@konneal/engine 0.1.0 → 0.1.2

package/workers/worker_public/src/codecs.ts

@@ -0,0 +1,77 @@
+// The serving-side identifier codec registry — the TS mirror of
+// ingest/codecs.py. Each publisher has an identifier grammar; the
+// profile declares the codec id and every publisher-specific parse
+// (how a question names a document, how a label reads, how graph node
+// ids are shaped) lives in the codec implementation, never in domain
+// code. The plain-slug codec is the generic floor: no grammar, no
+// doc-number steering — labels are the identifier as given.
+import { P } from "./profile.ts";
+
+export interface DocScope {
+  doc_number: string;
+  edition?: string;
+  label: string;
+}
+
+export interface RefCodec {
+  /** The explicit form: "R 60-1:2021", "urn:…" — null when unparseable. */
+  parse(doc: string, edition?: string): DocScope | null;
+  /** The gap-tolerant scan over question text — null when nothing names a document. */
+  scanQuestion(query: string): DocScope | null;
+  /** The graph's node id → document number — null for other shapes. */
+  graphDocNumber(nodeId: string): string | null;
+  /** A docidentifier's family key ("R-60") for edition steering — null when not of the grammar. */
+  familyOf(docidentifier: string): string | null;
+}
+
+/** OIML's grammar: type letter (R/D/B/G/E) + 1–3 digits, optional part,
+ *  optional edition year; the URN provenance form; part numbers are
+ *  significant (R 60-1), the edition is never part of the number. */
+export const oimlPubid: RefCodec = {
+  parse(doc, edition) {
+    const m =
+      doc.match(/^urn:oiml:pub:([rdbge]):(\d{1,3})(?:-[0-9A-Za-z]+)?(?::(\d{4}))?$/i) ??
+      doc.match(/^(?:OIML\s+)?([RDBGE])\s*(\d{1,3})(?:-[0-9A-Za-z]+)?(?::(\d{4}))?$/i);
+    if (!m) return null;
+    const type = m[1].toUpperCase();
+    const ed = edition ?? m[3] ?? undefined;
+    return { doc_number: m[2], ...(ed ? { edition: ed } : {}), label: `OIML ${type} ${m[2]}${ed ? `:${ed}` : ""}` };
+  },
+  scanQuestion(query) {
+    const re = /\b(OIML\s+)?([RDBGE])(\s*)0*(\d{1,3})(?:\s*[-–]\s*\d+)?(?:\s*:\s*(\d{4}))?/gi;
+    for (const m of query.matchAll(re)) {
+      const [, oimlPrefix, letter, gap, digits, edition] = m;
+      // a glued single digit is a class/designation ("E2 weights"), never a naming
+      if (digits!.length === 1 && !oimlPrefix && !gap) continue;
+      const num = String(Number(digits));
+      const type = letter!.toUpperCase();
+      return { doc_number: num, ...(edition ? { edition } : {}), label: `OIML ${type} ${num}${edition ? `:${edition}` : ""}` };
+    }
+    return null;
+  },
+  graphDocNumber(nodeId) {
+    const m = nodeId.match(/^doc:OIML-[A-Z]-(\d+)-/);
+    return m ? m[1] : null;
+  },
+  familyOf(di) {
+    const m = /^(?:OIML\s+)?([A-Z])\s?(\d{1,3})(?:[-–]([0-9A-Za-z]+))?/.exec(di);
+    return m ? `${m[1]}-${m[2]}` : null;
+  },
+};
+
+/** The generic floor: the identifier is whatever string it is. */
+export const plainSlug: RefCodec = {
+  parse: () => null,
+  scanQuestion: () => null,
+  graphDocNumber: () => null,
+  familyOf: () => null,
+};
+
+const REGISTRY: Record<string, RefCodec> = {
+  "oiml-pubid": oimlPubid,
+  "plain-slug": plainSlug,
+};
+
+export function refCodec(): RefCodec {
+  return REGISTRY[P().publisher.codec] ?? plainSlug;
+}

package/workers/worker_public/src/config.ts

@@ -212,7 +212,7 @@ export function datasetsFor(session: unknown): unknown[] {
     description: d.description,
     enabled: datasetAllowed(d, session),
     ...(d.session
-      ? { requires: `the ${d.permission ?? "ai-preview"} permission (id.oimlsmart.org)`, authenticated: !!session }
+      ? { requires: `the ${d.permission ?? "ai-preview"} permission (${P().publisher.identity.issuer.replace(/^https?:\/\//, "")})`, authenticated: !!session }
       : {}),
   }));
 }

package/workers/worker_public/src/context.ts

@@ -73,28 +73,16 @@ export function parseContext(body: any): DeclaredContext | null {
   return { kind: c.kind, label, ...(route ? { route } : {}), ...(doc ? { doc } : {}), ...(edition ? { edition } : {}) };
 }
 
