Night-Shift Postmortem — 2026-06-09 (parity campaign: affix gated, country → deterministic)

Sketched live during the shift (04:45→15:00 UTC), finalized at wrap-up.

What shipped

• v0.9.8-affix gated + int8 ship-ready (PR #461 merged). street_prefix 0→78 (P100/R64), street_suffix 0→67 (P100/R50) on the real-affix OOD eval; US golden street +2.1 (negative space); unit retained 93.8. FR postcode −3.9 (US-shard dilution) trips the >2pp gate → promote deferred to operator (#462); DeepSeek recommends promote-with-annotation.
• Affix-aware eval tooling (scripts/eval/score-affix.ts, probe-affix-decode.ts) — per-locale-f1's foldToComponents joins affixes into street, so it can't measure the split; these score the unfolded decode. Caught the gate's false 0% (the model splits 7/10, perfect precision).
• Parity scorecard (docs/articles/evals/parity-scorecard-2026-06-09.md, #375) — one authoritative per-tag-vs-v0 table, two lenses (arenas + per-tag), with the fold gotcha documented.
• Country lever resolved (PR #463 + #464): salvaged ISO 3166-1 from isp-nexus into codex/country/ + surface-form layer + matchCountry; multi-locale shard (US/DE/FR/IT/NL, v1/v2/v3); real-country OOD eval (34 cases). Key result: the model is the wrong tool. The v1 cumulative shard makes the model over-fire (country-real 49 F1, golden precision 23% — it learns "trailing token = country"), whereas a deterministic matchCountry on the trailing comma-segment scores P=R=F1=100. Lever moves to a post-parse ProposalClassifier (#464); the model path is kept on #463 as the exploration record.
• po_box + cedex reservoirs resolved deterministically — zero GPU, taxonomy applied up front (PR #463, #464). Applied the country lesson before spending any GPU. matchPOBox (codex/us/po-box.ts) per comma-segment = P=R=F1=100 on a curated real-OOD eval (n=25, 7 negatives incl. "Box Canyon Rd"/"Boxwood Lane" traps, 0 FP). CEDEX regex (FR, in-segment — a different locale and match shape) = P=R=F1=100 on cedex-real (n=15, 5 negatives). The taxonomy now holds across 3 tags, 2 locales, 2 match shapes — all where a retrain would dilute. Seeds a shared ClosedVocabTagger design.

What went well

• The proven recipe (codex synth + format-diverse + real-OOD gate + int8 value_info-strip) carried two more levers cleanly through build+gate.
• Caught the fold gotcha fast (the affix 0% was a measurement artifact, not a training failure) — built the right tool instead of chasing a phantom.
• The multi-locale country shard recovered the affix FR-postcode dilution (95.6→99.5) — the secondary design goal worked.
• Front-loaded the scorecard during the affix train window; no idle.
• Let a negative result reshape the strategy instead of forcing the lever. The country over-firing wasn't a bug to tune away — it revealed a lever-shape taxonomy: closed-vocab/fixed-position tags (country, po_box, cedex) want a deterministic matcher; open-vocab/boundary tags (affix, unit, locality) want a retrain. That taxonomy now guides the remaining levers (po_box likely goes deterministic too).

What could've gone better

• Jumped to a cumulative country run (base+unit+affix+country in one) rather than proving country solo first — it diluted (affix suffix 67→59) and country over-fired (golden precision 23%, fp high). The v2-vs-v3 unit lesson ("prove solo before combining") applied and I under-weighted it. Silver lining: the dilution is the evidence behind the consolidation rule (prove solo, then consolidate with a bigger step budget) now in the scorecard. But two retrains (v1 cumulative + v2 negatives) burned GPU on a lever that a 10-line deterministic probe should have pre-empted — I should have run the deterministic probe before the first country retrain, given country's obvious closed-vocab/trailing shape.
• The affix gate's first read was a false 0% (fold) — a per-tag affix scorer should have existed before the shard (now it does).

Decisions made autonomously

1. Held the affix promote (gate not cleanly passed; FR postcode −3.9 > 2pp) against DeepSeek's promote-rec — deferred to the operator via #462 rather than overriding their pre-registered gate while offline.
2. Country lever as cumulative consolidation (bakes unit+affix+country) — efficient in intent but it surfaced dilution + over-firing. The DeepSeek pro consult on the fork timed out (180s); I proceeded on the diagnosis. Resolution (autonomous): redirect country from retrain to a deterministic tagger. The deterministic probe (P=R=100) settled it — country is closed-vocab and trailing, so a matchCountry ProposalClassifier (#464) is the right tool, not another shard. Kept #463 as the exploration record rather than force-merging an over-firing model.
3. Held the v3 FR-fix as a committed-but-superseded artifact rather than launching a v3 retrain — once the deterministic result landed, spending more GPU on the model path was unjustified. v3's corpus correctness fix (FR number-street order) is still committed for whoever revisits.

Open questions for the operator

• #462: promote v0.9.8-affix → v4.2.0? (clean affix win, FR-postcode −3.9 trade). DeepSeek says yes-with-annotation. Still needs your call — I did not promote while offline against the pre-registered gate.
• #463 disposition: merge for the reusable codex/country + eval assets, or close in favor of #464 (deterministic)? Either is fine; the codex salvage is the durable part.
• #464 homograph guard: the deterministic country tagger needs a guard for state/country homographs (Georgia, Jordan) before it can default-on. Fire only when the trailing segment is unambiguously a country, or downrank when an upstream region already claims the span. (Architecture itself is decided — overlay, see next steps; this is the one remaining design detail.)

Concrete next steps

• #464 — build one ClosedVocabTagger (country + po_box + cedex; all three measured at P=R=F1=100 this shift). Architecture decided (DeepSeek pro consult): a post-parse ProposalClassifier overlay, default-off/byte-stable, not an in-pass classifiers/ entry — because it must correct the model's over-firing, so it runs after the decode and overwrites. Homograph guard (country Georgia/Jordan) lives in the overlay's apply logic; po_box/cedex have no homograph risk and can land first. ~½ day, zero GPU.
• intersection (gated — regressed before, needs care); FR venue/region (#330).
• Consolidation v1.0 (#466) once affix is promoted: make the affix shard multi-locale (clears its own FR-postcode gate without bundling the now-deterministic country), prove it solo, then weight-merge the affix delta into the v4.1.0 base rather than re-stacking shards (the dilution is step-budget-bound). DeepSeek-recommended; gated on #462 + operator go + GPU budget.

Numbers (running)

metric	value
shift window	04:45→15:00 UTC
models trained (Modal)	3 (affix v0.9.8, country v1 cumulative, country v2 negatives)
Modal spend	~$10–12 of $20
NaN incidents	0
CI failures	0 (corpus-cli flake re-ran green)
regressions shipped	0 (affix held experimental; nothing promoted)
PRs / issues	#461 (merged), #463 (open — country exploration), #462 (issue — affix promote), #464 (issue — deterministic country)
campaign tags	unit ✅ shipped · affix ✅ gated/deferred (#462) · country ✅ + po_box ✅ resolved → deterministic (#464)
reservoirs (zero-GPU)	po_box (100%) + cedex (100%) deterministic probes, ClosedVocabTagger design