@metaharness/weight-eft 0.1.0

package/dist/export.js

@@ -0,0 +1,249 @@
+// SPDX-License-Identifier: MIT
+//
+// export.ts — Archive → training-data exporter.
+//
+// Reads Darwin's trajectory archive (a DarwinTrajectory[]; reconstructable from
+// Firestore darwin_runs + local prediction/trajectory artifacts) and emits two
+// standard training sets:
+//
+//   SFT  = ALL gold-resolved trajectories (cheap-OWN *and* frontier-escalation).
+//          Frontier successes are included for OFF-POLICY-SAFE DISTILLATION —
+//          the cheap model learns to imitate a frontier success on issues it
+//          could not solve itself. SFT (max-likelihood) is off-policy-stable.
+//
+//   DPO  = ON-POLICY cheap-vs-cheap pairs ONLY. chosen = a resolved sample,
+//          rejected = an empty/failed sample by the SAME cheap model on the
+//          SAME instance (BoN-derived). We do NOT emit frontier-chosen-vs-cheap-
+//          rejected as DPO: an off-policy preference pair is unstable (the
+//          reference policy never produced the chosen completion). That signal
+//          goes to SFT instead.
+//
+// THE CONTAMINATION GUARD (ADR-198, the headline correctness property): strict
+// train/eval instance-ID disjointness. Any trajectory whose instance_id is in
+// the caller's evalHoldout is excluded; an overlap throws. Training on eval
+// instances is fake lift.
+//
+// Two further guards: a LONG-CONTEXT filter (drop/truncate over-budget
+// trajectories, never silently) and TOOL-CALL FIDELITY (tool_calls survive into
+// SFT messages, never stringified).
+import { detectRewardHack } from './reward-hack.js';
+const DEFAULT_MAX_TOKENS = 28000;
+/**
+ * Rough token estimate for a message list. We deliberately do NOT pull in a
+ * tokenizer dependency (the package is dependency-free, like darwin-mode). The
+ * ~4-chars-per-token heuristic is conservative and stable; tool_calls
+ * arguments count too (they cost context). This is a budget gate, not a billing
+ * figure — over-estimating is the safe direction.
+ */
+export function estimateTokens(messages) {
+    let chars = 0;
+    for (const m of messages) {
+        chars += (m.content ?? '').length;
+        chars += (m.name ?? '').length;
+        if (m.tool_calls) {
+            for (const tc of m.tool_calls) {
+                chars += tc.function.name.length + tc.function.arguments.length + tc.id.length;
+            }
+        }
+        chars += 8; // per-message role/structural overhead
+    }
+    return Math.ceil(chars / 4);
+}
+/**
+ * Truncate an over-length trajectory by dropping the OLDEST middle tool
+ * round-trips while preserving the load-bearing turns: the leading system+user
+ * (the issue) and the trailing assistant (the final patch). Returns a copy.
+ */
+function truncateTrajectory(messages, maxTokens) {
+    if (messages.length <= 3)
+        return messages;
+    // Keep a head (system + first user/issue) and a tail (final assistant).
+    let headEnd = 1;
+    while (headEnd < messages.length && messages[headEnd].role !== 'user')
+        headEnd++;
+    headEnd = Math.min(headEnd + 1, messages.length); // include the issue user turn
+    const head = messages.slice(0, headEnd);
+    const tail = [messages[messages.length - 1]];
+    const middle = messages.slice(headEnd, messages.length - 1);
+    // Greedily re-add middle turns from the END (most recent context) until budget.
+    const kept = [];
+    for (let i = middle.length - 1; i >= 0; i--) {
+        const candidate = [...head, ...kept.slice(), middle[i], ...tail];
+        if (estimateTokens(candidate) > maxTokens && kept.length > 0)
+            break;
+        kept.unshift(middle[i]);
+    }
+    return [...head, ...kept, ...tail];
+}
+/**
+ * Split a full trajectory into a DPO (prompt, completion) boundary. ReAct
+ * diverges at the first ACTION, so prompt = the shared leading system + user
+ * (issue) messages, and the completion is everything from the first assistant
+ * turn onward.
+ */
+function splitPromptCompletion(messages) {
+    let firstAssistant = messages.findIndex((m) => m.role === 'assistant');
+    if (firstAssistant < 0)
+        firstAssistant = messages.length; // no assistant turn → empty completion
+    return {
+        prompt: messages.slice(0, firstAssistant),
+        completion: messages.slice(firstAssistant),
+    };
+}
+/**
+ * THE CONTAMINATION GUARD. Throws if any trajectory's instance_id appears in
+ * the eval holdout. Call it the first thing the exporter does — fail loud, not
+ * silently filter, when train/eval disjointness is violated.
+ */
+export function assertTrainEvalDisjoint(trajectories, evalHoldout) {
+    const holdout = new Set(evalHoldout);
+    const overlap = new Set();
+    for (const t of trajectories) {
+        if (holdout.has(t.instance_id))
+            overlap.add(t.instance_id);
+    }
+    if (overlap.size > 0) {
+        const sample = [...overlap].slice(0, 10).join(', ');
+        throw new Error(`weight-eft contamination guard: ${overlap.size} training instance_id(s) overlap the eval holdout ` +
+            `(${sample}${overlap.size > 10 ? ', …' : ''}). Training on eval instances is fake lift — refusing to export. ` +
+            `Exclude these instance_ids from the training archive or remove them from evalHoldout.`);
+    }
+}
+/**
+ * Build the SFT and DPO sets from a Darwin trajectory archive.
+ *
+ * @param trajectories the input archive (already-excluded-of-holdout OR raw —
+ *   the exporter excludes holdout members itself, but ASSERTS disjointness on
+ *   what remains so a programming error can't slip eval data through).
+ */
