@konneal/engine 0.2.104 → 0.2.109

package/dist/{chunk-NCZHN6QW.js → chunk-G4KBK5XZ.js}

@@ -19,9 +19,10 @@ import {
   tableRetyped,
   toolNote,
   verdictNote
-} from "./chunk-ANUHRJ4Z.js";
+} from "./chunk-G56TYXNN.js";
 import {
   NO_CONTEXT,
+  STAGE_NAMES,
   answerQuality,
   appliedContext,
   buildMessages,
@@ -33,6 +34,7 @@ import {
   editionNote,
   estimateTokens,
   experimentalSourceLabels,
+  fill,
   graphExpand,
   identityNote,
   listwiseRerank,
@@ -43,6 +45,7 @@ import {
   opCfg,
   opTokenMember,
   parseContext,
+  promptVars,
   qualityNote,
   rawSessionToken,
   refCodec,
@@ -57,20 +60,20 @@ import {
   tokenBudget,
   understandQuery,
   usageTotal
-} from "./chunk-25ECO5M7.js";
+} from "./chunk-2R7OVYAO.js";
 import {
   embed,
   generateOnce,
   portModelRunner,
   portStore
-} from "./chunk-KNCRYAX3.js";
+} from "./chunk-TH5XGPD5.js";
 import {
   corsHeaders,
   err,
   json,
   readJson,
   validateQuery
-} from "./chunk-DTW4UW2E.js";
+} from "./chunk-DJ65EF5G.js";
 import {
   canonicalRefusal,
   refusalAnswer
@@ -80,7 +83,7 @@ import {
   requestSalt,
   resolveRequestScope,
   standardKeysFrom
-} from "./chunk-U4ERFUOK.js";
+} from "./chunk-4NXH5ZI6.js";
 import {
   LIMITS,
   MODELS,
@@ -91,7 +94,7 @@ import {
   requestEffort,
   roleModel,
   sha256Hex
-} from "./chunk-BOAITSVV.js";
+} from "./chunk-VT7DR6NQ.js";
 import {
   certificateLinks,
   registerNote,
@@ -1335,6 +1338,90 @@ function exactCacheKey(indexVersion, gen, ns, queryHash) {
   return `a:${indexVersion}:g${gen}:${ns}:${queryHash}`;
 }
 