-export interface DocScope {
-  /** the Vectorize doc_number filter value (the publication FAMILY —
-   *  an entity's clause provenance spans parts: R 60-1 requirements,
-   *  R 60-2 tests) */
-  doc_number: string;
-  edition?: string;
-  /** the canonical label form for the echo + the prompt note */
-  label: string;
-}
+import { refCodec, type DocScope } from "./codecs.ts";
+import { P } from "./profile.ts";
+export type { DocScope };
 
 /** Parse the two reference forms the estate speaks: the URN the SMART
  *  models carry as clause provenance (urn:oiml:pub:r:60-1:2021) and the
  *  plain docidentifier (OIML R 60-1:2021 / R 60). Part designations
  *  parse but do not narrow the scope (the family IS the scope). */
 export function parseDocRef(doc: string, edition?: string): DocScope | null {
-  const m =
-    doc.match(/^urn:oiml:pub:([rdbge]):(\d{1,3})(?:-[0-9A-Za-z]+)?(?::(\d{4}))?$/i) ??
-    doc.match(/^(?:OIML\s+)?([RDBGE])\s*(\d{1,3})(?:-[0-9A-Za-z]+)?(?::(\d{4}))?$/i);
-  if (!m) return null;
-  const type = m[1].toUpperCase();
-  const ed = edition ?? m[3] ?? undefined;
-  return { doc_number: m[2], ...(ed ? { edition: ed } : {}), label: `OIML ${type} ${m[2]}${ed ? `:${ed}` : ""}` };
+  return refCodec().parse(doc, edition);
 }
 
 /** Read the FIRST publication the question's own text names, in the
@@ -108,15 +96,7 @@ export function parseDocRef(doc: string, edition?: string): DocScope | null {
  *  A glued single digit is a class/designation ("E2 weights"), never a
  *  naming; part designations parse but do not narrow the family. */
 export function namedDocumentIn(query: string): DocScope | null {
-  const re = /\b(OIML\s+)?([RDBGE])(\s*)0*(\d{1,3})(?:\s*[-–]\s*\d+)?(?:\s*:\s*(\d{4}))?/gi;
-  for (const m of query.matchAll(re)) {
-    const [, oimlPrefix, letter, gap, digits, edition] = m;
-    if (digits!.length === 1 && !oimlPrefix && !gap) continue;
-    const num = String(Number(digits));
-    const type = letter!.toUpperCase();
-    return { doc_number: num, ...(edition ? { edition } : {}), label: `OIML ${type} ${num}${edition ? `:${edition}` : ""}` };
-  }
-  return null;
+  return refCodec().scanQuestion(query);
 }
 
 /** Resolve the declared document against the publications registry: the
@@ -202,7 +182,7 @@ export function contextNote(declared: DeclaredContext | null, scope: DocScope |
     return undefined;
   }
   if (declared.kind === "page") {
-    return `Context note: the user is viewing ${declared.label || "a page"}${declared.route ? ` (${declared.route})` : ""} in the OIML SMART platform. The passages come from the general corpus; frame procedural guidance for that page when relevant.`;
+    return `Context note: the user is viewing ${declared.label || "a page"}${declared.route ? ` (${declared.route})` : ""} in the ${P().publisher.product_name} platform. The passages come from the general corpus; frame procedural guidance for that page when relevant.`;
   }
   if (declared.kind === "entity") {
     return scope

package/workers/worker_public/src/graph.ts

@@ -3,15 +3,14 @@
 // the graph says are relevant. Mirrors graph.py's node-id format
 // (doc:OIML-R-60-1-2017, concept:<id>).
 import type { Env } from "./env";
+import { refCodec } from "./codecs.ts";
 
