@mailwoman/neural-weights-fr-fr 10.0.0 → 10.2.0

package/LICENSE.md

@@ -0,0 +1,36 @@
+# License — @mailwoman/neural-weights-fr-fr
+
+Generated by `mwops release write-rights-files` from package.json and model-card.json. Edit the generator or those files, never this one.
+
+## The terms Sister Software grants
+
+This package is distributed under `AGPL-3.0-only OR LicenseRef-Commercial`.
+
+Under the AGPL-3.0-only branch, three obligations attach: keep the copyright notices, license a
+modified copy under the same terms, and offer corresponding source to people who use a modified copy
+over a network (section 13). Under a commercial agreement, the source offer and the share-alike
+condition do not apply to the code and model artifacts Sister Software authors.
+
+- Open-source terms: https://github.com/sister-software/mailwoman/blob/main/LICENSE.md
+- Commercial terms, as executed at checkout: https://mailwoman.ai/license/terms/commercial-2026-10
+- Commercial terms, as a reference template that does not itself execute:
+  https://github.com/sister-software/mailwoman/blob/main/COMMERCIAL-LICENSE.md
+
+## What the commercial grant does not cover
+
+Section 2 of the commercial agreement limits the grant to the code and model artifacts Sister
+Software authors. It does not extend to third-party components distributed with Mailwoman, which
+stay under their own licenses. Any attribution or share-alike obligation attaching to those
+components continues to apply and is not waived or relicensed by a commercial agreement.
+
+## Where the sources are recorded
+
+`PROVENANCE.json` in this package records, per artifact, what the repository holds about its
+inputs and which questions it cannot answer. Read it as the record it is: a package with no recorded
+training attribution is a package whose model card records none, which is not the same as a package
+with no attributable inputs. No artifact in this package is recorded as cleared for any particular
+use.
+
+`THIRD_PARTY_NOTICES.md` in the repository lists the components Mailwoman distributes and the terms
+each one carries. Reference data an operator downloads separately carries the terms of whoever
+published it, and nothing here states what those are.

