@konneal/engine 0.2.155 → 0.2.158

package/dist/{ask-MRVXR7CY.js → ask-7GRQQEBN.js}

@@ -1,8 +1,8 @@
 import {
   handleAsk
-} from "./chunk-TXF5VMNM.js";
+} from "./chunk-STDXSVLX.js";
 import "./chunk-ECOYO7EX.js";
-import "./chunk-VYEMFSU5.js";
+import "./chunk-PBHMQZKA.js";
 import "./chunk-G2RE4XRV.js";
 import "./chunk-NI4CIK6W.js";
 import "./chunk-W7J5SFXJ.js";

package/dist/{chunk-VYEMFSU5.js → chunk-PBHMQZKA.js}

@@ -271,7 +271,7 @@ function parseNumber(raw) {
   return Number(s.replace(/,(?=\d{3}\b)/g, ""));
 }
 var TOL_SEP_SRC = "\\u00b1|\\bpm\\b|\\+/-";
-var NUM_SRC = "[0-9][0-9 ,.]*[0-9]";
+var NUM_SRC = "[0-9]+(?:[0-9 ,.]*[0-9])?";
 var UNIT_SRC = "%|\\u00b0?[A-Za-z\\u00b5\\u03a9]{0,6}";
 function statedUncertainty(value, tolRaw, tolUnitRaw, valueUnit) {
   const tolNum = parseNumber(tolRaw);
@@ -404,6 +404,42 @@ function evaluate(content, query, opts) {
           const decided = corners.filter((c2) => c2 !== null);
           if (decided.length && decided.some((c2) => c2 !== result)) {
             voidReason = voidReason ?? `the stated uncertainty (${un.stated}) on ${first.split(".").pop()} crosses the decision boundary \u2014 the check evaluates ${result ? "true" : "false"} at the stated value but ${decided.find((c2) => c2 !== result) ? "true" : "false"} at the band edge; the question cannot be decided as stated`;
+          } else if (decided.length === corners.length && un.tol > 0) {
+            const flipsAt = (scale) => {
+              for (const dir of [1, -1]) {
+                const corner = { ...params, [first]: params[first] + dir * scale * un.tol };
+                try {
+                  if (!!parseAndEval(expression, corner) !== result) return true;
+                } catch {
+                  return true;
+                }
+              }
+              return false;
+            };
+            if (flipsAt(20)) {
+              let lo = 1, hi = 20;
+              for (let i = 0; i < 12; i++) {
+                const mid = (lo + hi) / 2;
+                if (flipsAt(mid)) hi = mid;
+                else lo = mid;
+              }
+              const marginTol = hi * un.tol;
+              uncertainty = {
+                ...uncertainty,
+                margin_tol: marginTol,
+                margin_bounded: true,
+                coverage: "k=2",
+                safety_factor: marginTol / (2 * un.tol)
+              };
+            } else {
+              uncertainty = {
+                ...uncertainty,
+                margin_tol: 20 * un.tol,
+                margin_bounded: false,
+                coverage: "k=2",
+                safety_factor: 20 * un.tol / (2 * un.tol)
+              };
+            }
           }
         }
       }

