{
	"name": "neural-weights-en-us",
	"version": "9.1.0",
	"model_lineage": "9.1.0: v4.4.0-suffix-boundary-v2 step-60000 — THE SUFFIX-BOUNDARY CURE on the 9.0.0 anchor-cure base (same from-scratch one-recipe provenance; int8 md5 98a49b5ca2948cdbe1dd981d60ef637d, fp32 0f9273a37db14fdd86a6d2c8806f8494, quantized on Modal in the training image). 60k cosine s42 on the v0.19.0 overlay: the extract-v2 diversity-first suffix-boundary generator (positional licensing lifted ordinary-source label correctness ~21%→94%; real facility shells; repetition reduced) + loader exclusion of the suffix extracts from the augmentation pool. Closes #1569 (the absorbed-suffix class): B1 target 119/128, contrast 39/48 (scratchpad/v440-step-060000/score.json); all 17 aggregate fp32 campaign legs PASS incl. the French bare-street floor. Release grade 2026-08-11: package-shaped int8 promotion check PASS 17/17 on v9.0.0-base floors (us.postcode 97.2, us.street 75.1, us.country_homograph_f1 83.3, fr.cedex_real 98.8, de.native_locality 91.6, arena.perturb 64.0); fp32↔int8 deltas ≤0.3pp on all 17 legs vs the 1.5pp cap (dual-artifact check); eval-model presets byte-identical to the shipped 9.0.0 weights. Receipts scratchpad/check-v440-60k-int8 (dual) + check-v440-60k-int8-wc (package-shaped). Bookkeeping corrected at this reduce: files_md5['model.onnx'] had not been flipped since the 6.7.0 era — it still named the v4.0.1 base's c968c24ad1a3e59658ff1d92c5720575 through the 9.0.0 reduce (the published 9.0.0 tarball carries cd31cfc346075d7c044d53499f79f141 bytes under a c968c24a card record), and link-dev-weights DEFAULT_MODEL pointed at the same stale file; both now track the shipping artifact, and files_md5 records locality-surface v7 (the generation files[] has declared since 9.0.0) instead of v6. || Prior base: 9.0.0-base: v4.2.0-base-anchor-v2 step-60000 — THE ANCHOR-CURE BASE (from scratch, no init_from, one recipe, one provenance; int8 md5 cd31cfc346075d7c044d53499f79f141). 60k cosine s42 on v0.17.0-batch = v0.15.0-venue + synth-sub-venue (120k rows, designator×identifier venue-interior structure, #1486) + synth-gb (800k HM Land Registry PPD rows, OGL v3.0) at the doses the 2k probe cleared. THE named delta vs the 7.0.0 base recipe: anchor_lookup_path → anchor-lookup v2 (2,286,339 keys, 2,214,803 letter-bearing; DE/FR/US/GB/NL/ES/IT) — the first base whose anchor channel fed every letter-bearing LOCALE_ORDER slot (the 7.0.0-era pilot lookup carried 67,708 keys, zero letters, US/DE/FR only — #1467's root cause). Also corrected into this base: dependent_locality class weight 0.3→1.0 (the R5 diagnosis). Check record docs/records/evals/2026-08-05-v420-base-anchor-v2-run-b.md + diagnosis #1515; promotion check v9.0.0-base PASS on golden v0.1.3 (the split-canonical answer key — us.street 75.2 vs 7.0.0's 71.7 on the corrected instrument; the v0.1.2 miss was the folded key crediting non-splitting). || Prior base: 7.0.0-base: v4.0.1-base-60k step-60000 — THE FROM-SCRATCH BASE (the v8.3.0 golden master; no init_from, one recipe, one provenance). 60k cosine s42 on v0.15.0-venue (v0.13.0-latam + the FR-order-aware synth-house-venue extract, 116,244 rows — the run-2 contingency that fixed the venue-led class the fine-tune lineage had inherited invisibly). Evidence bundle native from step 0: street_type[1] + locality_surface[2] channels, absence + false-evidence (0.15) curricula, digit-guarded four-law lexicons (street v3 / locality-surface trained on v7, SHIPPED with v6 — the G8 DE-fold revert: DE evidence read net-negative on its own board, the pre-registered per-ingredient contingency), ordinal 0.3 + punct-drop 0.3 + directional/region/glue augmentation, gazetteer curriculum ON. FIRST model whose locality channel actually trained (#1349: every bundle-era fine-tune's locality tensors were dropped by the trainer; fixed #1350). Ships beside its FISHER ARTIFACT (fisher-diag-v1.npz, diagonal empirical Fisher over the final 2k steps, all 39.26M trainables — the EWC brake for fine-tunes; HF/R2-distributed, the runtime never reads it). Prior lineage record: 6.7.0-bundle: v3.24.0-bundle-ordinal step-4000 — the Option-A evidence bundle (5-run productionization arc v3.19→v3.23, 2026-07-28). init_from v385 (6.6.0) step-8000 on v385's OWN feed (v0.13.0-latam — the deploc probe passengers removed after the flip-census/feed attribution), + street_type[1] + locality_surface[2] evidence channels, absence curriculum + evidence-noise 0.15, four-law curated lexicons (street-type v3 / locality-surface v6) with rules.digit_guard (evidence never paints beside a Unicode-Nd digit — the guard that turned alnum-housenumber from −0.325 to +0.013 lower+heal). LR 1.0e-5 constant, 8k steps, best save 4000 (the reproducible mid-run peak; 8k over-trust decay observed in all five runs). | v3.24 (2026-07-29): +augment_ordinal_prob 0.3 (5th↔Fifth street swaps) — the 8.2.0 pre-ship metamorphic catch trained out; world-city fixtures clean under the street-context check.",
	"phase": "SHIPPED (9.1.0, 2026-08-11, npm 9.1.0) — the suffix-boundary cure. B1 119/128 target / 39/48 contrast; promotion check v9.0.0-base PASS 17/17 package-shaped int8 with fp32↔int8 ≤0.3pp on every leg (cap 1.5, restored to the spec at this reduce); presets byte-identical to 9.0.0. Ship contract unchanged from 9.0.0 (span_mode 'shaped', lexicons resolved from this card by name, locality-surface v7). || Prior shipped record: SHIPPED (9.0.0-base, 2026-08-06, npm 9.0.0) — the anchor-cure base. gb-golden exact postcode 318/318 ×3 registers with the channel FED; Fisher slot-4 11.28% w/ CA/JP exact-zero controls; GB dep-loc 0→205/207 under this base package; promotion check v9.0.0-base PASS on golden v0.1.3 (split-canonical key; us.street 75.2 vs baseline 71.7, us.micro 87.5 vs 85.7, fr.region 81.2 vs 50.5; one watch: us.unit_real 93.9 vs 97.0, n=34). Ship contract: span_mode 'shaped' (required — the alnum-run scan cannot produce GB/NL unit keys), lexicons resolved