package/PROVENANCE.json

@@ -0,0 +1,125 @@
+{
+	"$comment": "Generated by `mwops release write-rights-files` from package.json and model-card.json. Edit the generator or those files, never this one.",
+	"package": "@mailwoman/neural-weights-fr-fr",
+	"package_version": "10.0.0",
+	"license": "AGPL-3.0-only OR LicenseRef-Commercial",
+	"model_card_version": "6.5.0",
+	"version_series": "overlay",
+	"version_series_meaning": "This package is a data-only overlay, and model_card_version names this overlay's own artifacts. The model graph its rows decode through is the one base_weights names, at that package's own model_card_version. Neither number is package_version, which moves in lockstep across every mailwoman workspace on every release.",
+	"base_weights": "@mailwoman/neural-weights-en-us",
+	"inherited_lineage": {
+		"package": "@mailwoman/neural-weights-en-us",
+		"package_version": "10.0.0",
+		"model_card_version": "9.1.0",
+		"chain": ["@mailwoman/neural-weights-fr-fr", "@mailwoman/neural-weights-en-us"],
+		"note": "These entries are @mailwoman/neural-weights-en-us's record of what its model graph was trained on. This package ships no graph and contributed none of these rows. Read them as the base's lineage rather than as a claim about this package's own artifacts, which are listed under artifacts and attributed under training_attribution.",
+		"attribution": [
+			{
+				"text": "HM Land Registry — Price Paid Data (OGL v3.0): the `synth-gb-v1` training extract (800,000 rows) NEW in 6.7.0. Crown copyright, attribution required. The GB placetype-pair index is built from the same source and ships in `neural-weights-en-gb`, whose own card carries its attribution.",
+				"license_named": "OGL v3.0"
+			},
+			{
+				"text": "LINZ-derived OpenAddresses NZ (CC-BY 4.0): the `synth-nz-v2` training extract (800,000 rows) NEW in 6.7.0.",
+				"license_named": "CC-BY 4.0"
+			},
+			{
+				"text": "OpenAddresses ES tuples (CC-BY per OA source): the `synth-es-pedania-v1` training extract (800,000 rows, below-municipio pedanía place names) NEW in 6.7.0.",
+				"license_named": "CC-BY per OA source"
+			},
+			{
+				"text": "BAN — Base Adresse Nationale (DINUM / IGN, Licence Ouverte 2.0): the `synth-fr-lieudit-v1` training extract (`nom_ld` lieu-dit field) NEW in 6.7.0, additional to the existing `synth-fr-fragment` extract (144,865 rows, 6.4.0).",
+				"license_named": "DINUM / IGN, Licence Ouverte 2.0"
+			},
+			{
+				"text": "Overture Maps Foundation — Addresses theme (per-country register license; Overture declares no identifier of its own for this theme, so the terms are the upstream register's): the `overture-latam` CA/MX training extract (456,230 rows, 6.6.0). Attribution-only, not share-alike. Read 2026-09-25 from https://docs.overturemaps.org/attribution/.",
+				"license_named": "456,230 rows, 6.6.0"
+			},
+			{
+				"text": "OpenAddresses CZ — ČÚZK RÚIAN (CC-BY 4.0): tokenizer-splice training text + the oa-cz coord eval sets",
+				"license_named": "CC-BY 4.0"
+			},
+			{
+				"text": "OpenAddresses PL — GUGiK / PRG (public, BDOT-derived): tokenizer-splice training text + the oa-pl coord eval sets",
+				"license_named": null
+			},
+			{
+				"text": "OpenAddresses SK — MINV/ZBGIS (CC-0/CC-BY per OA source): oa-sk coord eval set (free-rider validation)",
+				"license_named": "CC-0/CC-BY per OA source"
+			},
+			{
+				"text": "OpenAddresses SI — GURS (CC-BY per OA source): oa-si coord eval set (free-rider validation)",
+				"license_named": "CC-BY per OA source"
+			},
+			{
+				"text": "See THIRD_PARTY_NOTICES.md for the standing OpenAddresses / GeoNames / WOF attribution.",
+				"license_named": null
+			}
+		],
+		"unresolved": null
+	},
+	"artifacts": [
+		{
+			"path": "postcode-fr.bin",
+			"role": "postcode-binary",
+			"md5": null,
+			"digest": "unrecorded"
+		},
+		{
+			"path": "pair-index-fr.bin",
+			"role": "placetype-pair-index",
+			"md5": null,
+			"digest": "unrecorded"
+		},
+		{
+			"path": "fst-fr-fr.bin",
+			"role": "gazetteer-fst",
+			"md5": null,
+			"digest": "unrecorded"
+		},
+		{
+			"path": "fst-street-morphology.bin",
+			"role": "gazetteer-fst",
+			"md5": null,
+			"digest": "unrecorded"
+		},
+		{
+			"path": "anchor-lexicon-v1.json",
+			"role": "lexicon",
+			"md5": null,
+			"digest": "unrecorded"
+		},
+		{
+			"path": "locality-surface-lexicon-v7.json",
+			"role": "lexicon",
+			"md5": null,
+			"digest": "unrecorded"
+		},
+		{
+			"path": "street-type-lexicon-v3.json",
+			"role": "lexicon",
+			"md5": null,
+			"digest": "unrecorded"
+		},
+		{
+			"path": "country-surface-lexicon-v1.json",
+			"role": "lexicon",
+			"md5": null,
+			"digest": "unrecorded"
+		}
+	],
+	"model_graph": null,
+	"training_attribution": {
+		"status": "none-recorded-in-this-package",
+		"entries": [],
+		"by_use": {}
+	},
+	"attribution_recorded_in_another_package": [],
+	"corpus_version": "v257 fragment-campaign corpus (extract-v5) re-tokenized through v0.9.0-multisplice; span-boundary aux head from BIO labels.",
+	"effective_training_sources": null,
+	"tokenizer_version": "0.9.0-multisplice",
+	"unresolved": [
+		"No digest is recorded for 8 of 8 declared artifacts: postcode-fr.bin, pair-index-fr.bin, fst-fr-fr.bin, fst-street-morphology.bin, anchor-lexicon-v1.json, locality-surface-lexicon-v7.json, street-type-lexicon-v3.json, country-surface-lexicon-v1.json. An unrecorded digest is not a verified one.",
+		"This package's model card records no training attribution of its own. It ships 8 artifacts and no model graph, so its own inputs are unattributed here while the 10 entries under inherited_lineage describe @mailwoman/neural-weights-en-us's graph.",
+		"Which records trained the model this package ships or inherits is unresolved. `buildCorpus` writes a frozen `TRAINING_SOURCES.json` naming every source that contributed and under which terms, `effective-manifest.run.ts` derives from it and an `audit_epoch_mixture` output what one config's epoch drew, and no released model was built from a corpus carrying either. The corpus named above is the recipe's corpus rather than a record of what reached the model."
+	]
+}