 /** Graph expansion (G8 query lane): map understanding's term / named
  *  document onto the D1 projection (graph_nodes / graph_edges) and return
  *  the doc_numbers the graph says are relevant. Mirrors graph.py's node-id
  *  format (doc:OIML-R-60-1-2017, concept:<id>). */
 function docNumberOf(nodeId: string): string | null {
-  // doc:OIML-R-60-1-2017 → "60" | doc:OIML-B-18-2025 → "18"
-  const m = nodeId.match(/^doc:OIML-[A-Z]-(\d+)-/);
-  return m ? m[1] : null;
+  return refCodec().graphDocNumber(nodeId);
 }
 
 export async function graphExpand(env: Env, u: { term?: string | null; defined_terms?: string[]; docidentifier?: string | null } | null): Promise<string[] | undefined> {

package/workers/worker_public/src/index.ts

@@ -19,7 +19,7 @@ export type { Env };
 import { json, err, corsHeaders, withCors, readJson, authenticate, type ApiKey } from "./lib/http";
 
 import { handleSearch } from "./search";
-import { handleEnrich, handleSectionUnit, handleCaption, handleVectors, handleJudge, handleCreateKey, handleListKeys } from "./admin";
+import { handleEnrich, handleSectionUnit, handleCaption, handleVectors, handleJudge, handleCreateKey, handleListKeys, handleRevokeKey } from "./admin";
 import { handleResearch } from "./research";
 import { handleAsk } from "./ask";
 
@@ -31,6 +31,7 @@ import { handleAsk } from "./ask";
 // handler's, derived from the path. Mirrored in docs/spec-api.md.
 
 import { matchRoute, type RouteContext, type Route } from "./lib/router";
+import { P } from "./profile.ts";
 
 async function serveIndexPage(c: RouteContext): Promise<Response> {
   // HTML pages are served through the worker with must-revalidate so a
@@ -112,7 +113,7 @@ async function tierFor(c: RouteContext): Promise<{ tier: "anon" | "key" | "membe
   let key: ApiKey | null = null;
   if (isApi) {
     key = await authenticate(c.env, c.req);
-    if (!key) return err(401, "unauthorized", "Provide a valid API key: Authorization: Bearer oiml_...");
+    if (!key) return err(401, "unauthorized", `Provide a valid API key: Authorization: Bearer ${P().publisher.id}_...`);
   }
   let tier: "anon" | "key" | "member" = isApi ? "key" : "anon";
   if (!isApi && c.env.SESSION_SECRET && (await sessionFrom(c.req, c.env as any))) tier = "member";
@@ -215,7 +216,7 @@ async function verifyRoute(c: RouteContext): Promise<Response> {
     const checks = [
       { name: "quote_anchors", deterministic: true, pass: anchors.violations.length === 0, detail: `${anchors.violations.length} of ${anchors.total} quoted spans absent from the retrieved passages` },
       { name: "unit_references", deterministic: true, pass: refs.length === validRefs.length, detail: refs.length ? `${validRefs.length}/${refs.length} unit references resolve to served units` : "no unit references" },
-      { name: "citations_present", deterministic: true, pass: /\[[^\]]*(OIML|ISO)[^\]]*\]/.test(answer), detail: "normative claims should carry a passage citation" },
+      { name: "citations_present", deterministic: true, pass: new RegExp(`\\[[^\\]]*(${P().publisher.name})[^\\]]*\\]`).test(answer), detail: "normative claims should carry a passage citation" },
     ];
     const faith = await scoreFaithfulness(env.AI, roleModel(env, "grader"), answer, retrieved.hits.map((h: Hit) => h.text));
     return json({
@@ -421,6 +422,7 @@ export const ROUTES: Route[] = [
   { method: "POST", pattern: "/v1/admin/judge", handler: (c) => handleJudge(c.env, c.req) },
   { method: "POST", pattern: "/v1/admin/keys", handler: (c) => handleCreateKey(c.env, c.req) },
   { method: "GET", pattern: "/v1/admin/keys", handler: (c) => handleListKeys(c.env, c.req) },
+  { method: "DELETE", pattern: "/v1/admin/keys/:id", handler: (c) => handleRevokeKey(c.env, c.req, c.params.id) },
 ];
 