from this card by name (locality-surface v7 = the trained generation, #1510), shaped keyer case-folds (lowercase register 0/120→106/120, #1512). Territory rows + truncation pair track as improvement_target (#1515, #1519). || Prior shipped record: SHIPPED (6.7.0-bundle, 2026-07-29, npm 8.2.0) — the Option-A evidence bundle: v3.24.0-bundle-ordinal step-4000 (init_from v385 on v385's own feed) + street_type/locality_surface channels + four-law digit-guarded lexicons (street v3 / locality-surface v6) + inputMode register blocking (Decision A: fragmented feeds channels, formatted runs the trained absence identity; per-endpoint defaults). Cleared 17/17 through the ratified v6.0.0-shipped-baseline check (fr.cedex under the noise-honest 82.2 floor, NAMED WATCH: re-anchors at the next from-scratch retrain); fragment register: homonym +0.245 asis / +0.765 lower+heal, ALL +0.261 lower+heal. Full verdict ledger in eval.check_2026_07_28_v3230_bundle_guard. || Prior shipped record: SHIPPED (6.6.4, 2026-07-26) — bundle revision on UNCHANGED model bytes (v385, model.onnx md5 700f3cf4c08f31536f0bc053c281b19e): the per-locale FST gazetteers (fst-en-us/fr-fr/en-gb.bin) are REBUILT with build-time degenerate-surface curation (`mailwoman gazetteer build fst`, policy `degenerate-surface-exclusion v1.1` recorded in each artifact trailer) — bare function words (incl. the case-folded 'LA'/'OP' initialism aliases), bare street-type words, and all-function-word compositions are pruned from the bias index (the ASR-contextual-biasing 'curate the bias list' discipline; receipts + battery in PR #1332: FR street-particle 290→298, bare-street to P0 parity, golden us/fr byte-equal, gauntlet PASS ×3 layers; residual 3 rows are true homographs — the Option-A class). Also code-side in the same release: the pipeline pins the #1315 street-context check at positiveScale 0 (D2, closed #1320 — FR admin-street-homonym at exact P0 parity). Prior record: SHIPPED (6.6.2, 2026-07-24) — bundle revision on UNCHANGED model bytes (v385, model.onnx md5 700f3cf4c08f31536f0bc053c281b19e, same files_md5 as 6.6.0): the en-gb overlay's pair-index-gb.bin is rebuilt at δ=10 and @mailwoman/neural 7.7.0's auto probe chain activates GB dependent_locality on this shipped model (see neural-weights-en-gb/model-card.json 6.6.1 for the capability record). The 6.7.0 staged-candidate lineage is CLOSED: no checkpoint of v3.10–v3.13 passed the Gauntlet's metamorphic layer; five mechanisms falsified (starvation, encoder drift, save anomaly, comma-share, LR-hold/anneal); stop rule executed three times; the break is established in the first ≤1k steps of encoder+head co-training and is stable. Model-side redesign is unhurried research — the dep-loc capability no longer depends on it. Falsification record: docs/superpowers/plans/2026-07-23-deploc-redesign-dossier.md + .superpowers/sdd/task-8-report.md. 6.6.2 adds the en-nz overlay (pair-index-nz.bin, LINZ census, δ=10) + the identical-adjacent-segment (x,x) rule in @mailwoman/neural 7.8.0 — NZ dependent_locality live (246/246, 100% as-written; the repeated-name convention handled as registry evidence). Model bytes STILL v385/unchanged.",
	"license": "AGPL-3.0-only OR LicenseRef-Commercial",
	"locale": "en-us",
	"training": {
		"corpus_version": "v0.13.0-latam (v385's own feed, VERBATIM data section incl. punct_drop 0.3 + all source weights except synth-po-box-cedex 1.5→3.0 — the fine-tune replay boost) — the deploc additions (synth-gb 10.0 / nz / es-pedania / fr-lieudit) are PROBE PASSENGERS removed after the v3.22 feed attribution (GB street-boundary transfer + CA po-box dilution).",
		"tokenizer_version": "0.9.0-multisplice",
		"steps": 8000,
		"best_step": 4000,
		"hardware": "NVIDIA A100-SXM4-40GB (Modal cloud)",
		"recipe": "v3.24.0-bundle-ordinal-8k.yaml — v3.23 + augment_ordinal_prob 0.3 (the ONE delta); — init_from v385 (6.6.0) step-8000, --resume none, LR 1.0e-5 constant, 8k. Channels: street_type (feature_dim 1) + locality_surface (feature_dim 2) painted by the SHARED painter (gazetteer_anchor.py ↔ gazetteer-inference.ts, parity fixture evidence-parity-v2.json) from the digit-guarded lexicons; evidence_curriculum true (ramped per-row zero-out, independent draws) + evidence_noise_prob 0.15 (false-evidence curriculum: rolled painted patterns / synthetic false positives, labels gold). Pre-registered checks G1-G6 in the recipe header; the candidate save is 4000 per the over-trust watch.",
		"tokenizer_splice": "tokenizer_splice.py build-tokenizer --locales fr --per-file-cap 2000000 over OA fr/countrywide.csv (seed 42, 350k-line corpus, unigram vocab 24000) -> splice 2406 non-ASCII FR pieces into v0.7.1-nsplice (English byte-identity asserted). UNCHANGED from v240 (md5 04995524).",
		"data_attribution": [
			"HM Land Registry — Price Paid Data (OGL v3.0): the `synth-gb-v1` training extract (800,000 rows) NEW in 6.7.0, and the `pair-index-gb.bin` placetype-pair index (`neural-weights-en-gb`, this same task). Crown copyright, attribution required.",
			"LINZ-derived OpenAddresses NZ (CC-BY 4.0): the `synth-nz-v2` training extract (800,000 rows) NEW in 6.7.0.",
			"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.",
			"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).",
			"Overture Maps Foundation — Addresses theme (CDLA-Permissive-2.0): the `overture-latam` CA/MX training extract (456,230 rows, 6.6.0). Attribution-only, not share-alike.",
			"OpenAddresses CZ — ČÚZK RÚIAN (CC-BY 4.0): tokenizer-splice training text + the oa-cz coord eval sets",
			"OpenAddresses PL — GUGiK / PRG (public, BDOT-derived): tokenizer-splice training text + the oa-pl coord eval sets",
			"OpenAddresses SK — MINV/ZBGIS (CC-0/CC-BY per OA source): oa-sk coord eval set (free-rider validation)",
			"OpenAddresses SI — GURS (CC-BY per OA source): oa-si coord eval set (free-rider validation)",
			"See THIRD_PARTY_NOTICES.md for the standing OpenAddresses / GeoNames / WOF attribution."
