--holdout-set', resolveTaskSet(options.holdoutSet, repo));
-        if (options.model)
-            expArgs.push('--model', options.model);
-        if (options.temperature)
-            expArgs.push('--temperature', options.temperature);
-        if (assign.strategy)
-            expArgs.push('--strategy', assign.strategy);
-        if (options.runCmd) {
-            expArgs.push('--run-cmd', options.runCmd, '--accept-exit', options.acceptExit ?? '0');
-            if (options.buildCmd)
-                expArgs.push('--build-cmd', options.buildCmd);
-        }
+        out.tag = tag;
+        const expArgs = experimentArgs({ tag, candDir, candRef: out.candRef, store, nFamily: width, holdout: true, strategy });
         emit({ event: 'experiment', round: r, arm: assign.arm, model: assign.model, candidateRef: out.candRef });
         const exp = await spawnAsync('node', [self, ...expArgs], { stdio: 'capture' });
         try {
@@ -265,6 +525,20 @@ export async function autoResearchLoop(options) {
         }
         const keep = out.res.verdict?.decision === 'keep';
         out.gateAccepted = options.holdoutSet ? !!out.res.mergeEligible : keep;
+        // B4 family guards (step + running total from the round-0 base).
+        const blocks = guardCheck(out.res.families).filter((g) => g.blocked);
+        if (blocks.length) {
+            out.guardBlocks = blocks;
+            emit({ event: 'guard', round: r, arm: assign.arm, mode: guardMode, blocks });
+            log(`r${r}${width > 1 ? ` a${assign.arm}` : ''}: family guard ${guardMode === 'block' ? 'BLOCKS' : 'warns'}: ${describeGuardBlocks(blocks)}`);
+            if (guardMode === 'block')
+                out.gateAccepted = false;
+        }
+        if (out.flaws?.length) {
+            emit({ event: 'flaws', round: r, arm: assign.arm, flaws: out.flaws, gateWouldAccept: out.gateAccepted });
+            log(`r${r}${width > 1 ? ` a${assign.arm}` : ''}: Fixer declared ${out.flaws.length} flaw(s) — held for review, not merged`);
+            out.gateAccepted = false;
+        }
         emit({ event: 'gate', round: r, arm: assign.arm, model: assign.model, decision: out.res.verdict?.decision, effect: out.res.verdict?.effect, mergeEligible: !!out.res.mergeEligible, accepted: out.gateAccepted });
         return out;
     };
@@ -286,7 +560,8 @@ export async function autoResearchLoop(options) {
             }
             // 2-5. Fan out the arms (width 1 = the classic single-candidate round).
             log(`r${r}: fixing${width > 1 ? ` ×${width} arms` : ''} → ${goal.slice(0, 80)}`);
-            const results = await Promise.all(plan.arms.map((a) => runArm(a, r, goal)));
+            const roundApproaches = approachesForRound(approaches, r, plan.arms.length);
+            const results = await Promise.all(plan.arms.map((a, i) => runArm(a, r, goal, roundApproaches[i])));
             for (const a of results.filter((x) => x.skipped)) {
                 log(`r${r}${width > 1 ? ` a${a.arm}` : ''}: ${a.skipped}${a.reason ? ` (${a.reason})` : ''} → skip`);
             }
@@ -308,6 +583,13 @@ export async function autoResearchLoop(options) {
                     emit({ event: 'round_done', round: r, merged: false, skipped: 'no-verdict', reason: failedNoVerdict[0].reason, exitCode: failedNoVerdict[0].exitCode ?? null });
                 }
                 else {
+                    const guarded = results.filter((a) => a.skipped === 'eval-guard');
+                    if (guarded.length) {
+                        rounds.push({
+                            round: r, accepted: false, skipped: 'eval-guard', reason: guarded[0].reason,
+                            ...(width > 1 ? { arms: results.map(armHistory) } : { violations: guarded[0].violations }),
+                        });
+                    }
                     emit({ event: 'round_done', round: r, merged: false, skipped: first.skipped });
                 }
                 for (const a of results) {
@@ -328,6 +610,7 @@ export async function autoResearchLoop(options) {
             // Fail-closed: a judge that errors or returns no parseable verdict blocks
             // that candidate (approve defaults to false).
             let winner;
+            const judgeRejected = new Set();
             if (options.requireJudge) {
                 for (const cand of acceptedArms) {
                     const baseShort = git(['rev-parse', '--short', 'HEAD'], baseDir);
@@ -358,15 +641,34 @@ export async function autoResearchLoop(options) {
                         winner = cand;
                         break;
                     }
+                    judgeRejected.add(cand.arm);
                 }
             }
             else {
                 winner = acceptedArms[0];
             }
+            // 5c. Confirmation on fresh data (default on). ONE confirmation per round: the
+            // selected winner only, so the confirmation is a single pre-registered test.
