@mailwoman/neural 4.9.0 → 4.11.0

package/out/soft-features.d.ts

@@ -0,0 +1,67 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   The SOFT-FEATURE channels (#718) — the per-piece anchor + gazetteer clues the model conditions on
+ *   alongside `input_ids`. This module is the single, PURE, browser-safe home for the channel
+ *   choreography that used to live inline in `NeuralAddressClassifier.#decode`: build the postcode
+ *   anchor, build the gazetteer clue, and (when paired with the matching train-time half) suppress
+ *   the gazetteer clue adjacent to a postcode-anchor hit.
+ *
+ *   It is the LOAD-BEARING contract surface for the ProductionScorer (#718): the scorer asserts which
+ *   channels are fed, this function decides HOW they are fed. Keeping it a pure function means both
+ *   the classifier and any harness build features identically — there is exactly one choreography.
+ *
+ *   No `fs`, no Node builtins: the caller hands in the already-parsed lookup/lexicon (mirrors
+ *   `anchor-inference.ts` / `gazetteer-inference.ts`, which are themselves pure for the same
+ *   reason).
+ */
+import { type AnchorLookup } from "./anchor-inference.js";
+import { type GazetteerLexicon } from "./gazetteer-inference.js";
+import type { TokenizedPiece } from "./tokenizer.js";
+/** A built soft-feature channel: per-piece feature rows + per-piece confidence. */
+export interface SoftFeatureChannel {
+    features: number[][];
+    confidence: number[];
+}
+/** The soft-feature channels fed to the runner. Each is present only when its source is configured. */
+export interface SoftFeatures {
+    /** Postcode-anchor channel (#239/#240) — present iff `postcodeAnchorLookup` was supplied. */
+    anchor?: SoftFeatureChannel;
+    /**
+     * Gazetteer-anchor channel (#464) — present iff `gazetteerLexicon` was supplied. Already
+     * choreographed: when `suppressGazetteerNearPostcode` is set AND an anchor channel exists, the
+     * clue is zeroed adjacent to postcode-anchor hits before it's returned here.
+     */
+    gazetteer?: SoftFeatureChannel;
+}
+/** Sources + choreography for {@link buildSoftFeatures}. Mirrors the classifier's config fields. */
+export interface SoftFeatureSources {
+    /** Postcode→anchor lookup (#239/#240). Omit to skip the anchor channel. */
+    postcodeAnchorLookup?: AnchorLookup;
+    /** Gazetteer-anchor lexicon (#464). Omit to skip the gazetteer channel. */
+    gazetteerLexicon?: GazetteerLexicon;
+    /**
+     * Channel choreography (#464, v0.9.13 postcode fix): zero the gazetteer clue on pieces adjacent
+     * to a postcode-anchor hit. Needs BOTH a `gazetteerLexicon` and a `postcodeAnchorLookup` to take
+     * effect (the suppression is keyed off the anchor's confidence). PAIRING IS LOAD-BEARING — enable
+     * this IFF the model was trained with the matching train-time choreography. See
+     * `suppressGazetteerNearPostcode` in `gazetteer-inference.ts`.
+     */
+    suppressGazetteerNearPostcode?: boolean;
+}
+/**
+ * Build the soft-feature channels for `text`/`pieces` from the configured sources — the EXACT
+ * choreography previously inlined in `NeuralAddressClassifier.#decode`:
+ *
+ * 1. Anchor channel from `postcodeAnchorLookup` (no-op when unset).
+ * 2. Gazetteer channel from `gazetteerLexicon` (no-op when unset).
+ * 3. If both channels exist AND `suppressGazetteerNearPostcode`, zero the gazetteer clue adjacent to
+ *    postcode-anchor hits.
+ *
+ * Pure + byte-stable: the returned channels are identical to the pre-#718 inline path, so wiring
+ * this into `#decode` is a behavior-preserving refactor.
+ */
+export declare function buildSoftFeatures(text: string, pieces: ReadonlyArray<TokenizedPiece>, sources: SoftFeatureSources): SoftFeatures;
+//# sourceMappingURL=soft-features.d.ts.map

package/out/soft-features.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"soft-features.d.ts","sourceRoot":"","sources":["../soft-features.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAEH,OAAO,EAAuB,KAAK,YAAY,EAAE,MAAM,uBAAuB,CAAA;AAC9E,OAAO,EAAyD,KAAK,gBAAgB,EAAE,MAAM,0BAA0B,CAAA;AACvH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,gBAAgB,CAAA;AAEpD,mFAAmF;AACnF,MAAM,WAAW,kBAAkB;IAClC,QAAQ,EAAE,MAAM,EAAE,EAAE,CAAA;IACpB,UAAU,EAAE,MAAM,EAAE,CAAA;CACpB;AAED,uGAAuG;AACvG,MAAM,WAAW,YAAY;IAC5B,6FAA6F;IAC7F,MAAM,CAAC,EAAE,kBAAkB,CAAA;IAC3B;;;;OAIG;IACH,SAAS,CAAC,EAAE,kBAAkB,CAAA;CAC9B;AAED,oGAAoG;AACpG,MAAM,WAAW,kBAAkB;IAClC,2EAA2E;IAC3E,oBAAoB,CAAC,EAAE,YAAY,CAAA;IACnC,2EAA2E;IAC3E,gBAAgB,CAAC,EAAE,gBAAgB,CAAA;IACnC;;;;;;OAMG;IACH,6BAA6B,CAAC,EAAE,OAAO,CAAA;CACvC;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,iBAAiB,CAChC,IAAI,EAAE,MAAM,EACZ,MAAM,EAAE,aAAa,CAAC,cAAc,CAAC,EACrC,OAAO,EAAE,kBAAkB,GACzB,YAAY,CAgBd"}

package/out/soft-features.js

@@ -0,0 +1,49 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   The SOFT-FEATURE channels (#718) — the per-piece anchor + gazetteer clues the model conditions on
+ *   alongside `input_ids`. This module is the single, PURE, browser-safe home for the channel
+ *   choreography that used to live inline in `NeuralAddressClassifier.#decode`: build the postcode
+ *   anchor, build the gazetteer clue, and (when paired with the matching train-time half) suppress
+ *   the gazetteer clue adjacent to a postcode-anchor hit.
+ *
+ *   It is the LOAD-BEARING contract surface for the ProductionScorer (#718): the scorer asserts which
+ *   channels are fed, this function decides HOW they are fed. Keeping it a pure function means both
+ *   the classifier and any harness build features identically — there is exactly one choreography.
+ *
+ *   No `fs`, no Node builtins: the caller hands in the already-parsed lookup/lexicon (mirrors
+ *   `anchor-inference.ts` / `gazetteer-inference.ts`, which are themselves pure for the same
+ *   reason).
+ */
+import { buildAnchorFeatures } from "./anchor-inference.js";
+import { buildGazetteerFeatures, suppressGazetteerNearPostcode } from "./gazetteer-inference.js";
+/**
+ * Build the soft-feature channels for `text`/`pieces` from the configured sources — the EXACT
+ * choreography previously inlined in `NeuralAddressClassifier.#decode`:
+ *
+ * 1. Anchor channel from `postcodeAnchorLookup` (no-op when unset).
+ * 2. Gazetteer channel from `gazetteerLexicon` (no-op when unset).
+ * 3. If both channels exist AND `suppressGazetteerNearPostcode`, zero the gazetteer clue adjacent to
+ *    postcode-anchor hits.
+ *
+ * Pure + byte-stable: the returned channels are identical to the pre-#718 inline path, so wiring
+ * this into `#decode` is a behavior-preserving refactor.
+ */
+export function buildSoftFeatures(text, pieces, sources) {
+    const anchor = sources.postcodeAnchorLookup
+        ? buildAnchorFeatures(text, pieces, sources.postcodeAnchorLookup)
+        : undefined;
+    const gazetteer = sources.gazetteerLexicon
+        ? buildGazetteerFeatures(text, pieces, sources.gazetteerLexicon)
+        : undefined;
+    const gazFed = gazetteer && anchor && sources.suppressGazetteerNearPostcode
+        ? suppressGazetteerNearPostcode(gazetteer, anchor.confidence)
+        : gazetteer;
+    return {
+        ...(anchor ? { anchor } : {}),
+        ...(gazFed ? { gazetteer: gazFed } : {}),
+    };
+}
+//# sourceMappingURL=soft-features.js.map