		]
	},
	"architecture": {
		"hidden_size": 384,
		"num_hidden_layers": 6,
		"num_attention_heads": 6,
		"intermediate_size": 1536,
		"max_position_embeddings": 128,
		"vocab_size": 73143,
		"num_labels": 33,
		"params": "~29M (6-layer/384-hidden encoder + 73,143-piece multisplice embedding; init_from v264, 8k fine-tune w/ a non-exported span-boundary aux head)",
		"crf_at_training": false,
		"crf_at_inference": true,
		"phrase_priors": true
	},
	"requires": {
		"$comment": "The SHIP-CONFIG this model was TRAINED against (#718): anchor+gazetteer+country+conventions PLUS the Option-A evidence bundle (street_type/locality_surface). createScorer FAILS CLOSED on a declared-but-unfed channel. inputMode=formatted runs both evidence channels deliberately OFF via the register check (Decision A) — a DECLARED ablation, not a missing feed. Channel slot metadata doubles as the explain-endpoint contract (GTM B7).",
		"anchor": {
			"required": true,
			"span_mode": "shaped",
			"$comment": "span_mode 'shaped' is the 9.0.0 SHIP OBLIGATION (recipe header + ROAD_TO_V9 A2): the train painter keys postcodes as shape-detected spans, space-stripped upper ('SW1A 2AA' → 'SW1A2AA'); the default alnum-run scan can NEVER produce that key, so omitting this leaves GB/NL postcodes unanchored at serve — the mirror image of #1467. createScorer fails closed on a shaped-declaring card paired with an artifact whose keys the scan cannot reach."
		},
		"gazetteer": {
			"required": true
		},
		"country": {
			"required": true
		},
		"conventions": {
			"required": true,
			"mode": "auto"
		},
		"bridge": {
			"required": false
		},
		"suppress_gazetteer_near_postcode": true,
		"street_type": {
			"required": true,
			"feature_dim": 1,
			"slots": ["street_type"],
			"lexicon": "street-type-lexicon-v3.json"
		},
		"locality_surface": {
			"required": true,
			"feature_dim": 2,
			"slots": ["locality", "locality_homograph"],
			"lexicon": "locality-surface-lexicon-v7.json"
		}
	},
	"labels": [
		"O",
		"B-country",
		"I-country",
		"B-region",
		"I-region",
		"B-locality",
		"I-locality",
		"B-dependent_locality",
		"I-dependent_locality",
		"B-postcode",
		"I-postcode",
		"B-subregion",
		"I-subregion",
		"B-cedex",
		"I-cedex",
		"B-venue",
		"I-venue",
		"B-street",
		"I-street",
		"B-house_number",
		"I-house_number",
		"B-street_prefix",
		"I-street_prefix",
		"B-street_suffix",
		"I-street_suffix",
		"B-unit",
		"I-unit",
		"B-po_box",
		"I-po_box",
		"B-intersection_a",
		"I-intersection_a",
		"B-intersection_b",
		"I-intersection_b"
	],
	"components_supported": [
		"country",
		"region",
		"locality",
		"dependent_locality",
		"postcode",
		"subregion",
		"cedex",
		"venue",
		"street",
		"house_number",
		"street_prefix",
		"street_suffix",
		"unit",
		"po_box",
		"intersection_a",
		"intersection_b"
	],
	"notes": "v3.23.0-bundle-guard step-4000, STAGED not promoted (6.7.0-bundle). The evidence bundle is a FRAGMENT-REGISTER capability shipped under the operator's Decision A (2026-07-28): inputMode ∈ {fragmented, formatted} — explicit on CLI/API, per-endpoint register defaults (validation/batch → formatted, autocomplete/demo → fragmented, plain parse → shape classifier). Formatted mode runs the channels OFF (the trained absence identity) and scores the ablated column: v6.0.0-shipped-baseline PASS 17/17 (us.micro 87.1 vs shipped 86.8; fr.cedex 83.3 under the noise-honest 82.2 floor — NAMED WATCH: plateaued fine-tune forgetting, ~10/430 CA rows, re-anchors at the next from-scratch retrain). Fragment mode feeds both channels: admin-street-homonym +0.245 asis / +0.765 lower+heal (ON 0.948), bare-street +0.583 lower, particle +0.407, ALL +0.261; house-number classes guard-protected (+0.013/+0.030). Full verdict ledger: .superpowers/sdd/progress.md 2026-07-28; canary fixture data/eval/external/evidence-collision-canary.jsonl.",
	"format": {
		"model": "ONNX int8 dynamic (quantized from fp32, opset 17)",
		"tokenizer": "SentencePiece unigram, byte_fallback=true, vocab_size=73143 (v0.9.0-multisplice; ships WITH the model, both runtimes load it from this bundle)",
		"max_sequence_length": 128,
		"opset": 17,
		"fp32_size_mb": 146.6,
		"int8_size_mb": 37.6,
		"pair_index_us_bin": "PIX1 binary (schema 3), 1,341,264 bytes, 47,878 pairs, δ=10, β=5, parentDelta=5"
	},
	"files": {
		"model": "model.onnx",
		"tokenizer": "tokenizer.model",
		"model_card": "model-card.json",
		"calibration": "calibration.json",
		"calibration_per_locale": "calibration-per-locale.json",
		"street_type_lexicon": "street-type-lexicon-v3.json",
		"locality_surface_lexicon": "locality-surface-lexicon-v7.json",
		"pair_index": "pair-index-us.bin",
		"$comment_postcode_anchor": "Added 2026-08-06 (#1516). The package has always shipped this binary (release.config.json softFeed.postcodeDBByCountry.us; link-dev-weights builds it locally) — it was simply never DECLARED here, and this block is now read: `readDeclaredArtifactFile` uses it to decide whether an unfed anchor channel is a broken package (declared, missing) or a supported posture (never declared, e.g. en-gb under #1476). Undeclared meant a missing postcode-us.bin degraded in silence, which is worth 3-4 baseline gauntlet cases.",
		"postcode_anchor": "postcode-us.bin"
	},
	"calibration": {
		"method": "isotonic-regression (PAVA) over per-span softmax confidence; OPT-IN via core/decoder createCalibrator",
		"held_out_ece_raw": 0.0698,
		"held_out_ece_calibrated": 0.0017,
		"note": "RE-FIT on v5.3.0 (2026-07-04): calibration.json + calibration-per-locale.json are now fitted on THIS model's confidence distribution (35,600 spans over the 8k held-out set), not carried from v4.x. The from-scratch model was under-confident (meanConf 0.913 vs acc 0.980, gap -0.067); the fit corrects held-out OA ECE 0.0698→0.0017 (overall 0.0677→0.0028, per-locale NL 0.166→0.000 / DE 0.114→0.003). Reproduce: collect-span-confidences.ts → fit-isotonic-calibration.py --model-version 5.3.0. Apply via @mailwoman/core/decoder's createCalibrator; default parse output is byte-stable when omitted."
