Build N SSI take-up banding: 65+ seed at 8.4% collapses aged baseline to 0.94M vs SSA 2.42M; reform deltas look anchor-adjacent for the wrong reason

[MaxGhenis]

Summary

Build N (buildn-sparse-rmloss100-c3e378a, certified default on bundle 4.22.1) regresses the SSI baseline surface: the take-up seeding is age-banded and the 65+ band landed at 8.4% seed-true (18–64: 42.8%; under-18: 30.1%; 33.4% overall vs Build M's flat 67.0%). Measured with pe-us 1.764.6, TY2026:

baseline SSI	Build N	Build M	SSA
under 18	0.12M	0.12M	0.98M
18–64	4.36M	3.90M	3.99M
65+	0.94M	2.37M	2.42M
total	5.42M / $51.4B	6.39M / $58.5B	7.40M / ~$61B

Build M hit both adult age bands within ~2%; Build N loses 1.44M aged recipients (−61% vs SSA) and overshoots 18–64 by ~9%. Lower aged take-up than disabled take-up is directionally real, but a seed that reconciles to 0.94M against an observed 2.42M caseload is baseline-refuted. (Implied simulated aged eligibles ≈ 0.94M/8.4% ≈ 11M — if the banding solve broadened aged eligibility and then crushed the seed to compensate, the #431-style cap/trajectory may have bound; the reconcile machinery from #448 should show which.)

Caution: reform deltas now look anchor-adjacent for the wrong reason

$10k/$20k on Build N: +$2.07B / +294k new (vs OACT $1.0B / CBPP 224k — closest PE has landed), age mix 256k adults / 38k elderly / 0 children vs CBPP's 154k/47k/23k — also the most CBPP-shaped ever. Removal: +$5.56B / +738k (vs +$15.4B on Build M). These improvements are substantially an artifact of the suppressed 65+ take-up (fewer aged eligibles counted as claimants), not of the #374 asset blend (asset columns are byte-identical in distribution to Build M: 64.3% zeros, mean $123k). Recommend not reading Build N as meeting #374's acceptance.

Suggested check

Re-band the 65+ seed against the observed aged caseload (2.42M) as a hard reconcile target alongside the national total, and gate baseline SSI by age band (the harness in ssi-asset-limits .local-results/populace-rerun-scripts/ produces the by-age table in ~5 min).

Relates #430 (banding), #431 (swap cap), #448 (reconcile trajectory), #374 (acceptance), #453 (under-18 row unchanged: no child disability signal). Runs executed by the sol lane; numbers independently verified against the artifacts.

[MaxGhenis]

Picking this up — implementation PR in flight from the SSI analysis loop (sol-built, harness-validated). Will link the PR here.

[MaxGhenis]

Root cause found and documented in #508 (draft): the band thresholds are computed on pre-calibration weights and frozen — Build N's 65+ threshold (5.9%) was set against 40.3M pre-calibration candidates, the refit left 4.0M, and #477's one-shot cutover had removed the reconcile loop that would have caught it (band misses demoted to scorecard-only). Production fix = bounded post-selection reassign→gate→replay→refit loop, blocked on a target-contract adjudication (SSA Table 1 exact cells sum 7,289,843 vs the 7,404,820 registry pin) and on #453 for the child band to be feasible (under-18 candidate capacity is 177,582 vs the ~1.0M target). Details in the PR.

[MaxGhenis]

The #524 gate works; the recompute-once basis diverges — two-attempt production evidence. Build O attempts 2 and 3 (base-O, v9.4, main 15e5b7a):

attempt	basis	18–64 delivered vs 3,905,779	65+ delivered vs 2,382,142	rc
2	current-frame (pre-calib)	4,413,343 (+13.0%)	(also missed; artifact on disk)	1 — gate FAILED the build ✅
3	attempt 2's delivered weights	5,030,781 (+28.8%)	2,547,971 (≈+7.0%)	1 — gate FAILED again ✅

The gate itself is vindicated — both attempts would have shipped silently under pre-#524 code; the 65+ collapse class is basically cured (0.94M → 2.55M, within 7%). But the delivered-weight recompute amplified the 18–64 miss instead of damping it: the flags↔weights coupling has gain > 1 in this direction (hypothesis: the solve upweights SSI-flagged 18–64 persons to satisfy the $63.1B SSI payment dollar target — fewer/lower-benefit flagged people than reality ⇒ upweighting ⇒ count overshoot ⇒ a lower recomputed threshold reduces flagged mass further ⇒ stronger upweighting; a one-step recompute cannot find this fixed point).

Design iteration needed (NOT another basis rerun — oscillation risk is the #463 class):

1. Stabilizer candidate A: make the three band-count rows BINDING calibration targets (role ssa_ssi_age_band_recipients exists) so the solve itself pins delivered counts, with the delivery gate as verification — the count and dollar targets then negotiate inside one objective instead of across attempts.
2. Candidate B: external fixed basis — thresholds from the previous CERTIFIED release's delivered weights (stationary reference, no within-cycle feedback).
3. Either way: the divergence receipts above are the test fixture.

O attempt 2/3 artifacts (us_ssi_take_up.json only — the gate fires before diagnostics write, which also deserves a fix: forensics want the full diagnostics on gate failure) under _buildo-runtime/out/buildo-run/sparse/releases/. O-2 certification is blocked on this design decision.