package/out/soft-features.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"soft-features.js","sourceRoot":"","sources":["../soft-features.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAEH,OAAO,EAAE,mBAAmB,EAAqB,MAAM,uBAAuB,CAAA;AAC9E,OAAO,EAAE,sBAAsB,EAAE,6BAA6B,EAAyB,MAAM,0BAA0B,CAAA;AAqCvH;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,iBAAiB,CAChC,IAAY,EACZ,MAAqC,EACrC,OAA2B;IAE3B,MAAM,MAAM,GAAG,OAAO,CAAC,oBAAoB;QAC1C,CAAC,CAAC,mBAAmB,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,CAAC,oBAAoB,CAAC;QACjE,CAAC,CAAC,SAAS,CAAA;IACZ,MAAM,SAAS,GAAG,OAAO,CAAC,gBAAgB;QACzC,CAAC,CAAC,sBAAsB,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,CAAC,gBAAgB,CAAC;QAChE,CAAC,CAAC,SAAS,CAAA;IACZ,MAAM,MAAM,GACX,SAAS,IAAI,MAAM,IAAI,OAAO,CAAC,6BAA6B;QAC3D,CAAC,CAAC,6BAA6B,CAAC,SAAS,EAAE,MAAM,CAAC,UAAU,CAAC;QAC7D,CAAC,CAAC,SAAS,CAAA;IAEb,OAAO;QACN,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC7B,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KACxC,CAAA;AACF,CAAC"}

package/out/weights.d.ts

@@ -36,6 +36,12 @@ export interface ResolveWeightsOpts {
      * checkpoint via explicit paths.
      */
     modelCardPath?: string;
+    /**
+     * Serving tier (#718 D1). `"server"` (default) = anchor + gazetteer channels; `"pocket"` =
+     * anchor-only (skip the gazetteer lexicon even when shipped). Selects which soft-feature sibling
+     * artifacts {@link resolveWeights} surfaces — the loader feeds only the resolved channels.
+     */
+    tier?: "server" | "pocket";
 }
 export interface ResolvedWeights {
     modelPath: string;
@@ -52,6 +58,26 @@ export interface ResolvedWeights {
      * exist (pre-v0.6.0 bundles or CE-only training).
      */
     crfTransitionsPath?: string;
+    /**
+     * Path to the postcode→anchor source shipped beside the resolved model (#718 D1) — the soft-feed
+     * `loadFromWeights` reads to feed the anchor channel without a callsite change. Prefer the
+     * compact PCB1 binary (`postcode-<cc>.bin`, decoded via
+     * `PostcodeBinaryResolver.toAnchorLookup()`), else a JSON anchor lookup (`anchor-lookup.json`,
+     * parsed via `parseAnchorLookup`). `undefined` when the package ships neither (a plain/pre-#718
+     * bundle) — the loader then runs anchor-OFF. The `binary` flag tells the loader which parser to
+     * use.
+     */
+    anchorLookupPath?: {
+        path: string;
+        binary: boolean;
+    };
+    /**
+     * Path to the gazetteer-anchor lexicon (`anchor-lexicon-v1.json`, #464) shipped beside the
+     * resolved model. `undefined` when the package doesn't ship it, OR when `opts.tier === "pocket"`
+     * (pocket is anchor-only — the gazetteer channel is deliberately skipped). Read by the
+     * `loadFromWeights` soft-feed via `parseGazetteerLexicon`.
+     */
+    gazetteerLexiconPath?: string;
     /** "explicit" if both paths came from opts; "package:<name>" if resolved via require.resolve. */
     source: string;
 }
@@ -68,6 +94,89 @@ export declare function resolveWeights(opts: ResolveWeightsOpts): ResolvedWeight
  * and should be loud, not silently re-defaulted.
  */
 export declare function readLabelsFromModelCard(modelCardPath: string | undefined): readonly string[] | undefined;
+/**
+ * The structured `requires` block of a `model-card.json` (#718) — the declared SHIP-CONFIG the
+ * model was trained against. The ProductionScorer reads this and FAILS CLOSED when a declared
+ * channel isn't actually fed (silent OOD is the #566/#685 trap). Each channel is optional; a
+ * missing channel means "not declared" (treated as not-required).
+ */
+export interface RequiredChannels {
+    /** Postcode-anchor channel (#239/#240). */
+    anchor?: {
+        required: boolean;
+    };
+    /** Gazetteer-anchor channel (#464). */
+    gazetteer?: {
+        required: boolean;
+    };
+    /** Address-system conventions (#511 Tier A). `mode` mirrors `ParseOpts.addressSystemConventions`. */
+    conventions?: {
+        required: boolean;
+        mode?: "auto" | string;
+    };
+    /** Punctuation-gap span bridge (v4.4.0 corrective). */
+    bridge?: {
+        required: boolean;
+    };
+    /** Near-postcode gazetteer choreography (#464, v0.9.13). */
+    suppress_gazetteer_near_postcode?: boolean;
+}
+/**
+ * Read the structured `requires` block from a `model-card.json` (#718). DEFENSIVE: returns
+ * `undefined` when the card is absent, unreadable, or has no `requires` field (callers then INFER
+ * the required channels from the ONNX graph — see `inferRequiredChannelsFromInputs`). Throws ONLY
+ * when the field is PRESENT but corrupt (not an object, or a channel entry with a non-boolean
+ * `required`) — a malformed declared contract is a loud artifact bug, not a silent re-default.
+ */
+export declare function readRequiredChannels(modelCardPath: string | undefined): RequiredChannels | undefined;
+/**
+ * Back-compat inference of the required soft-feature channels from an ONNX model's declared input
+ * names (#718). A model that exports `anchor_features` / `gazetteer_features` declared those
+ * channels mandatory at train time — feeding zeros is the channel-off identity, but a model TRAINED
+ * with the channel is OOD when scored without it. Cards without a `requires` block (every pre-#718
+ * bundle) route through here so the fail-closed guard still protects them. Conventions/bridge are
+ * NOT graph-observable (no dedicated input), so they're left undeclared here — only the card
+ * declares them.
+ */
+export declare function inferRequiredChannelsFromInputs(inputNames: readonly string[]): RequiredChannels;
+/**
+ * One tag's certified capability under a (tier × address-system) cell of the capability manifest
+ * (#718/#719). `maskOffF1` is the model's measured per-tag exact-match F1 with the conventions mask
+ * OFF; `maskOnF1` is the same with the mask ON — recorded ONLY for tags some codex `forbiddenTags`
+ * row suppresses, because that's the only place the loader's delta-gate consults it.
+ */
+export interface TagCapability {
+    /** Measured per-tag F1 (percent) with the conventions mask OFF — the model's real capability. */
+    maskOffF1: number;
+    /** Measured per-tag F1 (percent) with the mask ON. Present only for codex-forbidden tags. */
+    maskOnF1?: number;
+}
+/**
+ * The `capabilities` block of a `model-card.json` (#718/#719): per serving TIER (`server` =
+ * anchor+gazetteer; `pocket` = anchor-only) × per codex address-system × per tag, the model's
+ * certified per-tag capability. The `createScorer` loader reads this to FAIL CLOSED when a
+ * conventions mask would forbid a tag the model is certified to emit — the structural fix that
+ * makes the D2/#719 bug-class (a mask destroying a demonstrated capability) impossible.
+ *
+ * Shape: `capabilities[tier][system][tag] = { maskOffF1, maskOnF1? }`. A `$comment` provenance key
+ * may sit alongside the tier keys and is ignored by readers.
+ */
+export type CapabilityManifest = Record<string, Record<string, Record<string, TagCapability>>>;
+/**
+ * Read the `capabilities` block from a `model-card.json` (#718/#719). DEFENSIVE, mirroring
+ * `readRequiredChannels`: returns `undefined` when the card is absent, unreadable, or has no
+ * `capabilities` field (a pre-#718 card → the loader's delta-gate is skipped, back-compat). Throws
+ * ONLY when the field is PRESENT but not an object — a corrupt declared contract is a loud artifact
+ * bug, not a silent skip. Tier/system/tag sub-shapes are read leniently (a malformed cell simply
+ * yields no capability claim — `undefined` from `lookupTagCapability`).
+ */
+export declare function readCapabilityManifest(modelCardPath: string | undefined): CapabilityManifest | undefined;
+/**
+ * Resolve `capabilities[tier][system][tag]` to a `TagCapability`, returning `undefined` for any
+ * missing/malformed cell (a tag the model is NOT certified for — the loader treats that as legal:
+ * the model can't emit it, so a mask can't destroy it). Skips the `$comment` provenance key.
+ */
+export declare function lookupTagCapability(manifest: CapabilityManifest | undefined, tier: string, system: string, tag: string): TagCapability | undefined;
 export interface CrfTransitions {
     transitions: number[][];
     startTransitions: number[];