+export function exportTrainingData(trajectories, options) {
+    const maxTokens = options.maxTokens ?? DEFAULT_MAX_TOKENS;
+    const holdout = new Set(options.evalHoldout);
+    const notes = [];
+    const report = {
+        totalTrajectories: trajectories.length,
+        excludedByHoldout: 0,
+        droppedOverLength: 0,
+        truncatedOverLength: 0,
+        droppedRewardHacked: 0,
+        sftRows: 0,
+        dpoRows: 0,
+        sftInstanceIds: [],
+        dpoInstanceIds: [],
+        notes,
+    };
+    // 1) CONTAMINATION GUARD — exclude holdout members, then assert disjointness
+    //    on what's left so an exclusion bug can never leak eval data downstream.
+    const inDomain = [];
+    for (const t of trajectories) {
+        if (holdout.has(t.instance_id)) {
+            report.excludedByHoldout++;
+            continue;
+        }
+        inDomain.push(t);
+    }
+    assertTrainEvalDisjoint(inDomain, holdout);
+    // 2) LONG-CONTEXT FILTER — drop or truncate over-budget trajectories, never
+    //    silently. Applied to a working copy so the original archive is untouched.
+    const sized = [];
+    for (const t of inDomain) {
+        const tokens = estimateTokens(t.messages);
+        if (tokens <= maxTokens) {
+            sized.push(t);
+            continue;
+        }
+        if (options.truncateOverLength) {
+            const truncated = truncateTrajectory(t.messages, maxTokens);
+            report.truncatedOverLength++;
+            notes.push(`truncated ${t.instance_id} (${t.model}) ${tokens}→~${estimateTokens(truncated)} tok (budget ${maxTokens})`);
+            sized.push({ ...t, messages: truncated });
+        }
+        else {
+            report.droppedOverLength++;
+            notes.push(`dropped ${t.instance_id} (${t.model}) ${tokens} tok > budget ${maxTokens}`);
+        }
+    }
+    // 2b) REWARD-HACKING FILTER (Ornith-1.0 borrow). Drop any trajectory whose
+    //     deterministic monitor flags a gold/test read, verification tamper, or
+    //     sandbox escape — an archived "success" that reward-hacked would teach
+    //     the model to reward-hack. The training-data analog of the conformance
+    //     firewall. Default ON. Applied BEFORE SFT/DPO so neither set sees it.
+    const dropHack = options.dropRewardHacked ?? true;
+    const clean = [];
+    for (const t of sized) {
+        if (!dropHack) {
+            clean.push(t);
+            continue;
+        }
+        const findings = detectRewardHack(t);
+        if (findings.length === 0) {
+            clean.push(t);
+            continue;
+        }
+        report.droppedRewardHacked++;
+        const kinds = [...new Set(findings.map((f) => f.kind))].join(',');
+        const detail = findings[0].detail;
+        notes.push(`reward-hack drop ${t.instance_id} (${t.model}) [${kinds}] — e.g. "${detail}" via ${findings[0].tool}`);
+    }
+    // 3) SFT — ALL gold-resolved trajectories (cheap-own AND frontier-escalation).
+    const sft = [];
+    const sftIds = new Set();
+    for (const t of clean) {
+        if (!t.resolved)
+            continue;
+        if (t.messages.length === 0)
+            continue; // a resolved attempt must have a trajectory
+        // Tool-call fidelity: we copy messages through verbatim — tool_calls are
+        // structured objects on the assistant turns, never stringified.
+        sft.push({ messages: t.messages });
+        sftIds.add(t.instance_id);
+    }
+    // 4) DPO — ON-POLICY cheap-vs-cheap pairs ONLY. Group cheap-tier trajectories
+    //    by (model, instance); pair a resolved (chosen) with an empty/failed
+    //    (rejected) sample from the SAME model on the SAME instance.
+    const dpo = [];
+    const dpoIds = new Set();
+    const cheapGroups = new Map();
+    for (const t of clean) {
+        if (t.tier !== 'cheap')
+            continue; // ON-POLICY only — frontier never enters DPO
+        const key = `${t.model}${t.instance_id}`;
+        const arr = cheapGroups.get(key);
+        if (arr)
+            arr.push(t);
+        else
+            cheapGroups.set(key, [t]);
+    }
+    for (const group of cheapGroups.values()) {
+        const chosen = group.filter((t) => t.resolved && t.messages.length > 0);
+        const rejected = group.filter((t) => !t.resolved);
+        if (chosen.length === 0 || rejected.length === 0)
+            continue;
+        // Deterministic pairing: lowest-sample resolved vs lowest-sample failed.
+        const bySample = (a, b) => (a.sample ?? 0) - (b.sample ?? 0);
+        const c = [...chosen].sort(bySample)[0];
+        const r = [...rejected].sort(bySample)[0];
+        const { prompt, completion: chosenCompletion } = splitPromptCompletion(c.messages);
+        const { completion: rejectedCompletion } = splitPromptCompletion(r.messages.length > 0 ? r.messages : prompt);
+        dpo.push({
+            prompt,
+            chosen: chosenCompletion,
+            rejected: rejectedCompletion,
+        });
+        dpoIds.add(c.instance_id);
+    }
+    report.sftRows = sft.length;
+    report.dpoRows = dpo.length;
+    report.sftInstanceIds = [...sftIds].sort();
+    report.dpoInstanceIds = [...dpoIds].sort();
+    return { sft, dpo, report };
+}
+/** Serialize SFT rows to JSONL (one row per line). */
+export function sftToJsonl(rows) {
+    return rows.map((r) => JSON.stringify(r)).join('\n') + (rows.length ? '\n' : '');
+}
+/** Serialize DPO rows to JSONL (one row per line). */
+export function dpoToJsonl(rows) {
+    return rows.map((r) => JSON.stringify(r)).join('\n') + (rows.length ? '\n' : '');
+}
+//# sourceMappingURL=export.js.map