[MaxGhenis]

Claiming the design-iteration lane (branch ssi-507-takeup-stabilizer, fresh off origin/main 15e5b7a). I analyzed the attempt 2/3 fixtures plus O-1's certified calibration_diagnostics.json before choosing — and the evidence revises the gain>1 hypothesis. Three mechanisms, each now code/artifact-grounded:

1. The divergence is a target-frame checkpoint identity omission, not in-solve feedback

_target_frame_checkpoint_identity keys on base-dataset sha, pe-us version, seed, period, registry version, CD crosswalk, and selection-mass protections — not on the SSI take-up assignment. Attempts 2 and 3 shared --checkpoint-root …/buildo-sparse2 (per buildo_sparse.sh): attempt 2 missed and wrote target_frame_checkpoint.h5 at 20:16Z; attempt 3 (started 20:26Z) drew fresh flags from the delivered basis, then hit attempt 2's checkpoint — every identity field matched — and re-solved attempt 2's exact stale matrix. Deterministic solve, same start, same seed ⇒ bit-identical release weights: every per-band candidate_capacity / reporter_candidate_floor and the 340,064,716.0088204 weighted universe agree to the last float digit across the two artifacts (verified programmatically), and attempt 3's recorded basis sha 8b8eb6b5… is exactly attempt 2's artifact sha. Attempt 3's 21-minute wall time (vs 39 for attempt 2) is the skipped materialization.

Net effect: the checkpoint hit forks the flag lineage — the solve optimizes columns built from the previous attempt's flags while the export and the delivery measurement carry the new flags. The #524 retry recipe never reached the solve; no basis choice can converge under that fork. (The forensics fix below would have made this a one-grep diagnosis: the in-memory compilation dict records target_frame_checkpoint.status: "hit", but the gate raised before it wrote.)

2. Candidate A's premise needs revising: the band rows already bind

O-1's certified calibration_diagnostics.json shows the three by-age rows in the solve under role ssa_ssi_age_band_recipients, materialized as indicator sums of engine ssi (which is takes_up_ssi_if_eligible-gated — verified in pe-us source), and the solve moved them hard:

band	init	final	target	final vs target
under_18	26,515	122,352	1,001,922	−87.8%
18–64	1,303,578	4,481,852	3,905,779	+14.7%
65+	236,396	971,690	2,382,142	−59.2%

Solve final_estimate equals the take-up artifact's selected_recipient_weight to 1e-6 on every band — the solve saw the shipped flags and chose this optimum while the $63.1B dollar row sat at −21.9% (init $14.2B → final $49.2B). The dollar-chasing part of the hypothesis is real: with dollar-starved flags the solve overshot 18–64 and couldn't reach 65+ (only ~6% of aged non-reporters were flagged; it upweighted them 4.1× and ran out of room). The flags cap the reachable set — binding rows can't fix a bad draw, and they're already binding. Candidate A as stated is a no-op.

3. _band_prior ignores the reporter floor — a deterministic overshoot the fixtures price exactly

The prior is target/capacity, but the law delivers floor + p×(capacity−floor) (reporters unconditically true, non-reporters drawn at p) — expected overshoot floor×(1−p) whenever the band subsamples. On attempt 3's own basis numbers:

• 65+: predicted +7.87%, realized +6.96% — the entire gate miss. Floor-aware p=(target−floor)/(cap−floor)=0.5454 prices delivery at target ±0.00%.
• 18–64: predicted +22.11%, realized +28.80% — the residual (+261k; +453k at attempt 2's p=0.411) is the fixed-seed hash-draw realization gap on ~770 lumpy sparse-support identities (monotone-shrinking in p; measured both attempts, same sign).

So "+13% → +28.8%" was never a diverging fixed-point iteration: attempt 3's map actually converged — to the wrong point (floor arithmetic) measured against a stale solve (checkpoint fork).

The stabilizer (smallest sound design, in flight)

1. Checkpoint identity carries the take-up assignment: the per-band assignment priors + prior-basis provenance (kind/sha/schema) join _target_frame_checkpoint_identity. A retry with a new basis now misses and re-materializes+re-solves against its own flags; same-assignment reruns stay warm. One-time invalidation of existing checkpoints is correct — they're all pre-fix. This heals every flag-downstream materialized column (ssi dollars, band counts, SSI-sensitive health/SNAP columns) in one key.
2. Floor-aware unsaturated prior: p = (target−floor)/(capacity−floor), total by construction (capacity ≤ target → existing reporter-share saturation fallback unchanged; floor ≥ target → 0 with anchor_excess reporting). Diagnostics schema 3→4 (assignment_prior semantics change); basis loader accepts 2/3/4 so the O-1 and attempt-2/3 artifacts all remain valid seeds.
3. Receipts before the raise (the forensics gap): the final-measurement SSI gates move after _write_release_calibration_diagnostics, AST-order-pinned like the #437 dropped the release H5 write; smoke gate scored a stale artifact (attempts 13/14) #443 export-ordering test, so gate-failed builds leave calibration_diagnostics.json + us_ssi_take_up.json — the Table-1.4 fail-loud pattern.

Not doing: candidate B as primary (an external certified basis is strictly staler than the failed attempt's own delivered basis once the fork is fixed; it stays what it already is — the documented seed for a new lineage). No realization-aware feedback on measured delivery (a p×target/delivered map is the #463-class door). No loss knobs (#492).

Prediction for the owner rerun (fresh --checkpoint-root or this fix; basis = attempt 3's artifact — its capacities/floors are real solved-weight measurements): 65+ delivery prices at target exactly; 18–64 draw lands ≈ target + realization gap and the fresh solve — now seeing a dollar-rich flag set — negotiates it against the $63.1B row with far less tension than O-1's dollar-starved optimum. If it still misses ±5%, the now-written receipts adjudicate draw-lumpiness vs solve-negotiation as the next iteration, with evidence.

TDD with the attempt 2/3 receipts as fixtures; sol cross-family review on the committed diff; owner session gates the merge and reruns.