package/out/weights.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"weights.d.ts","sourceRoot":"","sources":["../weights.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAQH,MAAM,WAAW,kBAAkB;IAClC,wFAAwF;IACxF,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,4EAA4E;IAC5E,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,iFAAiF;IACjF,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB;;;;;;OAMG;IACH,aAAa,CAAC,EAAE,MAAM,CAAA;CACtB;AAED,MAAM,WAAW,eAAe;IAC/B,SAAS,EAAE,MAAM,CAAA;IACjB,aAAa,EAAE,MAAM,CAAA;IACrB;;;;;OAKG;IACH,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB;;;OAGG;IACH,kBAAkB,CAAC,EAAE,MAAM,CAAA;IAC3B,iGAAiG;IACjG,MAAM,EAAE,MAAM,CAAA;CACd;AAED,wBAAgB,cAAc,CAAC,IAAI,EAAE,kBAAkB,GAAG,eAAe,CAkDxE;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,uBAAuB,CAAC,aAAa,EAAE,MAAM,GAAG,SAAS,GAAG,SAAS,MAAM,EAAE,GAAG,SAAS,CAwBxG;AAED,MAAM,WAAW,cAAc;IAC9B,WAAW,EAAE,MAAM,EAAE,EAAE,CAAA;IACvB,gBAAgB,EAAE,MAAM,EAAE,CAAA;IAC1B,cAAc,EAAE,MAAM,EAAE,CAAA;CACxB;AAED;;;GAGG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,MAAM,GAAG,SAAS,GAAG,cAAc,GAAG,SAAS,CA0B1F"}
+{"version":3,"file":"weights.d.ts","sourceRoot":"","sources":["../weights.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAQH,MAAM,WAAW,kBAAkB;IAClC,wFAAwF;IACxF,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,4EAA4E;IAC5E,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,iFAAiF;IACjF,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB;;;;;;OAMG;IACH,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB;;;;OAIG;IACH,IAAI,CAAC,EAAE,QAAQ,GAAG,QAAQ,CAAA;CAC1B;AAED,MAAM,WAAW,eAAe;IAC/B,SAAS,EAAE,MAAM,CAAA;IACjB,aAAa,EAAE,MAAM,CAAA;IACrB;;;;;OAKG;IACH,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB;;;OAGG;IACH,kBAAkB,CAAC,EAAE,MAAM,CAAA;IAC3B;;;;;;;;OAQG;IACH,gBAAgB,CAAC,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,OAAO,CAAA;KAAE,CAAA;IACpD;;;;;OAKG;IACH,oBAAoB,CAAC,EAAE,MAAM,CAAA;IAC7B,iGAAiG;IACjG,MAAM,EAAE,MAAM,CAAA;CACd;AAED,wBAAgB,cAAc,CAAC,IAAI,EAAE,kBAAkB,GAAG,eAAe,CAqExE;AAsBD;;;;;;;;;;GAUG;AACH,wBAAgB,uBAAuB,CAAC,aAAa,EAAE,MAAM,GAAG,SAAS,GAAG,SAAS,MAAM,EAAE,GAAG,SAAS,CAwBxG;AAED;;;;;GAKG;AACH,MAAM,WAAW,gBAAgB;IAChC,2CAA2C;IAC3C,MAAM,CAAC,EAAE;QAAE,QAAQ,EAAE,OAAO,CAAA;KAAE,CAAA;IAC9B,uCAAuC;IACvC,SAAS,CAAC,EAAE;QAAE,QAAQ,EAAE,OAAO,CAAA;KAAE,CAAA;IACjC,qGAAqG;IACrG,WAAW,CAAC,EAAE;QAAE,QAAQ,EAAE,OAAO,CAAC;QAAC,IAAI,CAAC,EAAE,MAAM,GAAG,MAAM,CAAA;KAAE,CAAA;IAC3D,uDAAuD;IACvD,MAAM,CAAC,EAAE;QAAE,QAAQ,EAAE,OAAO,CAAA;KAAE,CAAA;IAC9B,4DAA4D;IAC5D,gCAAgC,CAAC,EAAE,OAAO,CAAA;CAC1C;AAED;;;;;;GAMG;AACH,wBAAgB,oBAAoB,CAAC,aAAa,EAAE,MAAM,GAAG,SAAS,GAAG,gBAAgB,GAAG,SAAS,CA8CpG;AAED;;;;;;;;GAQG;AACH,wBAAgB,+BAA+B,CAAC,UAAU,EAAE,SAAS,MAAM,EAAE,GAAG,gBAAgB,CAM/F;AAED;;;;;GAKG;AACH,MAAM,WAAW,aAAa;IAC7B,iGAAiG;IACjG,SAAS,EAAE,MAAM,CAAA;IACjB,6FAA6F;IAC7F,QAAQ,CAAC,EAAE,MAAM,CAAA;CACjB;AAED;;;;;;;;;GASG;AACH,MAAM,MAAM,kBAAkB,GAAG,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,aAAa,CAAC,CAAC,CAAC,CAAA;AAE9F;;;;;;;GAOG;AACH,wBAAgB,sBAAsB,CAAC,aAAa,EAAE,MAAM,GAAG,SAAS,GAAG,kBAAkB,GAAG,SAAS,CAwBxG;AAED;;;;GAIG;AACH,wBAAgB,mBAAmB,CAClC,QAAQ,EAAE,kBAAkB,GAAG,SAAS,EACxC,IAAI,EAAE,MAAM,EACZ,MAAM,EAAE,MAAM,EACd,GAAG,EAAE,MAAM,GACT,aAAa,GAAG,SAAS,CAQ3B;AAED,MAAM,WAAW,cAAc;IAC9B,WAAW,EAAE,MAAM,EAAE,EAAE,CAAA;IACvB,gBAAgB,EAAE,MAAM,EAAE,CAAA;IAC1B,cAAc,EAAE,MAAM,EAAE,CAAA;CACxB;AAED;;;GAGG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,MAAM,GAAG,SAAS,GAAG,cAAc,GAAG,SAAS,CA0B1F"}

