@konneal/engine 0.2.39 → 0.2.41

package/dist/{ask-UGTXNPPY.js → ask-3LIAXR4S.js}

@@ -1,7 +1,7 @@
 import {
   handleAsk
-} from "./chunk-OHYAHQQV.js";
-import "./chunk-NBE7BCQ7.js";
+} from "./chunk-JG5GRJCZ.js";
+import "./chunk-N2UOENDY.js";
 import "./chunk-R2V3X6SQ.js";
 import "./chunk-A3QHHUN5.js";
 import "./chunk-4GJGBGJK.js";

package/dist/{chunk-EVCERBIH.js → chunk-AKAFLMIB.js}

@@ -7,7 +7,7 @@ import {
   sessionFrom,
   telemetry,
   understandQuery
-} from "./chunk-NBE7BCQ7.js";
+} from "./chunk-N2UOENDY.js";
 import {
   corsHeaders,
   err,

package/dist/{chunk-OHYAHQQV.js → chunk-JG5GRJCZ.js}

@@ -9,6 +9,7 @@ import {
   contextNote,
   editionNote,
   embed,
+  experimentalSourceLabels,
   generateOnce,
   graphExpand,
   identityNote,
@@ -34,7 +35,7 @@ import {
   syntheticUnderstanding,
   telemetry,
   understandQuery
-} from "./chunk-NBE7BCQ7.js";
+} from "./chunk-N2UOENDY.js";
 import {
   corsHeaders,
   err,
@@ -2448,8 +2449,9 @@ Answer account questions from these records ONLY: name the record when you use i
   );
   await attachFigureImages(env, messages, usedHits, q.query);
   if (userImage) {
+    const wantsMarkings = /\b(mark(ing|ings)?|nameplate|label|inscription|engrav|sticker|plate)\b|certificat/i.test(q.query);
     const last = messages[messages.length - 1];
-    const note = "\n\n(The user attached an image with this question; interpret it directly when answering.)";
+    const note = wantsMarkings ? "\n\n(The user attached a photo with this question. First transcribe every inscription you can actually read in the image \u2014 names, model references, accuracy classes, numeric values with their units, certificate or approval numbers \u2014 quoting them verbatim. Then answer the question from the numbered passages, citing the requirement each interpretation rests on. If an inscription is unreadable, say so; never invent a marking.)" : "\n\n(The user attached an image with this question; interpret it directly when answering.)";
     if (Array.isArray(last.content)) {
       const textPart = last.content.find((p) => p.type === "text");
       if (textPart) textPart.text += note;
@@ -2475,7 +2477,7 @@ Answer account questions from these records ONLY: name the record when you use i
 `));
           send({ type: "read", read: readAs() });
           const sourceQuality = answerQuality(cites.map((c) => c.quality));
-          send({ type: "citations", citations: cites, context_applied: ctxApplied, ...sourceQuality ? { source_quality: sourceQuality, quality_note: qualityNote(sourceQuality) } : {}, ...liveRecords ? { records: liveRecords } : {}, quota });
+          send({ type: "citations", citations: cites, context_applied: ctxApplied, ...sourceQuality ? { source_quality: sourceQuality, confidence_note: qualityNote(sourceQuality), ...sourceQuality === "ocr" ? { experimental_sources: experimentalSourceLabels(cites) } : {} } : {}, ...liveRecords ? { records: liveRecords } : {}, quota });
           let full = "";
           try {
             for await (const tok of sseTokens(stream)) {
@@ -2609,7 +2611,7 @@ Answer account questions from these records ONLY: name the record when you use i
   }
   completionBlocks.push(...await completeFigures(env.DB, answer, [...c2ns.blocks, ...completionBlocks], used));
   const jsonQuality = answerQuality(finalCites.map((c) => c.quality));
-  const out = { answer, citations: finalCites, ...jsonQuality ? { source_quality: jsonQuality, quality_note: qualityNote(jsonQuality) } : {}, 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 out = { answer, citations: finalCites, ...jsonQuality ? { source_quality: jsonQuality, confidence_note: qualityNote(jsonQuality), ...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;
   if (cacheable) {
     const warmVec = await warmEmbed ?? null;

package/dist/{chunk-NBE7BCQ7.js → chunk-N2UOENDY.js}

@@ -77,9 +77,21 @@ function answerQuality(qualities) {
   return worst;
 }
 function qualityNote(q) {
-  if (q === "verified") return "Grounded in verified sources: this publication's requirements ride a machine-checkable data model.";
-  if (q === "curated") return "Grounded in the edited corpus: Metanorma-authored documents, chunked at clause boundaries.";
-  return "Partly grounded in experimental OCR text: verify quotations and table values against the official publication.";
+  if (q === "verified") return "High confidence: this publication's requirements ride a machine-checkable data model.";
+  if (q === "curated") return "Established sources: Metanorma-edited documents, chunked at clause boundaries.";
+  return "WARNING: Partly grounded in experimental data source that was derived from OCR content. Please verify content against official publications.";
+}
+function experimentalSourceLabels(cites, cap = 6) {
+  const labels = [];
+  for (const c of cites) {
+    if (c.quality !== "ocr") continue;
+    const id = String(c.docidentifier || c.doc_id || "source");
+    const anchor = String(c.clause_anchor ?? "");
+    const garbage = /^[0-9a-f]{8}-[0-9a-f]{4}-/i.test(anchor) || anchor.startsWith("_") && anchor.length > 12;
+    labels.push(garbage || !anchor || anchor === "overview" ? id : `${id} \xA7${anchor}`);
+    if (labels.length >= cap) break;
+  }
+  return labels;
 }
 