 export default {

package/workers/worker_public/src/lib/http.ts

@@ -2,6 +2,7 @@
 // the vocabulary every route handler speaks (TODO.impl/23).
 import { LIMITS, sha256Hex } from "../config";
 import type { Env } from "../env";
+import { isAllowedBubbleOrigin } from "../bubble.ts";
 
 export const json = (body: unknown, status = 200, extra: Record<string, string> = {}) =>
   new Response(JSON.stringify(body), {
@@ -14,14 +15,9 @@ export const err = (status: number, code: string, message: string) =>
 
 export function corsHeaders(req: Request): Record<string, string> {
   const origin = req.headers.get("origin") ?? "";
-  const allowed =
-    origin === "https://oimlsmart.org" ||
-    /^https:\/\/[a-z0-9-]+\.oimlsmart\.org$/.test(origin) ||
-    // the local dev posture: the platform and the minisites develop on
-    // localhost ports against the live service (the bubble bridge admits
-    // the same class; anon quota is per-IP, member auth needs the token)
-    /^http:\/\/localhost(:\d{1,5})?$/.test(origin) ||
-    /^http:\/\/127\.0\.0\.1(:\d{1,5})?$/.test(origin);
+  // the publisher origin family (profile-declared) + the local dev
+  // posture — the same rule the bubble bridge applies, from one place
+  const allowed = isAllowedBubbleOrigin(origin);
   return allowed
     ? {
         "access-control-allow-origin": origin,
@@ -52,7 +48,7 @@ export interface ApiKey {
   day_limit: number;
 }
 
-export async function authenticate(env: Env, req: Request): Promise<ApiKey | null> {
+export async function authenticate(env: Pick<Env, "DB">, req: Request): Promise<ApiKey | null> {
   const auth = req.headers.get("authorization") ?? "";
   const m = auth.match(/^Bearer\s+(.+)$/i);
   if (!m) return null;

package/workers/worker_public/src/livedata.ts

@@ -25,6 +25,7 @@
  *  dependency-free (context.ts's pattern: the unit tests load it on
  *  plain node's type stripping, which never resolves extensionless
  *  relative imports). */
+import { P } from "./profile.ts";
 async function sha256Hex(s: string): Promise<string> {
   const digest = await crypto.subtle.digest("SHA-256", new TextEncoder().encode(s));
   return [...new Uint8Array(digest)].map((b) => b.toString(16).padStart(2, "0")).join("");
@@ -174,7 +175,7 @@ export type LiveRead =
 /** The platform page for a record — the role family decides the console
  *  (the same page the user's own browser would open). */
 function recordUrl(cfg: LiveDataConfig, roleFamily: string, store: string, row: any): string {
-  const std = typeof row.standard_id === "string" ? row.standard_id.replace(/^oiml-/, "") : null;
+  const std = typeof row.standard_id === "string" ? row.standard_id.replace(new RegExp(`^${P().publisher.id}-`, "i"), "") : null;
   if (store === "certificates") {
     if (roleFamily === "applicant") return `${cfg.platformApi}/app/portal/certificates/${row.id}`;
     if (std) return `${cfg.platformApi}/app/standards/${std}/certificates/${row.id}`;
@@ -245,7 +246,7 @@ export async function readMyAccount(_env: any, cfg: LiveDataConfig, token: strin
       records.push({
         store: "applications",
         id: String(row.id),
-        label: `Application ${row.application_number ?? row.id}${row.standard_id ? ` — ${String(row.standard_id).replace(/^oiml-/, "").toUpperCase().replace(/^R(\d)/, "R $1")}` : ""}`,
+        label: `Application ${row.application_number ?? row.id}${row.standard_id ? ` — ${String(row.standard_id).replace(new RegExp(`^${P().publisher.id}-`, "i"), "").toUpperCase().replace(/^R(\d)/, "R $1")}` : ""}`,
         url: recordUrl(cfg, family, "applications", row),
         status: row.status,
         date: row.submitted_date ?? row.date_of_application,

package/workers/worker_public/src/modelplane.ts

@@ -29,6 +29,7 @@
  *  slug segments (the packages' identifier shapes — /req/<class>/<id>,
  *  /conf/<class>/<id>, /term/<id>, /constraint/<id>, /characteristic/<id>,
  *  /state-machine/<id>, /dimension/<id>). */
+import { P } from "./profile.ts";
 const NODE_RE = /(?:^|[\s("'`])\/(req|conf|term|constraint|characteristic|state-machine|dimension)\/([a-z0-9][a-z0-9_-]*(?:\/[a-z0-9][a-z0-9_-]*)?)(?=[\s)"'`,;:.]|$)/i;
 