package/dist/{chunk-TXF5VMNM.js → chunk-STDXSVLX.js}

@@ -32,7 +32,7 @@ import {
   tableRetyped,
   toolNote,
   verdictNote
-} from "./chunk-VYEMFSU5.js";
+} from "./chunk-PBHMQZKA.js";
 import {
   NO_CONTEXT,
   STAGE_NAMES,
@@ -2914,7 +2914,25 @@ Answer account questions from these records ONLY: name the record when you use i
       ...h.metadata.table_selection ? { sel: h.metadata.table_selection } : {}
     }));
     const tq = await chargeAnswerTokens(env, bucketId, void 0, messageChars(messages), answer.length, probeTokensUsed);
-    return json({ ...out, context: contextOut, read: readAs(), quota: { ...quota, ...tq, credits: await creditBudget(env, bucketId), credit_estimate: creditPreview } }, 200, { ...corsHeaders(req), "server-timing": serverTiming() });
+    const manifestOut = body?.manifest === true ? { manifest: {
+      index_version: env.INDEX_VERSION,
+      corpus_gen: gen,
+      model,
+      effort: answerEffort(env),
+      fresh,
+      citations: used.length,
+      retrieval: used.map((h, i) => ({
+        rank: i + 1,
+        doc_id: h.metadata.doc_id,
+        clause_anchor: h.metadata.clause_anchor ?? null,
+        score: h.score,
+        rerank_score: h.rerank_score ?? null
+      })),
+      stages: stageTiming,
+      generate_retries: generateRetries,
+      total_ms: Date.now() - tStart
+    } } : {};
+    return json({ ...out, context: contextOut, read: readAs(), ...manifestOut, quota: { ...quota, ...tq, credits: await creditBudget(env, bucketId), credit_estimate: creditPreview } }, 200, { ...corsHeaders(req), "server-timing": serverTiming() });
   };
   if (liveResponse) {
     void pipeline(live, q).catch(() => {

package/dist/enrichPolicy.d.ts

@@ -0,0 +1,29 @@
+/** The model's raw response → the situating context: trimmed, the
+ *  wrapping quotes stripped, capped at 400 chars. Empty string = the
+ *  model had nothing to say (the deterministic-empty class). */
+export declare function normalizeContext(raw: unknown): string;
+export type EnrichOutcome = {
+    action: "serve";
+    context: string;
+    cached: boolean;
+} | {
+    action: "serve-raw";
+} | {
+    action: "fail";
+    error: string;
+};
+/** The per-chunk decision, given the cache read and the generation.
+ *  - a context (cached or generated) serves enriched;
+ *  - an EMPTY generation in the production flow serves the chunk RAW
+ *    (the deterministic-empty class, 2026-10-20: real corpus content —
+ *    B 18's thin terms sections, two R 134-1 clauses — must not vanish
+ *    from the index because its enrichment has nothing to say; a later
+ *    forced re-enrich overwrites the vector in place);
+ *  - an empty generation in mode:context / mode:ab FAILS: those
+ *    builders want the signal, not a side-effecting write. */
+export declare function enrichOutcome(opts: {
+    cached: string | null;
+    generated: string;
+    contextOnly: boolean;
+    abMode: boolean;
+}): EnrichOutcome;

package/dist/{tools-CWPQQTBQ.js → tools-W5HGME7W.js}

@@ -4,7 +4,7 @@ import {
   parseToolCall,
   runTool,
   toolNote
-} from "./chunk-VYEMFSU5.js";
+} from "./chunk-PBHMQZKA.js";
 import "./chunk-NI4CIK6W.js";
 import "./chunk-4NXH5ZI6.js";
 import "./chunk-VT7DR6NQ.js";

package/dist/verdict.d.ts

@@ -4,7 +4,13 @@ export interface MachineCheck {
     values: Record<string, number>;
     result: boolean | null;
     /** present when the question states a tolerance on a symbol this
-     *  check reads: the band's corner evaluations beside the nominal */
+     *  check reads: the band's corner evaluations beside the nominal,
+     *  and — when the band is robust (every corner agrees with the
+     *  nominal) — the GUM-style budget: the tolerance magnitude that
+     *  would flip the decision (bisected up to 20× the stated band;
+     *  margin_bounded false when even 20× does not flip), and the safety
+     *  factor of that margin over the k=2 expanded uncertainty
+     *  (U = 2u, the ≈95% coverage a metrologist reports). */
     uncertainty?: {
         symbol: string;
         stated: string;
@@ -12,6 +18,10 @@ export interface MachineCheck {
         unit?: string;
         nominal: boolean;
         corners: (boolean | null)[];
+        margin_tol?: number;
+        margin_bounded?: boolean;
+        coverage?: "k=2";
+        safety_factor?: number;
     };
 }
 export interface Verdict {