+// workers/worker_public/src/ablate.ts
+function parseAblate(body, adminToken, presentedToken, knownStages) {
+  const rawStages = body?.ablate_stages;
+  const noVerdict = body?.ablate_no_verdict === true;
+  const rawRoute = body?.ablate_route;
+  const speculative = body?.ablate_speculative === true;
+  const namesAblation = rawStages !== void 0 && rawStages !== null;
+  const routeAblation = rawRoute !== void 0 && rawRoute !== null && rawRoute !== "adaptive";
+  if (!namesAblation && !noVerdict && !routeAblation && !speculative) return { ok: true, config: null };
+  if (!adminToken || presentedToken !== adminToken) {
+    return {
+      ok: false,
+      status: 403,
+      code: "ablation_forbidden",
+      message: "ablation fields require the admin credential (x-admin-token)"
+    };
+  }
+  let route = "adaptive";
+  if (routeAblation) {
+    if (rawRoute !== "fast" && rawRoute !== "deep") {
+      return { ok: false, status: 400, code: "invalid_ablation", message: 'ablate_route must be "fast", "deep" or "adaptive"' };
+    }
+    route = rawRoute;
+  }
+  let stages = null;
+  if (namesAblation) {
+    if (!Array.isArray(rawStages) || rawStages.length === 0) {
+      return { ok: false, status: 400, code: "invalid_ablation", message: "ablate_stages must be a non-empty array of stage names" };
+    }
+    if (rawStages.some((n) => typeof n !== "string")) {
+      return { ok: false, status: 400, code: "invalid_ablation", message: "ablate_stages must contain only stage-name strings" };
+    }
+    const known = new Set(knownStages);
+    const unknown = rawStages.filter((n) => !known.has(n));
+    if (unknown.length) {
+      return { ok: false, status: 400, code: "invalid_ablation", message: `unknown stage(s): ${unknown.join(", ")}` };
+    }
+    stages = [...new Set(rawStages)];
+  }
+  return { ok: true, config: { stages, noVerdict, route, speculative } };
+}
+
+// workers/worker_public/src/route.ts
+var ROUTE_FAST_WORD_CAP = 10;
+var ROUTE_FAST_DROPS = ["hyde", "multi-query", "sub-query"];
+function routeFor(u, query) {
+  const features = [];
+  const words = query.trim().split(/\s+/).filter(Boolean).length;
+  if (u?.complexity === "complex") features.push("complex");
+  if ((u?.sub_queries ?? []).length > 0) features.push("sub-queries");
+  if (u?.process_intent) features.push("process-intent");
+  if (words > ROUTE_FAST_WORD_CAP) features.push("long-question");
+  if (features.length) return { route: "deep", features };
+  if (u?.doc_number) features.push("doc-scoped");
+  if (u?.term) features.push("definitional");
+  if ((u?.defined_terms ?? []).length > 0) features.push("terminology");
+  features.push("short-question");
+  return { route: "fast", features };
+}
+function fastRouteStages(registry) {
+  return registry.filter((n) => !ROUTE_FAST_DROPS.includes(n));
+}
+
+// workers/worker_public/src/speculative.ts
+var SPECULATIVE_SUBSETS = 3;
+function partitionSubsets(hits, k = SPECULATIVE_SUBSETS) {
+  const groups = [];
+  const byDoc = /* @__PURE__ */ new Map();
+  for (const h of hits) {
+    const key = h.metadata.docidentifier || h.metadata.doc_id || "";
+    const arr = byDoc.get(key) ?? [];
+    arr.push(h);
+    if (arr.length === 1) groups.push(arr);
+    byDoc.set(key, arr);
+  }
+  if (groups.length < 2) return null;
+  const subsets = Array.from({ length: Math.min(k, groups.length) }, () => []);
+  groups.forEach((g, i) => subsets[i % subsets.length].push(...g));
+  return subsets;
+}
+
+// workers/worker_public/prompts/speculative-verify.md
+var speculative_verify_default = "You answer questions about legal-metrology publications for {{PUBLISHER_NAME}}. Candidate draft answers were prepared in parallel, each from a PARTIAL view of the evidence (a subset of the passages); you see every candidate beside the FULL passages. Produce the final answer.\n\nRules:\n1. Compare the candidates against the full passages. Adopt the candidate best supported by the evidence as your starting point; correct every claim the full passages contradict or fail to support, and fold in anything the other candidates got right that the adopted one missed.\n2. Ground every factual claim in the numbered passages and cite them inline as [1], [2], \u2026 exactly as the passages are numbered below. Never cite a passage that is not there.\n3. Quote verbatim \u2014 with quotation marks \u2014 when a clause's exact wording matters (definitions, thresholds, requirements); a quotation must appear word-for-word in the passages.\n4. Numeric and table values come from the passages exactly; never retype a table as markdown \u2014 refer to the table and cite it.\n5. If the passages do not answer the question, say so plainly instead of guessing.\n6. Be complete but compact: answer the question, cite, stop.\n";
+
 // workers/worker_public/src/operations.ts
 function isOperationIntent(query) {
   return /\b(operations?|endpoints?)\b/i.test(query);
@@ -1544,6 +1631,43 @@ function userImageDataUrl(body) {
   if (!m || !m[2]) return null;
   return img;
 }
+var DRAFT_TOKEN_BUDGET = 1600;
+function speculativeListing(hits) {
+  return hits.map((h, i) => {
+    const id = (h.metadata.docidentifier || h.metadata.doc_id || "source").replace(/\s*\(([A-Z])\)\s*$/, "").trim();
