Repository navigation
Bridge projected earnings histories into actual Axiom AIME/PIA execution - #450
Merged
Merged
Conversation
Add populace_dynamics.axiom_benefit_bridge, an opt-in transport from retained cohort histories to the real axiom-rules-engine `run-lifetime` binary. It builds lifetime-request/v2 requests in the retained SSA Case A runner's format, runs the pinned engine (commit 49f2225, the binary that reproduced Case A) as a subprocess against pinned compiled artifacts, parses and validates the typed response, and returns per-person results carrying provenance: engine, artifact and candidate-module SHA-256s, request and stdout digests, selected rule versions, and the fact that the RuleSpec candidates are unaccepted. The bridge performs no benefit arithmetic; indexing, top-N selection, averaging and the PIA formula run inside the engine. The bridge refuses rather than guesses: - every elapsed-window year through the year before entitlement must be supplied; gaps raise BenefitBridgeRefusal listing the years (a leading gap may be accepted only by an exact, recorded acknowledgement); - pre-2014 years come only from an explicit CallerSuppliedHistory whose synthetic flag travels onto every row; otherwise the gap is recorded; - projected labor income is used as creditable earnings only under an explicit ProjectedAmountDeclaration matching the history's unit; - missing birth years, wage-index years, contribution-base checks, and mortality removal before entitlement all refuse; - the PIA output is requested only for the candidate's 2026 first-eligibility cohort; other people get AIME from the AIME-only artifact compiled from the same 415(b) module. The engine parses decimals with rust_decimal's from_str_exact, which rejects exact expansions of most fractional binary64 values, so the binary64-to-decimal encoding is an explicit caller choice recorded per row. Tests: fake-engine unit tests for request construction, refusals, response validation, binding tampering and the subprocess path, plus cohort tests on invented people produced by actual engine steps. Integration tests (skipped without the engine) rebuild SSA's published Case A and Case B requests byte-for-byte through the bridge and reproduce AIME $5,825 / PIA $2,609.80 and AIME $11,463 with identical stdout. Register the module in POST_REVIEW_SOURCE_EXCLUSIONS, its pinned tuple and the reachability guard. Remove a byte-identical duplicate of test_source_reachability_includes_implicit_package_initializers left by the stack reconciliation (ruff F811; the second copy already shadowed the first, so collection is unchanged). Tier manifest recollected: unit 2,246, artifact 2,672, integration_psid 848, reproduction_legacy 520, oracle_policyengine 159; total 6,445. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Independent review of ef53849 found four defects; each now has a regression test. - OrdinaryRetirementDeterminations accepted any indexing year before entitlement. Under the bridge's scoped assumptions (no disability, no death before entitlement) 42 USC 415(b)(3)(A)(ii)(I) with 415(a)(3)(B)(i) fixes it at the year of attaining 62 minus 2. On the real pinned engine, INVENTED flat earnings for a 1953 birth year entitled in 2023 gave AIME 5,806 when indexed to 2013 and 7,835 when indexed to 2021, with no refusal. The constructor now refuses any other indexing year, an age-21 year other than birth year + 21 (or + 20 for a January 1 birth, per the retained POMS RS 00615.015 extract), and attaining 62 before 1979 (415(a)(3)(A)). - prepare_request sent rows dated before the birth year to the engine, where they entered the top-N selection. It now refuses them and lists the years under gaps["years_before_birth"]. - execute_request ran any PreparedRequest, so a hand-built or dataclasses.replace-d request could document rows, determinations or a wage index that were not in the wire next to the wire's digest. It now rebuilds the request from the recorded inputs and raises ValueError on any divergence in wire, request, outputs or artifact role. - AIME-only results listed the 415(a) and 2026 bend-point modules in provenance.module_sha256, although the AIME-only program holds only 415(b). AxiomEngineBinding gains artifact_sources; the reviewed binding attributes 415(a), 415(b) and the bend points to the AIME/PIA program and the 415(b) source alone to the AIME-only program. Bindings without recorded sources keep the previous list-everything behavior. Case A and Case B requests remain byte-identical to the retained requests, and the real-engine tests still reproduce AIME 5,825 / PIA 2,609.80 and AIME 11,463. Re-running the r3 fixture passes with this code gave the same status, reasons, gaps, AIME, request and stdout digests for all 60 person outcomes; only AIME-only module attribution changed. Tier manifest: unit 2,250; total 6,449. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
MaxGhenis
changed the base branch from
land-dynamics-stack-20260922
to
master
September 22, 2026 22:29
Tier manifest recollected on the merged tree: unit 2,341; total 6,540. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
First integrated earnings-to-benefit chain: projected Dynamics cohort histories go through the actual Axiom rules engine to AIME (and ordinary PIA where the candidate covers the cohort). Stacked on #449.
src/populace_dynamics/axiom_benefit_bridge.pybuilds lifetime requests in the retained SSA Case A runner's format. It runs the pinnedaxiom-rules-engine run-lifetimebinary (engine commit49f2225) against pinned compiled RuleSpec artifacts and parses the typed response. It does no benefit arithmetic itself: indexing, computation-year selection, averaging and the PIA formula all run inside the engine. Every result records the engine binary, artifact and rule-module SHA-256s, request and stdout digests, and the fact that the rule candidates are unaccepted.Calibration
SSA's published 2026 Case A and Case B, rebuilt through the bridge, produce requests and engine output byte-identical to the retained runs: Case A AIME $5,825 and PIA $2,609.80, Case B AIME $11,463. These are the same two published cases the earlier runners matched, so this adds no new benchmark.
Refusals instead of guesses
The bridge raises a refusal listing the years involved for:
Pre-projection history enters only as an explicit caller-supplied history whose synthetic flag appears on every row. Projected labor income becomes covered earnings only under an explicit caller declaration. The float-to-decimal conversion is an explicit, recorded choice, because the engine rejects exact decimal expansions of fractional floats.
Run on the r3 fixture (12 invented people)
With nothing added, all 12 are refused, with reasons recorded (missing pre-2014 and post-2022 years, unpublished future wage indexes). With explicit synthetic zero rows for pre-2014 years, the engine computes AIME for two people per entitlement scenario. These exercise the transport only; they are not estimates. Evidence:
microcosm-launch-evidence/dynasim-parity-20260909/axiom-benefit-bridge-20260922/RESULTS.md.Review
An independent Opus reviewer rebuilt the Case A/B requests outside the test suite and diffed them. It ran the real-engine integration tests, which executed rather than skipped, and fixed four defects in
8caf372f, each with a regression test: the indexing-year check, rows before birth, and two provenance and refusal gaps. Bridge tests and the birth-evidence identity tests pass, and ruff is clean.Not included
This PR adds no COLA, claiming adjustment, eligibility, payment, survivor benefit, benefit trajectory, or DYNASIM comparison. The Axiom candidates remain unaccepted until the encoder closure runs (blocked on a protected install and Codex quota).
🤖 Generated with Claude Code