package/dist/worker_public/src/index.js

@@ -4,13 +4,13 @@ import {
   handleMemories,
   scoreFaithfulness,
   scoreJudge
-} from "../../chunk-TXF5VMNM.js";
+} from "../../chunk-STDXSVLX.js";
 import "../../chunk-ECOYO7EX.js";
 import {
   TOOLS_REGISTRY,
   bindModelNode,
   modelGroundingBlock
-} from "../../chunk-VYEMFSU5.js";
+} from "../../chunk-PBHMQZKA.js";
 import {
   handleSearch
 } from "../../chunk-OL7KKGEO.js";
@@ -351,6 +351,18 @@ async function handleGetShared(env, slug) {
 // workers/worker_public/prompts/enrichment.md
 var enrichment_default = "You write a retrieval context for a passage from {{CORPUS_KIND}}. The context is prepended to the passage before embedding so a semantic search can locate the passage even when the query uses different vocabulary than the passage itself.\n\nWrite ONE concise sentence (at most 40 words) that situates the passage: name the publication by its exact {{PUBLISHER_NAME}} identifier (including part or annex when applicable) and what the passage covers \u2014 paraphrasing the topic in words DIFFERENT from the passage's own. Do not copy the passage verbatim, do not add facts that are not derivable from the passage or its header, do not answer or explain the content. Reply with the context sentence only \u2014 no quotes, no preamble.\n";
 
+// workers/worker_public/src/enrichPolicy.ts
+function normalizeContext(raw) {
+  return typeof raw === "string" ? raw.trim().replace(/^["\']|["']$/g, "").slice(0, 400) : "";
+}
+function enrichOutcome(opts) {
+  const fromCache = typeof opts.cached === "string" && opts.cached;
+  if (fromCache) return { action: "serve", context: fromCache, cached: true };
+  if (opts.generated) return { action: "serve", context: opts.generated, cached: false };
+  if (opts.contextOnly || opts.abMode) return { action: "fail", error: "empty enrichment" };
+  return { action: "serve-raw" };
+}
+
 // workers/worker_public/prompts/doc-summary.md
 var doc_summary_default = "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.\n";
 
@@ -408,10 +420,11 @@ ${c.text.slice(0, 1500)}` }
             reasoning_effort: abMode ? effort : "high"
           });
           const raw = typeof res?.response === "string" && res.response.trim() ? res.response : res?.choices?.[0]?.message?.content;
-          context = typeof raw === "string" ? raw.trim().replace(/^["\']|[\"']$/g, "").slice(0, 400) : "";
+          context = normalizeContext(raw);
           if (!context) {
             console.log("enrich raw keys:", Object.keys(res ?? {}), "sample:", JSON.stringify(res).slice(0, 300));
-            if (contextOnly || abMode) return { id: c.id, ok: false, error: "empty enrichment" };
+            const outcome = enrichOutcome({ cached: null, generated: "", contextOnly, abMode });
+            if (outcome.action === "fail") return { id: c.id, ok: false, error: outcome.error };
             const rawText = typeof c.metadata.chunk_text === "string" && c.metadata.chunk_text ? c.metadata.chunk_text : c.text;
             const rawVector = await embed(portModelRunner(env), MODELS.embed, rawText.slice(0, 6e3));
             await env.VECTORIZE.upsert([{ id: c.id, values: rawVector, metadata: { ...c.metadata, chunk_text: rawText } }]);
@@ -1105,13 +1118,13 @@ async function callToolOnce(env, ctx, _req, tier, key, name, args, t0) {
       // stream:false forces the JSON lane (anon defaults to SSE)
       body: JSON.stringify({ ...args, stream: false })
     });
-    const { runTool, TOOLS_REGISTRY: TOOLS_REGISTRY2 } = await import("../../tools-CWPQQTBQ.js");
+    const { runTool, TOOLS_REGISTRY: TOOLS_REGISTRY2 } = await import("../../tools-W5HGME7W.js");
     let payload;
     if (TOOLS_REGISTRY2.some((t) => t.name === name && t.audiences.includes("mcp"))) {
       const r = await runTool({ ...env, DB: env.DB }, { name, args }, "mcp");
       payload = r ? { tool: r.name, query: r.query, result: r.output } : { error: { message: `tool ${name} returned nothing for the given arguments` } };
     } else {
-      const res = name === "ask" ? await (await import("../../ask-MRVXR7CY.js")).handleAsk(env, ctx, inner, tier, key) : await (await import("../../search-GAU5SEZI.js")).handleSearch(env, ctx, inner, tier, key);
+      const res = name === "ask" ? await (await import("../../ask-7GRQQEBN.js")).handleAsk(env, ctx, inner, tier, key) : await (await import("../../search-GAU5SEZI.js")).handleSearch(env, ctx, inner, tier, key);
       payload = await res.json().catch(() => ({ error: { message: "tool transport failed", status: res.status } }));
     }
     telemetry(env, ctx, tier, `mcp:${name}`, null, !payload?.error, JSON.stringify(payload).length, await sha256Hex(`${name}:${JSON.stringify(args ?? {})}`), void 0, void 0, { durationMs: Date.now() - t0 });