package/dist/export.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"export.js","sourceRoot":"","sources":["../src/export.ts"],"names":[],"mappings":"AAAA,+BAA+B;AAC/B,EAAE;AACF,gDAAgD;AAChD,EAAE;AACF,gFAAgF;AAChF,+EAA+E;AAC/E,0BAA0B;AAC1B,EAAE;AACF,iFAAiF;AACjF,8EAA8E;AAC9E,6EAA6E;AAC7E,8EAA8E;AAC9E,EAAE;AACF,4EAA4E;AAC5E,4EAA4E;AAC5E,iFAAiF;AACjF,2EAA2E;AAC3E,+EAA+E;AAC/E,gCAAgC;AAChC,EAAE;AACF,+EAA+E;AAC/E,8EAA8E;AAC9E,4EAA4E;AAC5E,0BAA0B;AAC1B,EAAE;AACF,uEAAuE;AACvE,gFAAgF;AAChF,oCAAoC;AAWpC,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAC;AAEpD,MAAM,kBAAkB,GAAG,KAAK,CAAC;AAEjC;;;;;;GAMG;AACH,MAAM,UAAU,cAAc,CAAC,QAAuB;IACpD,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,CAAC,IAAI,QAAQ,EAAE,CAAC;QACzB,KAAK,IAAI,CAAC,CAAC,CAAC,OAAO,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC;QAClC,KAAK,IAAI,CAAC,CAAC,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC;QAC/B,IAAI,CAAC,CAAC,UAAU,EAAE,CAAC;YACjB,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,UAAU,EAAE,CAAC;gBAC9B,KAAK,IAAI,EAAE,CAAC,QAAQ,CAAC,IAAI,CAAC,MAAM,GAAG,EAAE,CAAC,QAAQ,CAAC,SAAS,CAAC,MAAM,GAAG,EAAE,CAAC,EAAE,CAAC,MAAM,CAAC;YACjF,CAAC;QACH,CAAC;QACD,KAAK,IAAI,CAAC,CAAC,CAAC,uCAAuC;IACrD,CAAC;IACD,OAAO,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;AAC9B,CAAC;AAED;;;;GAIG;AACH,SAAS,kBAAkB,CAAC,QAAuB,EAAE,SAAiB;IACpE,IAAI,QAAQ,CAAC,MAAM,IAAI,CAAC;QAAE,OAAO,QAAQ,CAAC;IAC1C,wEAAwE;IACxE,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,OAAO,OAAO,GAAG,QAAQ,CAAC,MAAM,IAAI,QAAQ,CAAC,OAAO,CAAC,CAAC,IAAI,KAAK,MAAM;QAAE,OAAO,EAAE,CAAC;IACjF,OAAO,GAAG,IAAI,CAAC,GAAG,CAAC,OAAO,GAAG,CAAC,EAAE,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,8BAA8B;IAChF,MAAM,IAAI,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,OAAO,CAAC,CAAC;IACxC,MAAM,IAAI,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC;IAC7C,MAAM,MAAM,GAAG,QAAQ,CAAC,KAAK,CAAC,OAAO,EAAE,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IAC5D,gFAAgF;IAChF,MAAM,IAAI,GAAkB,EAAE,CAAC;IAC/B,KAAK,IAAI,CAAC,GAAG,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC5C,MAAM,SAAS,GAAG,CAAC,GAAG,IAAI,EAAE,GAAG,IAAI,CAAC,KAAK,EAAE,EAAE,MAAM,CAAC,CAAC,CAAC,EAAE,GAAG,IAAI,CAAC,CAAC;QACjE,IAAI,cAAc,CAAC,SAAS,CAAC,GAAG,SAAS,IAAI,IAAI,CAAC,MAAM,GAAG,CAAC;YAAE,MAAM;QACpE,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;IAC1B,CAAC;IACD,OAAO,CAAC,GAAG,IAAI,EAAE,GAAG,IAAI,EAAE,GAAG,IAAI,CAAC,CAAC;AACrC,CAAC;AAED;;;;;GAKG;AACH,SAAS,qBAAqB,CAAC,QAAuB;IACpD,IAAI,cAAc,GAAG,QAAQ,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,WAAW,CAAC,CAAC;IACvE,IAAI,cAAc,GAAG,CAAC;QAAE,cAAc,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,uCAAuC;IACjG,OAAO;QACL,MAAM,EAAE,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,cAAc,CAAC;QACzC,UAAU,EAAE,QAAQ,CAAC,KAAK,CAAC,cAAc,CAAC;KAC3C,CAAC;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,uBAAuB,CACrC,YAAgC,EAChC,WAA6B;IAE7B,MAAM,OAAO,GAAG,IAAI,GAAG,CAAC,WAAW,CAAC,CAAC;IACrC,MAAM,OAAO,GAAG,IAAI,GAAG,EAAU,CAAC;IAClC,KAAK,MAAM,CAAC,IAAI,YAAY,EAAE,CAAC;QAC7B,IAAI,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,WAAW,CAAC;YAAE,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC;IAC7D,CAAC;IACD,IAAI,OAAO,CAAC,IAAI,GAAG,CAAC,EAAE,CAAC;QACrB,MAAM,MAAM,GAAG,CAAC,GAAG,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACpD,MAAM,IAAI,KAAK,CACb,mCAAmC,OAAO,CAAC,IAAI,oDAAoD;YACjG,IAAI,MAAM,GAAG,OAAO,CAAC,IAAI,GAAG,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,mEAAmE;YAC9G,uFAAuF,CAC1F,CAAC;IACJ,CAAC;AACH,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,kBAAkB,CAChC,YAAgC,EAChC,OAAsB;IAEtB,MAAM,SAAS,GAAG,OAAO,CAAC,SAAS,IAAI,kBAAkB,CAAC;IAC1D,MAAM,OAAO,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,WAAW,CAAC,CAAC;IAC7C,MAAM,KAAK,GAAa,EAAE,CAAC;IAE3B,MAAM,MAAM,GAAiB;QAC3B,iBAAiB,EAAE,YAAY,CAAC,MAAM;QACtC,iBAAiB,EAAE,CAAC;QACpB,iBAAiB,EAAE,CAAC;QACpB,mBAAmB,EAAE,CAAC;QACtB,mBAAmB,EAAE,CAAC;QACtB,OAAO,EAAE,CAAC;QACV,OAAO,EAAE,CAAC;QACV,cAAc,EAAE,EAAE;QAClB,cAAc,EAAE,EAAE;QAClB,KAAK;KACN,CAAC;IAEF,6EAA6E;IAC7E,6EAA6E;IAC7E,MAAM,QAAQ,GAAuB,EAAE,CAAC;IACxC,KAAK,MAAM,CAAC,IAAI,YAAY,EAAE,CAAC;QAC7B,IAAI,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,WAAW,CAAC,EAAE,CAAC;YAC/B,MAAM,CAAC,iBAAiB,EAAE,CAAC;YAC3B,SAAS;QACX,CAAC;QACD,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IACnB,CAAC;IACD,uBAAuB,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC;IAE3C,4EAA4E;IAC5E,+EAA+E;IAC/E,MAAM,KAAK,GAAuB,EAAE,CAAC;IACrC,KAAK,MAAM,CAAC,IAAI,QAAQ,EAAE,CAAC;QACzB,MAAM,MAAM,GAAG,cAAc,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC;QAC1C,IAAI,MAAM,IAAI,SAAS,EAAE,CAAC;YACxB,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;YACd,SAAS;QACX,CAAC;QACD,IAAI,OAAO,CAAC,kBAAkB,EAAE,CAAC;YAC/B,MAAM,SAAS,GAAG,kBAAkB,CAAC,CAAC,CAAC,QAAQ,EAAE,SAAS,CAAC,CAAC;YAC5D,MAAM,CAAC,mBAAmB,EAAE,CAAC;YAC7B,KAAK,CAAC,IAAI,CACR,aAAa,CAAC,CAAC,WAAW,KAAK,CAAC,CAAC,KAAK,KAAK,MAAM,KAAK,cAAc,CAAC,SAAS,CAAC,gBAAgB,SAAS,GAAG,CAC5G,CAAC;YACF,KAAK,CAAC,IAAI,CAAC,EAAE,GAAG,CAAC,EAAE,QAAQ,EAAE,SAAS,EAAE,CAAC,CAAC;QAC5C,CAAC;aAAM,CAAC;YACN,MAAM,CAAC,iBAAiB,EAAE,CAAC;YAC3B,KAAK,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC,WAAW,KAAK,CAAC,CAAC,KAAK,KAAK,MAAM,iBAAiB,SAAS,EAAE,CAAC,CAAC;QAC1F,CAAC;IACH,CAAC;IAED,2EAA2E;IAC3E,4EAA4E;IAC5E,4EAA4E;IAC5E,4EAA4E;IAC5E,2EAA2E;IAC3E,MAAM,QAAQ,GAAG,OAAO,CAAC,gBAAgB,IAAI,IAAI,CAAC;IAClD,MAAM,KAAK,GAAuB,EAAE,CAAC;IACrC,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;QACtB,IAAI,CAAC,QAAQ,EAAE,CAAC;YACd,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;YACd,SAAS;QACX,CAAC;QACD,MAAM,QAAQ,GAAG,gBAAgB,CAAC,CAAC,CAAC,CAAC;QACrC,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAC1B,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;YACd,SAAS;QACX,CAAC;QACD,MAAM,CAAC,mBAAmB,EAAE,CAAC;QAC7B,MAAM,KAAK,GAAG,CAAC,GAAG,IAAI,GAAG,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QAClE,MAAM,MAAM,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC;QAClC,KAAK,CAAC,IAAI,CACR,oBAAoB,CAAC,CAAC,WAAW,KAAK,CAAC,CAAC,KAAK,MAAM,KAAK,aAAa,MAAM,SAAS,QAAQ,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CACvG,CAAC;IACJ,CAAC;IAED,+EAA+E;IAC/E,MAAM,GAAG,GAAa,EAAE,CAAC;IACzB,MAAM,MAAM,GAAG,IAAI,GAAG,EAAU,CAAC;IACjC,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;QACtB,IAAI,CAAC,CAAC,CAAC,QAAQ;YAAE,SAAS;QAC1B,IAAI,CAAC,CAAC,QAAQ,CAAC,MAAM,KAAK,CAAC;YAAE,SAAS,CAAC,4CAA4C;QACnF,yEAAyE;QACzE,gEAAgE;QAChE,GAAG,CAAC,IAAI,CAAC,EAAE,QAAQ,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC;QACnC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC;IAC5B,CAAC;IAED,8EAA8E;IAC9E,yEAAyE;IACzE,iEAAiE;IACjE,MAAM,GAAG,GAAa,EAAE,CAAC;IACzB,MAAM,MAAM,GAAG,IAAI,GAAG,EAAU,CAAC;IACjC,MAAM,WAAW,GAAG,IAAI,GAAG,EAA8B,CAAC;IAC1D,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;QACtB,IAAI,CAAC,CAAC,IAAI,KAAK,OAAO;YAAE,SAAS,CAAC,6CAA6C;QAC/E,MAAM,GAAG,GAAG,GAAG,CAAC,CAAC,KAAK,IAAI,CAAC,CAAC,WAAW,EAAE,CAAC;QAC1C,MAAM,GAAG,GAAG,WAAW,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QACjC,IAAI,GAAG;YAAE,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;;YAChB,WAAW,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;IACjC,CAAC;IACD,KAAK,MAAM,KAAK,IAAI,WAAW,CAAC,MAAM,EAAE,EAAE,CAAC;QACzC,MAAM,MAAM,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,IAAI,CAAC,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;QACxE,MAAM,QAAQ,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC;QAClD,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;YAAE,SAAS;QAC3D,yEAAyE;QACzE,MAAM,QAAQ,GAAG,CAAC,CAAmB,EAAE,CAAmB,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,CAAC;QACjG,MAAM,CAAC,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;QACxC,MAAM,CAAC,GAAG,CAAC,GAAG,QAAQ,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;QAC1C,MAAM,EAAE,MAAM,EAAE,UAAU,EAAE,gBAAgB,EAAE,GAAG,qBAAqB,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC;QACnF,MAAM,EAAE,UAAU,EAAE,kBAAkB,EAAE,GAAG,qBAAqB,CAC9D,CAAC,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,MAAM,CAC5C,CAAC;QACF,GAAG,CAAC,IAAI,CAAC;YACP,MAAM;YACN,MAAM,EAAE,gBAAgB;YACxB,QAAQ,EAAE,kBAAkB;SAC7B,CAAC,CAAC;QACH,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC;IAC5B,CAAC;IAED,MAAM,CAAC,OAAO,GAAG,GAAG,CAAC,MAAM,CAAC;IAC5B,MAAM,CAAC,OAAO,GAAG,GAAG,CAAC,MAAM,CAAC;IAC5B,MAAM,CAAC,cAAc,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,EAAE,CAAC;IAC3C,MAAM,CAAC,cAAc,GAAG,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,EAAE,CAAC;IAE3C,OAAO,EAAE,GAAG,EAAE,GAAG,EAAE,MAAM,EAAE,CAAC;AAC9B,CAAC;AAED,sDAAsD;AACtD,MAAM,UAAU,UAAU,CAAC,IAAc;IACvC,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;AACnF,CAAC;AAED,sDAAsD;AACtD,MAAM,UAAU,UAAU,CAAC,IAAc;IACvC,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;AACnF,CAAC"}