	},
	"base_relpath": "models/quantized/model-v3110-deploc-feed-step-008000-int8.onnx under $MAILWOMAN_DATA_ROOT; training output on Modal volume mailwoman-training:/output-v3110-deploc-feed-s42/checkpoints/step-008000",
	"eval": {
		"check_2026_07_23_v3110_deploc_feed_8k": {
			"$comment": "Task 8 (placetype-pair-prior arc, ship-prep) full battery + Gauntlet, feed-8k vs v385 (6.6.0), all measured fresh this session (ONNX CPU int8). Full transcript: .superpowers/sdd/task-8-prep-report.md; the calibration + checkpoint-choice battery this check builds on: .superpowers/sdd/task-7-report.md.",
			"verdict": "SHIP-BLOCKED — every pre-registered guard PASSES; the Gauntlet (a standing ship-check requirement, not one of the six pre-registered guards) FAILS on its metamorphic layer. Not promoted pending operator triage.",
			"golden_us_fr": {
				"us_micro_f1": {
					"v385_fresh": 0.867,
					"feed_8k": 0.862,
					"delta_pp": -0.5,
					"bar": "±0.7pp of v385",
					"result": "PASS"
				},
				"fr_micro_f1": {
					"v385_fresh": 0.899,
					"feed_8k": 0.911,
					"delta_pp": 1.2,
					"bar": "±0.7pp of v385",
					"result": "outside magnitude, favorable direction"
				}
			},
			"fr_fragment_bare_locality": {
				"v385_fresh": [0.983, [0.964, 0.991]],
				"feed_8k": [0.988, [0.971, 0.995]],
				"feed_2k_for_contrast": [0.665, [0.617, 0.709]],
				"bar": ">= 0.90",
				"result": "PASS (feed-8k); feed-2k FAILS outright — the deciding reason feed-8k is the ratified ship checkpoint over feed-2k"
			},
			"digit_bare_street_hn_check_revision": {
				"$comment": "Operator-ratified 2026-07-23 (Teffen Ellis). The pre-registered bar (0.902) was a same-lineage GB-probe checkpoint's own number, not a real ship baseline for a v385-forked checkpoint. Restated against v385's own fresh baseline, measured same session same board.",
				"v385_fresh_baseline": 0.755,
				"feed_8k": [0.868, [0.831, 0.897]],
				"delta_pp": 11.3,
				"old_bar_0_902_probe_anchor": "historical context only, not the active bar — CI [0.831,0.897] excludes it",
				"result": "PASS (bar restated against 0.755; +11.3pp)"
			},
			"val_macro_f1": {
				"v385": 0.7047,
				"feed_8k": 0.7176,
				"delta_pp": 1.29,
				"bar": "±1.0pp of v385",
				"result": "outside magnitude, favorable direction"
			},
			"dependent_locality_boards": {
				"gb_prior_on_at_delta_5": {
					"n": 69,
					"emit": 69,
					"tag_correct": 67,
					"tag_correct_pct": 97.1
				},
				"nz_checkpoint_only": {
					"n": 246,
					"emit": 0,
					"tag_correct": 0
				},
				"es_checkpoint_only": {
					"n": 65,
					"emit": 9,
					"tag_correct": 7,
					"tag_correct_pct": 10.8
				},
				"fr_lieudit_checkpoint_only": {
					"n": 80,
					"emit": 2,
					"tag_correct": 1,
					"tag_correct_pct": 1.25
				}
			},
			"dependent_locality_ablation": {
				"$comment": "The prior — not the checkpoint's own resurrected classifier row — carries nearly all of GB's measured recall. NZ/ES/FR-lieudit have no equivalent prior and show the checkpoint's own resurrection is weak-to-absent.",
				"v385_plus_prior_at_delta_6_rung3_control": {
					"n": 69,
					"emit_tag_correct": "3/69",
					"note": "weights matter"
				},
				"feed_checkpoint_prior_off": {
					"n": 69,
					"emit_tag_correct": "0/69",
					"note": "prior matters (either checkpoint)"
				},
				"feed_checkpoint_prior_on_at_delta_star": {
					"n": 69,
					"emit": "69/69",
					"tag_correct": "67/69 (97.1%)",
					"note": "together: +67pp"
				}
			},
			"venue_confound_fp": {
				"delta_star": 5,
				"fp": "48/6500",
				"fp_pct": 0.738,
				"floor": "<= 122/6500 (the retired δ=6.0 working number)"
			},
			"demo_presets": "6/6 byte-identical (both weights worlds — en-us control path AND en-gb weights with the prior wired)",
			"error_analysis_2pp_rule": {
				"worst_mover": "street -1.5pp",
				"no_tag_regresses_gt_2pp": true,
				"result": "PASS"
			},
			"gauntlet": {
				"regression": "PASS (33/33 conditional; 1 tracked known_fail now passes and should be promoted to status=pass — fr-chat-qui-peche-street-swallows-locality)",
				"metamorphic": "FAIL — NEW check-failing violation: INV[comma-drop] '1600 Pennsylvania Ave NW, Washington DC 20500' -> comma-dropped form loses the rooftop resolution entirely (38.8977,-77.0365 -> 0,0). v385 HOLDS this exact case (confirmed same session). feed-2k independently FAILS the same layer with DIFFERENT violations on '350 Fifth Avenue, New York, NY' (283.5km off) — reads as a lineage-wide v3.11.0-deploc-feed robustness regression on NY/DC-class admin resolution, not a single-checkpoint fluke. Neither Task 6 nor Task 7 ran the Gauntlet; this is a NEW finding from Task 8.",
				"held_out": "PASS — z=0.00 (candidate not significantly worse), n=300 fresh FR/BAN draw",
				"combined_verdict": "FAIL — do not ship (until the metamorphic regression above is triaged)"
			}
		},
		"check_2026_07_18_v710_camx": "v6.0.0-shipped-baseline check + golden 2pp check (v385 vs shipped v381, channels-on, same run): PASS. Golden: exact-match 24.2% (1105 vs 1104, +1 row), country 89.8 vs 91.4 (−1.6, small-N 220/245, sub-threshold — the only mover >0.5pp down), region 78.1 vs 78.6 (−0.5), locality/postcode/street/house_number flat. NO tag >2pp down. Native-order held-out board: MX locality 74.0→82.7 (+8.7), CA neutral. Gauntlet (eval gauntlet --candidate): regression 33/33 PASS; metamorphic INV 62/63, DIR 3/3, BAND 16/21 — identical known-xfail profile to shipped v381 (the 1 INV[comma-drop] xfail is the pre-existing FR #1101 case; US comma-drop held). No new regression. Grade: docs/articles/evals/2026-07-18-v710-camx-latam-grade.md.",