package/package.json

@@ -27,7 +27,7 @@
     "playwright": "^1.62.1",
     "yaml": "^2.9.1"
   },
-  "version": "0.2.155",
+  "version": "0.2.158",
   "description": "The Konneal engine: the publisher-agnostic build pipeline and API plane for standards intelligence (retrieval, answer contract, verdicts, evaluation).",
   "license": "BSD-3-Clause",
   "type": "module",

package/workers/worker_public/src/admin.ts

@@ -6,6 +6,7 @@ import { embed } from "./ai";
 import { err, json, corsHeaders, readJson, authenticate } from "./lib/http";
 import type { Env } from "./env";
 import enrichmentPrompt from "../prompts/enrichment.md";
+import { normalizeContext, enrichOutcome } from "./enrichPolicy.ts";
 import docSummaryPrompt from "../prompts/doc-summary.md";
 import familySummaryPrompt from "../prompts/family-summary.md";
 import sectionSummaryPrompt from "../prompts/section-summary.md";
@@ -75,19 +76,14 @@ export async function handleEnrich(env: Env, ctx: ExecutionContext, req: Request
           const raw = typeof res?.response === "string" && res.response.trim()
             ? res.response
             : res?.choices?.[0]?.message?.content;
-          context = typeof raw === "string" ? raw.trim().replace(/^["\']|[\"']$/g, "").slice(0, 400) : "";
+          context = normalizeContext(raw);
           if (!context) {
             console.log("enrich raw keys:", Object.keys(res ?? {}), "sample:", JSON.stringify(res).slice(0, 300));
-            // The deterministic-empty class (observed 2026-10-20 on B 18's
-            // thin terms sections and two R 134-1 clauses — empty under
-            // BOTH the default and the alternate enrichment model): the
-            // chunk's own text is real corpus content, so the production
-            // flow serves it RAW rather than leaving a hole in the index
-            // — embed the original text without the context band, upsert
-            // without the ctx marker; a later forced re-enrich overwrites
-            // the vector in place. mode:"context" and "ab" keep failing:
-            // their builders want the signal, not a side-effecting write.
-            if (contextOnly || abMode) return { id: c.id, ok: false, error: "empty enrichment" };
+            // serve-raw: the chunk's own text is real corpus content —
+            // embed it without the context band, upsert without the ctx
+            // marker; a later forced re-enrich overwrites in place.
+            const outcome = enrichOutcome({ cached: null, generated: "", contextOnly, abMode });
+            if (outcome.action === "fail") return { id: c.id, ok: false, error: outcome.error };
             const rawText = typeof c.metadata.chunk_text === "string" && c.metadata.chunk_text ? c.metadata.chunk_text : c.text;
             const rawVector = await embed(portModelRunner(env), MODELS.embed, rawText.slice(0, 6000));
             await env.VECTORIZE.upsert([{ id: c.id, values: rawVector, metadata: { ...c.metadata, chunk_text: rawText } }]);