package/dist/genome.d.ts

@@ -0,0 +1,38 @@
+/**
+ * The value of the `weightAdapter` gene. `null` (or absent) === BASE: the cheap
+ * tier runs with no LoRA adapter — the pre-gene default, byte-identical key.
+ * A string is an adapter id (e.g. "glm5.2-sft", "glm5.2-sft-dpo").
+ */
+export type WeightAdapterGene = string | null;
+/** The canonical "no adapter" sentinel. Absent gene === BASE === this. */
+export declare const BASE_ADAPTER: WeightAdapterGene;
+/**
+ * The adapter variants Darwin evolution may choose among. BASE is always first
+ * (the control). The tuned variants reference adapters the runner produced:
+ * SFT-only and SFT+DPO are the two recipes; a ratio variant blends adapter
+ * strength. ALL are inert until an actual adapter is trained (GPU job) — the
+ * gene only NAMES an adapter; it does not create one.
+ */
+export declare const WEIGHT_ADAPTERS: WeightAdapterGene[];
+/**
+ * Normalize a raw gene value. Absent / '' / 'base' / 'none' all coerce to BASE
+ * so an unset gene is indistinguishable from an explicit base choice — that's
+ * what keeps pre-gene genomes byte-identical.
+ */
+export declare function normalizeWeightAdapter(v: unknown): WeightAdapterGene;
+/**
+ * Stable key fragment for a genome's adapter choice. Empty string for BASE so
+ * the gene contributes NOTHING to a genome key unless an adapter is selected —
+ * the backward-compatibility invariant (a base-adapter genome keys identically
+ * to a pre-gene genome).
+ */
+export declare function weightAdapterSuffix(gene: WeightAdapterGene): string;
+/**
+ * Map an adapter gene to the CLI flag the cheap-tier solver forwards. BASE
+ * emits NO flag (runs the stock model). A tuned adapter emits `--lora-adapter
+ * <id>` (the solver loads the adapter onto the cheap base before solving).
+ */
+export declare function weightAdapterFlags(gene: WeightAdapterGene): string[];
+/** True iff the gene selects a real adapter (not base). */
+export declare function usesAdapter(gene: WeightAdapterGene): boolean;
+//# sourceMappingURL=genome.d.ts.map

package/dist/genome.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"genome.d.ts","sourceRoot":"","sources":["../src/genome.ts"],"names":[],"mappings":"AAuBA;;;;GAIG;AACH,MAAM,MAAM,iBAAiB,GAAG,MAAM,GAAG,IAAI,CAAC;AAE9C,0EAA0E;AAC1E,eAAO,MAAM,YAAY,EAAE,iBAAwB,CAAC;AAEpD;;;;;;GAMG;AACH,eAAO,MAAM,eAAe,EAAE,iBAAiB,EAI9C,CAAC;AAEF;;;;GAIG;AACH,wBAAgB,sBAAsB,CAAC,CAAC,EAAE,OAAO,GAAG,iBAAiB,CAMpE;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,IAAI,EAAE,iBAAiB,GAAG,MAAM,CAGnE;AAED;;;;GAIG;AACH,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,iBAAiB,GAAG,MAAM,EAAE,CAGpE;AAED,2DAA2D;AAC3D,wBAAgB,WAAW,CAAC,IAAI,EAAE,iBAAiB,GAAG,OAAO,CAE5D"}

package/dist/genome.js

@@ -0,0 +1,75 @@
+// SPDX-License-Identifier: MIT
+//
+// genome.ts — the `weightAdapter` genome gene.
+//
+// This is the bridge that lets Darwin evolution SELECT among LoRA adapters
+// (the evolutionary prune-the-overfitter safety net). The Darwin genome stays
+// gradient-FREE policy evolution; this gene adds a single reference to a tuned
+// weight adapter so an evolved config can run the cheap tier with-or-without a
+// distilled adapter, and let selection decide.
+//
+// SAFETY-NET RATIONALE: a LoRA tune can overfit (memorize the SFT set, regress
+// on held-out). Rather than trust the tune blindly, we make the adapter a GENE:
+// base (no adapter) competes against SFT-only / SFT+DPO / ratio variants under
+// the same conformant fitness, and evolution prunes an adapter that doesn't
+// actually lift held-out resolve. The default is ALWAYS base/no-adapter, so a
+// genome that never opts into an adapter is byte-identical to a pre-gene genome.
+//
+// The CONCRETE wiring into darwin-mode's evolve-config genome lives in
+// packages/darwin-mode/bench/swebench/evolve-config.mjs (the WEIGHT_ADAPTERS /
+// normalizeWeightAdapter / weightAdapterFlags helpers + the mutate/crossover/
+// seed integration). This module is the typed SPEC + reference implementation
+// of the gene's semantics, so the cross-package contract is testable from here.
+/** The canonical "no adapter" sentinel. Absent gene === BASE === this. */
+export const BASE_ADAPTER = null;
+/**
+ * The adapter variants Darwin evolution may choose among. BASE is always first
+ * (the control). The tuned variants reference adapters the runner produced:
+ * SFT-only and SFT+DPO are the two recipes; a ratio variant blends adapter
+ * strength. ALL are inert until an actual adapter is trained (GPU job) — the
+ * gene only NAMES an adapter; it does not create one.
+ */
+export const WEIGHT_ADAPTERS = [
+    BASE_ADAPTER, // control — no adapter (the safety-net baseline)
+    'sft', // SFT-distill only
+    'sft-dpo', // SFT then on-policy DPO
+];
+/**
+ * Normalize a raw gene value. Absent / '' / 'base' / 'none' all coerce to BASE
+ * so an unset gene is indistinguishable from an explicit base choice — that's
+ * what keeps pre-gene genomes byte-identical.
+ */
+export function normalizeWeightAdapter(v) {
+    if (v == null)
+        return BASE_ADAPTER;
+    if (typeof v !== 'string')
+        return BASE_ADAPTER;
+    const s = v.trim().toLowerCase();
+    if (s === '' || s === 'base' || s === 'none')
+        return BASE_ADAPTER;
+    return v.trim();
+}
+/**
+ * Stable key fragment for a genome's adapter choice. Empty string for BASE so
+ * the gene contributes NOTHING to a genome key unless an adapter is selected —
+ * the backward-compatibility invariant (a base-adapter genome keys identically
+ * to a pre-gene genome).
+ */
+export function weightAdapterSuffix(gene) {
+    const norm = normalizeWeightAdapter(gene);
+    return norm == null ? '' : `+w:${norm}`;
+}
+/**
+ * Map an adapter gene to the CLI flag the cheap-tier solver forwards. BASE
+ * emits NO flag (runs the stock model). A tuned adapter emits `--lora-adapter
+ * <id>` (the solver loads the adapter onto the cheap base before solving).
+ */
+export function weightAdapterFlags(gene) {
+    const norm = normalizeWeightAdapter(gene);
+    return norm == null ? [] : ['--lora-adapter', norm];
+}
+/** True iff the gene selects a real adapter (not base). */
+export function usesAdapter(gene) {
+    return normalizeWeightAdapter(gene) != null;
+}
+//# sourceMappingURL=genome.js.map