		"check_2026_07_17_v650_punct_digit": "v6.0.0-shipped-baseline check: PASS, 0 tags >2pp below shipped v310, +2.3pp us.country_homograph_f1 (only >2pp mover). Gauntlet: PASS — clear to ship (regression 33/33; metamorphic INV[comma-drop] 9/10 held — the US '1600 Pennsylvania Ave NW Washington DC' comma-free rooftop RESTORED, the FR 'Rue du Chevaleret' remains the tracked #1101 xfail; held-out FR/BAN n=300 not significantly worse). Digit ownership (NO board 3): bare-street-hn 0.693->0.733 (+4pp), bare-pc guard held 1.000. Grade: package-shaped --weights-cache (#718), int8 candidate vs int8 shipped, on 2026-07-17.",
		"check_2026_07_16_v640_fr_fragment": {
			"promotion_eval": "v6.0.0-shipped-baseline check: PASS 18/18, no revision, no relaxation. This spec was reduce from the SHIPPED v264's own readings BEFORE v310 was graded, with per-metric noise margins (floor = shipped − max(1.0pp, 2 × the 95% binomial half-width)); it tightened 15 of 18 floors over the retired v5.3.0-family spec. The FR fragment board's targets and guards were pre-registered in the v3.1.0-fr-fragment config header and in baselines.json (profile fragment-fr-v264), registered against the shipped model before the run. 2pp pre-publish bar: clear on every tag. No trade was measured and none is claimed.",
			"promotion_eval_docs": [
				"docs/articles/evals/2026-07-16-t2-fr-fragment-verdict.md (the read)",
				"docs/articles/evals/2026-07-16-t1c-fragment-board-verdict.md (the diagnosis this fix confirms)",
				"docs/articles/evals/2026-07-16-span-head-closed.md (the pre-registered follow-up: the span head does NOT compose with this extract)"
			],
			"fr_fragment_board": {
				"$note": "Board 2 — BAN-sampled, 2800 fixtures, street surfaces disjoint from the training extract. Wilson 95% CIs. No target's interval overlaps its baseline.",
				"bare_street": {
					"v264": 0.215,
					"v310": 0.715,
					"ci95_v310": [0.669, 0.757]
				},
				"street_particle": {
					"v264": 0.273,
					"v310": 0.855,
					"ci95_v310": [0.817, 0.886]
				},
				"admin_street_homonym": {
					"v264": 0.087,
					"v310": 0.517,
					"ci95_v310": [0.469, 0.566]
				},
				"date_name": {
					"v264": 0.055,
					"v310": 0.158,
					"ci95_v310": [0.125, 0.196]
				},
				"street_housenumber_guard": {
					"v264": 0.925,
					"v310": 0.948
				},
				"alnum_housenumber_guard": {
					"v264": 0.925,
					"v310": 0.96
				},
				"bare_locality_counter_guard": {
					"v264": 0.98,
					"v310": 0.98,
					"$note": "PRE-REGISTERED as the cell that might collapse by design. It held exactly, same interval — the model learned the distinction, not a new default."
				},
				"overall": {
					"v264": 0.494,
					"v310": 0.733
				}
			},
			"global_parity_floor": {
				"$note": "Board 1 — the guard. The FAIL verdicts eval parity prints on street/house_number are the v7 CAMPAIGN floors (0.90 / 0.97); shipped v264 fails them too. They are a target, not a regression check.",
				"street": {
					"v264": "154/267 = 0.577",
					"v310": "162/267 = 0.6067"
				},
				"house_number": {
					"v264": "117/146 = 0.8014",
					"v310": "117/146 = 0.8014"
				},
				"postcode": {
					"v264": "71/72 = 0.9861",
					"v310": "71/72 = 0.9861"
				}
			},
			"known_open_not_addressed": "Digit ownership — which tag owns a digit-bearing token (house_number vs postcode vs route number). Untouched by this extract and explicitly out of its scope. Parity PRECISION (new in this cycle — the half no floor could previously see, since every floor did `if (!goldValues?.length) continue`) reads postcode 25/249 = 0.100, street 12/54 = 0.222, house_number 6/175 = 0.034 emitted where the gold has none. v310 moves this by −1 row. Tracked as its own workstream; whether these become formal floors is an open operator decision."
		},
		"check_2026_07_04_v530_full_family": {
			"promotion_eval": "v5.3.0-family check (clone of v5.2.0-nordic + the fr.postcode 99.3→99.2 revision): PASS with the three operator-approved trades stated above. The promote decision ran on the 2026-07-03 RE-CHECK of the archived candidate under the #942-floored stack — fresh dumps BOTH arms, same stack (the #942-D1 stack-drift lesson): 6/6 ni legs PASS; the SI wrong-city re-check bar (≤+2pp) failed at +2.7pp and was adjudicated at the fork, not silently absorbed. Original 100k battery: 15/17 floors (the two above), mask-regression PASS, zero NaN.",
			"promotion_eval_docs": [
				"check-v193/RESULTS.md (pre-registered salvage read + re-check, 2026-07-03)",
				"#901 (the campaign + the promote fork), #949 (the FR fragmentation this fixes)"
			],
			"fr_bare_street_40row": {
				"v5_2_0": {
					"bare": "16/40",
					"anchored": "13/40"
				},
				"v5_3_0": {
					"bare": "34/40",
					"anchored": "37/40"
				}
			},
			"wrong_city_resolved_gt20km_n1k": {
				"cz": {
					"v5_2_0": 13.7,
					"v5_3_0": 6.6
				},
				"sk": {
					"v5_2_0": 13.3,
					"v5_3_0": 6.6
				},
				"pl": {
					"v5_2_0": 7.1,
					"v5_3_0": 6.2
				},
				"si": {
					"v5_2_0": 8.7,
					"v5_3_0": 11.4,
					"$note": "the stated trade — floored by #942"
				}
			},
			"ni_legs": "US-2k p50 3.27 identical, resolve 1.000 both; FR-3k p50 1.30 identical (deconfounded same-stack); SE +4.6pp resolve; Abo 0.12km vs 859km"
		},
		"check_2026_07_02_v510_vocab_splice": {
			"promotion_eval": "#884 checks ALL PASS, pre-registered before measurement (check log: /mnt/playpen/mailwoman-data/scratch-825/check-v193/RESULTS.md; evidence comments on #884). US-2k non-inferiority passed at the strongest possible level — the int8 candidate's per-row dump is byte-identical to shipped v4.15.0's. CZ/PL/SK/SI improve checks: wrong-city diff CIs wholly negative, resolved-p50 improve CIs wholly negative, resolve-rate floors held. Functional eyeball: CZ churn (103 fixed / 43 newly-wrong) diagnosed as a pre-existing locality-truncation→namesake residual exposed where the baseline won by garbage-parse fluke — NOT a parse regression (10/10 rows read). The int8 SLO overage (+3.8 MB) was an explicit operator accept (four locales for the same bytes).",