+    const edition = h.metadata.edition && !id.includes(h.metadata.edition) ? ":" + h.metadata.edition : "";
+    const anchor = h.metadata.clause_anchor ? ` \xA7${h.metadata.clause_anchor}` : "";
+    return `[${i + 1}] ${id}${edition}${anchor} ${h.text.replace(/\s+/g, " ").slice(0, 700)}`;
+  }).join("\n\n");
+}
+async function speculativeDraftVerify(env, opts) {
+  const subsets = partitionSubsets(opts.hits);
+  if (!subsets) return null;
+  const draftSets = subsets.map(
+    (subset) => buildMessages(opts.query, subset, opts.lang ?? void 0, [], void 0, void 0, DRAFT_TOKEN_BUDGET).messages
+  );
+  const drafts = await Promise.all(
+    draftSets.map((m) => generateOnce(env, opts.draftModel, m, "low").catch(() => null))
+  );
+  const candidates = drafts.map((d) => (d ?? "").trim()).filter((d) => d.length > 0);
+  if (!candidates.length) return null;
+  const candidateText = candidates.map((d, i) => `<candidate ${i + 1} (prepared from a partial view)>
+${d}
+</candidate ${i + 1}>`).join("\n\n");
+  const system = fill(speculative_verify_default.trimEnd(), promptVars({}));
+  const answer = await generateOnce(env, opts.verifierModel, [
+    { role: "system", content: system },
+    { role: "user", content: `Question: ${opts.query}
+
+Candidate draft answers:
+${candidateText}
+
+Full context passages:
+${speculativeListing(opts.hits)}` }
+  ], "low").catch(() => null);
+  if (!answer || !answer.trim()) return null;
+  return { answer, chosen: candidates.length, nDrafts: drafts.length };
+}
 async function cacheGet(env, gen, ns, query, lang, salt) {
   const key = exactCacheKey(env.INDEX_VERSION, gen, ns, await sha256Hex(cacheKeyMaterial(query, lang, salt)));
   const hit = await env.CACHE.get(key, "json");
@@ -1706,6 +1830,9 @@ async function handleAsk(env, ctx, req, tier, key) {
   const body = await readJson(req);
   const q = validateQuery(body);
   if (!q) return err(400, "invalid_input", `query is required (1-${LIMITS.maxInputChars} chars)`);
+  const ablated = parseAblate(body, env.ADMIN_TOKEN, req.headers.get("x-admin-token"), STAGE_NAMES);
+  if (!ablated.ok) return err(ablated.status, ablated.code, ablated.message);
+  const ablate = ablated.config;
   const declaredCtx = parseContext(body);
   const draftAct = P().publisher.features?.drafts ? detectDraftIntent(q.query) : null;
   const apiCallIntent = !draftAct && P().publisher.features?.api_call_drafts ? detectApiCallIntent(q.query, declaredCtx) : null;
@@ -1720,6 +1847,8 @@ async function handleAsk(env, ctx, req, tier, key) {
     scope: opMember.scope
   } : null);
   const effort = requestEffort(env, member, body?.effort);
+  const declaredEffort = body?.effort;
+  const routeEffort = (route) => route === "fast" && !declaredEffort ? "low" : effort;
   const limit = tier === "key" ? key.day_limit : tier === "member" || member ? num(env, "MEMBER_DAY_ASK", 300) : num(env, "ANON_DAY_ASK", 20);
   const bucketId = tier === "key" ? `key:${key.id}` : member ? `sub:${member.sub}` : clientIp(req);
   const quota = await checkQuota(env, "ask", bucketId, limit, effort === "low" ? 1 : 2);
@@ -1792,7 +1921,7 @@ async function handleAsk(env, ctx, req, tier, key) {
   const fresh = freshRequested(body);
   const gen = await corpusGen(env.CACHE);
   const withImage = effectiveImage || storedImage;
-  const cached = fresh || contextual || declaredCtx || draftAct || apiCallIntent || withImage ? null : await cacheGet(env, gen, ns, q.query, q.lang, salt);
+  const cached = fresh || contextual || declaredCtx || draftAct || apiCallIntent || withImage || ablate ? null : await cacheGet(env, gen, ns, q.query, q.lang, salt);
   const wantsStream = body?.stream === true || tier === "anon" && body?.stream !== false;
   if (cached) {
     telemetry(env, ctx, tier, "ask", null, true, (cached.value.answer ?? "").length, cached.value.query_hash, q.lang, "exact", telemetryMeta());
@@ -2066,7 +2195,7 @@ ${summary}` }] : [],
     console.log("model plane: bound", boundModel.node_id, `[${boundModel.standard}]`, boundModel.clause?.urn ?? "no-clause", boundModel.gated ? "(gated: license)" : "");
   }
   const modelNote = boundModel && !boundModel.gated ? modelGroundingBlock(boundModel) : void 0;
-  const machineVerdict = boundModel && !boundModel.gated ? evaluate(boundModel.content, q.query) : null;
+  const machineVerdict = boundModel && !boundModel.gated && !ablate?.noVerdict ? evaluate(boundModel.content, q.query) : null;
   const machineNote = machineVerdict && boundModel ? verdictNote(machineVerdict, boundModel) : void 0;
   const modelNodeRows = async (kind, docNum) => {
     const attempt = (num3) => {
@@ -2081,7 +2210,7 @@ ${summary}` }] : [],
   };
   let conditionVerdict = null;
   let conditionStandard = null;
-  if (!machineVerdict && !boundModel && P().publisher.features?.model_plane) {
+  if (!machineVerdict && !boundModel && !ablate?.noVerdict && P().publisher.features?.model_plane) {
     const ql = q.query.toLowerCase();
     const severityWord = /\b(severity|test|valid|tolerance|condition|within)\b/.test(ql);
     const stated = quantitiesIn(q.query);
@@ -2195,6 +2324,10 @@ ${summary}` }] : [],
       boundaryNote = boundaryNoteText(match, citing);
     }
   }
+  const routeMode = ablate?.route ?? "adaptive";
+  const routed = routeMode === "adaptive" ? routeFor(understanding, q.query) : { route: routeMode, features: [`forced-${routeMode}`] };