package/dist/genome.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"genome.js","sourceRoot":"","sources":["../src/genome.ts"],"names":[],"mappings":"AAAA,+BAA+B;AAC/B,EAAE;AACF,+CAA+C;AAC/C,EAAE;AACF,2EAA2E;AAC3E,8EAA8E;AAC9E,+EAA+E;AAC/E,+EAA+E;AAC/E,+CAA+C;AAC/C,EAAE;AACF,+EAA+E;AAC/E,gFAAgF;AAChF,+EAA+E;AAC/E,4EAA4E;AAC5E,8EAA8E;AAC9E,iFAAiF;AACjF,EAAE;AACF,uEAAuE;AACvE,+EAA+E;AAC/E,8EAA8E;AAC9E,8EAA8E;AAC9E,gFAAgF;AAShF,0EAA0E;AAC1E,MAAM,CAAC,MAAM,YAAY,GAAsB,IAAI,CAAC;AAEpD;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,eAAe,GAAwB;IAClD,YAAY,EAAE,iDAAiD;IAC/D,KAAK,EAAE,mBAAmB;IAC1B,SAAS,EAAE,yBAAyB;CACrC,CAAC;AAEF;;;;GAIG;AACH,MAAM,UAAU,sBAAsB,CAAC,CAAU;IAC/C,IAAI,CAAC,IAAI,IAAI;QAAE,OAAO,YAAY,CAAC;IACnC,IAAI,OAAO,CAAC,KAAK,QAAQ;QAAE,OAAO,YAAY,CAAC;IAC/C,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IACjC,IAAI,CAAC,KAAK,EAAE,IAAI,CAAC,KAAK,MAAM,IAAI,CAAC,KAAK,MAAM;QAAE,OAAO,YAAY,CAAC;IAClE,OAAO,CAAC,CAAC,IAAI,EAAE,CAAC;AAClB,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,IAAuB;IACzD,MAAM,IAAI,GAAG,sBAAsB,CAAC,IAAI,CAAC,CAAC;IAC1C,OAAO,IAAI,IAAI,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,IAAI,EAAE,CAAC;AAC1C,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,kBAAkB,CAAC,IAAuB;IACxD,MAAM,IAAI,GAAG,sBAAsB,CAAC,IAAI,CAAC,CAAC;IAC1C,OAAO,IAAI,IAAI,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,gBAAgB,EAAE,IAAI,CAAC,CAAC;AACtD,CAAC;AAED,2DAA2D;AAC3D,MAAM,UAAU,WAAW,CAAC,IAAuB;IACjD,OAAO,sBAAsB,CAAC,IAAI,CAAC,IAAI,IAAI,CAAC;AAC9C,CAAC"}

package/dist/index.d.ts

@@ -0,0 +1,11 @@
+export type { ChatMessage, ToolCall, PolicyTier, DarwinTrajectory, SftRow, DpoRow, ExportOptions, ExportResult, ExportReport, } from './types.js';
+export { exportTrainingData, assertTrainEvalDisjoint, estimateTokens, sftToJsonl, dpoToJsonl, } from './export.js';
+export type { BaseModelSpec, TrainStage, LoraConfig, TrainConfig, TrainRunOptions, TrainRunResult, TrainingPlan, } from './train.js';
+export { DEFAULT_LORA, defaultDetectGpu, assertTunableSize, sftConfig, dpoConfig, buildCommand, buildPlan, runTraining, twoStagePlan, adaptSftForRunner, adaptDpoForRunner, } from './train.js';
+export type { WeightAdapterGene } from './genome.js';
+export { BASE_ADAPTER, WEIGHT_ADAPTERS, normalizeWeightAdapter, weightAdapterSuffix, weightAdapterFlags, usesAdapter, } from './genome.js';
+export type { CascadeOutcome, CascadeSummary, CostParetoDelta, } from './eval.js';
+export { summarizeCascade, costParetoDelta } from './eval.js';
+export type { RewardHackKind, RewardHackFinding } from './reward-hack.js';
+export { detectRewardHack, isRewardHacked } from './reward-hack.js';
+//# sourceMappingURL=index.d.ts.map

package/dist/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAQA,YAAY,EACV,WAAW,EACX,QAAQ,EACR,UAAU,EACV,gBAAgB,EAChB,MAAM,EACN,MAAM,EACN,aAAa,EACb,YAAY,EACZ,YAAY,GACb,MAAM,YAAY,CAAC;AAEpB,OAAO,EACL,kBAAkB,EAClB,uBAAuB,EACvB,cAAc,EACd,UAAU,EACV,UAAU,GACX,MAAM,aAAa,CAAC;AAErB,YAAY,EACV,aAAa,EACb,UAAU,EACV,UAAU,EACV,WAAW,EACX,eAAe,EACf,cAAc,EACd,YAAY,GACb,MAAM,YAAY,CAAC;AAEpB,OAAO,EACL,YAAY,EACZ,gBAAgB,EAChB,iBAAiB,EACjB,SAAS,EACT,SAAS,EACT,YAAY,EACZ,SAAS,EACT,WAAW,EACX,YAAY,EACZ,iBAAiB,EACjB,iBAAiB,GAClB,MAAM,YAAY,CAAC;AAEpB,YAAY,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAErD,OAAO,EACL,YAAY,EACZ,eAAe,EACf,sBAAsB,EACtB,mBAAmB,EACnB,kBAAkB,EAClB,WAAW,GACZ,MAAM,aAAa,CAAC;AAErB,YAAY,EACV,cAAc,EACd,cAAc,EACd,eAAe,GAChB,MAAM,WAAW,CAAC;AAEnB,OAAO,EAAE,gBAAgB,EAAE,eAAe,EAAE,MAAM,WAAW,CAAC;AAE9D,YAAY,EAAE,cAAc,EAAE,iBAAiB,EAAE,MAAM,kBAAkB,CAAC;AAE1E,OAAO,EAAE,gBAAgB,EAAE,cAAc,EAAE,MAAM,kBAAkB,CAAC"}