package/out/weights.js

@@ -66,7 +66,42 @@ export function resolveWeights(opts) {
     const modelCardPath = existsSync(modelCardCandidate) ? modelCardCandidate : undefined;
     const crfCandidate = resolve(packageDir, "crf-transitions.json");
     const crfTransitionsPath = existsSync(crfCandidate) ? crfCandidate : undefined;
-    return { modelPath, tokenizerPath, modelCardPath, crfTransitionsPath, source: `package:${packageName}` };
+    // Soft-feature sibling artifacts (#718 D1): the anchor + gazetteer sources the package ships so
+    // `loadFromWeights` can feed the channels the model was trained against — without a callsite
+    // change. Resolved package-dir-relative via the same `existsSync → undefined` pattern as the CRF
+    // transitions above. The locale tag's region subtag (`en-us` → `us`) names the PCB1 binary.
+    const country = locale.split("-")[1] ?? "";
+    const anchorLookupPath = resolveAnchorLookupSibling(packageDir, country);
+    // Tier `"pocket"` is anchor-only — never surface the gazetteer lexicon (the loader then skips it).
+    const gazetteerCandidate = resolve(packageDir, "anchor-lexicon-v1.json");
+    const gazetteerLexiconPath = opts.tier === "pocket" ? undefined : existsSync(gazetteerCandidate) ? gazetteerCandidate : undefined;
+    return {
+        modelPath,
+        tokenizerPath,
+        modelCardPath,
+        crfTransitionsPath,
+        ...(anchorLookupPath ? { anchorLookupPath } : {}),
+        ...(gazetteerLexiconPath ? { gazetteerLexiconPath } : {}),
+        source: `package:${packageName}`,
+    };
+}
+/**
+ * Locate the package's postcode→anchor source for the soft-feed (#718 D1), preferring the compact
+ * PCB1 binary (`postcode-<cc>.bin`, ~0.66 MB) over the much larger JSON lookup
+ * (`anchor-lookup.json`, the 3.2 MB pilot dump). Returns the path + a `binary` flag so the loader
+ * picks the right parser (`PostcodeBinaryResolver.toAnchorLookup()` vs `parseAnchorLookup`).
+ * `undefined` when neither ships.
+ */
+function resolveAnchorLookupSibling(packageDir, country) {
+    if (country) {
+        const binary = resolve(packageDir, `postcode-${country}.bin`);
+        if (existsSync(binary))
+            return { path: binary, binary: true };
+    }
+    const json = resolve(packageDir, "anchor-lookup.json");
+    if (existsSync(json))
+        return { path: json, binary: false };
+    return undefined;
 }
 /**
  * Read the `labels` array from a `model-card.json` file. Returns `undefined` when the file is
@@ -107,6 +142,127 @@ export function readLabelsFromModelCard(modelCardPath) {
     }
     return Object.freeze(labels.slice());
 }
+/**
+ * Read the structured `requires` block from a `model-card.json` (#718). DEFENSIVE: returns
+ * `undefined` when the card is absent, unreadable, or has no `requires` field (callers then INFER
+ * the required channels from the ONNX graph — see `inferRequiredChannelsFromInputs`). Throws ONLY
+ * when the field is PRESENT but corrupt (not an object, or a channel entry with a non-boolean
+ * `required`) — a malformed declared contract is a loud artifact bug, not a silent re-default.
+ */
+export function readRequiredChannels(modelCardPath) {
+    if (!modelCardPath || !existsSync(modelCardPath))
+        return undefined;
+    let raw;
+    try {
+        raw = readFileSync(modelCardPath, "utf8");
+    }
+    catch {
+        return undefined;
+    }
+    let parsed;
+    try {
+        parsed = JSON.parse(raw);
+    }
+    catch {
+        return undefined;
+    }
+    if (typeof parsed !== "object" || parsed === null)
+        return undefined;
+    const requires = parsed.requires;
+    if (requires === undefined)
+        return undefined;
+    if (typeof requires !== "object" || requires === null || Array.isArray(requires)) {
+        throw new Error(`model-card.json at ${modelCardPath} has a malformed \`requires\` field — ` +
+            `expected an object, got ${JSON.stringify(requires)}.`);
+    }
+    const obj = requires;
+    // Channel entries must be `{ required: boolean, ... }`; a present-but-shapeless entry is corrupt.
+    for (const channel of ["anchor", "gazetteer", "conventions", "bridge"]) {
+        const entry = obj[channel];
+        if (entry === undefined)
+            continue;
+        if (typeof entry !== "object" ||
+            entry === null ||
+            typeof entry.required !== "boolean") {
+            throw new Error(`model-card.json at ${modelCardPath} has a malformed \`requires.${channel}\` entry — ` +
+                `expected { required: boolean }, got ${JSON.stringify(entry)}.`);
+        }
+    }
+    if (obj.suppress_gazetteer_near_postcode !== undefined && typeof obj.suppress_gazetteer_near_postcode !== "boolean") {
+        throw new Error(`model-card.json at ${modelCardPath} has a malformed \`requires.suppress_gazetteer_near_postcode\` ` +
+            `field — expected a boolean, got ${JSON.stringify(obj.suppress_gazetteer_near_postcode)}.`);
+    }
+    return requires;
+}
+/**
+ * Back-compat inference of the required soft-feature channels from an ONNX model's declared input
+ * names (#718). A model that exports `anchor_features` / `gazetteer_features` declared those
+ * channels mandatory at train time — feeding zeros is the channel-off identity, but a model TRAINED
+ * with the channel is OOD when scored without it. Cards without a `requires` block (every pre-#718
+ * bundle) route through here so the fail-closed guard still protects them. Conventions/bridge are
+ * NOT graph-observable (no dedicated input), so they're left undeclared here — only the card
+ * declares them.
+ */
+export function inferRequiredChannelsFromInputs(inputNames) {
+    const names = new Set(inputNames);
+    return {
+        ...(names.has("anchor_features") ? { anchor: { required: true } } : {}),
+        ...(names.has("gazetteer_features") ? { gazetteer: { required: true } } : {}),
+    };
+}
+/**
+ * Read the `capabilities` block from a `model-card.json` (#718/#719). DEFENSIVE, mirroring
+ * `readRequiredChannels`: returns `undefined` when the card is absent, unreadable, or has no
+ * `capabilities` field (a pre-#718 card → the loader's delta-gate is skipped, back-compat). Throws
+ * ONLY when the field is PRESENT but not an object — a corrupt declared contract is a loud artifact
+ * bug, not a silent skip. Tier/system/tag sub-shapes are read leniently (a malformed cell simply
+ * yields no capability claim — `undefined` from `lookupTagCapability`).
+ */
+export function readCapabilityManifest(modelCardPath) {
+    if (!modelCardPath || !existsSync(modelCardPath))
+        return undefined;
+    let raw;
+    try {
+        raw = readFileSync(modelCardPath, "utf8");
+    }
+    catch {
+        return undefined;
+    }
+    let parsed;
+    try {
+        parsed = JSON.parse(raw);
+    }
+    catch {
+        return undefined;
+    }
+    if (typeof parsed !== "object" || parsed === null)
+        return undefined;
+    const capabilities = parsed.capabilities;
+    if (capabilities === undefined)
+        return undefined;
+    if (typeof capabilities !== "object" || capabilities === null || Array.isArray(capabilities)) {
+        throw new Error(`model-card.json at ${modelCardPath} has a malformed \`capabilities\` field — ` +
+            `expected an object, got ${JSON.stringify(capabilities)}.`);
+    }
+    return capabilities;
+}
+/**
+ * Resolve `capabilities[tier][system][tag]` to a `TagCapability`, returning `undefined` for any
+ * missing/malformed cell (a tag the model is NOT certified for — the loader treats that as legal:
+ * the model can't emit it, so a mask can't destroy it). Skips the `$comment` provenance key.
+ */
+export function lookupTagCapability(manifest, tier, system, tag) {
+    const tierCell = manifest?.[tier];
+    if (!tierCell || typeof tierCell !== "object")
+        return undefined;
+    const systemCell = tierCell[system];
+    if (!systemCell || typeof systemCell !== "object")
+        return undefined;
+    const cap = systemCell[tag];
+    if (!cap || typeof cap !== "object" || typeof cap.maskOffF1 !== "number")
+        return undefined;
+    return cap;
+}
 /**
  * Read learned CRF transition parameters from `crf-transitions.json`. Returns `undefined` when the
  * file is missing or malformed — callers fall back to the structural BIO mask only.

