Skip to main content

Releases & capabilities

Two version series exist and used to be easy to confuse:

  • npm version โ€” what npm install gives you. All workspaces release in lockstep (e.g. mailwoman@4.1.0 pairs with @mailwoman/neural-weights-en-us@4.1.0).
  • Training series (v0.x.y) โ€” internal model lineage used by configs, the eval ledger, and the docs. A training run only becomes an npm release if it passes its pre-registered gates.

Since 4.0.0 the npm version is the one that matters to consumers; the shipped weights' model-card.json records the exact training lineage, corpus, and tokenizer. The full per-model score record is evals/scores-by-version.json (schema: eval-ledger.schema.json).

The matrixโ€‹

npmDateModel lineageWhat it addedPer-tag truth
4.2.0 (current)2026-06-10v1.0.2-consolidation-runB @ 20k (init_from consolidation v1.0.0 @ 40k)The v1.0 parity consolidation: locality +12.8, region +10.7, country 0โ†’89.8 (gazetteer soft anchor), affixes exist (0โ†’64.9/48.8), FR house_number 94.6 best-ever, DE 90.9. Stated re-baselines: US street โˆ’2.3, unit โˆ’1.7 (measured 29M ceiling, #492). 4/4 ship gate.Parity scorecard 2026-06-10
4.1.02026-06-09v0.9.7-unit-v3 @ 20k (off v0.9.3, off v0.7.2)Secondary-unit coverage (unit 0 โ†’ 92.3 real-OOD) โ€” the first negative-space parity win. Int8 quant toolchain pinned + fixed (Safari WebGPU graph stability). Two-backend artifact release (npm + demo pointer).Parity scorecard 2026-06-09
4.0.02026-06-06v0.9.3-de-regiontailFirst fully-synced release across all workspaces (mailwoman installable again, end to end). Multi-locale model (both-order German synth, region-tail), live postcode-anchor inference, self-conditioning. New scoped packages published (codex, locale-gate, resolvers, neural-web).ledger v0.9.3 rows
2.x / 3.0.0 (weights)2026-05v0.4.0 (8.87M params, 21 labels)The Phase-3 integration era: 6-stage runtime pipeline, CRF decode, hybrid policy registry. Weights versioned independently of the runtime packages (3.0.0 weights + 2.1.0 runtime) โ€” the misalignment this page exists to prevent.ledger 0.1.0โ€“3.0.0 rows

Earlier training-series numbering (v0.1.0, v0.2.0, the realigned 3.0.0) predates the lockstep policy; treat the eval ledger as the record there.

In flightโ€‹

SeriesTargetStatus
architecture escalation (post-4.2.0)TBD#492: dedicated affix head (probe) vs ~48M widening โ€” the measured 29M stability ceiling; operator GO required.

Capability quick-referenceโ€‹

CapabilitySince
6-stage runtime pipeline (normalize โ†’ โ€ฆ โ†’ resolve)2.x era
Hybrid rule/neural policy registry2.x era
WOF SQLite resolver (exact-match tiering, coordinate-first postcode disambiguation)4.0.0
Postcode anchor at inference (PCB1)4.0.0
Japanese postcode-route resolution (Geographic Rule Engine)4.0.0
Isotonic confidence calibration (per-locale calibration.json in the weights bundle)4.0.0
Secondary-unit (unit) recognition4.1.0
Street affixes, country soft anchor, multi-locale balance4.2.0

When a release ships, this page gains a row and the status page gets re-verified โ€” they change together or not at all.