+  const routeSubset = ablate?.stages ?? (routed.route === "fast" ? fastRouteStages(STAGE_NAMES) : null);
+  if (routeSubset) console.log("route:", routed.route, routed.features.join("+"), "\u2014", STAGE_NAMES.length - routeSubset.length, "lanes dropped");
   try {
     const tR = Date.now();
     if (declaredCtx?.kind === "account") {
@@ -2232,7 +2365,8 @@ Answer account questions from these records ONLY: name the record when you use i
       datasetScope: narrowed ? corpora : null,
       standardKeys,
       lexicalBoost,
-      licensedDocNumbers
+      licensedDocNumbers,
+      ...routeSubset ? { stageSubset: routeSubset } : {}
     });
     stageTiming["retrieve-core"] = Date.now() - tR;
     console.log("stage: retrieve", Date.now() - tR, "ms");
@@ -2256,7 +2390,7 @@ Answer account questions from these records ONLY: name the record when you use i
     if (grade === "weak" && understanding?.docidentifier) {
       const broaden = `${understanding.standalone_query || q.query} ${understanding.docidentifier}`.trim();
       const tc = Date.now();
-      const second = await retrieve(env, q.query, { prev, understanding, queryOverride: broaden, federate, datasetScope: narrowed ? corpora : null, standardKeys, sealScope: declaredScoped ? docScope : null, lexicalBoost });
+      const second = await retrieve(env, q.query, { prev, understanding, queryOverride: broaden, federate, datasetScope: narrowed ? corpora : null, standardKeys, sealScope: declaredScoped ? docScope : null, lexicalBoost, ...routeSubset ? { stageSubset: routeSubset } : {} });
       const grade2 = await gradeRetrieval(env.AI, roleModel(env, "grader"), q.query, second.hits.map((h) => h.text));
       stageTiming.corrective = Date.now() - tc;
       if (grade2 === "good") retrieved = second;
@@ -2269,7 +2403,7 @@ Answer account questions from these records ONLY: name the record when you use i
   const { hits } = retrieved;
   if (hits.length === 0 && !liveRecords?.length && (!boundModel || boundModel.gated)) {
     const answer2 = licenseBoundaryRefusal(modelDocHint?.doc_number ?? understanding?.doc_number ?? null, standardKeys) ?? refusalAnswer();
-    const out2 = { answer: answer2, citations: [], model, query_hash: await sha256Hex(q.query), context_applied: ctxApplied };
+    const out2 = { answer: answer2, citations: [], model, query_hash: await sha256Hex(q.query), context_applied: ctxApplied, ...ablate ? { ablate } : {}, route: { route: routed.route, features: routed.features } };
     telemetry(env, ctx, tier, "ask", model, true, answer2.length, out2.query_hash, q.lang, void 0, telemetryMeta());
     return json({ ...out2, quota });
   }
@@ -2359,7 +2493,7 @@ ${q.query}`.replace("\n\n\n\n", "\n\n") }];
   const queryHash = await sha256Hex(q.query);
   const cites = boundModel ? [modelCitation(boundModel), ...citations(usedHits)] : citations(usedHits);
   if (wantsStream) {
-    const stream = await generateStream(env, model, messages, effort);
+    const stream = await generateStream(env, model, messages, routeEffort(routed.route));
     if (stream) {
       const encoder = new TextEncoder();
       const sse = new ReadableStream({
@@ -2401,7 +2535,7 @@ ${q.query}`.replace("\n\n\n\n", "\n\n") }];
               ...messages.slice(0, -1),
               { role: "user", content: "The photograph was attached and the register note lists actual matching certificate rows. Present those rows \u2014 number, holder, model, status, and every document link \u2014 as the answer. Do not claim that nothing was provided." },
               messages[messages.length - 1]
-            ], effort);
+            ], routeEffort(routed.route));
             if (regen) {
               full = "";
               try {
@@ -2451,7 +2585,7 @@ ${q.query}`.replace("\n\n\n\n", "\n\n") }];
           if (streamed.violations.length > 0) {
             console.log("anchors:", streamed.violations.length, "of", streamed.total, "unverified \u2014 not caching");
           }
-          if (streamed.violations.length === 0 && canonical.length > 0 && !contextual && !declaredCtx && !canonical.includes(refusalAnswer())) {
+          if (streamed.violations.length === 0 && canonical.length > 0 && !contextual && !declaredCtx && !canonical.includes(refusalAnswer()) && !ablate) {
             ctx.waitUntil(
               env.CACHE.put(exactCacheKey(env.INDEX_VERSION, gen, ns, await sha256Hex(cacheKeyMaterial(q.query, q.lang, salt))), JSON.stringify({ answer: canonical, citations: cites, model, query_hash: queryHash, ...sourceQuality ? { source_quality: sourceQuality, confidence_note: streamedNote || qualityNote(sourceQuality) } : {}, ...streamedEntailment ? { entailment: streamedEntailment } : {} }), { expirationTtl: LIMITS.cacheTtlSec })
             );
@@ -2470,14 +2604,32 @@ ${q.query}`.replace("\n\n\n\n", "\n\n") }];
     }
   }
   const tGen = Date.now();
-  let answer = await generateOnce(env, model, messages, effort);
+  let answer = null;
+  if (ablate?.speculative) {
+    const spec = await speculativeDraftVerify(env, {
+      query: q.query,
+      hits: usedHits,
+      lang: q.lang,
+      draftModel: MODELS.member,
+      verifierModel: roleModel(env, "verifier")
+    }).catch((e) => {