package/out/weights.js.map

@@ -1 +1 @@
-{"version":3,"file":"weights.js","sourceRoot":"","sources":["../weights.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,OAAO,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,SAAS,CAAA;AAClD,OAAO,EAAE,aAAa,EAAE,MAAM,aAAa,CAAA;AAC3C,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAA;AAE5C,MAAM,GAAG,GAAG,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;AAsC1C,MAAM,UAAU,cAAc,CAAC,IAAwB;IACtD,MAAM,KAAK,GAAa,EAAE,CAAA;IAE1B,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,aAAa,EAAE,CAAC;QAC1C,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,SAAS,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,sCAAsC,IAAI,CAAC,SAAS,EAAE,CAAC,CAAA;QACxG,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,aAAa,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,0CAA0C,IAAI,CAAC,aAAa,EAAE,CAAC,CAAA;QACpH,6FAA6F;QAC7F,8FAA8F;QAC9F,iGAAiG;QACjG,MAAM,aAAa,GAAG,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,iBAAiB,CAAC,CAAA;QACzE,MAAM,aAAa,GAAG,IAAI,CAAC,aAAa,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,SAAS,CAAC,CAAA;QACnG,OAAO,EAAE,SAAS,EAAE,IAAI,CAAC,SAAS,EAAE,aAAa,EAAE,IAAI,CAAC,aAAa,EAAE,aAAa,EAAE,MAAM,EAAE,UAAU,EAAE,CAAA;IAC3G,CAAC;IAED,oFAAoF;IACpF,2FAA2F;IAC3F,4DAA4D;IAC5D,MAAM,MAAM,GAAG,CAAC,IAAI,CAAC,MAAM,IAAI,OAAO,CAAC,CAAC,WAAW,EAAE,CAAA;IACrD,MAAM,WAAW,GAAG,6BAA6B,MAAM,EAAE,CAAA;IACzD,IAAI,UAAkB,CAAA;IACtB,IAAI,CAAC;QACJ,MAAM,WAAW,GAAG,GAAG,CAAC,OAAO,CAAC,GAAG,WAAW,eAAe,CAAC,CAAA;QAC9D,UAAU,GAAG,OAAO,CAAC,WAAW,CAAC,CAAA;IAClC,CAAC;IAAC,MAAM,CAAC;QACR,MAAM,IAAI,KAAK,CACd,qBAAqB,WAAW,iCAAiC,WAAW,IAAI;YAC/E,oDAAoD,CACrD,CAAA;IACF,CAAC;IAED,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,OAAO,CAAC,UAAU,EAAE,YAAY,CAAC,CAAA;IACrE,MAAM,aAAa,GAAG,IAAI,CAAC,aAAa,IAAI,OAAO,CAAC,UAAU,EAAE,iBAAiB,CAAC,CAAA;IAClF,KAAK,CAAC,IAAI,CAAC,SAAS,EAAE,aAAa,CAAC,CAAA;IAEpC,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC,EAAE,CAAC;QAC1D,MAAM,IAAI,KAAK,CACd,mBAAmB,WAAW,gBAAgB,UAAU,gCAAgC;YACvF,aAAa,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI;YACnC,iFAAiF;YACjF,oDAAoD,CACrD,CAAA;IACF,CAAC;IAED,MAAM,kBAAkB,GAAG,OAAO,CAAC,UAAU,EAAE,iBAAiB,CAAC,CAAA;IACjE,MAAM,aAAa,GAAG,UAAU,CAAC,kBAAkB,CAAC,CAAC,CAAC,CAAC,kBAAkB,CAAC,CAAC,CAAC,SAAS,CAAA;IAErF,MAAM,YAAY,GAAG,OAAO,CAAC,UAAU,EAAE,sBAAsB,CAAC,CAAA;IAChE,MAAM,kBAAkB,GAAG,UAAU,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,SAAS,CAAA;IAE9E,OAAO,EAAE,SAAS,EAAE,aAAa,EAAE,aAAa,EAAE,kBAAkB,EAAE,MAAM,EAAE,WAAW,WAAW,EAAE,EAAE,CAAA;AACzG,CAAC;AAED;;;;;;;;;;GAUG;AACH,MAAM,UAAU,uBAAuB,CAAC,aAAiC;IACxE,IAAI,CAAC,aAAa,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC;QAAE,OAAO,SAAS,CAAA;IAClE,IAAI,GAAW,CAAA;IACf,IAAI,CAAC;QACJ,GAAG,GAAG,YAAY,CAAC,aAAa,EAAE,MAAM,CAAC,CAAA;IAC1C,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,MAAe,CAAA;IACnB,IAAI,CAAC;QACJ,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAA;IACzB,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,MAAM,KAAK,IAAI;QAAE,OAAO,SAAS,CAAA;IACnE,MAAM,MAAM,GAAI,MAA+B,CAAC,MAAM,CAAA;IACtD,IAAI,MAAM,KAAK,SAAS;QAAE,OAAO,SAAS,CAAA;IAC1C,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,KAAK,QAAQ,CAAC,EAAE,CAAC;QAClG,MAAM,IAAI,KAAK,CACd,sBAAsB,aAAa,sCAAsC;YACxE,8CAA8C,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,GAAG,CACxE,CAAA;IACF,CAAC;IACD,OAAO,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,KAAK,EAAE,CAAsB,CAAA;AAC1D,CAAC;AAQD;;;GAGG;AACH,MAAM,UAAU,kBAAkB,CAAC,OAA2B;IAC7D,IAAI,CAAC,OAAO,IAAI,CAAC,UAAU,CAAC,OAAO,CAAC;QAAE,OAAO,SAAS,CAAA;IACtD,IAAI,GAAW,CAAA;IACf,IAAI,CAAC;QACJ,GAAG,GAAG,YAAY,CAAC,OAAO,EAAE,MAAM,CAAC,CAAA;IACpC,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,MAAe,CAAA;IACnB,IAAI,CAAC;QACJ,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAA;IACzB,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,MAAM,KAAK,IAAI;QAAE,OAAO,SAAS,CAAA;IACnE,MAAM,GAAG,GAAG,MAAiC,CAAA;IAC7C,MAAM,WAAW,GAAG,GAAG,CAAC,WAAW,CAAA;IACnC,MAAM,KAAK,GAAG,GAAG,CAAC,iBAAiB,CAAA;IACnC,MAAM,GAAG,GAAG,GAAG,CAAC,eAAe,CAAA;IAC/B,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,WAAW,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IACjG,IAAI,WAAW,CAAC,MAAM,KAAK,CAAC,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC,IAAI,GAAG,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,SAAS,CAAA;IACxF,OAAO;QACN,WAAW,EAAE,WAAyB;QACtC,gBAAgB,EAAE,KAAiB;QACnC,cAAc,EAAE,GAAe;KAC/B,CAAA;AACF,CAAC"}
