Night Shift 7 — 2026-06-06 (order-robustness arc, concluded — plus a demo-refinement tier)

Final. Numbers self-emitted. The shift opened on a single question — does both-order training close the German international-order gap? — and closed it across three retrains; the back half turned to the demo-refinement tier the operator asked for once DeepSeek (delegated authority) called "no more GPU."

What shipped

• #323 — both-order German corpus (merged). synthesizeLocaleRow({ order }) renders German in native AND international layout; build-german-shard.mjs --intl-fraction; a reproducible native-order eval asset. The fix the whole shift is built on.
• #324 — parse --default-country (merged). The CLI resolved globally, so a bare NY hit a Scottish homonym (IL→France, WA→Spain); now inferred from --locale (the demo already passed country:"US"). 10/10 tests, the NY→NYC integration guard included.
• #325 — order-robustness eval & diag tooling (merged). scripts/eval/de-order-eval.sh (the 2×2 native/intl × anchor on/off + US/FR gate harness), --model-anchor-lookup on per-locale-f1.ts (anchor-aware per-tag F1), and the three diag scripts the finding/memories reference.
• v0.9.2 both-order retrain — trained clean (20k, no NaN), evaluated, not shipped (research result, not a production regression). Full report: 2026-06-06-v0.9.2-eval.md. Headline: both-order training moved the model's intrinsic international parsing (anchor-off intl 35.9 → 48.4, +12.5pp) but the production-anchor config is flat (intl 45.5 → 44.5) — the postcode anchor, injected at the trailing postcode, fights the word order on international layout. Native held (82.1, beats Pelias), US/FR held (97.5 / 84.9), locality F1 0.82 (no collapse). The residual gap traces to the synth dropping the region tail the eval feeds — the v0.9.3 fix.
• #326 — v0.9.2 eval report (merged). The 2×2 + the anchor-vs-word-order finding + the region-tail mechanism. Issue #327 filed for the v0.9.3 direction.
• DeepSeek consult (3 turns, delegated authority) — turn 1–2 pressure-tested the anchor-vs-word-order mechanism, signed off region-tail as the v0.9.3 single variable, and prescribed a posterior-ablation diagnostic (force DE=1.0) that left v0.9.2 intl at 44.5% → the anchor harm is structural-positional, not posterior; it designed v0.9.4 (dual-injection) as the fallback. Turn 3, after v0.9.4 also failed, made the strategic call (the operator was away and had delegated authority for the shift): accept the asymmetry, no more GPU, and the multi-locale plan survives via a country-conditioned anchor. See #337.
• #328 / #329 — v0.9.3 region-tail (merged, trained, evaluated, not shipped). International German renders City, Region Postcode now. Clean negative: ≈ v0.9.2 on every locality metric (native off/on 48.3/83.6 beats Pelias, intl off/on 47.1/44.7, US 97.2 / FR 84.9 held). The region tail did exactly its job — international region-match 0% → ~40%, Berlin intl-loc 34.4→38.7 — but locality stayed flat. The anchor-on intl gap (38.9pp) is the anchor's positional harm, not the corpus (the posterior diagnostic already ruled out feature content). The corpus lever is exhausted for the anchor-on intl gap. → v0.9.4.
• #334 / #336 — v0.9.4 dual-injection (merged, trained, evaluated, not shipped). DeepSeek's signed-off architectural fix: pool the anchor and inject it also at position 0, an order-independent cue (no new params, c=0 identity holds, 8/8 anchor tests, ONNX-export-equivalent). It did nothing to the intl gap (native off/on 48.3/83.5, intl off/on 47.3/43.7, US 96.4 / FR 84.9 held). Three retrains, one immovable number (~44%). The anchor's native-help (+35) / intl-hurt (−4) asymmetry is fundamental — not corpus, not region, not injection position.
• #337 — the anchor fork, RESOLVED (merged). With the operator away, the third DeepSeek turn (delegated authority) made the call: (b) accept the asymmetry, burn no more GPU. Its mechanistic read reframes the multi-locale plan — the anchor's direction is learned per locale from the dominant training order, so the next thread is a country-conditioned anchor vector, not another tuning pass. Decision record in 2026-06-06-v0.9.2-eval.md §"accept the asymmetry".
• Multi-locale tooling (#332 / #333 / #335) — build-locale-shard.mjs rebuilt registry-driven, with a streaming Algorithm-R reservoir so FR/US-scale OA CSVs (2.5 GB+) sample without OOM; NL postcode normalization; and Italy + Spain OA data acquired (IT 468 MB, ES 451 MB, durable backups on /mnt/playpen). DE / NL / FR / IT / ES are all shard-ready. (Correction: an earlier note here called ES "no clean countrywide aggregate." That was wrong — I'd only checked the results.../latest/run/es path, which 404s. The es/countrywide source points to a cached upstream CSV on data.openaddresses.io that downloads fine; it's the raw CNIG schema rather than OA-conformed, so the builder gained a per-part conform map — street = join(tipo_vial, nombre_via), region = comunidad_autonoma. A 2k smoke build renders both orders with the region tail carried, e.g. 2 CALLE JACINTO, Lepe, Andalucía 21440.)
• Demo-refinement tier (#338 / #339 / #340 / #343, + the S38 design doc #342) — the operator's "visualizer breakdown + polish" ask, shipped as four additive demo PRs once GPU was off the table:
  • #338 (S34) — span-highlight visualizer. The raw input rendered as a displaCy-style ribbon, each tagged span tinted by its confidence (the table's red→amber→green tiering) with the tag labelled beneath; dropped spans read as literal colour gaps. The decoder already carried char offsets — flattenTree was dropping them.
  • #339 (S37) — per-stage timing breakdown. A stacked bar (shape+kind / classify / resolve) sized by each stage's wall-clock; the neural inference visibly dominates. Measured in the demo via performance.now(), the one-time DB load excluded so resolve isn't skewed.
  • #340 (S40) — copy-result-as-JSON. A "Copy JSON" button that yields a clean paste-into-an-issue object (input + components with offsets + resolved place). Enter-submit and the loading splash already existed.
  • #343 (S35) — containment hierarchy tree. The parse as the tree it actually is — region ⊃ locality ⊃ {street ⊃ house_number, postcode} — in a collapsed <details>, read straight from result.tree.roots (the nesting the flat table throws away).
  • #342 (S38 design doc) — the service-worker cache groundwork. We did NOT ship a SW: the heavy assets are fetched cross-origin from HF by version, HF 302s to an hourly-expiring signed CDN URL, so a blind cache would store a dead redirect. Documented the gotchas + a hand-rolled Cache-API sketch + a 5-behavior Playwright gate, so a focused session starts fast instead of half-shipping at the shift tail.