 /** The first model-node id a text names (the declared chip label first,
@@ -46,7 +47,9 @@ export function modelNodeRefIn(text: string | undefined | null): string | null {
  *  carry a plane; anything else resolves null (honest: no model to bind). */
 export function standardForDocNumber(docNumber: string | undefined): string | null {
   if (!docNumber) return null;
-  return ["60", "91", "129", "144"].includes(docNumber) ? `oiml-r${docNumber}` : null;
+  const models = P().sources?.models;
+  if (!models?.standards?.length || !models?.standard_prefix) return null;
+  return (models.standards as string[]).includes(docNumber) ? `${models.standard_prefix}${docNumber}` : null;
 }
 
 export interface BoundModelNode {
@@ -139,7 +142,7 @@ export function modelGroundingBlock(node: BoundModelNode): string {
   const c = node.content ?? {};
   const lines: string[] = [];
   lines.push(
-    `Model grounding — the OIML SMART model plane (the platform's machine-readable Recommendation model, derived from the Primmel packages, the models' single source of truth):`,
+    P().prompts.vars.model_grounding_intro ?? "Model grounding — the model plane's own statement:",
   );
   lines.push(`Node: ${node.node_id} (${node.kind.replace(/_/g, " ")}) — ${node.name} [${node.standard}]`);
   if (node.clause) lines.push(`Provenance: ${node.clause.urn}`);
@@ -183,7 +186,11 @@ export function modelGroundingBlock(node: BoundModelNode): string {
 export function modelCitation(node: BoundModelNode) {
   return {
     doc_id: `model:${node.standard}`,
-    docidentifier: `OIML SMART model (${node.standard.replace(/^oiml-r/, "R ")})`,
+    docidentifier: `${P().publisher.name} SMART model (${(() => {
+      const prefix = P().sources?.models?.standard_prefix ?? "";
+      const letter = prefix.replace(/^.*-/, "").toUpperCase();
+      return String(node.standard).replace(new RegExp(`^${prefix}`, "i"), `${letter} `);
+    })()})`,
     edition: "",
     language: "en",
     clause_anchor: node.clause?.ref || "model",
@@ -209,5 +216,6 @@ export function modelEcho(node: BoundModelNode) {
 
 /** The per-corpus guidance note (config.ts's DATASETS pattern — every
  *  retrieved model-plane chunk carries it, chip or no chip). */
-export const MODEL_CORPUS_NOTE =
-  "Some passages are the OIML SMART model plane (labeled OIML SMART model) — the platform's machine-readable Recommendation models derived from the Primmel packages. Treat their machine limits, applicability rules and acceptance criteria as the model's own statement of them (quote machine limits verbatim); where a model passage and a prose passage disagree, say so explicitly and cite both.";
+export function modelCorpusNote(): string {
+  return P().prompts.vars.model_passage_note ?? "";
+}

package/workers/worker_public/src/pipeline.ts

@@ -12,9 +12,24 @@ import { tableContext } from "./tablecontext";
 // here so the existing import surface keeps working
 export { refusalAnswer } from "./refusal";
 
