[MaxGhenis]

Launch-day prep for d1386, the next certified US build (a Route A rerun on main). Max ruled on 2026-10-15 to start it with whatever had merged. The 10-18 start didn't fire. On 2026-10-19 Max confirmed it launches that day on current main, without #1275 (d1556 moves to the following build).

Launched 2026-10-19 from this branch on main 8b0c1950b, which carries the #1186 Social Security fix and the #1227 EITC fix:

• The base on Modal, run route-a-rerun-base-20261018-8b0c1950b65f, after check mode passed.

  • The first attempt was cancelled after 822 s, when the launching Mac slept. That led to --run --spawn, below.
  • The relaunch COMPLETED in 5,988 s. Both attempts together listed at $6.18.

• Local stage 000/001 lineage evidence, under supervision.

  • compare-lineage matched everything except the stage-001 digest.
  • The d1712 check found only person.pre_subsidy_rent differing: 313 rows, 6.08e-15 relative at most, with identical sums. That is the 9/30 class.
  • The base was accepted on 2026-10-19.

• The guard sweep passed: 107 of 107 reform passes.

• The gate dry run ended AT-RISK only (exit 2). At base weights, the QRF register and export input mass both depend on the solve.

• The release started 2026-10-19 22:21Z from a read-only export of this branch's driver. After its solve, 8.3 h in, it failed four terminal gates, and the next build needs Max's rulings on them:

  • the Social Security total came out 6.4% under SSA's: retirement benefits were 8.5% under, against 3.9% in the last certified release;
  • three new QRF tail columns came out over 0.75;
  • bond_assets is now stale.

  None of these is in what this PR changes, and the release's checkpoints are kept for a rerun.

Commits after the first review (of 4010c4b84):

• f053b3eaa and 54f6dbc34 answer it;

• 0779d4280 adds --run --spawn;

• c0fb05094 clones the fetched base instead of copying it;

• fd557a50d gives the gate dry run a populace-us- release id;

• f3341e507 answers an Opus 5.5 review of those five, which asked for changes:

  • Spawn: the "receipt either way" claim was false. Step 5 now names all four end states, and base.fc blocks a second paid launch.
  • Step 6 reuses a recorded check only when its config matches, including W's HEAD and every file argument's sha256.
  • tools/route_a/lineage_class_check.py is the d1712 check that ran. The provenance copy has the same syntax tree before formatting; it has example and Hypothesis tests.
  • Step 6b renews the QRF register under d1711.
  • Step 7 starts the release from a read-only export of the driver.

• 4e1392430 answers the GPT-6.1 Sol review of f3341e507, which asked for changes:

  • Step 6 hashed output files as inputs, so the second pass refused a finished sweep. It now hashes named inputs only.

  • Step 5 matches the attempt by call id and the receipt by attempt_id.

  • Step 3 gives each launch its own log and waits a bounded time.

  • Four more lineage-check tests; deleting the total-tolerance check now makes one of them fail.

  • Left open: the step 6 checks run without the driver preflight's HF_HUB_OFFLINE=1. They ran this way, and changing their environment would invalidate their recorded results.

  • b11330202 and 60a75bf88 answer two more Sol rounds. The last one approved with nits, and the nits are fixed in 60a75bf88.

What

• Base plan. docs/us-modal-stage-route-a-rerun-base-plan.json is a Modal base plan for a Route A rerun on main:

  • the Connecticut-fixed block ladder (d713: 6840b990…, 18,991,218 bytes), now on the inputs volume;
  • main's corrected congressional-district crosswalk (a347303f…), as a separate us-puf-support-base-rerun registration, so the historical us-puf-support-base and its local reference stay pinned to 4b57d15;
  • the commit, PICK_MAIN_COMMIT_ON_2026_10_18, a placeholder that validate refuses.

  The list-price ceiling is unchanged at $14.90.

• Lineage evidence. tools/route_a/make_local_reference.py produces fresh local stage 000/001 digests for the chosen commit, so compare-lineage can run strictly.

  • The reference's argv and environment are written from the plan's registry. The masked run_config, source and dependency identity, hash seed, pipeline and checkpoint digests carry the evidence.
  • The only local measurement is a whole --stage all attempt that the disk floor killed: 2,055 s and 11.66 GiB peak RSS. It likely ran past stage 001, and the supervisor misses processes that had already exited, so it is a rough upper estimate for stages 000/001, not their measured cost.

• Driver. tools/route_a/route_a.sh keeps its published feed and crosswalk pins as defaults, but takes FEED_SHA, FEED_BYTES and CROSSWALK_SHA from settings when the chosen commit changed them. Main's release feed is now 965a29ac…, and the base feed is kept separately.

• Docs. The launch-day sequence is docs/us-route-a-rerun-20261018.md, with a settings template, a runbook delta and the prep report. The sequence:

  • records Max's two rulings of 2026-10-19:
    • d1711: d490 covers this build's QRF tail register, although the d177/d450 waivers expire at it;
    • d1712: a repeat of the 9/30 stage-001 float-rounding difference may be accepted, and anything else stops for him;
  • runs the release guard checks under supervise.py.

Inputs checked against main

• Unchanged: the ASEC 2022–2024 pins, the education archives, the PUF and the ACS.
• Not pooled: main added ASEC 2025, but Route A's explicit pooling excludes it.
• Changed: the release feed and the crosswalk, both handled above.
• Since the prep base (412f9165):
  • The base builder, release tool and gate preflight/dry-run tools are unchanged.
  • The new US runtime modules are reached only from the ACS-local tools (ACS local release: supply source-backed SPM independence (is_spm_independent_minor_role) so every SPM unit has a classified adult #1263, Store the ACS local donor's retired inputs under the live ones #1309).
  • The lock bump is UK-side.
  • The shared release contract in microcosm-data (contract.py, release.py, publish_cli.py) changed with the UK UK local-area release: publish a K=15, 150,000-household dense dataset in at most two builds #1301 disclosure work. The release's own gates and the offline publisher preflight are the check on the US side.
  • The driver's resolve check passes on 8b0c1950b.

Known gaps

• validate omits measured_locally for the rerun tool because MEASURED is keyed by tool name. This is left as is: those measurements are of the 4b57d15 base, and the cost-cap figures don't use them.

Tests

377 isolated tests pass. Four environment checks were deferred to the launch-day env. Lint, formatting of the changed files, Bash syntax of every launch-doc block and tools/ci_test_plan.py verify pass.

🤖 Generated with Claude Code