+      console.log("speculative: failed, falling through \u2014", String(e).slice(0, 120));
+      return null;
+    });
+    if (spec) {
+      answer = spec.answer;
+      stageTiming.speculative = Date.now() - tGen;
+      console.log("speculative:", spec.nDrafts, "drafts,", spec.chosen, "viable \u2014 verified");
+    }
+  }
+  if (!answer) answer = await generateOnce(env, model, messages, routeEffort(routed.route));
   if (answer === null) {
     generateRetries += 1;
     const isFigureAttachMessage = (m) => Array.isArray(m.content) && m.content.some((part) => part?.type === "text" && /^The original image of figure unit /.test(part.text ?? ""));
     const flat = messages.filter((m) => !isFigureAttachMessage(m)).map(
       (m) => typeof m.content === "string" ? m : { ...m, content: m.content.filter((p) => p?.type === "text").map((p) => (p?.text ?? "").replace(/\n?\(The user attached an image with this question; interpret it directly when answering\.\)/, "")).join("\n") }
     );
-    answer = await generateOnce(env, MODELS.fallback, flat, effort);
+    answer = await generateOnce(env, MODELS.fallback, flat, routeEffort(routed.route));
   }
   if (answer) answer = canonicalRefusal(answer);
   stageTiming.generate = Date.now() - tGen;
@@ -2501,7 +2653,7 @@ ${q.query}`.replace("\n\n\n\n", "\n\n") }];
       const tableUnitId = unreferenced ? used.find((h) => h.metadata.unit_id && h.metadata.block === "table")?.metadata.unit_id : void 0;
       const note = retyped || unreferenced ? `Correction notice: your draft reproduced a table as markdown or presented a served table's data without its reference. Rewrite the answer: describe the table in prose, cite the clause, and write the reference token [[u:${tableUnitId ?? "<unit id>"}]] exactly where the table belongs. Do not render any table as markdown.` : ANCHOR_CORRECTION_NOTE;
       generateRetries += 1;
-      const corrected = await generateOnce(env, model, [...messages, { role: "system", content: note }], effort);
+      const corrected = await generateOnce(env, model, [...messages, { role: "system", content: note }], routeEffort(routed.route));
       if (corrected) {
         const correctedAnswer = canonicalRefusal(corrected);
         const retryAnchors = checkQuoteAnchors(correctedAnswer, used.map((h) => h.text));
@@ -2521,11 +2673,12 @@ ${q.query}`.replace("\n\n\n\n", "\n\n") }];
         understanding: { ...understanding, standalone_query: `${understanding?.standalone_query || q.query} ${reflection.missing_info}` },
         sealScope: declaredScoped ? docScope : null,
         datasetScope: narrowed ? corpora : null,
-        standardKeys
+        standardKeys,
+        ...routeSubset ? { stageSubset: routeSubset } : {}
       });
       if (retryRetrieve.hits.length > 0) {
         const { messages: retryMessages, usedHits: retryUsed } = buildMessages(q.query, retryRetrieve.hits, q.lang, keptHistory, void 0, summary, budget);
-        const retryAnswer = await generateOnce(env, model, retryMessages, effort);
+        const retryAnswer = await generateOnce(env, model, retryMessages, routeEffort(routed.route));
         if (retryAnswer) {
           answer = canonicalRefusal(retryAnswer);
           used = retryUsed;
@@ -2578,8 +2731,8 @@ ${q.query}`.replace("\n\n\n\n", "\n\n") }];
       if (v.support !== "supported") console.log("entailment:", v.support, raced.score, raced.ungrounded_claims?.slice(0, 2));
     }
   }
-  const out = { answer, citations: finalCites, ...jsonQuality ? { source_quality: jsonQuality, confidence_note: entailmentNote || qualityNote(jsonQuality), ...entailment ? { entailment } : {}, ...jsonQuality === "ocr" ? { experimental_sources: experimentalSourceLabels(finalCites) } : {} } : {}, model: MODELS.member, query_hash: queryHash, follow_ups: understanding?.follow_ups ?? [], blocks: [...c2ns.blocks, ...verdictBlock ? [verdictBlock] : [], ...conditionBlock ? [conditionBlock] : [], ...aggregationBlock ? [aggregationBlock] : [], ...completionBlocks], context_applied: ctxApplied, ...liveRecords ? { records: liveRecords } : {} };
-  const cacheable = !contextual && !declaredCtx && !answer.includes(refusalAnswer()) && finalAnchors.violations.length === 0;
+  const out = { answer, citations: finalCites, ...jsonQuality ? { source_quality: jsonQuality, confidence_note: entailmentNote || qualityNote(jsonQuality), ...entailment ? { entailment } : {}, ...jsonQuality === "ocr" ? { experimental_sources: experimentalSourceLabels(finalCites) } : {} } : {}, model: MODELS.member, query_hash: queryHash, follow_ups: understanding?.follow_ups ?? [], blocks: [...c2ns.blocks, ...verdictBlock ? [verdictBlock] : [], ...conditionBlock ? [conditionBlock] : [], ...aggregationBlock ? [aggregationBlock] : [], ...completionBlocks], context_applied: ctxApplied, ...liveRecords ? { records: liveRecords } : {}, ...ablate ? { ablate } : {}, route: { route: routed.route, features: routed.features } };
+  const cacheable = !contextual && !declaredCtx && !ablate && !answer.includes(refusalAnswer()) && finalAnchors.violations.length === 0;
   if (cacheable) {
     const ck = exactCacheKey(env.INDEX_VERSION, gen, ns, await sha256Hex(cacheKeyMaterial(q.query, q.lang, salt)));
     ctx.waitUntil(env.CACHE.put(ck, JSON.stringify(out), { expirationTtl: LIMITS.cacheTtlSec }));