+/** The interpolation source for every prompt: the profile's declared
+ *  vars plus the derived publisher tokens. Call sites never build
+ *  their own var map. */
+export function promptVars(extra: Record<string, string> = {}): Record<string, string> {
+  // profile vars are snake_case keys; template tokens are UPPER_SNAKE —
+  // the map is built here, once, so no call site can spread the raw
+  // keys again (the pre-varianlization bug: the identity token sat
+  // unmatched and rendered empty in the system prompt)
+  const out: Record<string, string> = { PUBLISHER_NAME: P().publisher.name };
+  for (const [k, v] of Object.entries(P().prompts?.vars ?? {})) {
+    if (typeof v === "string") out[k.toUpperCase()] = v;
+  }
+  return { ...out, ...extra };
+}
+
 /** Fill {{TOKEN}} placeholders in a prompt data file. Unknown/empty tokens
  *  resolve to "" so optional lines vanish cleanly. */
-function fill(template: string, vars: Record<string, string>): string {
+export function fill(template: string, vars: Record<string, string>): string {
   return template.replace(/\{\{(\w+)\}\}/g, (_m, k: string) => (k in vars ? vars[k] : ""));
 }
 import { QueryFilters, toVectorizeFilter } from "./selfquery";
@@ -168,7 +183,7 @@ export function identityNote(member: boolean): string {
   const upsell = locked.length
     ? `Signed-in members additionally search: ${locked.map((d) => `${d.label} (${d.description})`).join("; ")}.`
     : "";
-  return fill(conversationalPromptText, { CORPORA: corpora, UPSELL: upsell })
+  return fill(conversationalPromptText, promptVars({ CORPORA: corpora, UPSELL: upsell }))
     .split("\n")
     .filter((l) => l.trim())
     .join("\n");
@@ -260,14 +275,13 @@ export function buildMessages(
 
   // the prompt itself is data (prompts/system.md); one rule per line,
   // joined with spaces exactly as the original array form
-  const system = fill(systemPromptText, {
-    ...P().prompts.vars,
+  const system = fill(systemPromptText, promptVars({
     HISTORY_CONTEXT: history.length
       ? " Earlier turns of this conversation are provided for context — answer the LATEST question, treating the passages below as the source of truth for facts and citations."
       : "",
     CORPUS_NOTES: corpusNotes,
     LANG_CLAUSE: lang ? ` (explicitly requested: ${lang})` : "",
-  })
+  }))
     .split("\n")
     .map((l) => l.trim())
     .filter(Boolean)
@@ -348,12 +362,13 @@ export function buildMessages(
   };
 }
 
-function oimlPublicationUrl(meta: ChunkMeta): string | undefined {
-  if (!meta.doctype || !meta.doc_number) return undefined;
-  const typeMap: Record<string, string> = { R: "r", D: "d", B: "b", G: "g", E: "e" };
-  const t = typeMap[meta.doctype];
-  if (!t) return undefined;
-  return `https://www.oiml.org/en/publications/${t}${meta.doc_number}`;
+/** The publisher's catalog page for a publication, from the profile's
+ *  URL template ({type} = lowercase doctype, then the number); no
+ *  template, no catalog link. */
+function publicationUrl(meta: ChunkMeta): string | undefined {
+  const tpl = P().publisher.catalog_url_template;
+  if (!tpl || !meta.doctype || !meta.doc_number) return undefined;
+  return tpl.replace("{type}", meta.doctype.toLowerCase()) + meta.doc_number;
 }
 
 export function citations(hits: Hit[]) {
@@ -368,8 +383,8 @@ export function citations(hits: Hit[]) {
       clause_title: h.metadata.clause_title,
       status: h.metadata.status ?? "unknown",
       superseded_by: h.metadata.superseded_by || undefined,
-      corpus: h.metadata.corpus || "oiml",
-      url: oimlPublicationUrl(h.metadata),
+      corpus: h.metadata.corpus || P().publisher.id,
+      url: publicationUrl(h.metadata),
       snippet: h.text.slice(0, 400),
       score: h.rerank_score ?? h.score,
     }))

package/workers/worker_public/src/profile.gen.ts

@@ -8,12 +8,21 @@ export const PROFILE = {
     "product_name": "Fixture Answers",
     "description": "A minimal publisher profile exercising every declared surface: an open dataset, a permission-gated dataset, production and lane corpora, prompt vars and retrieval vocabulary.",
     "domains": {
-      "public": "fixture.example.org"
+      "public": "fixture.example.org",
+      "origin_suffix": "fixture.example.org"
     },
     "identity": {
       "issuer": "https://id.fixture.example.org"
     },
-    "codec": "plain-slug"
+    "codec": "plain-slug",
+    "session_cookie": "fixture-session",
+    "references": {
+      "label_prefix": ""
+    },
+    "features": {
+      "drafts": false,
+      "model_plane": false
+    }
   },
   "datasets": [
     {
@@ -90,12 +99,22 @@ export const PROFILE = {
     ]
   },
   "retrieval": {
-    "process_expansion": " fixture certification system framework application evaluation"
+    "process_expansion": " fixture certification system framework application evaluation",
+    "process_note": "Retrieval note: these passages come from the fixture certification system documents because they govern application procedures for fixture publications."
   },
   "prompts": {
     "vars": {
       "assistant_identity": "the fixture assistant — a public service answering questions about the fixture publisher's documents",
-      "refusal_sentence": "I don't have information on this in the indexed fixture documents."
+      "refusal_sentence": "I don't have information on this in the indexed fixture documents.",
+      "account_note_source": "the user's own fixture account",
+      "corpus_kind": "a fixture corpus publication",
+      "corpus_kind_plural": "fixture corpus publications",
+      "cite_example": "FIXTURE 1:2024 §2.1",
+      "cite_quote_example": "FIXTURE 1:2024 §2.1: \"the limit shall not exceed one interval\"",
+      "parts_example": "FIXTURE 1-1, FIXTURE 1-A",
+      "docid_example": "FIXTURE 1-2",
+      "spelling_examples": "\"f1\", \"FIXTURE 1\"",
+      "process_vocab": "the fixture certification system framework"
     }
   }
 } as const;