+{"version":3,"file":"weights.js","sourceRoot":"","sources":["../weights.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,OAAO,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,SAAS,CAAA;AAClD,OAAO,EAAE,aAAa,EAAE,MAAM,aAAa,CAAA;AAC3C,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAA;AAE5C,MAAM,GAAG,GAAG,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;AA6D1C,MAAM,UAAU,cAAc,CAAC,IAAwB;IACtD,MAAM,KAAK,GAAa,EAAE,CAAA;IAE1B,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,aAAa,EAAE,CAAC;QAC1C,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,SAAS,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,sCAAsC,IAAI,CAAC,SAAS,EAAE,CAAC,CAAA;QACxG,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,aAAa,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,0CAA0C,IAAI,CAAC,aAAa,EAAE,CAAC,CAAA;QACpH,6FAA6F;QAC7F,8FAA8F;QAC9F,iGAAiG;QACjG,MAAM,aAAa,GAAG,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,iBAAiB,CAAC,CAAA;QACzE,MAAM,aAAa,GAAG,IAAI,CAAC,aAAa,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,SAAS,CAAC,CAAA;QACnG,OAAO,EAAE,SAAS,EAAE,IAAI,CAAC,SAAS,EAAE,aAAa,EAAE,IAAI,CAAC,aAAa,EAAE,aAAa,EAAE,MAAM,EAAE,UAAU,EAAE,CAAA;IAC3G,CAAC;IAED,oFAAoF;IACpF,2FAA2F;IAC3F,4DAA4D;IAC5D,MAAM,MAAM,GAAG,CAAC,IAAI,CAAC,MAAM,IAAI,OAAO,CAAC,CAAC,WAAW,EAAE,CAAA;IACrD,MAAM,WAAW,GAAG,6BAA6B,MAAM,EAAE,CAAA;IACzD,IAAI,UAAkB,CAAA;IACtB,IAAI,CAAC;QACJ,MAAM,WAAW,GAAG,GAAG,CAAC,OAAO,CAAC,GAAG,WAAW,eAAe,CAAC,CAAA;QAC9D,UAAU,GAAG,OAAO,CAAC,WAAW,CAAC,CAAA;IAClC,CAAC;IAAC,MAAM,CAAC;QACR,MAAM,IAAI,KAAK,CACd,qBAAqB,WAAW,iCAAiC,WAAW,IAAI;YAC/E,oDAAoD,CACrD,CAAA;IACF,CAAC;IAED,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,OAAO,CAAC,UAAU,EAAE,YAAY,CAAC,CAAA;IACrE,MAAM,aAAa,GAAG,IAAI,CAAC,aAAa,IAAI,OAAO,CAAC,UAAU,EAAE,iBAAiB,CAAC,CAAA;IAClF,KAAK,CAAC,IAAI,CAAC,SAAS,EAAE,aAAa,CAAC,CAAA;IAEpC,IAAI,CAAC,UAAU,CAAC,SAAS,CAAC,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC,EAAE,CAAC;QAC1D,MAAM,IAAI,KAAK,CACd,mBAAmB,WAAW,gBAAgB,UAAU,gCAAgC;YACvF,aAAa,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI;YACnC,iFAAiF;YACjF,oDAAoD,CACrD,CAAA;IACF,CAAC;IAED,MAAM,kBAAkB,GAAG,OAAO,CAAC,UAAU,EAAE,iBAAiB,CAAC,CAAA;IACjE,MAAM,aAAa,GAAG,UAAU,CAAC,kBAAkB,CAAC,CAAC,CAAC,CAAC,kBAAkB,CAAC,CAAC,CAAC,SAAS,CAAA;IAErF,MAAM,YAAY,GAAG,OAAO,CAAC,UAAU,EAAE,sBAAsB,CAAC,CAAA;IAChE,MAAM,kBAAkB,GAAG,UAAU,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,SAAS,CAAA;IAE9E,gGAAgG;IAChG,6FAA6F;IAC7F,iGAAiG;IACjG,4FAA4F;IAC5F,MAAM,OAAO,GAAG,MAAM,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAA;IAC1C,MAAM,gBAAgB,GAAG,0BAA0B,CAAC,UAAU,EAAE,OAAO,CAAC,CAAA;IACxE,mGAAmG;IACnG,MAAM,kBAAkB,GAAG,OAAO,CAAC,UAAU,EAAE,wBAAwB,CAAC,CAAA;IACxE,MAAM,oBAAoB,GACzB,IAAI,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,UAAU,CAAC,kBAAkB,CAAC,CAAC,CAAC,CAAC,kBAAkB,CAAC,CAAC,CAAC,SAAS,CAAA;IAErG,OAAO;QACN,SAAS;QACT,aAAa;QACb,aAAa;QACb,kBAAkB;QAClB,GAAG,CAAC,gBAAgB,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACjD,GAAG,CAAC,oBAAoB,CAAC,CAAC,CAAC,EAAE,oBAAoB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACzD,MAAM,EAAE,WAAW,WAAW,EAAE;KAChC,CAAA;AACF,CAAC;AAED;;;;;;GAMG;AACH,SAAS,0BAA0B,CAClC,UAAkB,EAClB,OAAe;IAEf,IAAI,OAAO,EAAE,CAAC;QACb,MAAM,MAAM,GAAG,OAAO,CAAC,UAAU,EAAE,YAAY,OAAO,MAAM,CAAC,CAAA;QAC7D,IAAI,UAAU,CAAC,MAAM,CAAC;YAAE,OAAO,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,CAAA;IAC9D,CAAC;IACD,MAAM,IAAI,GAAG,OAAO,CAAC,UAAU,EAAE,oBAAoB,CAAC,CAAA;IACtD,IAAI,UAAU,CAAC,IAAI,CAAC;QAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,CAAA;IAC1D,OAAO,SAAS,CAAA;AACjB,CAAC;AAED;;;;;;;;;;GAUG;AACH,MAAM,UAAU,uBAAuB,CAAC,aAAiC;IACxE,IAAI,CAAC,aAAa,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC;QAAE,OAAO,SAAS,CAAA;IAClE,IAAI,GAAW,CAAA;IACf,IAAI,CAAC;QACJ,GAAG,GAAG,YAAY,CAAC,aAAa,EAAE,MAAM,CAAC,CAAA;IAC1C,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,MAAe,CAAA;IACnB,IAAI,CAAC;QACJ,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAA;IACzB,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,MAAM,KAAK,IAAI;QAAE,OAAO,SAAS,CAAA;IACnE,MAAM,MAAM,GAAI,MAA+B,CAAC,MAAM,CAAA;IACtD,IAAI,MAAM,KAAK,SAAS;QAAE,OAAO,SAAS,CAAA;IAC1C,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,KAAK,QAAQ,CAAC,EAAE,CAAC;QAClG,MAAM,IAAI,KAAK,CACd,sBAAsB,aAAa,sCAAsC;YACxE,8CAA8C,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,GAAG,CACxE,CAAA;IACF,CAAC;IACD,OAAO,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,