 // workers/worker_public/src/lexical.ts
@@ -24594,6 +24606,7 @@ export {
   generateOnce,
   answerQuality,
   qualityNote,
+  experimentalSourceLabels,
   ftsMatchQuery,
   refCodec,
   NO_CONTEXT,

package/dist/quality.d.ts

@@ -2,5 +2,16 @@ import type { ChunkMeta } from "../../shared/chunk.ts";
 export type SourceQuality = "verified" | "curated" | "ocr";
 export declare function hitQuality(m: ChunkMeta): SourceQuality;
 export declare function answerQuality(qualities: (SourceQuality | undefined)[]): SourceQuality | null;
-/** The one-line posture the interface renders under the answer. */
+/** The one-line posture the interface renders under the answer. The
+ *  OCR rung is the owner's 2026-09-26 wording, verbatim. */
 export declare function qualityNote(q: SourceQuality): string;
+/** The citations standing on the OCR rung, named for the reader — the
+ *  confidence line says WHICH sources are experimental, not just that
+ *  some are. Producer-UUID anchors say nothing and are dropped; a cap
+ *  keeps the line readable on citation-heavy answers. */
+export declare function experimentalSourceLabels(cites: {
+    quality?: string;
+    docidentifier?: string;
+    doc_id?: string;
+    clause_anchor?: string;
+}[], cap?: number): string[];

package/dist/{search-G23CY5I6.js → search-QZDQVB57.js}

@@ -1,7 +1,7 @@
 import {
   handleSearch
-} from "./chunk-EVCERBIH.js";
-import "./chunk-NBE7BCQ7.js";
+} from "./chunk-AKAFLMIB.js";
+import "./chunk-N2UOENDY.js";
 import "./chunk-R2V3X6SQ.js";
 import "./chunk-A3QHHUN5.js";
 import "./chunk-4GJGBGJK.js";

package/dist/worker_public/src/index.js

@@ -1,6 +1,6 @@
 import {
   handleSearch
-} from "../../chunk-EVCERBIH.js";
+} from "../../chunk-AKAFLMIB.js";
 import {
   bindModelNode,
   checkQuoteAnchors,
@@ -9,7 +9,7 @@ import {
   modelGroundingBlock,
   scoreJudge,
   standardForDocNumber
-} from "../../chunk-OHYAHQQV.js";
+} from "../../chunk-JG5GRJCZ.js";
 import {
   buildMessages,
   citations,
@@ -35,7 +35,7 @@ import {
   sessionFrom,
   telemetry,
   understandQuery
-} from "../../chunk-NBE7BCQ7.js";
+} from "../../chunk-N2UOENDY.js";
 import {
   authenticate,
   corsHeaders,
@@ -999,7 +999,7 @@ async function handleMcp(env, ctx, req, tier, key) {
       // stream:false forces the JSON lane (anon defaults to SSE)
       body: JSON.stringify({ ...args, stream: false })
     });
-    const res = name === "ask" ? await (await import("../../ask-UGTXNPPY.js")).handleAsk(env, ctx, inner, tier, key) : await (await import("../../search-G23CY5I6.js")).handleSearch(env, ctx, inner, tier, key);
+    const res = name === "ask" ? await (await import("../../ask-3LIAXR4S.js")).handleAsk(env, ctx, inner, tier, key) : await (await import("../../search-QZDQVB57.js")).handleSearch(env, ctx, inner, tier, key);
     return res.json().catch(() => ({ error: { message: "tool transport failed", status: res.status } }));
   });
   if (out.ok && "accepted" in out) return new Response(null, { status: 202 });

package/package.json

@@ -27,7 +27,7 @@
     "playwright": "^1.62.1",
     "yaml": "^2.9.1"
   },
-  "version": "0.2.39",
+  "version": "0.2.41",
   "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/ask.ts