package/dist/{chunk-ANUHRJ4Z.js → chunk-G56TYXNN.js}

@@ -4,13 +4,13 @@ import {
   portIndex,
   portModelRunner,
   portStore
-} from "./chunk-KNCRYAX3.js";
+} from "./chunk-TH5XGPD5.js";
 import {
   standardKeysFrom
-} from "./chunk-U4ERFUOK.js";
+} from "./chunk-4NXH5ZI6.js";
 import {
   THRESHOLDS
-} from "./chunk-BOAITSVV.js";
+} from "./chunk-VT7DR6NQ.js";
 import {
   P
 } from "./chunk-3FYJM7LH.js";

package/dist/{chunk-TFGJ6NYS.js → chunk-HAU5INEJ.js}

@@ -6,27 +6,27 @@ import {
   sessionFrom,
   telemetry,
   understandQuery
-} from "./chunk-25ECO5M7.js";
+} from "./chunk-2R7OVYAO.js";
 import {
   portModelRunner
-} from "./chunk-KNCRYAX3.js";
+} from "./chunk-TH5XGPD5.js";
 import {
   corsHeaders,
   err,
   json,
   readJson,
   validateQuery
-} from "./chunk-DTW4UW2E.js";
+} from "./chunk-DJ65EF5G.js";
 import {
   entitlementScope,
   standardKeysFrom
-} from "./chunk-U4ERFUOK.js";
+} from "./chunk-4NXH5ZI6.js";
 import {
   LIMITS,
   MODELS,
   num,
   sha256Hex
-} from "./chunk-BOAITSVV.js";
+} from "./chunk-VT7DR6NQ.js";
 
 // workers/worker_public/src/search.ts
 async function handleSearch(env, ctx, req, tier, key) {

package/dist/{chunk-KNCRYAX3.js → chunk-TH5XGPD5.js}

@@ -1,7 +1,7 @@
 import {
   answerEffort,
   effortBudget
-} from "./chunk-BOAITSVV.js";
+} from "./chunk-VT7DR6NQ.js";
 
 // workers/worker_public/src/ai.ts
 var delay = (ms) => new Promise((r) => setTimeout(r, ms));

package/dist/{chunk-BOAITSVV.js → chunk-VT7DR6NQ.js}

@@ -28,7 +28,12 @@ var MODELS = {
   // cross-encoder prunes, listwise orders jointly)
   listwise: "@cf/zai-org/glm-4.7-flash",
   // deep-research loop (G10): bounded agentic iterations, members-only
-  research: "@cf/zai-org/glm-5.3-flash"
+  research: "@cf/zai-org/glm-5.3-flash",
+  // speculative draft-verify (TODO.sota/02 row 2): the strong verifier
+  // over the cheap drafts — the research-tier model verifies; the drafter
+  // is the standard answer model (MODELS.member). VERIFIER_MODEL swaps it
+  // live, same lever as every role.
+  verifier: "@cf/zai-org/glm-5.2"
 };
 var LIMITS = {
   /** Attachment quotas (TODO.new-era/8): per-member ceilings so a member's
@@ -65,6 +70,11 @@ var THRESHOLDS = {
    *  pool below the primary dense lane; they must earn their window
    *  slot under the cross-encoder, not by graph membership alone. */
   graphLaneDiscount: 0.75,
+  /** The PPR lane's merge discount (TODO.sota/06, catalog row 4): graph-
+   *  propagated candidates arrive at the same half-discount the narrowed
+   *  graph lane pays (rag#137's law — flooding dilutes the pool), and
+   *  only the cross-encoder's preference survives the cut. */
+  pprDiscount: 0.375,
   licensedLaneScore: 0.5,
   /** The entailment gate (TODO.new-era/10): the faithfulness scorer
    *  promoted from eval-only onto the serving path. The confidence line

