Night-10 postmortem — 2026-06-10 (the flag-plant shift)

Autonomous shift, 08:08–15:00 UTC equivalent (operator offline, extended-trust delegation). Plan: 2026-06-10-NIGHT-SHIFT-PLAN.md. Headline mandate: ship Run B as v4.2.0 iff the ship gate passes; merge the flag-plant set; then eat the queue. All three happened; the queue is substantially eaten.

What shipped

• v4.2.0 — the v1.0 parity flag-plant — live on every surface, byte-verified — and in true ship config it beats v0 on the CLEAN arena (41 vs 29), the first time ever. Ship gate 4/4 (2026-06-10-night-10-ship-gate.md): honest-eval VT identical to baseline, presets clean (+ the intended affix split), int8 ≤0.1pp of fp32 (deterministic quant, md5 9eb4a99f…), DE native 90.9. Merge sequence #468 → #469 → #491 (epic #466 closed), zero conflicts (the operator's squash choreography worked exactly as designed), zero merge-wall blocks. Bookkeeping #494 (cards w/ honest init_from lineage, ledger row, scorecard re-emit, status/releases contract pages). HF staged + default; the R2 leg (10 objects, served-model md5 = the gated artifact); publish.yml → npm 4.2.0 (registry-direct verified), tag + release object.
• promotion-gate.sh (#479, closed via #495): gate-specs-as-contracts; validated by reproducing the manual ship verdict bit-for-bit (12/12 floors, max int8 delta 0.1pp).
• US source-independent holdout (#472, closed via #496): 6,453 NAD-only rows, 44 states. No memorization cliff — v4.2.0 within noise of v4.1.0 on never-seen-lineage data; the consolidation's gains are real.
• Address-point tier (#476, closed via #497): VT prototype. Coord p50/p90 3.4/7.4 km → 0.0/0.0 km, p99 277 → 6.2 km, 93.1% hit rate; admin attribution untouched by construction. The 6.9% misses are #483's interpolation population; the shard is its gold.
• postal_city alias table (#475 asset, via #499): built (19,880 pairs, 51.3% divergent) — and its measurement REFUTED my own published attribution (below).
• Parser hardening part 1 (#481, via #500): one #decode (a third drift surface had already grown), repairs-in-both recorded, +2 policy sharp-case tests, loud lexicon validation. Scorer numbers reproduced exactly post-extraction.
• Docs/infra: RELEASING.md R2-leg correction (#501); REPRODUCIBILITY.md + the pinned- toolchain verifier (#480 safe parts); stale worktrees removed.
• Issues filed from findings: #492 (architecture escalation, evidence-backed, cheap- probe-first), #493 (lossless decomposition spec), #498 (census-designation name-credit — 54% of NAD locality misses, the #386 class US edition). Design notes posted on #478 (config surface, routing overlay, pre-registered v0-only→0 metric) and #487 (format-gap audit + TIGER-EDGES recipe + gate draft).

What went well

• The gate stack carried the night. Pre-registered ship gate → binary verdict → no 3am judgment calls. The promotion-gate runner then reproduced the verdict mechanically — the night validated its own tooling on its own decision.
• Probe-first kept paying. The fill-rate probe's heir (the alias-table join) killed a wrong attribution in 20 minutes; the smoke test surfaced Barre City ≠ Barre Town before it became a wrong-answer class.
• Background CI watchers + interleaving: seven PRs merged with zero idle waiting; every wait window held a reservoir item.

What could've gone better

• I published a wrong finding and corrected it hours later. The NAD note's Finding 2 attributed the locality gap to vanity cities; the measurement said census designations (1/461 vs 54%). Correction block published same-night, #498 filed with the real numbers. Lesson: classify the misses BEFORE naming the cause — the attribution was plausible, available, and wrong.
• The arena harness graded the wrong model silently (env vars, not flags; the identical-to-baseline numbers were the tell). Caught before publication; the env-var interface is a footgun worth a flags PR.
• The release pipeline had an undocumented mandatory leg (R2). Cost ~20 minutes of cache-vs-origin diagnosis mid-release; now documented with verification commands.
• My benchmark hygiene: the byte-stability stash-check for #500 compared the same compiled tree twice (worthless); replaced with scorer re-runs against known numbers.

Decisions made autonomously (alternatives considered)

1. Ran honest-eval with zero-filled gaz clues (harness lacks the flag) — accepted as conservative-valid since it PASSED degraded; alternative was harness surgery mid-gate (rejected: gate integrity over completeness).
2. Repairs-in-both for parseWithLogits — reconcile must see user-path tokens; the opts were silently ignored before, so no default change. Alternative (document the skip) rejected as preserving a latent divergence.
3. Arena dip (−2/3pp whole-parse) reported with caveats, NOT treated as a gate — the arenas were never pre-registered as ship criteria; adding one retroactively is gate drift in the other direction. Flagged for morning eyes instead.
4. #487 eval build deferred (census downloads + shapefile parsing at hour 7) — recipe posted instead; risk-ordering guardrail applied.
5. #475 consumption deferred after the measurement showed the NAD eval can't see the vanity-city mode — the claim waits for an eval whose inputs carry postal surfaces.

Open questions for the operator

1. The arena whole-parse dip RESOLVED before shift end: gaz/anchor support added to the arena harness; in TRUE ship config v4.2.0 scores 41/71/18 vs v4.1.0's 22/62/11 and v0's 29/39/26 — every prior anchor-era arena number was handicapped, the "dip" was the harness, and v4.2.0 beats v0 on the clean arena for the first time. Scorecard updated with the correction block.
2. #397: close as resolved-by-process, or re-title to the test-restores-symlink ask?
3. #492 architecture escalation: the cheap probe (dedicated affix head) is specced and gated — fund it when?
4. The night-9 → night-10 arc (silent gate drift caught → canonical bars restored → stated re-baseline shipped) is a strong blog story; draft deliberately not written tonight — want it?

Concrete next steps

• Queue head (epic #488): #478 implementation (config surface first — design reviewed on-issue), then #498 (designation credit, NJ bellwether), #481 part 2 (TLA, repo.ts), #487 eval build (recipe on-issue), #483 interpolation (gold standard now exists).
• #480 remaining: loader strict-mode, curriculum stamping, snapshot publishing.
• Eval follow-ups: gaz-fed arena harness; a postal-surface eval for #475.

Numbers

shift span	08:08–15:00 UTC (operator offline)
models trained	0 (treadmill guard held; 1 export + 2 quant runs, ~3 min A100)
released	v4.2.0: npm (13 pkgs, registry-verified), HF + R2 (md5-verified), tag, release object
PRs merged	10 — #491, #494, #495, #496, #497, #499, #500, #501, #502, #503 (gaz-fed arenas)
issues closed	#466, #472, #476, #479 (+ #475/#481/#487/#478 advanced with scoped completions)
issues filed	#492, #493, #498
evals run	ship gate ×2 artifacts, arenas ×2 (one invalid, caught), NAD holdout ×2 models, VT tier on/off, gate-runner validation
findings corrected by measurement	3 (NAD Finding 2; the stash byte-check; the arena 'dip' — handicapped harness, resolved to a +19/+9/+7 sweep)
CI failures	1 (MDX raw-angle, fixed in 8 min)
merge-wall blocks	0