Registration: the first estimates report (forecast ledger entry 8)

[MaxGhenis]

REGISTRATION — the first estimates report (forecast ledger entry 8)

Registration

Class: registered estimates report (the class chartered by
docs/design/first_estimates_report.md, ratified at 6586b92,
amendment 1 at ee1221d, amendment 2 ratified at the implementation
merge 070cbb720cb2f10da7ffae8122a2269e58b39e6e — PR #285 nine
referee rounds; PR #286 nine referee rounds, final verdict SHIP with
the hardened-guard exploits replayed and held).

Output artifact: runs/first_estimates_v1.json with its paired
.env.json sidecar, integrity-bound by the primary's recorded sidecar
hash, written append-only by artifacts.write_new.

The canonical execution rule (§11, restated verbatim as the sole
normative wording): one registered run; publishes_regardless;
no_self_rescue; at most one coordinator-adjudicated, report-first
retry solely for an external pre-output failure yielding no
estimate-bearing information. A published v1, any changed
configuration byte, or a second failure of any kind requires fresh
registration — there is no same-ceremony v2 path. Preparation,
invariant, compute, or publication aborts publish an append-only
incident record (first_estimates_incident.v1, the frozen nine-key
schema) before any further step. The durable attempt claim and the
retry-consumption marker enforce these terms across crashes.

The sealed invocation (amendment 2, restated verbatim): before any
populace_dynamics import, the runner self-re-executes the same
interpreter as python -I -B -X pycache_prefix=<fresh-empty-sentinel-directory> scripts/run_first_estimates.py <registered-arguments>; the launcher
procedure is the launchd user-domain topology of the candidate-3
record (one-shot LaunchAgent, RunAtLoad, AbandonProcessGroup,
ProcessType=Interactive, no KeepAlive, launcher spawns-and-exits,
runner-held caffeinate -sim, bootstrap → health check → bootout,
runner ending as a pid-1 orphan in no application coalition;
publication performed by the coordinator after exit).

Registered arguments:
--registration-reference <this issue's number as populace-dynamics#<n>> (the reference string is this registration's
identity; the coordinator validates it against the canonical byte
bound and binds the durable claim to it).

Frozen configuration (the coordinator hard-asserts or records)

• Source: master
  070cbb720cb2f10da7ffae8122a2269e58b39e6e — a clean checkout; the
  guards refuse ANY tracked or untracked drift, ignored executable
  artifacts under the code roots, assume-unchanged/skip-worktree
  entries, redirected git routing, and an unsealed interpreter.
• Design identity: 6586b92 + amendments ee1221d and the §11
  sealed-invocation restatement ratified at 070cbb7.
• Candidate-3 binding: the gate runner's normative
  fit/preflight/materialization sequence with both CandidateSpec
  sha256s asserted against the registration-12 pins — family
  734a5b04f347c5d4904bbc6d5ab9a1c2876272d35284eedd2f450518acf1cec5,
  engine
  c9be28a28d6fcc3911723872386906af559f6d0e0d5c89a87f741a5b2c3eacd6
  (a mismatch aborts); unsplit population; draw indices 0-19
  (root seeds 5200+k per the RNG registry law).
• Statutory parameters: policyengine-us 1.752.2 (path and content
  hashes recorded in the artifact); rate legs read from
  gov/irs/payroll/social_security/rate/{employee,employer}.yaml
  with the 6.2%/6.2%/12.4% fail-closed assertions and no fallback;
  the committed SSA COLA-history extraction
  data/external/ssa_cola_history.json, sha256
  12da7f0e0d33fc53eaa31567d86c3cc035a49feefc2e4695bceea3379c2a38db,
  determination-year keyed 1975-2022 (verified against the official
  SSA series at the PR Entry-8 implementation: the first-estimates statutory pipeline (design 6586b92) #286 round-2 review, all 48 rows), with
  coverage asserted for 1979-2022 exactly.
• The staged inclusion law as ratified: Stage A DI screen over all
  persons; Stage B any-slice candidate universe; Stage C origins with
  the strict-< truncated-renormalized 6.B5.1 imputation under the
  code-frozen stock-imputation namespace; Stage D's five ordered
  predicates with the frozen excluded_* keys; the eligibility-era
  bound ≥ 1979; the 0.80 coverage rule; cutoff-before-imputation.
• Ledgers: as-of AIME; eligibility-year PIA; months-based
  benefit_factor; stepwise COLA with dime flooring per the SSA
  application convention; no recomputation (affected count
  published); the payment-window triple intersection; nominal
  in-window revenue with wage-base capping and no expectation layers.
• PSID staging: ~/PolicyEngine/psid-data (the label-verified
  readers' documented staging path).

Candid expectations (house rule)

Every dollar is frame-relative, pre-alignment, labor-income-proxy
— the design's three labels appear on every table, and the artifact's
closing claim is what the statutory formulas yield on the candidate-3
GATED_REALIZED reproduction panel, not what Social Security pays and
collects. No level anchor exists in-repo; no comparison to any
administrative total is made (the context ratio is
deferred_to_anchor_extraction by amendment 1). The registration
names what would look pathological rather than predicting magnitudes:
an empty or near-empty modeled_award universe, an exclusion class
absorbing the overwhelming majority of candidates, a coverage-ratio
distribution collapsing at the 0.80 boundary, or across-draw SDs
rivaling means. The odd-year earnings carry law and its affected-share
diagnostics accompany every annual table, with biennial companions.

Runtime and publication

Twenty draws of a 2014→2022 unsplit projection plus the statutory
pipeline — hours-scale. On completion (any verdictless outcome: this
class has no verdict), the coordinator commits the artifact pair on a
branch and opens the publication PR; forecast ledger entry 8
resolves at that PR's merge (p50 2026-08-02). Incidents, if any,
publish to runs/first_estimates_incident_<n>.json and are
cross-referenced.

[MaxGhenis]

Registered-configuration bytes frozen. The coordinator's --registered-configuration document — the code's own registered_configuration_echo over the loaded parameter bundle, canonical JSON, 6,376 bytes, sha256 3f66e516d82d433450c7cfd0ec95f85183fa6b45ceeb1e4aa66e336ca0a1249c — generated under a no-bytecode interpreter with the run's exact parameter stack and validated byte-for-byte by the coordinator at preparation. Invocation #1 (procedural pre-import refusal: the coordinator's own verification import had minted __pycache__ under src — the guard caught precisely what it exists to catch; no incident, no claim, one-shot untouched) and invocation #2 (argument-parser refusal pre-guard: the configuration document requirement surfaced) are disclosed; the caches are purged and invocation #3 launches under the registered topology with the frozen document.

[MaxGhenis]

Invocation #3 — preparation incident 1; this registration is CONSUMED. The coordinator refused the registered-configuration path as cross-root (the document sat in the coordinator's lane directory, outside the sealed repository root — the round-3 enforcement working as designed), published runs/first_estimates_incident_1.json (schema-valid, timestamped 2026-07-25T05:00:33Z), and created the durable attempt claim bound to this registration. The reason class is not external_, so per the canonical rule this failure requires fresh registration — no retry. Disposition (coordinator adjudication, report-first): incident 1 is committed to the repository as the ceremony record; the frozen configuration is regenerated against the fresh registration's reference and committed IN-ROOT as a tracked file (resolving the delivery contradiction between the cross-root rule and the total-cleanliness guard — a committed configuration satisfies both); the stale #288-bound attempt claim is removed under this disclosed adjudication exactly as the design provides ("removal only by fresh-registration adjudication"); the fresh registration supersedes this issue and cites this comment.