package/workers/worker_public/src/ask.ts

@@ -1969,7 +1969,33 @@ let convEntities: Array<{ entity: string; kind: string }> = [];
     ...((h.metadata as any).table_selection ? { sel: (h.metadata as any).table_selection } : {}),
   }));
   const tq = await chargeAnswerTokens(env, bucketId, undefined, messageChars(messages!), answer!.length, probeTokensUsed);
-  return json({ ...out, context: contextOut, read: readAs(), quota: { ...quota, ...tq, credits: await creditBudget(env, bucketId), credit_estimate: creditPreview } }, 200, { ...corsHeaders(req), "server-timing": serverTiming() });
+  // the reproducibility manifest (TODO.sota/14.6, v1): an opt-in, one-object
+  // derivation record — the served snapshot identity, the answer model, the
+  // retrieval anatomy (rank, corpus score, rerank score per used passage),
+  // the stage timings, and the enforcement counters. Everything here is
+  // already computed; the flag only ASSEMBLES it for the harness and for
+  // external auditors. v2 names the CRAG grades per passage.
+  const manifestOut = body?.manifest === true
+    ? { manifest: {
+        index_version: env.INDEX_VERSION,
+        corpus_gen: gen,
+        model,
+        effort: answerEffort(env),
+        fresh,
+        citations: used.length,
+        retrieval: used.map((h: Hit, i: number) => ({
+          rank: i + 1,
+          doc_id: h.metadata.doc_id,
+          clause_anchor: h.metadata.clause_anchor ?? null,
+          score: h.score,
+          rerank_score: h.rerank_score ?? null,
+        })),
+        stages: stageTiming,
+        generate_retries: generateRetries,
+        total_ms: Date.now() - tStart,
+      } }
+    : {};
+  return json({ ...out, context: contextOut, read: readAs(), ...manifestOut, quota: { ...quota, ...tq, credits: await creditBudget(env, bucketId), credit_estimate: creditPreview } }, 200, { ...corsHeaders(req), "server-timing": serverTiming() });
   }
 
   // THE DISPATCH: the streamed lane's response went out at the top — the

package/workers/worker_public/src/enrichPolicy.ts

@@ -0,0 +1,40 @@
+// The enrichment route's decision policy, dependency-free so plain node
+// can load it for unit tests (admin.ts imports .md prompts that only
+// the bundler resolves — the serve-raw policy rode the live
+// verification alone until this extraction). The handler keeps the
+// I/O: cache, generation, embedding, upsert; this module owns the
+// DECISION table — what a chunk's outcome is, given what came back.
+
+/** The model's raw response → the situating context: trimmed, the
+ *  wrapping quotes stripped, capped at 400 chars. Empty string = the
+ *  model had nothing to say (the deterministic-empty class). */
+export function normalizeContext(raw: unknown): string {
+  return typeof raw === "string" ? raw.trim().replace(/^["\']|["']$/g, "").slice(0, 400) : "";
+}
+
+export type EnrichOutcome =
+  | { action: "serve"; context: string; cached: boolean }
+  | { action: "serve-raw" }
+  | { action: "fail"; error: string };
+
+/** The per-chunk decision, given the cache read and the generation.
+ *  - a context (cached or generated) serves enriched;
+ *  - an EMPTY generation in the production flow serves the chunk RAW
+ *    (the deterministic-empty class, 2026-10-20: real corpus content —
+ *    B 18's thin terms sections, two R 134-1 clauses — must not vanish
+ *    from the index because its enrichment has nothing to say; a later
+ *    forced re-enrich overwrites the vector in place);
+ *  - an empty generation in mode:context / mode:ab FAILS: those
+ *    builders want the signal, not a side-effecting write. */
+export function enrichOutcome(opts: {
+  cached: string | null;
+  generated: string;
+  contextOnly: boolean;
+  abMode: boolean;
+}): EnrichOutcome {
+  const fromCache = typeof opts.cached === "string" && opts.cached;
+  if (fromCache) return { action: "serve", context: fromCache, cached: true };
+  if (opts.generated) return { action: "serve", context: opts.generated, cached: false };
+  if (opts.contextOnly || opts.abMode) return { action: "fail", error: "empty enrichment" };
+  return { action: "serve-raw" };
+}