package/workers/worker_public/src/profile2.gen.ts

@@ -0,0 +1,122 @@
+// GENERATED from profile2/*.yaml — regenerate: node scripts/gen_profile.mjs
+// (never edit; the drift test compares this file to the sources)
+export const PROFILE = {
+  "publisher": {
+    "id": "atlas",
+    "name": "Atlas",
+    "full_name": "The Atlas Standards Institute",
+    "product_name": "Atlas Answers",
+    "description": "The second fixture publisher of the reference matrix: a technical-standards institute issuing specifications and errata, with a working-group corpus behind a session permission.",
+    "domains": {
+      "public": "answers.atlas.example",
+      "origin_suffix": "atlas.example"
+    },
+    "identity": {
+      "issuer": "https://identity.atlas.example"
+    },
+    "codec": "plain-slug",
+    "session_cookie": "atlas-session",
+    "references": {
+      "label_prefix": "ATLAS"
+    },
+    "production": [
+      "spec",
+      "errata",
+      "model"
+    ],
+    "lanes": {
+      "review": [
+        "review"
+      ],
+      "glossary": [
+        "glossary"
+      ]
+    },
+    "features": {
+      "drafts": false,
+      "model_plane": false
+    }
+  },
+  "datasets": [
+    {
+      "id": "spec",
+      "label": "Atlas Specifications",
+      "description": "The institute's published specifications and errata",
+      "corpora": [
+        "spec",
+        "errata"
+      ],
+      "note": "Some passages come from the Atlas errata corpus — cite them the same way as every other passage."
+    },
+    {
+      "id": "wg",
+      "label": "Working-group corpus",
+      "description": "An access-restricted corpus proving the permission gate",
+      "session": true,
+      "permission": "committee",
+      "corpora": [
+        "wg"
+      ]
+    }
+  ],
+  "corpora": {
+    "production": [
+      "spec",
+      "errata",
+      "model"
+    ],
+    "lanes": {
+      "review": [
+        "review"
+      ],
+      "glossary": [
+        "glossary"
+      ]
+    },
+    "corpora": {
+      "spec": {
+        "repo": "fixtures/matrix2",
+        "note": "the second fixture corpus"
+      }
+    },
+    "bibliography": {},
+    "terminology": {},
+    "models": {}
+  },
+  "sources": {
+    "corpora": {},
+    "bibliography": {},
+    "terminology": {},
+    "models": {}
+  },
+  "ui": {
+    "suggestions": [
+      "What does ATLAS 12 specify?",
+      "Which errata are open?"
+    ],
+    "models_disclosure": [
+      {
+        "role": "Answers",
+        "model": "atlas-answer-model"
+      }
+    ],
+    "smoke": [
+      {
+        "label": "sanity",
+        "query": "What does ATLAS 12 specify?",
+        "expect": "ATLAS"
+      }
+    ]
+  },
+  "retrieval": {
+    "process_expansion": " atlas review procedure errata committee specification",
+    "process_note": "Retrieval note: these passages come from the Atlas review-procedure documents because they govern the institute's publication process."
+  },
+  "prompts": {
+    "vars": {
+      "assistant_identity": "the Atlas Answers assistant — a public service answering questions about the Atlas Standards Institute's specifications",
+      "refusal_sentence": "I don't have information on this in the indexed Atlas specifications.",
+      "account_note_source": "the user's own Atlas Answers account"
+    }
+  }
+} as const;