package/dist/config.d.ts

@@ -10,6 +10,7 @@ export declare const MODELS: {
     readonly enrich: "@cf/deepseek-ai/deepseek-v4-pro-0813";
     readonly listwise: "@cf/zai-org/glm-4.7-flash";
     readonly research: "@cf/zai-org/glm-5.3-flash";
+    readonly verifier: "@cf/zai-org/glm-5.2";
 };
 export declare const LIMITS: {
     /** Attachment quotas (TODO.new-era/8): per-member ceilings so a member's
@@ -37,6 +38,11 @@ export declare const THRESHOLDS: {
      *  pool below the primary dense lane; they must earn their window
      *  slot under the cross-encoder, not by graph membership alone. */
     readonly graphLaneDiscount: 0.75;
+    /** The PPR lane's merge discount (TODO.sota/06, catalog row 4): graph-
+     *  propagated candidates arrive at the same half-discount the narrowed
+     *  graph lane pays (rag#137's law — flooding dilutes the pool), and
+     *  only the cross-encoder's preference survives the cut. */
+    readonly pprDiscount: 0.375;
     readonly licensedLaneScore: 0.5;
     /** The entailment gate (TODO.new-era/10): the faithfulness scorer
      *  promoted from eval-only onto the serving path. The confidence line

package/dist/openapi-types.d.ts

@@ -871,6 +871,24 @@ export interface components {
              * @default false
              */
             fresh: boolean;
+            /** @description Research-only (TODO.sota/09): run the retrieval registry projected to these stage names (registry order preserved; unknown names are a 400). Requires the x-admin-token header — a request naming a configuration without it is a 403. Ablated answers bypass every answer cache read and write and echo the active configuration in the response's ablate field. */
+            ablate_stages?: string[];
+            /**
+             * @description Research-only: skip the verdict engine (machine verdicts and condition-set evaluation); the answer model narrates from passages alone. Same admin credential and cache law as ablate_stages.
+             * @default false
+             */
+            ablate_no_verdict: boolean;
+            /**
+             * @description Research-only (TODO.sota/02 row 2): arm speculative draft-verify — the cheap model drafts an answer per diversified subset of the window in parallel, one strong call verifies every candidate against the full passages and produces the final answer. Same admin credential and cache law as ablate_stages.
+             * @default false
+             */
+            ablate_speculative: boolean;
+            /**
+             * @description Research-only: force the route (TODO.sota/05). fast drops the candidate-expansion lanes (hyde, multi-query, sub-query) and floors generation effort at low; deep keeps the full registry. adaptive (the default) lets the router decide from understanding's features. Same admin credential and cache law as ablate_stages.
+             * @default adaptive
+             * @enum {string}
+             */
+            ablate_route: "fast" | "deep" | "adaptive";
             /** @description Forces the answer language, for example en or fr. */
             lang?: string;
             /** @description The previous question, for conversational context. */
@@ -913,6 +931,12 @@ export interface components {
                 intent?: string;
                 doc?: string | null;
                 edition?: string | null;
+            };
+            /** @description The adaptive route the ask rode (TODO.sota/05): fast (the registry minus the candidate-expansion lanes, generation effort floored at low) or deep (the full registry), with the features that fired — the telemetry the routing-accuracy analysis reads. */
+            route?: {
+                /** @enum {string} */
+                route?: "fast" | "deep";
+                features?: string[];
                 term?: string | null;
                 terms?: string[];
                 lang?: string | null;

package/dist/prompts/doc-summary.md

@@ -0,0 +1 @@
+You write the document-level summary for a legal-metrology publication in {{PUBLISHER_NAME}}'s retrieval index. The section summaries below cover the publication clause by clause. Write one compact abstract, three to five sentences: what the publication regulates or provides, its scope and applicability, and the structure a reader should know (parts, annexes, where the normative core sits). Use the publication's own terms exactly. No preamble, no headings.