			"promotion_eval_docs": [
				"docs/articles/evals/2026-07-01-day-825-tokenizer-fix.md",
				"docs/articles/evals/2026-07-02-lab-session-postmortem.md",
				"docs/articles/evals/parity-scorecard-2026-07-02.md (the pre-splice baseline, 17/17 PASS)"
			],
			"wrong_city_resolved_gt20km_n1k": {
				"cz": {
					"v4_15_0": 22.4,
					"v5_1_0": 14.8,
					"diff_ci95_pp": [-10.3, -5.1]
				},
				"pl": {
					"v4_15_0": 27.9,
					"v5_1_0": 7.9,
					"diff_ci95_pp": [-22.7, -17.4]
				},
				"sk": {
					"v4_15_0": 22.2,
					"v5_1_0": 13.2,
					"diff_ci95_pp": [-11.5, -6.4]
				},
				"si": {
					"v4_15_0": 18.9,
					"v5_1_0": 10.6,
					"diff_ci95_pp": [-10.4, -6.4]
				},
				"$note": "reservoir-sampled 1k-row OA coord sets (oa-{cz,pl,sk,si}-coord-1k.jsonl), graded parse→resolve→great-circle, int8 ship artifact"
			},
			"us_2k_non_inferiority": {
				"row_dump": "BYTE-IDENTICAL to shipped v4.15.0 int8 (diff 0 everywhere)",
				"coord_p50_km": 3.31,
				"resolve_rate": 1,
				"region_correct": 0.999
			},
			"tokenization_examples": "Vysoká 4→1 piece, Grudziądz 6→1, Świętokrzyska 9→1, Čistá 5→2; SK fertility −51.9%, SI −35.2% on diacritic-bearing names with zero SK/SI-specific pieces"
		},
		"check_2026_06_25_v415_anchor_absorption": {
			"promotion_eval": "v4.15.0-boundary check PASS after a stated, operator-approved revision of two postcode floors (us.postcode 97→95, fr.postcode 99.5→99.3). The #723 fix (SLICE-H anchor-ON 20→100) trades −2pp us.postcode LABEL-F1 on the rare leading-5-digit-postcode (VT E911) case. The trade is COORDINATE-INVISIBLE — proven by the non-circular oa-resolver-eval below, not asserted. Every other floor holds with margin (us.locality 76.7/62.2, us.street 82.3/74, us.region 88.6/80.1, de.native_locality 89.3/83.8, arena.perturb 78/71); mask-regression check (#718) PASS; int8↔fp32 delta 0. No silent drift: the revision is in scripts/eval/checks/v4.15.0-boundary.json with the reason, mirrored here and in the promotion doc.",
			"promotion_eval_doc": "docs/articles/evals/2026-06-25-v4.15.0-promotion.md",
			"slice_h_anchor_on": {
				"v192": 20,
				"v193a3": 100,
				"$note": "leading-5-digit house# correctly labeled house_number (not postcode) with the anchor ON — the #723 production bug"
			},
			"oa_resolver_eval_us": {
				"n": 2000,
				"$note": "non-circular: real OA gov lat/lon vs resolved admin centroid, full ship-config (anchor+gazetteer+suppress)",
				"locality_match": {
					"v192": 0.977,
					"v193a3": 0.978
				},
				"region_match": {
					"v192": 0.999,
					"v193a3": 0.999
				},
				"coord_p50_km": {
					"v192": 3.4,
					"v193a3": 3.4
				},
				"coord_p90_km": {
					"v192": 10.7,
					"v193a3": 10.7
				},
				"coord_p99_km": {
					"v192": 275.3,
					"v193a3": 259.1
				},
				"vt_locality_match": {
					"v192": 0.926,
					"v193a3": 0.926,
					"$note": "VT = the E911 leading-postcode state where the postcode label trade lives — FLAT at the coordinate"
				},
				"sd_locality_match": {
					"v192": 0.965,
					"v193a3": 0.975
				}
			},
			"label_deltas_coordinate_invisible": [
				"us.postcode -2.0 (95.0) — rare leading-5-digit-postcode VT E911 case; VT coordinate flat (92.6=92.6)",
				"fr.postcode -0.2 (99.3) — eval noise"
			]
		},
		"check_2026_06_19_fr_admin_split": {
			"promotion_eval": "FR centroid check PASS — surpasses v1.5.0 on the SHIPPED assembled anchor-ON coordinate (held-out FR golden, 3000 disjoint communes): coord p50 42.5→2.2 km, mean 175→105.5 km (−40%), région-emit 39.6→99.6%. US assembled coordinate FLAT (oa-resolver-eval, 2000 OA-US rows: locality-match 97.5→97.6, p50 3.3=3.3). Two label deltas trip the 2pp pre-publish check, BOTH coordinate-invisible and operator-approved (#728): us.locality −2.2 (precision-only, 0 recall misses) and fr.country −3.5 (resolver given --default-country FR). No silent drift — both stated here + on #728 + the night postmortem.",
			"promotion_eval_doc": "docs/articles/evals/2026-06-19-night-postmortem.md",
			"bridge_retired": true,
			"flagged_deltas_coordinate_invisible": [
				"us.locality -2.2 (precision-only — 0 recall misses, spurious-FP on gold-locality-absent rows)",
				"fr.country -3.5 (resolver given --default-country FR, never reaches the coordinate; abandoned v1.8.1 country-extract targeted this)"
			]
		},
		"fr_centroid_check_heldout": {
			"n": 3000,
			"coord_mean_km": {
				"v1_5_0": 174.99,
				"v1_8_0": 105.51
			},
			"coord_p50_km": {
				"v1_5_0": 42.52,
				"v1_8_0": 2.17
			},
			"coord_p90_km": {
				"v1_5_0": 389.37,
				"v1_8_0": 334.15
			},
			"resolve_rate": {
				"v1_5_0": 0.595,
				"v1_8_0": 0.785
			},
			"region_emit_rate": {
				"v1_5_0": 0.396,
				"v1_8_0": 0.996
			},
			"region_correct_rate": {
				"v1_5_0": 0.364,
				"v1_8_0": 0.964
			},
			"diacritic_727_breaks": {
				"v1_5_0": 23,
				"v1_8_0": 24
			}
		},
		"us_assembled_coord_oa": {
			"n": 2000,
			"locality_match": {
				"v1_5_0": 0.975,
				"v1_8_0": 0.976
			},
			"coord_p50_km": {
				"v1_5_0": 3.3,
				"v1_8_0": 3.3
			},
			"coord_p90_km": {
				"v1_5_0": 10.8,
				"v1_8_0": 10.7
			},
			"region_match": {
				"v1_5_0": 0.999,
				"v1_8_0": 0.999
			}
		},
		"per_component_int8_shipconfig": {
			"$comment": "per-locale-f1 anchor-ON matched pair vs v1.5.0 (postmortem guardrail tables). The unfolded per-tag F1 (split affix, intersections, etc.) is the authoritative `capabilities` block below. Tags off the FR-extract path (us intersections/unit, de native_locality, arena.perturb) were not re-measured for v1.8.0 — the single training variable was the FR admin-split extract and the architecture is identical to v1.5.0.",