package/workers/worker_public/src/refusal.ts

@@ -8,27 +8,24 @@ export function refusalAnswer(): string {
 
 // the model occasionally paraphrases the refusal sentence ("...information
 // on how to make lasagna in the indexed..."); the API contract is the
-// exact canonical sentence — normalize variants, keep the redirect tail
-const REFUSAL_VARIANT = /^\s*I don[’']?t have information on .{1,120}? in the indexed OIML(?: \w+){0,2} (?:publications|passages|documents|corpus)\.?/i;
-
-// Upstream wording drift (rag#88, the golden refusal pin recalibrated
-// 2026-09-01): the model now also refuses with no pinned sentence at all —
-// a "can't help" redirect naming the OIML publications, or a "no real
-// answer to give" paraphrase. The runtime and the harness pins
-// (scripts/eval.mjs, tests/e2e.mjs, tests/eval-suite.mjs) carry the SAME
-// family: every shape the tests accept canonicalizes to the pinned
-// sentence here, and nothing else does.
-const REFUSAL_DRIFT: RegExp[] = [
-  /\b(can'?t|cannot|couldn'?t|unable)\b[^.]{0,120}?\b(indexed )?OIML(?: \w+){0,2} publications\b/i,
-  /\bno real answer to give\b[^.]{0,120}?\bOIML\b/i,
-  /\b(?:falls|well) outside\b[^.]{0,120}?\b(?:what I can answer|my scope|the scope of)\b/i,
-  /\boutside (?:of )?what (?:I|this service) can answer\b/i,
-  // "I can't answer that — weather forecasting is outside my scope":
-  // requires the refusal verb, so a scope DISCUSSION inside a real answer
-  // ("this exemption is outside the scope of R 60") never matches
-  /\bI can[’']?t answer\b[^.]{0,100}?\bscope\b/i,
-  /^\s*I don[’']?t have any indexed OIML \w+(?:s)? (?:covering|about|on)\b/im,
-];
+// exact canonical sentence — normalize variants, keep the redirect tail.
+// The patterns are built per call: the publisher token is profile data.
+function refusalPatterns(publisher: string): { variant: RegExp; drift: RegExp[] } {
+  return {
+    variant: new RegExp(`^\\s*I don[’']?t have information on .{1,120}? in the indexed ${publisher}(?: \\w+){0,2} (?:publications|passages|documents|corpus)\\.?`, "i"),
+    drift: [
+      new RegExp(`\\b(can'?t|cannot|couldn'?t|unable)\\b[^.]{0,120}?\\b(indexed )?${publisher}(?: \\w+){0,2} (?:publications|passages|documents|corpus)\\b`, "i"),
+      new RegExp(`\\bno real answer to give\\b[^.]{0,120}?\\b${publisher}\\b`, "i"),
+      /\b(?:falls|well) outside\b[^.]{0,120}?\b(?:what I can answer|my scope|the scope of)\b/i,
+      /\boutside (?:of )?what (?:I|this service) can answer\b/i,
+      // "I can't answer that — weather forecasting is outside my scope":
+      // requires the refusal verb, so a scope DISCUSSION inside a real answer
+      // ("this exemption is outside the scope of R 60") never matches
+      /\bI can[’']?t answer\b[^.]{0,100}?\bscope\b/i,
+      new RegExp(`^\\s*I don[’']?t have any indexed ${publisher} \\w+(?:s)? (?:covering|about|on)\b`, "im"),
+    ],
+  };
+}
 
 /** Start of the sentence containing offset `i` (after the nearest ". ",
  *  "! ", "? ", or newline before it, else the string start). */
@@ -54,6 +51,7 @@ function sentenceEnd(answer: string, i: number): number {
 export function canonicalRefusal(answer: string): string {
   const CANON = refusalAnswer();
   if (answer.includes(CANON)) return answer;
+  const { variant: REFUSAL_VARIANT, drift: REFUSAL_DRIFT } = refusalPatterns(P().publisher.name);
   const variant = answer.match(REFUSAL_VARIANT);
   if (variant) return answer.replace(variant[0], CANON);
   for (const drift of REFUSAL_DRIFT) {