What went well

• Measure-first paid off twice. The whole shift exists because we re-measured German in its native order and found the "collapse" was largely an eval artifact. Tonight's S1 probe (below) did the same to the anchor's role.
• The corpus pipeline mirrored v0.4.1-de exactly — overlay manifest, base shards by reference, no rebuild of 677M rows. Clean, fast, verified on the volume before launch.
• Pivoting cleanly when GPU was called off. DeepSeek's "(b), no more GPU" verdict could have read as a dead end; instead it freed the back half for the demo tier the operator had explicitly asked for. Three small additive PRs, each typechecked + logic-validated + e2e-safe, none touching the v0.6.0 production path.
• The demo work degraded gracefully by design. SpanHighlight and TimingPanel both render null when their data is absent (older models with no offsets, an all-O parse), so the table alone always tells the story — no new failure surface.

What could've gone better

• A branch-base trap: the committed native-order asset read 0.0% because that branch predated #322 (the anchor inference code); parseAnchorLookup wasn't compiled in. Cost ~15 min of confusion before I traced it to the branch base. Lesson logged: an anchor eval needs #322 on the branch.
• The generalized builder (S3) hit real walls I should have anticipated: FR's OA CSV is 2.5 GB (over the 1 GB buffer → needs streaming + reservoir sampling), NL's template reformats the postcode. DE-parity is proven; the rest is held, honestly documented, not half-shipped.
• The demo tier serialized on merges. S34/S37/S40 each touch ResultPanel.tsx, so each waited on the previous to merge before it could branch cleanly (the last one rebased onto the prior). It worked, but a single demo PR bundling all three would have spent less wall-clock on merge-poll loops. Worth batching same-file UI changes next time.
• I couldn't visually verify the demo in-shift. The ribbon/timing/copy UI only renders after a real parse (60 MB of model + WOF assets), so I leaned on tsc + the docs build + standalone segmentation tests rather than an actual screenshot. Sound, but the operator should eyeball the deployed demo — flagged below.

Decisions made autonomously

• Did NOT bump the critical dependabot alert (vitest, dev-scoped). The exploit needs the Vitest UI server listening (never run); bumping the test runner mid-shift risked destabilizing the CI gate I depend on. Logged for the operator instead. (Alternative: bump it — rejected as higher-risk-than-reward.)
• Held build-locale-shard.mjs rather than ship a generalizer that doesn't yet beat the German builder.
• Reused the pilot model card (33 labels, identical schema) for the v0.9.2 eval rather than block on a fresh card export.

Surprises (captured fresh)

• S1 — the anchor is load-bearing in BOTH orders, not a US-order band-aid. v0.9.1 anchor-on model, DE locality, 2×2: US/off 35.9 · US/on 45.8 · native/off 48.4 · native/on 83.8. The anchor adds +35.4pp on native order (vs +9.9 on US). Revises the #321 guess that the anchor merely rescued a US-order symptom. Caveat: c=0 deprives an anchor-trained model of a learned signal; the clean test is whether v0.9.2 reaches native perf WITHOUT the anchor (the gate's 2×2).
• S2 — the order artifact is systemic but the anchor is the bigger lever. Re-measured the PR3 v0.9.0-selfcond model (no anchor): US-order 35.5 → native 48.2 (+12.7pp, Sachsen-driven 36.8→62.2; Berlin flat 34.2 both). So the earlier "collapses" were partly the eval order — but the no-anchor native ceiling is only ~48% (= v0.9.1 anchor-off). Two separable levers: order-training lifts the anchor-OFF intl ceiling; the anchor unlocks native 48→83 (and Berlin needs it). Resolved the "not yet re-measured" caveat in memory.
• S-diagnostic — the anchor's intl harm is structural, not posterior. Forcing the German postcode posterior to DE=1.0 on v0.9.2's intl eval left it at 44.5% (= uniform). The additive anchor at the trailing postcode is the harm, independent of its feature content. → v0.9.4 dual-injection design.