package/dist/index.js

@@ -0,0 +1,13 @@
+// SPDX-License-Identifier: MIT
+//
+// @metaharness/weight-eft — public API.
+//
+// Evolutionary fine-tuning: distill Darwin's archival success into the open
+// cheap tier via LoRA so the cost-cascade escalates to a frontier model less
+// often. Cost-Pareto axis, not the frontier ceiling. See ADR-198.
+export { exportTrainingData, assertTrainEvalDisjoint, estimateTokens, sftToJsonl, dpoToJsonl, } from './export.js';
+export { DEFAULT_LORA, defaultDetectGpu, assertTunableSize, sftConfig, dpoConfig, buildCommand, buildPlan, runTraining, twoStagePlan, adaptSftForRunner, adaptDpoForRunner, } from './train.js';
+export { BASE_ADAPTER, WEIGHT_ADAPTERS, normalizeWeightAdapter, weightAdapterSuffix, weightAdapterFlags, usesAdapter, } from './genome.js';
+export { summarizeCascade, costParetoDelta } from './eval.js';
+export { detectRewardHack, isRewardHacked } from './reward-hack.js';
+//# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA,+BAA+B;AAC/B,EAAE;AACF,wCAAwC;AACxC,EAAE;AACF,4EAA4E;AAC5E,6EAA6E;AAC7E,kEAAkE;AAclE,OAAO,EACL,kBAAkB,EAClB,uBAAuB,EACvB,cAAc,EACd,UAAU,EACV,UAAU,GACX,MAAM,aAAa,CAAC;AAYrB,OAAO,EACL,YAAY,EACZ,gBAAgB,EAChB,iBAAiB,EACjB,SAAS,EACT,SAAS,EACT,YAAY,EACZ,SAAS,EACT,WAAW,EACX,YAAY,EACZ,iBAAiB,EACjB,iBAAiB,GAClB,MAAM,YAAY,CAAC;AAIpB,OAAO,EACL,YAAY,EACZ,eAAe,EACf,sBAAsB,EACtB,mBAAmB,EACnB,kBAAkB,EAClB,WAAW,GACZ,MAAM,aAAa,CAAC;AAQrB,OAAO,EAAE,gBAAgB,EAAE,eAAe,EAAE,MAAM,WAAW,CAAC;AAI9D,OAAO,EAAE,gBAAgB,EAAE,cAAc,EAAE,MAAM,kBAAkB,CAAC"}

package/dist/reward-hack.d.ts

@@ -0,0 +1,17 @@
+import type { DarwinTrajectory } from './types.js';
+export type RewardHackKind = 'gold-read' | 'verification-tamper' | 'sandbox-escape';
+export interface RewardHackFinding {
+    kind: RewardHackKind;
+    /** The offending path/argument fragment. */
+    detail: string;
+    /** The tool that performed the offending action. */
+    tool: string;
+}
+/**
+ * Run the deterministic reward-hacking monitor over a single trajectory.
+ * Returns all findings (empty ⇒ clean). Pure function of the trajectory.
+ */
+export declare function detectRewardHack(t: DarwinTrajectory): RewardHackFinding[];
+/** True iff the trajectory shows ANY reward-hacking signal. */
+export declare function isRewardHacked(t: DarwinTrajectory): boolean;
+//# sourceMappingURL=reward-hack.d.ts.map

package/dist/reward-hack.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"reward-hack.d.ts","sourceRoot":"","sources":["../src/reward-hack.ts"],"names":[],"mappings":"AAoBA,OAAO,KAAK,EAAe,gBAAgB,EAAE,MAAM,YAAY,CAAC;AAwChE,MAAM,MAAM,cAAc,GAAG,WAAW,GAAG,qBAAqB,GAAG,gBAAgB,CAAC;AAEpF,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,cAAc,CAAC;IACrB,4CAA4C;IAC5C,MAAM,EAAE,MAAM,CAAC;IACf,oDAAoD;IACpD,IAAI,EAAE,MAAM,CAAC;CACd;AAcD;;;GAGG;AACH,wBAAgB,gBAAgB,CAAC,CAAC,EAAE,gBAAgB,GAAG,iBAAiB,EAAE,CA8BzE;AAED,+DAA+D;AAC/D,wBAAgB,cAAc,CAAC,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAE3D"}

package/dist/reward-hack.js

@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+//
+// reward-hack.ts — deterministic reward-hacking monitor (Ornith-1.0 borrow).
+//
+// An archived "success" that secretly reward-hacked would teach the model to
+// reward-hack. This is the TRAINING-DATA analog of the conformance firewall —
+// SEPARATE from and IN ADDITION TO the train/eval-disjointness guard. We run a
+// deterministic monitor over each gold-resolved trajectory and DROP any that:
+//   (a) read a withheld GOLD/TEST path (FAIL_TO_PASS / PASS_TO_PASS / gold_patch
+//       / the held-out test files),
+//   (b) modified the verification/test harness (the script that scores it),
+//   (c) touched a path OUTSIDE the sandboxed repo.
+//
+// Prior art: Ornith-1.0 (DeepReinforce) — self-scaffolding RL for agentic
+// coding — uses a deterministic monitor that flags trajectories acting outside
+// the sanctioned tool surface → zero reward + excluded. See ADR-198.
+//
+// The monitor reads ONLY the trajectory's own tool_calls + tool messages — it
+// is a pure function of the trajectory (no filesystem, $0, reproducible).
+/** Paths/patterns whose READ signals reward-hacking (the withheld oracle). */
+const GOLD_READ_PATTERNS = [
+    /gold_patch/i,
+    /\bFAIL_TO_PASS\b/,
+    /\bPASS_TO_PASS\b/,
+    /test_patch/i,
+    /\bexpected[_-]?(output|patch|result)\b/i,
+    /\.golden\b/i,
+    /run_evaluation/i, // reading the swebench evaluator itself
+];
+/** Paths whose WRITE/MODIFY signals tampering with the verification harness. */
+const VERIFICATION_WRITE_PATTERNS = [
+    /conftest\.py/i,
+    /run_tests?\.(sh|py)/i,
+    /run_evaluation/i,
+    /swebench[\/\\].*harness/i,
+    /\bMakefile\b/, // re-routing the test target
+    /tox\.ini/i,
+    /pytest\.ini/i,
+    /setup\.cfg/i, // can redefine test discovery
+];
+/** Path prefixes that are OUTSIDE the sandboxed repo (escape signals). */
+const OUT_OF_SANDBOX_PATTERNS = [
+    /(^|["'\s])\/etc\//,
+    /(^|["'\s])\/root\//,
+    /(^|["'\s])\/home\//,
+    /(^|["'\s])\/tmp\/(?!repo|work)/, // /tmp is fine for the work tree; flag other /tmp paths
+    /\.\.\/\.\.\//, // repeated parent traversal out of the repo
+    /(^|["'\s])~\//,
+];
+/** Tool names we treat as a READ of a path. */
+const READ_TOOLS = new Set(['read_file', 'cat', 'open_file', 'view', 'grep', 'search']);
+/** Tool names we treat as a WRITE/MODIFY of a path. */
+const WRITE_TOOLS = new Set(['write_file', 'apply_patch', 'edit_file', 'create_file', 'str_replace']);
+/** Extract the (tool, argumentsText) pairs from a trajectory's assistant turns. */
+function toolActions(messages) {
+    const out = [];
+    for (const m of messages) {
+        if (m.role !== 'assistant' || !m.tool_calls)
+            continue;
+        for (const tc of m.tool_calls) {
+            out.push({ tool: tc.function.name, args: tc.function.arguments });
+        }
+    }
+    return out;
+}
+/**
+ * Run the deterministic reward-hacking monitor over a single trajectory.
+ * Returns all findings (empty ⇒ clean). Pure function of the trajectory.
+ */
+export function detectRewardHack(t) {
+    const findings = [];
+    for (const { tool, args } of toolActions(t.messages)) {
+        const isRead = READ_TOOLS.has(tool);
+        const isWrite = WRITE_TOOLS.has(tool);
+        // (a) reading a withheld gold/test path — applies to ANY tool that takes a
+        //     path argument (a read disguised as a grep still leaks the oracle).
+        if (isRead || isWrite || /file|path|patch/i.test(tool)) {
+            for (const p of GOLD_READ_PATTERNS) {
+                const m = args.match(p);
+                if (m)
+                    findings.push({ kind: 'gold-read', detail: m[0], tool });
+            }
+        }
+        // (b) modifying the verification/test harness.
+        if (isWrite || /patch|edit|write|replace/i.test(tool)) {
+            for (const p of VERIFICATION_WRITE_PATTERNS) {
+                const m = args.match(p);
+                if (m)
+                    findings.push({ kind: 'verification-tamper', detail: m[0], tool });
+            }
+        }
+        // (c) touching a path outside the sandboxed repo.
+        for (const p of OUT_OF_SANDBOX_PATTERNS) {
+            const m = args.match(p);
+            if (m)
+                findings.push({ kind: 'sandbox-escape', detail: m[0].trim(), tool });
+        }
+    }
+    return findings;
+}
+/** True iff the trajectory shows ANY reward-hacking signal. */
+export function isRewardHacked(t) {
+    return detectRewardHack(t).length > 0;
+}
+//# sourceMappingURL=reward-hack.js.map