			"us": {
				"house_number": 98.6,
				"locality": 75.7,
				"region": 89.8,
				"street": 80.8,
				"postcode": 98.6,
				"country": 68.9
			},
			"fr": {
				"house_number": 99.5,
				"locality": 87.6,
				"po_box": 83.3,
				"region": 43.7,
				"street": 89.4,
				"postcode": 99.7,
				"country": 59.2
			}
		},
		"vs_1_5_0": {
			"fr_coord_p50": "42.5 → 2.2 km (−95%) — the win; first model in the arc to beat v1.5.0 on the shipped coordinate",
			"fr_region_emit": "39.6 → 99.6% (+60pp) — learned the locality↔département split",
			"us_assembled_coord": "FLAT (locality-match +0.1, p50 identical)",
			"us_locality_label": "−2.2pp (precision-only, coordinate-invisible; check-firing, operator-approved #728)",
			"fr_country_label": "−3.5pp (resolver given country, coordinate-invisible; check-firing, operator-approved #728)",
			"fr_street": "+1.9 (87.5 → 89.4)",
			"fr_po_box": "+10.6 (72.7 → 83.3)",
			"fr_house_number": "99.3 → 99.5 (held — v1.5.0's raison d'être)"
		},
		"check_2026_07_28_v3230_bundle_guard": {
			"$comment": "The Option-A promotion battery on the STAGED 6.7.0-bundle candidate (v3.23-4k int8). G5 graded the ABLATED column (= formatted-mode production behavior under Decision A); fragment bars graded fed. Ledger + censuses: .superpowers/sdd/progress.md 2026-07-28.",
			"g1_fragment_bars": "PASS — homonym +0.245 asis / +0.765 lower+heal; alnum-hn −0.007 asis / +0.013 lower; street-hn −0.015 / +0.030; ALL lower+heal +0.261",
			"g2_ablation": "borderline-pass (OFF bare-locality −0.018 vs ±0.015, ≈1 fixture row; all else at/above v385 reference)",
			"g3_invariance": "PASS — zero new violation classes vs v385",
			"g4_gauntlet": "PASS ×3 (regression/metamorphic/held-out)",
			"g5_golden_ablated": "PASS 17/17 (v6.0.0-shipped-baseline; fr.cedex 83.3 under the ratified 82.2 noise-honest floor — NAMED WATCH)",
			"g6_canary": "formatted-mode clean by construction; fed-mode 4/20 flips recorded as misroute-cost input"
		},
		"check_2026_07_30_v700_base_promote": "PROMOTED as model 7.0.0 (npm 8.3.0). Ship-config battery (int8, locality v6): v6.0.0-shipped-baseline 16/17 + fr.bare_street_intact + mask-regression PASS — the one nominal miss is us.street_prefix 91.3 vs the 96.0 fallback-margin floor (n=25; four spelled-out-directional recall misses, precision 100; clears the support-aware v7.0.0-base floor). fr.cedex_real 90.5 — the v6-era NAMED WATCH RETIRED (above the unwaived 85.6-class reading). us.street 90.5 / us.locality 81.0 / us.micro 89.8 — multiple floors up vs the shipped 6.7.0. P0 fragment grid FLAT 40k→60k both registers (the over-trust decay proven a fine-tune-recipe artifact): asis ALL 0.803, lower+heal 0.983. Gauntlet: regression 3 venue must-clears CLEARED + 3 tracked cases promoted to pass + the 9-probe directional-homograph set 9/9 (zero prefix-over-splitting cost); metamorphic 5 tracked xfails (one fewer than shipped); held-out PASS. ONE NAMED WATCH rides this promote: si-sentinel-ucakar (SI diacritic digit-split; raw split byte-identical to shipped, which passes only via the word-consistency heal; designed fix = the char-path lineage). New check spec: v7.0.0-base.json (support-aware Wilson margins, reduce from this candidate's ship-config readings).",
		"r5_us_dependent_locality": {
			"$comment": "Pre-registered bars from docs/superpowers/plans/2026-08-01-r5-us-deploc-preregistration.md, all measured before the ship decision.",
			"gauntlet_with_index": "PASS — verdict and xfail set identical to the baseline without the index (B-R5.1)",
			"venue_confound_board": {
				"n": 60,
				"false_positives": 0,
				"fp_pct": 0.0,
				"bar_pct": 2.0,
				"note": "US law-1 board: neighbourhood surfaces opening venue names (B-R5.2)"
			},
			"positive_board": {
				"n": 60,
				"emit": 60,
				"tag_correct": 60,
				"tag_correct_pct": 100.0,
				"locality_correct": 60,
				"note": "Sampled from 31,474 state-resolved WOF pairs. Measures mechanism efficacy where the gazetteer has data, NOT generalization to unknown surfaces (B-R5.3)"
			},
			"browser_fallback_disclosure": "detectPairIndexCountry routes bare postcode-less Latin queries to us; the US index misses all five GB canonical pairs, so the fallback does not manufacture GB bias (D-R5.4)"
		}
	},
	"files_md5": {
		"$comment": "The SHIPPED artifact md5s — the identity link-dev-weights.ts (#397 guard) and the gauntlet self-check (#1024) verify dev symlinks / materialized tarballs against, repo-wide. 9.1.0 = v4.4.0-suffix-boundary-v2 step-60000 int8 ($MAILWOMAN_DATA_ROOT/models/quantized/model-v440-suffix-boundary-v2-step-060000-int8.onnx); tokenizer UNCHANGED (v0.9.0-multisplice). Flip this block, requires, and link-dev-weights DEFAULT_MODEL in lockstep at every model reduce — the 9.0.0 reduce missed the flip and shipped a card whose model.onnx entry named the prior base.",
		"model.onnx": "98a49b5ca2948cdbe1dd981d60ef637d",
		"tokenizer.model": "5c01cdcd4ae25849c5cb26b69fd3dde9",
		"street-type-lexicon-v3.json": "d0194a730805f9e6a886a98e8421061a",
		"locality-surface-lexicon-v7.json": "08b0305aab6e65fd3556be23b01b7756"
	},
	"capabilities": {
		"server": {
			"us": {
				"street_prefix": {
					"maskOffF1": 1.5
				},
				"street": {
					"maskOffF1": 24.4
				},
				"street_suffix": {
					"maskOffF1": 0.3,
					"maskOnF1": 0.3
				},
				"house_number": {
					"maskOffF1": 98.3
				},
				"locality": {
					"maskOffF1": 79.6
				},
				"region": {
					"maskOffF1": 91
				},
				"postcode": {
					"maskOffF1": 98.5
				},
				"country": {
					"maskOffF1": 69.4
				},
				"unit": {
					"maskOffF1": 6.3
				},
				"intersection_a": {
					"maskOffF1": 66.7
				},
				"intersection_b": {
					"maskOffF1": 20
				},
				"po_box": {
					"maskOffF1": 72
