Registration 7: the first estimates report (forecast ledger entry 8)

REGISTRATION 7 — 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, the §11 sealed-invocation restatement at
070cbb7, amendment 2 at 4104d3d with corrective revision 10.1 at
f771b49 (the current design identity), implemented at 1c7f354.
The design is at revision 10.1.

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 (unchanged): 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 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 populace-dynamics#<this issue> and --registered-configuration docs/registrations/first_estimates_registration_7_configuration.json.

What changed since registration 5

Registration 5's incident opened a chain that is now closed, every
link referee-gated and on the record (#299):

• PR #301 scoped the conflicting-exact-birth-year guard to the
  report population (the incident-5 defect; the eight conflicted
  persons touch no report surface).
• Amendment 2 (PR #302, ratified 4104d3d) resolved the deeper
  defect #301 exposed — 6,392 of 30,482 population members undated by
  every ratified source: §3.1's seed-coordinate clause 3 and
  unresolved class, the Stage-C split around Stage C.5, repeal of
  the whole-file numeric-birth guards for the counting law, and the
  frozen §10 birth-timing sensitivity (coherent ±1 stress scenarios:
  −$30.3B/−0.92% and −$312.6B/−9.47% on the $3,301.7B baseline;
  recomputed per draw; traveling with every publication until a
  ratified birth-timing resolution retires it).
• Revision 10.1 (PR #304, f771b49) corrected the §10 row's
  candidate share to 2,892/3,083 = 93.8% after the implementation
  referee caught the ratified text contradicting its own artifact.
• The implementation (PR #303, 1c7f354) landed after three
  referee rounds (ten findings, then three, then one, then zero) with
  the committed evidence reducer's artifact as its oracle — including
  the draw-0 dollar totals to the cent.

Rehearsal evidence (new; a precondition of this registration)

A full out-of-ceremony rehearsal of the coordinator's compute and
invariant phases — projection of all twenty draws, per-draw preparation,
batch preparation, and artifact assembly, stopping before publication —
completed on this code before this registration was created. Its log
and the assembled artifact's canonical byte count are recorded in the
launch comment. Registrations 3, 4, and 5 each died on a failure a
rehearsal would have surfaced in minutes; rehearsing before spending a
registration is now standing practice for this class.

The rehearsal is not a substitute for the ceremony and confers no
authority: it publishes nothing, its artifact is discarded, and the
registered run recomputes everything under the sealed interpreter.

Frozen configuration (the coordinator hard-asserts or records)

• Source: master 1c7f354 — 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.
• 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; rate legs 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 data/external/ssa_cola_history.json
  (sha256 12da7f0e…a38b, determination-year keyed, coverage asserted
  for 1979-2022 exactly).
• The staged inclusion law as ratified at revision 10.1: Stage A
  DI screen over all persons; Stage B any-slice candidate universe;
  Stage C origin classification (birth-free) for every candidate;
  Stage C.5 birth-year resolvability with the frozen
  excluded_birth_year_unresolved key (published even when empty;
  origins = C.5 + Stage-D + included); Stage C completed — the
  strict-< truncated-renormalized 6.B5.1 imputation under the
  code-frozen stock-imputation namespace — for C.5 survivors; Stage
  D's five ordered predicates with the frozen excluded_* keys;
  eligibility era ≥ 1979; the 0.80 coverage rule;
  cutoff-before-imputation. Birth years resolve under §3.1's four
  clauses with the five frozen source classes and the per-draw §10
  birth-timing sensitivity published alongside.
• Ledgers: as-of AIME; eligibility-year PIA; months-based
  benefit_factor; stepwise COLA with dime flooring; 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.

Candid expectations (house rule)

Every dollar is frame-relative, pre-alignment, labor-income-proxy.
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). What would look pathological: 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.

Measured before this registration (the committed evidence artifact),
and stated so the run cannot be credited with a surprise: the report
population is 30,482 persons; birth sources 13,727 inferred + 9,673
exact + 4,077 derived + 690 synthetic + 2,315 unresolved; canonical
candidates 3,083 (2,784 domain-resident; origins 1,495 modeled /
1,588 backfill); excluded_birth_year_unresolved expected empty;
included claimants 1,514 by source 0 derived / 74 exact / 1,440
inferred. The rehearsal's full Stage A-D funnel is recorded in the
launch comment, and the registered run is expected to reproduce it
exactly; any divergence is itself a finding about determinism and
will be disclosed as one.

Runtime and publication

Twenty draws of a 2014→2022 unsplit projection plus the statutory
pipeline; the sealed registration-5 run reached the invariant phase in
~40 minutes at the pinned eight-thread topology, so the compute is
hours-scale at most. 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 1-4 are committed; incident 5's
record publishes in the records PR that carries this registration's
configuration document (the append-only debt named on #299); any
further incident publishes to runs/first_estimates_incident_6.json.

Lineage from registration 6

Registration 6 (#305) was consumed by incident 6 — a
refusal-to-start on registration 5's stale durable attempt claim,
raised before any claim was bound to #305 and before any input was
loaded. The claim was adjudicated and removed with disclosure on #299;
the incident record commits with this registration's document. Nothing
about the code, the design, the document construction, or the
rehearsal evidence changed between the two registrations. The
pre-launch checklist now verifies: attempt claim absent, retry marker
absent, artifact pair absent, worktree pristine at the registered
commit.

Rehearsal evidence (the precondition, met)

Full out-of-ceremony rehearsal at 1c7f354 (worktree fresh, pristine;
launchd vehicle, nice 0, pinned eight-thread topology), completed
2026-07-26 02:19 ET, REHEARSAL PASSED:

• fit + preflights + materialization + twenty draws: 3,096.9 s
  (draws 104-136 s each after the first);
• batch preparation, including the per-draw §10 birth-timing
  sensitivity: 539.4 s;
• artifact assembly: 0.9 s; canonical artifact 19,513,372 bytes,
  schema first_estimates.v1, all thirteen top-level keys present
  (tables: modeled_award_flow / opening_stock / revenue; diagnostics
  include birth_timing_sensitivity and common_support_agreement).

Funnel spot-checks against the committed evidence artifact: the four
real birth-source classes draw-invariant at exactly
13,727/9,673/4,077/2,315 (sample SD 0.0 across all twenty draws);
draw-0 synthetic natives 690; excluded_birth_year_unresolved = 0 in
every draw (published, as the law requires); draw-0 included claimants
1,514 with included_birth_source__derived_projection_age = 0.0 in
all twenty draws — the ratified nil-contribution expectation is now
a measured fact across the whole batch, not only draw 0. Across draws:
included 1,518.1 ± 16.8.

The rehearsal publishes nothing and confers no authority; the
registered run recomputes everything under the sealed interpreter and
is expected to reproduce the draw-0 funnel exactly.

Start with docs/design/first_estimates_report.md and the registered configuration at docs/registrations/first_estimates_registration_7_configuration.json, then inspect scripts/run_first_estimates.py and the implementation at commit 1c7f354. Done means completing the sealed twenty-draw run and producing runs/first_estimates_v1.json with its .env.json sidecar, or recording the required incident before publication.