@@ -25,7 +25,7 @@ import { evaluate as machineEvaluate, verdictNote } from "./verdict";
 import { evaluateConditionSets, quantitiesIn, type ConditionVerdict } from "./conditions";
 import { evaluateAggregation, type AggregationVerdict } from "./aggregation";
 import { matchLicensedTopic, boundaryNoteText } from "./boundary";
-import { answerQuality, qualityNote } from "./quality";
+import { answerQuality, experimentalSourceLabels, qualityNote } from "./quality";
 import { detectDraftIntent, prepareDraft } from "./drafts";
 import { detectApiCallIntent, prepareApiCall } from "./apicalls";
 import { memoryNote } from "./memories";
@@ -1115,9 +1115,16 @@ async function handleAsk(
   await attachFigureImages(env, messages, usedHits, q.query);
   if (userImage) {
     // the user's own image rides on the question message — retrieval stays
-    // text-driven; the answer model reads the image as question context
+    // text-driven; the answer model reads the image as question context.
+    // A markings/nameplate/certificate photo gets a TRANSCRIPTION note:
+    // read the inscriptions verbatim first, then interpret from the
+    // passages (the marking requirements are corpus content) — never
+    // invent a marking that cannot be read.
+    const wantsMarkings = /\b(mark(ing|ings)?|nameplate|label|inscription|engrav|sticker|plate)\b|certificat/i.test(q.query);
     const last = messages[messages.length - 1];
-    const note = "\n\n(The user attached an image with this question; interpret it directly when answering.)";
+    const note = wantsMarkings
+      ? "\n\n(The user attached a photo with this question. First transcribe every inscription you can actually read in the image — names, model references, accuracy classes, numeric values with their units, certificate or approval numbers — quoting them verbatim. Then answer the question from the numbered passages, citing the requirement each interpretation rests on. If an inscription is unreadable, say so; never invent a marking.)"
+      : "\n\n(The user attached an image with this question; interpret it directly when answering.)";
     if (Array.isArray(last.content)) {
       const textPart = last.content.find((p: any) => p.type === "text");
       if (textPart) textPart.text += note;
@@ -1147,7 +1154,7 @@ async function handleAsk(
           // retrieval, before a single token of the answer
           send({ type: "read", read: readAs() });
           const sourceQuality = answerQuality(cites.map((c: any) => c.quality));
-          send({ type: "citations", citations: cites, context_applied: ctxApplied, ...(sourceQuality ? { source_quality: sourceQuality, quality_note: qualityNote(sourceQuality) } : {}), ...(liveRecords ? { records: liveRecords } : {}), quota, });
+          send({ type: "citations", citations: cites, context_applied: ctxApplied, ...(sourceQuality ? { source_quality: sourceQuality, confidence_note: qualityNote(sourceQuality), ...(sourceQuality === "ocr" ? { experimental_sources: experimentalSourceLabels(cites) } : {}) } : {}), ...(liveRecords ? { records: liveRecords } : {}), quota, });
           let full = "";
           try {
             for await (const tok of sseTokens(stream)) {
@@ -1342,7 +1349,7 @@ async function handleAsk(
   completionBlocks.push(...(await completeFigures(env.DB, answer, [...c2ns.blocks, ...completionBlocks], used)));
 
   const jsonQuality = answerQuality(finalCites.map((c: any) => c.quality));
-  const out = { answer, citations: finalCites, ...(jsonQuality ? { source_quality: jsonQuality, quality_note: qualityNote(jsonQuality) } : {}), 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 out = { answer, citations: finalCites, ...(jsonQuality ? { source_quality: jsonQuality, confidence_note: qualityNote(jsonQuality), ...(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;
   if (cacheable) {
     const warmVec = (await warmEmbed) ?? null;

package/workers/worker_public/src/quality.ts

@@ -40,9 +40,30 @@ export function answerQuality(qualities: (SourceQuality | undefined)[]): SourceQ
   return worst;
 }
 
-/** The one-line posture the interface renders under the answer. */
+/** The one-line posture the interface renders under the answer. The
+ *  OCR rung is the owner's 2026-09-26 wording, verbatim. */
 export function qualityNote(q: SourceQuality): string {
-  if (q === "verified") return "Grounded in verified sources: this publication's requirements ride a machine-checkable data model.";
-  if (q === "curated") return "Grounded in the edited corpus: Metanorma-authored documents, chunked at clause boundaries.";
-  return "Partly grounded in experimental OCR text: verify quotations and table values against the official publication.";
+  if (q === "verified") return "High confidence: this publication's requirements ride a machine-checkable data model.";
+  if (q === "curated") return "Established sources: Metanorma-edited documents, chunked at clause boundaries.";
+  return "WARNING: Partly grounded in experimental data source that was derived from OCR content. Please verify content against official publications.";
+}
+
+/** The citations standing on the OCR rung, named for the reader — the
+ *  confidence line says WHICH sources are experimental, not just that
+ *  some are. Producer-UUID anchors say nothing and are dropped; a cap
+ *  keeps the line readable on citation-heavy answers. */
+export function experimentalSourceLabels(
+  cites: { quality?: string; docidentifier?: string; doc_id?: string; clause_anchor?: string }[],
+  cap = 6,
+): string[] {
+  const labels: string[] = [];
+  for (const c of cites) {
+    if (c.quality !== "ocr") continue;
+    const id = String(c.docidentifier || c.doc_id || "source");
+    const anchor = String(c.clause_anchor ?? "");
+    const garbage = /^[0-9a-f]{8}-[0-9a-f]{4}-/i.test(anchor) || (anchor.startsWith("_") && anchor.length > 12);
+    labels.push(garbage || !anchor || anchor === "overview" ? id : `${id} §${anchor}`);
+    if (labels.length >= cap) break;
+  }
+  return labels;
 }