package/README.md

@@ -1,5 +1,5 @@
 ---
-license: agpl-3.0
+license: AGPL-3.0
 language:
   - fr
 library_name: onnx
@@ -50,7 +50,7 @@ Mailwoman is a **calibrated, retrieval-augmented sequence labeler over a
 microlanguage** — coupled to a gazetteer that resolves its output to
 coordinates. This bundle is the sequence-labeler half: a small transformer
 encoder doing BIO token classification over the address schema. It is not an LLM
-and nothing about it is generative; for a closed label set over short strings,
+and none of it is generative; for a closed label set over short strings,
 boring NER is a feature.
 
 The design splits the problem in two: **the model learns the grammar, the
@@ -65,27 +65,27 @@ Knowledge informs; it never overrides.
 
 ## Intended use
 
-Parsing free-text French postal addresses into structured components
+The model parses free-text French postal addresses into structured components
 (country, region, locality, dependent_locality, postcode, subregion, cedex,
 venue, street, house_number) for **geocoding** — resolving a parsed address to
 coordinates via a gazetteer/resolver. The model is the parsing front-end of that
-pipeline, not a standalone geocoder.
+pipeline rather than a standalone geocoder.
 
 ## Ship-config requirement (read before using)
 
 The Mailwoman model expects the soft anchor + gazetteer channels fed at
-inference. Running it with those channels off is out-of-distribution and silently
+inference. A run with those channels off is out-of-distribution and silently
 collapses the admin tags (country/region/locality/postcode) — an anchor-off
 metric on an anchor-trained model is systematically misleading. Construct the
 scorer through `@mailwoman/neural`'s `createScorer` (the canonical
-`ProductionScorer`), which reads the bundle's `requires`/channel contract and
+`ProductionScorer`), which reads the bundle's `requires`/channel interface and
 **fails closed** if a declared channel isn't fed. Do not hand-wire the raw ONNX
 session with the anchor input zero-filled. Every eval here is
 **anchor-on**.
 
 ## Evaluation
 
-We grade the **assembled coordinate** (the resolved place), not raw label-F1 in
+We grade the **assembled coordinate** (the resolved place) rather than raw label-F1 in
 isolation. For FR, the model resolves addresses well when given the correct
 default country (the gazetteer holds ~114k FR places); on a real FR sample the
 neural front-end out-parses the rules baseline it replaces on locality-match.
@@ -115,7 +115,7 @@ Mailwoman confidences are isotonic-calibrated (PAVA) against held-out data and
 applied **opt-in** via `@mailwoman/core`'s `createCalibrator`; default parse
 output is byte-stable when calibration is omitted.
 
-## Training
+## Train
 
 From-scratch (no pretrained base) on an NVIDIA A100 (Modal cloud), CE loss only
 (the dual CRF loss diverged and was retired; CRF is inference-only Viterbi). The
@@ -136,8 +136,8 @@ card carries the full corpus + recipe provenance.
 
 ## License & links
 
-- **License:** [AGPL-3.0-only](https://www.gnu.org/licenses/agpl-3.0.html).
+- **License:** [AGPL-3.0-only](https://www.gnu.org/licenses/AGPL-3.0.html).
 - **Docs & concepts:** https://mailwoman.ai
 - **Loader / scorer API:** `@mailwoman/neural` (`createScorer`).
-- The functional contract for this bundle is `model-card.json`; this `README.md`
+- The functional interface for this bundle is `model-card.json`; this `README.md`
   is the HuggingFace-facing card.