package/workers/worker_public/src/verdict.ts

@@ -20,7 +20,13 @@ export interface MachineCheck {
   values: Record<string, number>;
   result: boolean | null; // null: could not evaluate (missing params)
   /** present when the question states a tolerance on a symbol this
-   *  check reads: the band's corner evaluations beside the nominal */
+   *  check reads: the band's corner evaluations beside the nominal,
+   *  and — when the band is robust (every corner agrees with the
+   *  nominal) — the GUM-style budget: the tolerance magnitude that
+   *  would flip the decision (bisected up to 20× the stated band;
+   *  margin_bounded false when even 20× does not flip), and the safety
+   *  factor of that margin over the k=2 expanded uncertainty
+   *  (U = 2u, the ≈95% coverage a metrologist reports). */
   uncertainty?: {
     symbol: string;
     stated: string;
@@ -28,6 +34,10 @@ export interface MachineCheck {
     unit?: string;
     nominal: boolean;
     corners: (boolean | null)[];
+    margin_tol?: number;
+    margin_bounded?: boolean;
+    coverage?: "k=2";
+    safety_factor?: number;
   };
 }
 
@@ -223,7 +233,10 @@ export interface StatedUncertainty {
 }
 
 const TOL_SEP_SRC = "\\u00b1|\\bpm\\b|\\+/-";
-const NUM_SRC = "[0-9][0-9 ,.]*[0-9]";
+// single digits bind too — a metrologist's "± 5 v" is a first-class
+// tolerance (the two-digit minimum silently dropped every one-digit
+// value and tolerance; found by the GUM-budget tests, 2026-10-20)
+const NUM_SRC = "[0-9]+(?:[0-9 ,.]*[0-9])?";
 const UNIT_SRC = "%|\\u00b0?[A-Za-z\\u00b5\\u03a9]{0,6}";
 
 function statedUncertainty(
@@ -392,6 +405,43 @@ export function evaluate(content: unknown, query: string, opts?: { dimensions?:
               `the check evaluates ${result ? "true" : "false"} at the stated value but ` +
               `${decided.find((c) => c !== result) ? "true" : "false"} at the band edge; the question cannot be decided as stated`
             );
+          } else if (decided.length === corners.length && un.tol > 0) {
+            // the band is robust: report the GUM-style budget — bisect
+            // the tolerance scaling for the flip point (both directions,
+            // up to 20× the stated band), then state the safety factor
+            // over the k=2 expanded uncertainty (U = 2u)
+            const flipsAt = (scale: number): boolean => {
+              for (const dir of [1, -1]) {
+                const corner = { ...params, [first]: params[first]! + dir * scale * un.tol };
+                try {
+                  if (!!parseAndEval(expression, corner) !== result) return true;
+                } catch { return true; }
+              }
+              return false;
+            };
+            if (flipsAt(20)) {
+              let lo = 1, hi = 20;
+              for (let i = 0; i < 12; i++) {
+                const mid = (lo + hi) / 2;
+                if (flipsAt(mid)) hi = mid; else lo = mid;
+              }
+              const marginTol = hi * un.tol;
+              uncertainty = {
+                ...uncertainty!,
+                margin_tol: marginTol,
+                margin_bounded: true,
+                coverage: "k=2",
+                safety_factor: marginTol / (2 * un.tol),
+              };
+            } else {
+              uncertainty = {
+                ...uncertainty!,
+                margin_tol: 20 * un.tol,
+                margin_bounded: false,
+                coverage: "k=2",
+                safety_factor: (20 * un.tol) / (2 * un.tol),
+              };
+            }
           }
         }
       }