package/dist/prompts/speculative-verify.md

@@ -0,0 +1,9 @@
+You answer questions about legal-metrology publications for {{PUBLISHER_NAME}}. Candidate draft answers were prepared in parallel, each from a PARTIAL view of the evidence (a subset of the passages); you see every candidate beside the FULL passages. Produce the final answer.
+
+Rules:
+1. Compare the candidates against the full passages. Adopt the candidate best supported by the evidence as your starting point; correct every claim the full passages contradict or fail to support, and fold in anything the other candidates got right that the adopted one missed.
+2. Ground every factual claim in the numbered passages and cite them inline as [1], [2], … exactly as the passages are numbered below. Never cite a passage that is not there.
+3. Quote verbatim — with quotation marks — when a clause's exact wording matters (definitions, thresholds, requirements); a quotation must appear word-for-word in the passages.
+4. Numeric and table values come from the passages exactly; never retype a table as markdown — refer to the table and cite it.
+5. If the passages do not answer the question, say so plainly instead of guessing.
+6. Be complete but compact: answer the question, cite, stop.

package/dist/route.d.ts

@@ -0,0 +1,29 @@
+export type Route = "fast" | "deep";
+/** Minimal structural shape of the understanding the router reads —
+ *  the full contract lives in understandContract.ts; this subset keeps
+ *  the module self-contained. */
+export interface RouteFeatures {
+    complexity?: "simple" | "complex" | null;
+    doc_number?: string | null;
+    term?: string | null;
+    defined_terms?: string[];
+    sub_queries?: string[];
+    process_intent?: boolean;
+}
+/** THRESHOLDS-adjacent policy numbers (declared here until the grid
+ *  measures them; every steering number lives in one place — this is
+ *  the router's place). */
+export declare const ROUTE_FAST_WORD_CAP = 10;
+/** The candidate-expansion lanes the fast route drops (the stage
+ *  names in the registry). */
+export declare const ROUTE_FAST_DROPS: string[];
+/** Route one question from its understanding features. Returns the
+ *  route and the fired features — the response echoes both, which is
+ *  the telemetry the routing-accuracy analysis reads (TODO.sota/05). */
+export declare function routeFor(u: RouteFeatures | null, query: string): {
+    route: Route;
+    features: string[];
+};
+/** The fast route's stage list: the registry minus the expansion lanes.
+ *  projectStages owns the ordering and validation semantics. */
+export declare function fastRouteStages(registry: readonly string[]): string[];

package/dist/search-HKGN7F3O.js

@@ -0,0 +1,13 @@
+import {
+  handleSearch
+} from "./chunk-HAU5INEJ.js";
+import "./chunk-2R7OVYAO.js";
+import "./chunk-TH5XGPD5.js";
+import "./chunk-DJ65EF5G.js";
+import "./chunk-A3QHHUN5.js";
+import "./chunk-4NXH5ZI6.js";
+import "./chunk-VT7DR6NQ.js";
+import "./chunk-3FYJM7LH.js";
+export {
+  handleSearch
+};

package/dist/speculative.d.ts

@@ -0,0 +1,9 @@
+export declare const SPECULATIVE_SUBSETS = 3;
+/** Minimal structural shape the partition reads. */
+export interface Partitionable {
+    metadata: {
+        docidentifier?: string;
+        doc_id?: string;
+    };
+}
+export declare function partitionSubsets<T extends Partitionable>(hits: T[], k?: number): T[][] | null;

package/dist/{tools-EXDADNWX.js → tools-WSWEGJVQ.js}

@@ -4,10 +4,10 @@ import {
   parseToolCall,
   runTool,
   toolNote
-} from "./chunk-ANUHRJ4Z.js";
-import "./chunk-KNCRYAX3.js";
-import "./chunk-U4ERFUOK.js";
-import "./chunk-BOAITSVV.js";
+} from "./chunk-G56TYXNN.js";
+import "./chunk-TH5XGPD5.js";
+import "./chunk-4NXH5ZI6.js";
+import "./chunk-VT7DR6NQ.js";
 import "./chunk-3FYJM7LH.js";
 export {
   TOOLS_REGISTRY,

package/dist/worker_mcp/src/index.js

@@ -1,7 +1,7 @@
 import {
   authenticate
-} from "../../chunk-DTW4UW2E.js";
-import "../../chunk-BOAITSVV.js";
+} from "../../chunk-DJ65EF5G.js";
+import "../../chunk-VT7DR6NQ.js";
 import {
   P,
   setProfile

package/dist/worker_public/src/config.js

@@ -16,7 +16,7 @@ import {
   roleModel,
   sha256Hex,
   today
-} from "../../chunk-BOAITSVV.js";
+} from "../../chunk-VT7DR6NQ.js";
 import "../../chunk-3FYJM7LH.js";
 export {
   DATASETS,