KAAK,EAAE,CAAsB,CAAA;AAC1D,CAAC;AAqBD;;;;;;GAMG;AACH,MAAM,UAAU,oBAAoB,CAAC,aAAiC;IACrE,IAAI,CAAC,aAAa,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC;QAAE,OAAO,SAAS,CAAA;IAClE,IAAI,GAAW,CAAA;IACf,IAAI,CAAC;QACJ,GAAG,GAAG,YAAY,CAAC,aAAa,EAAE,MAAM,CAAC,CAAA;IAC1C,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,MAAe,CAAA;IACnB,IAAI,CAAC;QACJ,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAA;IACzB,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,MAAM,KAAK,IAAI;QAAE,OAAO,SAAS,CAAA;IACnE,MAAM,QAAQ,GAAI,MAAiC,CAAC,QAAQ,CAAA;IAC5D,IAAI,QAAQ,KAAK,SAAS;QAAE,OAAO,SAAS,CAAA;IAC5C,IAAI,OAAO,QAAQ,KAAK,QAAQ,IAAI,QAAQ,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC;QAClF,MAAM,IAAI,KAAK,CACd,sBAAsB,aAAa,wCAAwC;YAC1E,2BAA2B,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,GAAG,CACvD,CAAA;IACF,CAAC;IACD,MAAM,GAAG,GAAG,QAAmC,CAAA;IAC/C,kGAAkG;IAClG,KAAK,MAAM,OAAO,IAAI,CAAC,QAAQ,EAAE,WAAW,EAAE,aAAa,EAAE,QAAQ,CAAU,EAAE,CAAC;QACjF,MAAM,KAAK,GAAG,GAAG,CAAC,OAAO,CAAC,CAAA;QAC1B,IAAI,KAAK,KAAK,SAAS;YAAE,SAAQ;QACjC,IACC,OAAO,KAAK,KAAK,QAAQ;YACzB,KAAK,KAAK,IAAI;YACd,OAAQ,KAAgC,CAAC,QAAQ,KAAK,SAAS,EAC9D,CAAC;YACF,MAAM,IAAI,KAAK,CACd,sBAAsB,aAAa,+BAA+B,OAAO,aAAa;gBACrF,uCAAuC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,GAAG,CAChE,CAAA;QACF,CAAC;IACF,CAAC;IACD,IAAI,GAAG,CAAC,gCAAgC,KAAK,SAAS,IAAI,OAAO,GAAG,CAAC,gCAAgC,KAAK,SAAS,EAAE,CAAC;QACrH,MAAM,IAAI,KAAK,CACd,sBAAsB,aAAa,iEAAiE;YACnG,mCAAmC,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,gCAAgC,CAAC,GAAG,CAC3F,CAAA;IACF,CAAC;IACD,OAAO,QAA4B,CAAA;AACpC,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,+BAA+B,CAAC,UAA6B;IAC5E,MAAM,KAAK,GAAG,IAAI,GAAG,CAAC,UAAU,CAAC,CAAA;IACjC,OAAO;QACN,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,iBAAiB,CAAC,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACvE,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,oBAAoB,CAAC,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC7E,CAAA;AACF,CAAC;AA2BD;;;;;;;GAOG;AACH,MAAM,UAAU,sBAAsB,CAAC,aAAiC;IACvE,IAAI,CAAC,aAAa,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC;QAAE,OAAO,SAAS,CAAA;IAClE,IAAI,GAAW,CAAA;IACf,IAAI,CAAC;QACJ,GAAG,GAAG,YAAY,CAAC,aAAa,EAAE,MAAM,CAAC,CAAA;IAC1C,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,MAAe,CAAA;IACnB,IAAI,CAAC;QACJ,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAA;IACzB,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,MAAM,KAAK,IAAI;QAAE,OAAO,SAAS,CAAA;IACnE,MAAM,YAAY,GAAI,MAAqC,CAAC,YAAY,CAAA;IACxE,IAAI,YAAY,KAAK,SAAS;QAAE,OAAO,SAAS,CAAA;IAChD,IAAI,OAAO,YAAY,KAAK,QAAQ,IAAI,YAAY,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,CAAC;QAC9F,MAAM,IAAI,KAAK,CACd,sBAAsB,aAAa,4CAA4C;YAC9E,2BAA2B,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,GAAG,CAC3D,CAAA;IACF,CAAC;IACD,OAAO,YAAkC,CAAA;AAC1C,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,mBAAmB,CAClC,QAAwC,EACxC,IAAY,EACZ,MAAc,EACd,GAAW;IAEX,MAAM,QAAQ,GAAG,QAAQ,EAAE,CAAC,IAAI,CAAC,CAAA;IACjC,IAAI,CAAC,QAAQ,IAAI,OAAO,QAAQ,KAAK,QAAQ;QAAE,OAAO,SAAS,CAAA;IAC/D,MAAM,UAAU,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAA;IACnC,IAAI,CAAC,UAAU,IAAI,OAAO,UAAU,KAAK,QAAQ;QAAE,OAAO,SAAS,CAAA;IACnE,MAAM,GAAG,GAAG,UAAU,CAAC,GAAG,CAAC,CAAA;IAC3B,IAAI,CAAC,GAAG,IAAI,OAAO,GAAG,KAAK,QAAQ,IAAI,OAAQ,GAAqB,CAAC,SAAS,KAAK,QAAQ;QAAE,OAAO,SAAS,CAAA;IAC7G,OAAO,GAAoB,CAAA;AAC5B,CAAC;AAQD;;;GAGG;AACH,MAAM,UAAU,kBAAkB,CAAC,OAA2B;IAC7D,IAAI,CAAC,OAAO,IAAI,CAAC,UAAU,CAAC,OAAO,CAAC;QAAE,OAAO,SAAS,CAAA;IACtD,IAAI,GAAW,CAAA;IACf,IAAI,CAAC;QACJ,GAAG,GAAG,YAAY,CAAC,OAAO,EAAE,MAAM,CAAC,CAAA;IACpC,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,MAAe,CAAA;IACnB,IAAI,CAAC;QACJ,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAA;IACzB,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAA;IACjB,CAAC;IACD,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,MAAM,KAAK,IAAI;QAAE,OAAO,SAAS,CAAA;IACnE,MAAM,GAAG,GAAG,MAAiC,CAAA;IAC7C,MAAM,WAAW,GAAG,GAAG,CAAC,WAAW,CAAA;IACnC,MAAM,KAAK,GAAG,GAAG,CAAC,iBAAiB,CAAA;IACnC,MAAM,GAAG,GAAG,GAAG,CAAC,eAAe,CAAA;IAC/B,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,WAAW,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IACjG,IAAI,WAAW,CAAC,MAAM,KAAK,CAAC,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC,IAAI,GAAG,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,SAAS,CAAA;IACxF,OAAO;QACN,WAAW,EAAE,WAAyB;QACtC,gBAAgB,EAAE,KAAiB;QACnC,cAAc,EAAE,GAAe;KAC/B,CAAA;AACF,CAAC"}