+            let confirm;
+            let notConfirmedArm;
+            if (winner && options.confirm !== false) {
+                log(`r${r}${width > 1 ? ` a${winner.arm}` : ''}: confirming on fresh train runs`);
+                confirm = await confirmWinner(winner, r);
+                const ok = confirm.confirmed;
+                emit({ event: 'confirm', round: r, arm: winner.arm, mode: confirm.mode, decision: confirm.decision, effect: confirm.effect, pValue: confirm.pValue, ...(confirm.combinedP != null ? { combinedP: confirm.combinedP } : {}), reason: confirm.reason, confirmed: ok });
+                log(`r${r}: confirm ${ok ? theme.colors.success('CONFIRMED') : theme.colors.error('NOT CONFIRMED')} (${confirm.reason})`);
+                if (!ok) {
+                    recordNearMiss(winner, r, confirm);
+                    notConfirmedArm = winner.arm;
+                    winner = undefined;
+                }
+            }
             const accept = !!winner;
+            // Success metric = TRAIN score (the confirmation's when it ran), never the holdout.
             const candScore = (() => {
-                const src = winner ?? acceptedArms[0] ?? verdictArms[0];
-                const s = src?.res?.benchSummaries?.candidate?.holdout ?? src?.res?.benchSummaries?.candidate?.train;
+                if (!winner)
+                    return undefined;
+                const s = confirm?.trainSummary ?? winner.res?.benchSummaries?.candidate?.train;
                 const t = s?.tasks?.find((x) => x.taskId === successMetric?.taskId);
                 return t?.meanScore;
             })();
@@ -374,9 +676,42 @@ export async function autoResearchLoop(options) {
             rounds.push({
                 round: r, candRef: lead.candRef, decision: lead.res?.verdict?.decision, effect: lead.res?.verdict?.effect,
                 mergeEligible: lead.res?.mergeEligible, accepted: accept, candScore,
+                ...(confirm ? { confirm: { confirmed: confirm.confirmed, mode: confirm.mode, decision: confirm.decision, effect: confirm.effect, pValue: confirm.pValue, ...(confirm.combinedP != null ? { combinedP: confirm.combinedP } : {}), reason: confirm.reason, runs: confirm.runs } } : {}),
+                ...(lead.res?.runPlan ? { runPlan: lead.res.runPlan } : {}),
+                ...(lead.guardBlocks ? { guardBlocks: lead.guardBlocks } : {}),
                 ...(width > 1 ? { winnerArm: winner?.arm, arms: results.map(armHistory) } : {}),
             });
             log(`r${r}: ${lead.res?.verdict?.decision?.toUpperCase()} effect ${lead.res?.verdict?.effect}  ${accept ? theme.colors.success('ACCEPT') : theme.colors.error('reject')}${options.holdoutSet ? '' : ' (train-only, UNVERIFIED — pass --holdout-set)'}${width > 1 && winner ? ` (arm ${winner.arm}${winner.model ? ` · ${winner.model}` : ''})` : ''}`);
+            // Run planning: warn when the screening runs cannot see --target-effect; with
+            // --auto-runs, raise the runs for LATER rounds (never mid-round) up to --max-runs.
+            const roundPlan = lead.res?.runPlan;
+            if (roundPlan && !roundPlan.adequate) {
+                const msg = `r${r}: underpowered — ${roundPlan.runsPerArm} runs/arm detect ≥ ${roundPlan.detectableEffect} pts at power ${roundPlan.power}; +${roundPlan.targetEffect} needs ${roundPlan.requiredRunsPerArm}/arm`;
+                log(msg);
+                emit({ event: 'power_warning', round: r, runPlan: roundPlan });
+                if (options.autoRuns && runsPerArm < maxRuns) {
+                    const next = Math.min(maxRuns, roundPlan.requiredRunsPerArm);
+                    if (next > runsPerArm) {
+                        log(`r${r}: --auto-runs → ${next} runs/arm from round ${r + 1}${next === maxRuns && roundPlan.requiredRunsPerArm > maxRuns ? ` (capped by --max-runs; ${roundPlan.requiredRunsPerArm} needed)` : ''}`);
+                        emit({ event: 'runs_changed', round: r, from: runsPerArm, to: next, required: roundPlan.requiredRunsPerArm });
+                        runsPerArm = next;
+                    }
+                }
+            }
+            // B6: keep every arm that did not merge (its commit under a ref + a row with why).