				},
				"venue": {
					"maskOffF1": 94.6
				},
				"dependent_locality": {
					"maskOffF1": 0
				}
			},
			"fr": {
				"street_prefix": {
					"maskOffF1": 80
				},
				"street": {
					"maskOffF1": 54.5
				},
				"house_number": {
					"maskOffF1": 100
				},
				"locality": {
					"maskOffF1": 100
				},
				"postcode": {
					"maskOffF1": 100
				}
			}
		},
		"pocket": {
			"us": {
				"street_prefix": {
					"maskOffF1": 1.5
				},
				"street": {
					"maskOffF1": 24.3
				},
				"street_suffix": {
					"maskOffF1": 0.3,
					"maskOnF1": 0.3
				},
				"house_number": {
					"maskOffF1": 97.7
				},
				"locality": {
					"maskOffF1": 58.6
				},
				"region": {
					"maskOffF1": 71
				},
				"postcode": {
					"maskOffF1": 98.6
				},
				"country": {
					"maskOffF1": 0
				},
				"unit": {
					"maskOffF1": 6.8
				},
				"intersection_a": {
					"maskOffF1": 18.2
				},
				"intersection_b": {
					"maskOffF1": 5.6
				},
				"po_box": {
					"maskOffF1": 66.7
				},
				"venue": {
					"maskOffF1": 94.1
				},
				"dependent_locality": {
					"maskOffF1": 0
				}
			},
			"fr": {
				"street_prefix": {
					"maskOffF1": 80
				},
				"street": {
					"maskOffF1": 54.5
				},
				"house_number": {
					"maskOffF1": 100
				},
				"locality": {
					"maskOffF1": 100
				},
				"postcode": {
					"maskOffF1": 100
				}
			}
		},
		"$comment": "Per-tier (server=anchor+gazetteer; pocket=anchor-only) × address-system × tag capability manifest (#718/#719). maskOffF1 = measured per-tag exact-match F1 with the conventions mask OFF; maskOnF1 = the same with mask ON (recorded only for tags some codex forbiddenTags row suppresses — the loader's delta-check consults only those). createScorer FAILS CLOSED when a conventions row forbids a tag with maskOffF1 − maskOnF1 > 0.05 (the mask provably destroys a real capability). Generated by `mailwoman eval capability-manifest` against the v1.5.0 int8."
	},
	"fisher_artifact": {
		"$comment": "The consolidation artifact (v8.3.0 D4/#1354): diagonal empirical Fisher over the base run's final 2,000 steps, covering all 39,261,743 trainable parameters. The EWC brake input for every fine-tune off this base (train.ewc_fisher_path). HF/R2-distributed ONLY — the runtime never reads it, so it stays out of the npm tarball.",
		"file": "fisher-diag-v1-model-9.1.0.npz",
		"sidecar": "fisher-diag-v1-model-9.1.0.json",
		"md5": "e9d77f7d8b785109a07df0ee05f5e3ee",
		"captured_at_step": 60000,
		"count_batches": 2000,
		"distribution": "hf+r2"
	},
	"us_artifacts": {
		"$comment": "Hierarchy campaign R5 (2026-08-01). The US instance of the placetype-pair prior — the first non-postal-register one.",
		"pair_index_us_bin": {
			"format": "PIX1 placetype-pair index — a retrieval-augmented (child, parent) dependent_locality prior, e.g. (\"Park Slope\", \"Brooklyn\")",
			"source": "Who's On First admin hierarchy ($MAILWOMAN_DATA_ROOT/wof/admin-global-priority.db). NO postal register: USPS routes city/state/ZIP, so unlike GB's PPD there is no mail source carrying US dependent localities. Children: borough + neighbourhood. Parents: locality, localadmin AND borough — WOF parents US neighbourhoods to the locality (\"Astoria\" hangs off New York, not Queens), so a locality-only parent rule drops the borough-level pairs this instance exists for.",
			"pairs": 47878,
			"bytes": 1341264,
			"md5": "31e7cd899dab5af0b8d7ec03fe5cb548",
			"$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. 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). δ=5 is the smallest magnitude that saturates bar B-2's brooklyn-class sub-board and is flat through 20, at 0.00% parent-side FP on B-3: docs/records/evals/2026-08-04-pix1-whole-edge-verdict.md. The δ/β pair is inherited from the GB calibration rather than re-swept, because the R5 finding is that the classifier's dependent_locality deficit is UNIFORM across locales (~7.0 logits mean, measured on the en-gb card), not GB-specific. The R5 bars were measured at exactly this pair; changing either number invalidates them.",
			"country_check": "hard-blocked on the header's country (us). Verified inert on GB input: the index misses all five GB canonical pairs (Shoreditch/London, Nine Elms/London, Clapham/London, Camden/London, Didsbury/Manchester) — see neural/test/weights.test.ts.",
			"not_committed": "derived artifact, .gitignore'd (same posture as postcode-us.bin) — rebuilt via `mailwoman gazetteer pair-index --out neural-weights-en-us --country us --delta 10 --transition-beta 5 --borough-db <wof/admin-global-priority.db>`",
			"behavior_change": "This changes DEFAULT en-us parse output. A US address carrying a neighbourhood or borough line previously emitted no dependent_locality AND SILENTLY DROPPED the second admin level (\"123 Main St, Park Slope, Brooklyn, NY\" → locality=Park Slope, Brooklyn discarded). It now emits dependent_locality=Park Slope, locality=Brooklyn.",
			"$comment_md5": "REPRODUCIBILITY: a PIX1 header embeds `buildDate`, so a rebuild from identical sources produces different bytes and a different md5. This md5 therefore identifies the artifact STAGED TO HUGGING FACE for this release — it is a provenance record of the shipped bytes, NOT a check a local rebuild can reproduce. A mismatch after rebuilding is expected and is not corruption; verify pair COUNT and header delta/beta instead.",
			"$comment_currency": "CURRENCY FILTER (2026-08-02): both ends of every pair are now required LIVE — `is_current != 0 AND is_deprecated = 0`. The WOF ingest hard-rejects `wof:superseded_by` records but RECORDS `edtf:deprecated` and `mz:is_current` as columns rather than rejecting them, so filtering is the reader's job and this extraction was not doing it. Pair count 49033 → 47878."
		}
	}
}