package/out/word-consistency.d.ts

@@ -0,0 +1,60 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Per-word BIO tag-consistency repair (#727 + the fr.country / admin-token fragmentation class).
+ *
+ *   The model emits per-PIECE BIO labels. On rows where an admin token is adjacent to a non-latin
+ *   (byte-fallback) locality, carries diacritics, or is all-caps/reordered, the per-piece labels
+ *   can DISAGREE within a single word — `VERMONT` → `VER`[B-locality] + `MONT`[B-region], `Lozère`
+ *   → `Loz`[locality] + `ère`[B-region] — and the span decoder reads that as a tag-change
+ *   mid-token, fracturing the admin component. The model already KNOWS the word's tag (99.7% of
+ *   normal rows are unanimous); the defect is the lack of a word-level consistency constraint.
+ *
+ *   Fix (DeepSeek-Pro consult `contested-frag`, 2026-06-19): a SentencePiece word — a `▁`-started
+ *   piece + its non-`▁` continuations — must carry ONE tag. The tag is chosen by a
+ *   CONFIDENCE-WEIGHTED vote, NOT first-piece-wins: sum each piece's softmax mass per TAG TYPE
+ *   (B-X
+ *
+ *   - I-X collapsed; `O` included) across the word, argmax the type, and force `B-<type>` then
+ *       `I-<type>` (or all `O`). The near-certain `ère`→region then pulls the whole word to region,
+ *       healing both the fragment and the `Loz` bleed. Operates ONLY within a `▁`-delimited word,
+ *       so cross-word multi-token names ("Saint Paul" → two words) are untouched and the decoder's
+ *       existing cross-word merge still joins them.
+ *
+ *   Safety: a word whose pieces already agree is left byte-identical (no change). The vote includes
+ *   `O`, so a genuinely-`O` word stays `O` (no spurious spans).
+ *
+ *   Gate outcome (2026-06-19, fr-admin-split-gate + per-locale-f1, MAILWOMAN_WORD_CONSISTENCY=1): NOT
+ *   a clean win, so this ships DEFAULT-OFF. It heals clean-latin fragments (`PRUNIÈRES LOZÈRE
+ *   FRANCE` → region=`LOZÈRE`), but on OOD byte-soup rows where the per-piece confidence is itself
+ *   unreliable the vote AMPLIFIES that noise — e.g. `VERMONT, ウェストミンスター` lets the region word
+ *   absorb the adjacent CJK locality — and it net-regressed street −12.6 on the adversarial golden.
+ *   The confidence-weighted vote out-votes a stray mis-tagged piece only when the surviving pieces
+ *   are themselves trustworthy; on byte-fallback pieces that premise breaks. A confidence-gated
+ *   variant (skip the heal when the word's mean p(bestType) is below a floor, or when any piece is
+ *   a raw byte-fallback piece) is the path to a clean win — tracked on #727.
+ */
+export interface WordConsistencyResult {
+    /** A new per-piece label-index array, word-consistent (input is not mutated). */
+    labelIndices: number[];
+    /** PieceIndex → mean p(chosen type) across the word, for pieces in a word that was HEALED. */
+    healedConfidence: Map<number, number>;
+    /** Count of words whose labels were rewritten (0 = byte-identical to the input). */
+    healedWords: number;
+}
+/**
+ * Rewrite per-piece label indices so every `▁`-delimited word carries one tag, chosen by a
+ * confidence-weighted vote over the post-prior `emissions`. See the module docstring.
+ *
+ * @param pieces SentencePiece pieces (the `▁`-marked surface is the word-boundary signal).
+ * @param emissions Per-piece × per-label scores AFTER all priors/masks (the distribution the argmax
+ *   would see). Softmaxed per piece for the vote so each piece's confidence carries its weight.
+ * @param labels The BIO label vocabulary (index ↔ label).
+ * @param labelIndices The current per-piece decision (viterbi path or argmax). Not mutated.
+ */
+export declare function enforceWordConsistency(pieces: ReadonlyArray<{
+    piece: string;
+}>, emissions: ReadonlyArray<ReadonlyArray<number>>, labels: readonly string[], labelIndices: readonly number[]): WordConsistencyResult;
+//# sourceMappingURL=word-consistency.d.ts.map

package/out/word-consistency.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"word-consistency.d.ts","sourceRoot":"","sources":["../word-consistency.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqCG;AAKH,MAAM,WAAW,qBAAqB;IACrC,iFAAiF;IACjF,YAAY,EAAE,MAAM,EAAE,CAAA;IACtB,8FAA8F;IAC9F,gBAAgB,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IACrC,oFAAoF;IACpF,WAAW,EAAE,MAAM,CAAA;CACnB;AASD;;;;;;;;;GASG;AACH,wBAAgB,sBAAsB,CACrC,MAAM,EAAE,aAAa,CAAC;IAAE,KAAK,EAAE,MAAM,CAAA;CAAE,CAAC,EACxC,SAAS,EAAE,aAAa,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC,EAC/C,MAAM,EAAE,SAAS,MAAM,EAAE,EACzB,YAAY,EAAE,SAAS,MAAM,EAAE,GAC7B,qBAAqB,CA+EvB"}