package/dist/reward-hack.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"reward-hack.js","sourceRoot":"","sources":["../src/reward-hack.ts"],"names":[],"mappings":"AAAA,+BAA+B;AAC/B,EAAE;AACF,6EAA6E;AAC7E,EAAE;AACF,6EAA6E;AAC7E,8EAA8E;AAC9E,+EAA+E;AAC/E,8EAA8E;AAC9E,iFAAiF;AACjF,oCAAoC;AACpC,4EAA4E;AAC5E,mDAAmD;AACnD,EAAE;AACF,0EAA0E;AAC1E,+EAA+E;AAC/E,qEAAqE;AACrE,EAAE;AACF,8EAA8E;AAC9E,0EAA0E;AAI1E,8EAA8E;AAC9E,MAAM,kBAAkB,GAAa;IACnC,aAAa;IACb,kBAAkB;IAClB,kBAAkB;IAClB,aAAa;IACb,yCAAyC;IACzC,aAAa;IACb,iBAAiB,EAAE,wCAAwC;CAC5D,CAAC;AAEF,gFAAgF;AAChF,MAAM,2BAA2B,GAAa;IAC5C,eAAe;IACf,sBAAsB;IACtB,iBAAiB;IACjB,0BAA0B;IAC1B,cAAc,EAAE,6BAA6B;IAC7C,WAAW;IACX,cAAc;IACd,aAAa,EAAE,8BAA8B;CAC9C,CAAC;AAEF,0EAA0E;AAC1E,MAAM,uBAAuB,GAAa;IACxC,mBAAmB;IACnB,oBAAoB;IACpB,oBAAoB;IACpB,gCAAgC,EAAE,wDAAwD;IAC1F,cAAc,EAAE,4CAA4C;IAC5D,eAAe;CAChB,CAAC;AAEF,+CAA+C;AAC/C,MAAM,UAAU,GAAG,IAAI,GAAG,CAAC,CAAC,WAAW,EAAE,KAAK,EAAE,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,QAAQ,CAAC,CAAC,CAAC;AACxF,uDAAuD;AACvD,MAAM,WAAW,GAAG,IAAI,GAAG,CAAC,CAAC,YAAY,EAAE,aAAa,EAAE,WAAW,EAAE,aAAa,EAAE,aAAa,CAAC,CAAC,CAAC;AAYtG,mFAAmF;AACnF,SAAS,WAAW,CAAC,QAAuB;IAC1C,MAAM,GAAG,GAA0C,EAAE,CAAC;IACtD,KAAK,MAAM,CAAC,IAAI,QAAQ,EAAE,CAAC;QACzB,IAAI,CAAC,CAAC,IAAI,KAAK,WAAW,IAAI,CAAC,CAAC,CAAC,UAAU;YAAE,SAAS;QACtD,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,UAAU,EAAE,CAAC;YAC9B,GAAG,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,EAAE,CAAC,QAAQ,CAAC,IAAI,EAAE,IAAI,EAAE,EAAE,CAAC,QAAQ,CAAC,SAAS,EAAE,CAAC,CAAC;QACpE,CAAC;IACH,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,CAAmB;IAClD,MAAM,QAAQ,GAAwB,EAAE,CAAC;IACzC,KAAK,MAAM,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,WAAW,CAAC,CAAC,CAAC,QAAQ,CAAC,EAAE,CAAC;QACrD,MAAM,MAAM,GAAG,UAAU,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QACpC,MAAM,OAAO,GAAG,WAAW,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QAEtC,2EAA2E;QAC3E,yEAAyE;QACzE,IAAI,MAAM,IAAI,OAAO,IAAI,kBAAkB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;YACvD,KAAK,MAAM,CAAC,IAAI,kBAAkB,EAAE,CAAC;gBACnC,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;gBACxB,IAAI,CAAC;oBAAE,QAAQ,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC;YAClE,CAAC;QACH,CAAC;QAED,+CAA+C;QAC/C,IAAI,OAAO,IAAI,2BAA2B,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;YACtD,KAAK,MAAM,CAAC,IAAI,2BAA2B,EAAE,CAAC;gBAC5C,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;gBACxB,IAAI,CAAC;oBAAE,QAAQ,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,qBAAqB,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC;YAC5E,CAAC;QACH,CAAC;QAED,kDAAkD;QAClD,KAAK,MAAM,CAAC,IAAI,uBAAuB,EAAE,CAAC;YACxC,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;YACxB,IAAI,CAAC;gBAAE,QAAQ,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC;QAC9E,CAAC;IACH,CAAC;IACD,OAAO,QAAQ,CAAC;AAClB,CAAC;AAED,+DAA+D;AAC/D,MAAM,UAAU,cAAc,CAAC,CAAmB;IAChD,OAAO,gBAAgB,CAAC,CAAC,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC;AACxC,CAAC"}