+            for (const a of results) {
+                if (!a.candRef || a === winner)
+                    continue;
+                const status = a.skipped === 'eval-guard' ? 'eval-guard'
+                    : a.skipped === 'no-verdict' ? 'no-verdict'
+                        : a.flaws?.length ? 'declared-flaws'
+                            : a.guardBlocks?.length && guardMode === 'block' ? 'guard-blocked'
+                                : a.arm === notConfirmedArm ? 'not-confirmed'
+                                    : judgeRejected.has(a.arm) ? 'judge-rejected'
+                                        : !a.gateAccepted ? `gate-${a.res?.verdict?.decision ?? 'none'}`
+                                            : 'outranked';
+                recordAttempt(a, r, status);
+            }
             // 6. Advance or discard.
             if (accept && winner) {
                 git(['branch', '-f', branch, winner.candRef]); // anchor the accepted commit
@@ -388,6 +723,8 @@ export async function autoResearchLoop(options) {
                 baseDir = winner.candDir;
                 baseRef = winner.candRef;
                 stale = 0;
+                // The drift ledger records the unbiased measurement: the confirmation's when it ran.
+                recordDrift(drift, confirm?.families ?? winner.res?.families ?? []);
                 if (defId) {
                     if (options.holdoutSet)
                         backlog.markVerified(defId, `loop-r${r}`);
@@ -424,10 +761,12 @@ export async function autoResearchLoop(options) {
         catch { /* */ }
         for (const w of worktrees)
             removeWorktree(w);
+        rmSync(sealDir, { recursive: true, force: true });
     }
     emit({ event: 'stop', reason: stop });
     const failedArms = rounds.filter(r => r.failed).map(r => ({ round: r.round, reason: r.reason, exitCode: r.exitCode }));
-    const out = { repo, branch, finalRef: baseRef, rounds: rounds.length, merges: rounds.filter(r => r.accepted).length, failures: failedArms.length, failedArms, stop, ...(width > 1 ? { width } : {}), history: rounds };
+    const out = { repo, branch, finalRef: baseRef, rounds: rounds.length, merges: rounds.filter(r => r.accepted).length, failures: failedArms.length, failedArms, nearMisses, attempts, runsPerArm,
+        drift: Object.fromEntries([...drift].map(([f, d]) => [f, { effect: d.effect, se: Math.round(Math.sqrt(d.variance) * 100) / 100, steps: d.steps }])), stop, ...(width > 1 ? { width } : {}), history: rounds };
     if (json) {
         console.log(JSON.stringify(out, null, 2));
         return;
@@ -435,6 +774,8 @@ export async function autoResearchLoop(options) {
     console.log();
     console.log(` ${theme.colors.highlight('Loop done')}  stop=${stop}  rounds=${rounds.length}  merges=${out.merges}`);
     console.log(theme.colors.muted(` Result on branch ${branch} → ${baseRef.slice(0, 8)}. Merge it when satisfied:  git -C ${repo} merge ${branch}`));
+    if (nearMisses.length)
+        console.log(theme.colors.muted(` ${nearMisses.length} near-miss(es) kept for re-testing: refs/autoresearch/near-miss/${branch}/*  (rows: ${nearMissLog})`));
     console.log();
 }
 /** Compact per-arm history entry for the round record / final JSON. */
@@ -445,6 +786,10 @@ function armHistory(a) {
         ...(a.candRef ? { candRef: a.candRef } : {}),
         ...(a.skipped ? { skipped: a.skipped } : {}),
         ...(a.reason ? { reason: a.reason } : {}),
+        ...(a.violations ? { violations: a.violations } : {}),
+        ...(a.guardBlocks ? { guardBlocks: a.guardBlocks } : {}),
+        ...(a.approach ? { approach: a.approach } : {}),
+        ...(a.flaws ? { flaws: a.flaws } : {}),
         ...(a.res ? { decision: a.res.verdict?.decision, effect: a.res.verdict?.effect, mergeEligible: a.res.mergeEligible } : {}),
         ...(a.gateAccepted != null ? { gateAccepted: a.gateAccepted } : {}),
     };