[MaxGhenis]

WIP. Do not merge. The coordinator lands this after independent review and the tier-count slot.

This is milestone F0 of the mortality, groups and financing lane: the build of Registration 18's A-COLA trust-fund test. It does two things:

• puts the stage-1 build on master;
• adds stage 2 and the close-out.

It runs nothing on real data and opens no comparator.

Sources

• Registration: #42, comment 5919983573, authorized under decision d658.
• Lane design: docs/design/mortality_groups_financing.md (Design: mortality, groups and financing lane (priced milestones, approved under d1922) #558), approved by Max on 2026-10-23 (d1922) at 2c19f5b6.
• Builder briefs: the cleared builder brief (2026-09-30) and its stage-2 addendum, EV/acola-build-20260930/stage2-brief-20261024.md, both kept outside the repository. The builder was a fresh blind lane. Nobody exposed to an A-COLA comparator shaped this build.

Commits

1. b1b2561f: replays stage 1 (9de858d0) onto master with git cherry-pick.
   • The two tier files conflicted and keep master's version. They are recounted in the last commit.
   • The other 24 files are byte-identical to 9de858d0. The stage-2 builder audited them and found no bug.
2. 75ef143a: scripts/compare_acola.py, the standard-library comparison program for the validation lane after the seal opens. Its tests use wholly invented source values.
3. d17d1d00: mechanism-L equivalence.
   • scripts/acola_mechanism_l_equivalence.py runs rows C0, C1 and M0–M4 on a seeded invented population: 69 people per row, 483 records.
   • The golden file was produced at the registered code commit dd222b9 in a temporary git worktree, and the HEAD test reproduces it bit for bit.
   • The SHA-256s are 07c59d69…87de for the golden file and 5d7eef14…7a27 for the script.
   • No REG18_MECHANISM_DRIFT.
4. 1b4923b3: stage 1's TR 2026 Table V.C1 capture and master's data/tr2026 accessor agree on the intermediate COLAs effective December 2027–2029.
5. e422ad55: a test on invented late claimants born 1917–1942 pins mechanism L's legacy 48-month delayed-credit cap. It ships with the design note docs/design/acola_trust_fund_implementation.md, whose section "Disclosures for the second registration comment" goes to Max for the second Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42 comment.
6. f78efbc6: the five acola_track sources join POST_REVIEW_SOURCE_EXCLUSIONS, and the tiers are recounted.

Invariants

The design note's "Arithmetic and invariants" section states these, and the tests check them. Every input is invented.

• Accounting:
  • component totals sum to each scenario's total;
  • the baseline total equals its OASI and DI parts;
  • P equals the baseline-share-weighted OASI and DI P wherever both denominators are positive;
  • a zero component denominator makes that P undefined, not zero.
• Zero reform: every change statistic, falling share and null is zero, and benefit-level medians stay equal and positive.
• Equal memberships: equal scenario memberships imply that P equals the ratio of means, within floating tolerance.
• Weighted median:
  • it is the exact half-weight midpoint, using Fraction weights;
  • it is invariant to permutation and bounded by its range;
  • ties and zero weights are handled exactly.
• Determinism: fixed seeds give the same output, shared draw states are used, and the legacy source pins are exact. No missing draw becomes an empty completed draw.
• Comparison arithmetic:
  • it is done in Decimal;
  • the four-corner tests check containment, monotonicity and positive denominators;
  • interval subtraction is [a,b] − [c,d] = [a−d, b−c];
  • seal files are checked before any comparison;
  • no acceptance threshold or verdict is computed.
• Intended exceptions:
  • Dual entitlement can raise a spouse's excess by more than the claimant's own cut, so individual benefits and P can rise under a cut. That exception is tested explicitly, with no universal tolerance.
  • Worker bounds allow the dime-rounding losses each case accrues.

Checks

Run by the lane session at f78efbc6's tree, with an empty PSID root:

Selection	Result
tests/acola_track	288 passed
Legacy selection (the cleared brief's section 7 list) plus tests/track_a_v2/test_dry_run_protocol.py	490 passed, 2 skipped
Birth-evidence exclusion and reachability (-k post_review)	2 passed
Full collection, -k tiers_match_policy_manifest	1 passed

The builder also ran Black at 79 columns, Ruff and git diff --check over the 25 Python files it added or changed.

The dd222b9..origin/master diff of the mechanism-L paths adds two files, engine/forward_earnings_s1_v1.py and ss/statutory_delayed_credit.py, and changes no existing file.

Tiers (recount-tiers.py):

Tier	Before	After
unit	18,088	18,351
artifact	4,253	4,278
Total	24,619	24,907

This PR edits tests/tier_counts.json, so its merge waits for the coordinator's tier-count slot.

Coordination

These PRs overlap this one in the exclusion tuple: #562, #561, #560, #230 and #223. No protected benefit module overlaps.

Provenance

• Builder: a Subfleet build lane (GPT-6.1 Sol, job 20261024-073819-mgf-f0-acola-stage2). Its sandbox could not write .git, so the lane session:
  • replayed and committed the work in the brief's steps;
  • reran the worktree equivalence the sandbox had blocked, recording it in the golden file's provenance;
  • reran the checks above.
• Data: no PSID data file was opened, no real-data statistic was computed, and no d1408 script was run.

Plan

• Stage 1 on master
• Stage 2: comparison script, equivalence, cross-check, disclosure, design note, housekeeping
• Independent blind review
• The coordinator's tier-count slot and landing
• F0R: Max's second Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42 comment, then the one-shot run

axiom: n/a: microsim-only registered test build

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
social-security-model	Ready	Preview	Oct 24, 2026 2:52pm UTC

[MaxGhenis]

Superseded by #585, a blind replay of this build. The builder tree stays on preserve/acola-f0-builder-tree-20261024 (0556744) until #585 lands. The blob-by-path comparison of the two heads (7 differing paths, all explained) is recorded in the lane's evidence folder.