Open questions for the operator

• THE fork (#327) — RESOLVED by DeepSeek (delegated authority), ~11:50 UTC → (b) accept the asymmetry. Three retrains — both-order corpus (v0.9.2), region tail (v0.9.3), and the architectural fix, dual-injection (v0.9.4) — ALL left the anchor-on international number immovable at ~44% (gap to native ~40pp). The anchor's native-help (+35) / international-hurt (−4) asymmetry is fundamental — not corpus, not region, not injection position. With the operator away, the third DeepSeek turn made the call: (b) — the native gain is large/robust/generalizing (US 96.4, FR 84.9 hold), the intl penalty small/stable, and anchor-off intl already ~48% (the c=0 path serves international order), so burn no more GPU. The turn's mechanistic read reframes the multi-locale plan: the anchor's direction is learned per locale from the dominant training order (US trains postcode-after-city → anchor looks left → never hurt US; German native is postcode-before-city → looks right; one shared vector can't serve both). The plan is constrained, not dead — the documented next thread is a country-conditioned anchor vector (a future iteration with its own gate), plus a no-GPU operational lever: route international-order inputs to the c=0 path via a lightweight order check. None of v0.9.2/3/4 promote; v0.6.0 stays production. The operator can revisit the next-thread choice; the shift's experimental arc is closed. (Decision record: 2026-06-06-v0.9.2-eval.md, §"accept the asymmetry"; #327.)
• The next anchor thread is the country-conditioned vector — pick when you're ready. The shift mapped the ground and #327 now tracks it; it gates the multi-locale retrain program, so it's the lever before any ES/IT/NL/FR retrain. A cheap no-GPU interim (route international-order inputs to the c=0 path via a lightweight order check) is filed there too.
• Eyeball the deployed demo. The span ribbon / timing bar / copy button were validated by tsc + the docs build + standalone tests, but not by an in-shift screenshot (they only render after a real 60 MB parse). A 30-second look at the live demo confirms the visuals land.
• The demo's service-worker cache (S38) is the operator's other explicit ask, not yet started. Today the ~60 MB bundle is HTTP-cache-only ("the browser caches everything" — fragile, no offline, no version awareness). Next session: evaluate @docusaurus/plugin-pwa vs a hand-rolled Cache API, precache model.onnx + tokenizer.model + wof-hot.db keyed by selectedVersion (evict old blobs on switch). Deserves its own focused session with real browser testing — too easy to half-ship at end-of-shift.
• FR (#330): the FR gap is venue (0%) + region (19%) — convention gaps, not basics. Worth a targeted FR venue+region corpus supplement on the next multi-locale push?
• The 19 high + 1 critical dependabot alerts (vitest dev-scope) — bump in a dedicated chore(deps) pass?

Concrete next steps

• Country-conditioned anchor (#327): a per-locale anchor direction (DE specializes on postcode-before-city, US on postcode-after-city), so no shared vector is forced to compromise. Its own gate via de-order-eval.sh. The interim no-GPU lever (order-check → c=0 route) is the cheap first move.
• Multi-locale shards: DE / NL / FR / IT / ES are all shard-ready via build-locale-shard.mjs (ES via a per-part conform map on the raw CNIG upstream). The gate before any retrain is the country-conditioned anchor (#327), not the data.
• Demo S38 — service-worker precache for the ~60 MB bundle, keyed by selectedVersion (see above).
• FR (#330): assemble a venue+region FR shard (real OSM/WOF POIs + FR région forms), measure with per-locale-f1.ts.

Numbers

metric	value
Shift window	04:16 UTC → 14:00 UTC
PRs merged	22 — #321-326, #328-329, #331-344 (4 demo: #338 S34, #339 S37, #340 S40, #343 S35 + #342 S38 doc)
Issues filed / resolved	#327 anchor fork RESOLVED (b, DeepSeek delegated authority), #330 FR gap / #239, #241
Models trained	3 (v0.9.2 both-order, v0.9.3 region-tail, v0.9.4 dual-injection — all 20k, all negative)
Modal spend	~$12 of $15 (3 runs; GPU called off after v0.9.4)
DeepSeek consults	1 session, 3 turns (turn 3 = the delegated-authority decision) + the posterior diagnostic
Data acquired	Italy (468 MB) + Spain (451 MB) OA countrywide; DE/NL/FR/IT/ES all shard-ready
NaN incidents	0
CI failures	0
Demo regressions	0 (v0.9.x research line + additive demo tier; v0.6.0 production untouched)