v0.5.0 — as shipped

A single landing page for anyone catching up on the v0.5.0 fresh-slate work.

The v0.5.0 plan (PHASE_8_v0_5_0_fresh_slate.md) split the work into six threads (A through F). All six landed in main by 2026-05-24. This article catalogues what is now in the tree, how the pieces fit together, and what is still cooking on the training side.

Why v0.5.0 exists

The v0.4.0 ablation campaign retrospective ended with three structural ceilings that small recipe tweaks could not clear:

1. The v0.1.0 tokenizer's byte-fallback rate on non-Latin scripts (the cause of 92% of country FN in the golden eval).
2. The classifier doing boundary discovery and type classification at the same time via BIO labels (the cause of the bio_slip slice on ", 22220" for 22220).
3. Stage 5 reconcile only sorting spans — it could not catch joint-incoherent parses like NY-NY Steakhouse, Houston, TX.

Patching each ceiling individually meant three medium iterations. v0.5.0 bundles them into one fresh-slate ship — pay the cost once, clear all three at once.

The "fresh-slate" framing is described in the operator's sharpen-the-axe note — it is the right shape when several improvements share a common cost (a retrain, an interface change) and bundling avoids paying that cost three times.

The six threads

Thread	Name	What it adds	PR
A0	Tokenizer harness + A0 weights	SentencePiece training pipeline; A0 trained on corpus-v0.3.0 as a byte-fallback baseline. A1 will retrain on corpus-v0.4.0 once B + B2 are both in.	#129
B	Kryptonite catalogue	4,771 adversarial address rows generated via DeepSeek — NY-NY Steakhouse, Paris, Texas, Saint Petersburg, FL, etc. Each annotated with the correct parse.	#130
B2	Transliteration pairs	~73K US/FR addresses transliterated into CJK / Cyrillic / Hangul / Han / Armenian script via DeepSeek. Validated by substring-match aligner (~1.1% reject rate). Included in corpus-v0.4.0.	#119
C-s	Classifier code path	Top-k inference (returns ranked sequences, not just argmax) + phrase-prior input features that condition on Stage 2.7's proposed spans. No training run in this PR — that lands once A1 + corpus-v0.4.0 are ready.	#128
D-s	Stage 5 reconcile	reconcile.ts joint decoder — beam search over (span × tag × resolver candidate) with concordance scoring via WOF parent-id chains. Catches the kryptonite cases at runtime.	#131
E	Phrase grouper	New @mailwoman/phrase-grouper workspace — Stage 2.7 proposes coherent input spans (street, postcode, locality, …) before Stage 3 runs. Rule-based v1 only; learned span proposer scoped for v0.5.1 (PHASE_8_E).	#126
F	Verdict-smoke discipline	New VERDICT_SMOKES.md + --smoke-mode {constant,long-tail} flag, so the cosine-LR mask that hid v0.4.0's divergence cannot reoccur.	#125

How the threads compose

The threads were designed to be parallel where possible and stacked where unavoidable. The actual data dependencies look like this:

Solid boxes are merged. The A1 tokenizer was retrained on corpus-v0.4.0 and halved byte-fallback (36.7% → 18.2%). The C-train (full classifier training run) is the remaining item — CE-only training in progress as of 2026-05-25.

The single biggest schedule reframe was A0 — the plan originally said "A is blocked by B" because the tokenizer trains on the same corpus as the classifier. We split that into:

• A0 trains on corpus-v0.3.0 immediately to validate the harness and establish a byte-fallback baseline.
• A1 retrains on corpus-v0.4.0 (= v0.3.0 + B + B2) when the corpus is complete.

That way A0 could ship in parallel with B + E + the scaffolds, instead of waiting for B + B2 to finish.

The other parallel-friendly move was C-s and D-s as scaffolds. The plan had C and D as serial training/integration work that needed real classifier output to validate. We pulled the code portion to t0 (forward-pass tests against stub data, mocked top-k for Stage 5) and deferred only the actual training run. That left the only mandatory serial portion as B → A1 → C-train.

What is still cooking

Item	State	ETA
CE-only full 50K C-train	In progress. Smoke passed: val_macro_f1=0.444, no divergence past step 2000. Full 50K run targeting first stable v0.5.0 weights.	~6-8h from launch.
C-train full classifier run	Blocked on A1 + corpus-v0.4.0. Estimated single H100-day on rented compute per the plan.	After A1 lands.

Three smaller follow-ups carried over from postmortems:

• Wire reconcileSpans as the default Stage 5 in runPipeline (currently behind an explicit call). Lands once the classifier emits real top-k. Tracked in D-s's PR body.
• Learned phrase-grouper variant (1-2M-param span proposer). Scope written in PHASE_8_E_learned_span_proposer.md. v0.5.1 stretch.
• Reservoir-sampling parallelisation in the tokenizer harness (~5× speedup possible per A0's session notes). Out-of-scope until A1 turnaround time becomes a real cost.

Where to read next

If you are catching up on the architecture, read in this order:

1. The knowledge ladder — the conceptual framing for why the two new layers (Stage 2.7 phrase grouper, expanded Stage 5 reconcile) exist.
2. The staged pipeline — the runtime composition end-to-end.
3. STAGES.md — formal per-stage type contracts.
4. VERDICT_SMOKES.md — the process discipline that kept v0.5.0 from repeating v0.4.0's cosine-LR mask bug.

If you want the line-by-line technical history, the v0.4.0 retrospective (retrospectives/v0-4-0-ablation-campaign.md) is what motivated v0.5.0; this article is the bookend.