package/model-card.json

@@ -280,7 +280,7 @@
 			"md5": "(staged at release time — see $comment_md5)",
 			"$comment_pix2": "PIX2 (schema 3, 2026-08-04): every record now carries the PARENT's ComponentTag alongside the child's — +1 byte per pair, pair COUNT unchanged. For FR the recorded parent tag is always `locality`: BAN's parent surface is `nom_commune`, the commune, which is the French postal locality line. Schema 1 and 2 binaries are refused; the artifact must be rebuilt.",
 			"delta_calibration": "δ=10, transitionBeta=5, parentDelta=5 (whole-edge parent bias, #46 — the record's own parentTag gets +5 over the parent window). FR is the leg that CAUGHT that mechanism's one real defect: biasing the parent's whole comma segment put the postcode inside `locality` and took fr-lieudit-golden.jsonl from 96.3% to 0.0% at δ≥6, invisible at δ=4. Fixed by scoping the parent write to the probe key's pieces; FR then held 77/80 at every δ through 20. See docs/records/evals/2026-08-04-pix1-whole-edge-verdict.md. The δ/β pair is inherited from the GB calibration, per R5's finding that the classifier's dependent_locality deficit is uniform across locales rather than GB-specific. The R6 bars were measured at exactly that pair.",
-			"country_check": "hard-blocked on the header's country (fr).",
+			"country_check": "hard-enforced: applies only when the header's country matches (fr).",
 			"not_committed": "derived artifact, .gitignore'd — rebuilt via `mailwoman gazetteer pair-index --out neural-weights-fr-fr --country fr --delta 10 --transition-beta 5 --ban-dir <corpus/sources/ban>`",
 			"required_decode_fixes": "This artifact is INERT without two locale-general probe-chain fixes that shipped alongside it: a LEADING postcode strip on parent-candidate keys (France writes '12210 Montpeyroux'; the probe previously stripped only trailing postcodes, the Anglo/NZ convention) and NEWLINE as a segment boundary (computeGroupSegments scanned for commas only, so multi-line addresses collapsed to one segment and the segment path went structurally inert). With the artifact but without them the golden board stayed at 0/80.",
 			"$comment_md5": "REPRODUCIBILITY: a PIX1 header embeds `buildDate`, so a rebuild from identical sources produces different bytes and a different md5. This md5 therefore identifies the artifact STAGED TO HUGGING FACE for this release — it is a provenance record of the shipped bytes, NOT a check a local rebuild can reproduce. A mismatch after rebuilding is expected and is not corruption; verify pair COUNT and header delta/beta instead.",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mailwoman/neural-weights-fr-fr",
-  "version": "10.0.0",
+  "version": "10.2.0",
   "description": "Mailwoman neural-classifier weights for locale 'fr-fr'. Data-only overlay — shares the base model.onnx + tokenizer.model from @mailwoman/neural-weights-en-us (they are byte-identical); ships only the fr-specific data siblings. Loaded by @mailwoman/neural at runtime.",
   "license": "AGPL-3.0-only OR LicenseRef-Commercial",
   "repository": {
@@ -10,6 +10,8 @@
   },
   "files": [
     "model-card.json",
+    "LICENSE.md",
+    "PROVENANCE.json",
     "postcode-fr.bin",
     "pair-index-fr.bin",
     "fst-fr-fr.bin",
@@ -30,13 +32,6 @@
     "!scripts/**",
     "!test/**"
   ],
-  "imports": {
-    "#*": {
-      "types": "./out/*.d.ts",
-      "node": "./out/*.js",
-      "default": "./out/*.js"
-    }
-  },
   "exports": {
     "./*": "./*"
   },
@@ -44,21 +39,14 @@
     "access": "public",
     "exports": {
       "./*": "./*"
-    },
-    "imports": {
-      "#*": {
-        "types": "./out/*.d.ts",
-        "node": "./out/*.js",
-        "default": "./out/*.js"
-      }
     }
   },
   "dependencies": {
-    "@mailwoman/neural-weights-en-us": "10.0.0"
+    "@mailwoman/neural-weights-en-us": "10.2.0"
   },
   "devDependencies": {
-    "@mailwoman/core": "10.0.0",
-    "@mailwoman/resolver-wof-sqlite": "10.0.0"
+    "@mailwoman/core": "10.2.0",
+    "@mailwoman/resolver-wof-sqlite": "10.2.0"
   },
   "mailwoman": {
     "baseWeights": "@mailwoman/neural-weights-en-us"