Repository navigation
Build N epic: CD target surface + base rebuild bundle (quartet, assets, QRF fix, feed regen, gate hardening) #449
Description
Max greenlit (2026-07-17) the congressional-district target surface as the next build after the Build M certification: "yes do it."
What
Turn on --include-congressional-district-targets for the next certified US build (Build N), expanding the calibration surface from the national+state class (~5,667 compiled targets in Build M) to the full geographic feed surface: the v8 Ledger feed carries 23,968 congressional-district facts alongside 11,191 state and 2,101 national — a ~30k-target-class surface.
Mechanics already in place
- The builder flag exists and is opt-in-off; the compile path translates CD facts via
--congressional-district-vintage-crosswalk(source vintage → current districts) and hard-requireshousehold.congressional_district_geoidon the support frame — which the Build M frame already carries (export parity tracks it at +2.3% drift). gate_congressional_district_targetscurrently records False in build manifests; Build N flips it to a gated surface.
Work items
- Crosswalk provenance: pin the source-to-119th crosswalk artifact (sha-pinned input, Build the 117th->119th CD vintage crosswalk from Census sources #288 alignment); the checkpoint identity and materialization-cache context already carry its sha.
- Preflight sizing (no solve): compile the CD-included registry, run the Add release-gate preflight: catch statically-knowable gate failures without a solve #436 preflight zero-support/support-expressibility pass over the 57k sparse selection — CD cells are exactly where a frozen sparse support goes thin (the Build G Build E: certification-candidate staging build from post-#287 main (fix target_compilation OOM first) #299 exclusion class). Expect a batch of per-cell reviewed exclusions or a reselection decision; adjudicate BEFORE burning a run (the campaign's attempt-9/10/11 lesson, now tooled).
- Solve-cost calibration: ~5.4× more constraint rows; measure a bounded-epoch dense probe before committing the full ladder. Loss-weighting review for the CD family (11k state rows already dominate counts; 24k CD rows will need family-level weighting the same way state rows got it).
- Gate adjudication pass: within-10% and zero-support baselines will move on a 30k surface; set expectations from the probe, not Build M's 90.2%.
- Scorecard: report Build N vs Build M on the shared national+state subset AND the CD surface separately, so the flip decision compares like with like.
Sequencing
After the Build M cert (policyengine.py#475) lands. Interacts with the geography-ladder/ACS lane (#288, Build L plan): the CD vintage story should be settled once and shared between calibration targets and the geography assignment.
Activity
MaxGhenis commentedon Jul 17, 2026
Max (2026-07-17): time the CD surface's cost, and "if not too much we'll remove the flag and make it default behavior." So #449's deliverable includes a decision-grade timing comparison, and the flag's fate is explicit:
Timing protocol (decision evidence)
Run the same commit, same base, same machine, no other compute, and record wall-clock per stage from the builder's own stage telemetry plus
/usr/bin/time -lpeak RSS:Also record: peak RSS per stage (jetsam headroom on 128 GB is a live constraint), and loss/within-10% on the shared national+state subset so quality regression is visible next to the cost.
Decision rule (Max's)
If the CD-on delta is acceptable ("not too much"), remove
--include-congressional-district-targetsentirely and make CD targets default behavior — no permanent flag, no config fork; the surface becomes the surface (consistent with the no-codified-tiers stance). The vintage crosswalk then becomes a required pinned input of every US build rather than an optional pairing.If the delta is material, keep the flag and record the measured cost here for the next revisit.
[-]Build N: turn on the congressional-district target surface (~30k targets; 23,968 CD facts in feed)[/-][+]Build N epic: CD target surface + base rebuild bundle (quartet, assets, QRF fix, feed regen, gate hardening)[/+]MaxGhenis commentedon Jul 17, 2026
Scope ratified by Max (2026-07-17). Build N's theme: one base rebuild + one feed regeneration + the CD surface + gate hardening. Build O's theme: the ACS multispine merge.
In (Build N)
Conditional: #424 (SSI candidate undercount) jumps in if the dense #447 trajectory adjudication points at the candidate universe.
Out (Build O)
Multispine program (#390, #395, #403, #346) and modeling-layer items: #372, #341, #321, #327, #417, #320, #360, #416, #412.
Parallel tracks (not build-gated)
Architecture/executor: #378–#382, #326, #388. UK/BE/NZ lanes run independently.
Hygiene
Verify-and-close sweep of superseded issues (#377, #386, #359, #356, #434; #367 stays open — unexercised, not fixed) running now with per-issue evidence.
MaxGhenis commentedon Jul 18, 2026
Scope addition from the SSI asset-limit analysis loop (cross-checked on the certified Build M):
is_disabled ≡ 0for ages 0–14 (CPS battery starts at 15); SIPP pu2023 carries verified-populated parent-reported child disability items (ECOGNIT 100% answered / 9.2% yes for 5–14, plus hearing/seeing/ambulatory/self-care) and RSSI_YRYN as a take-up anchor, with 0–4 banded to the SSA age curve.~/PolicyEngine/ssi-asset-limits/.local-results/populace-rerun-scripts/(pass_c_scenario.py+pass_f_combine.py, ~5 min/scenario on any 57k artifact;pass_h_*= the 40k-chunked variant for dense frames). Anchors: baseline SSI by age vs SSA 3.99M/2.42M/0.98M; $10k/$20k vs OACT $1.0B / CBPP 224k (23k children / 154k adults / ~47k elderly).MaxGhenis commentedon Jul 18, 2026
Scope addition from the IARIW OBBBA paper loop: #454 (ESI employer premium contributions) joins the Build N bundle, same workstream as the #451 quartet and the #340/#361 residuals — it is that exact class at ~100× the mass: dense f0af251 imputes ~$1.13T (right scale vs KFF), sparse exports drop the column, nothing anchors it, and the frozen eCPS computed $0 (documented as a levels-omission in the IARIW paper's CBO-style resource measure). Fix ladder is the standard motion: input-coverage requirement across release classes, Ledger fact for aggregate employer contributions (NHEA employer share or KFF EHBS × covered lives), calibration target + release gate against the anchor. The per-column disposition pass over the #361 94-column inventory should treat ESI as its highest-mass member.
MaxGhenis commentedon Jul 19, 2026
CD crosswalk revival done: PR #288 rebased onto main tip (86871cf, post-#455), conflicts resolved (builder import list; CD test now asserts the packaged-default semantics), 14 crosswalk + 116 builder tests green locally, CI green on 3.13/3.14. Marked ready for review — this unblocks the CD target surface (#314) and the timed CD A/B.
MaxGhenis commentedon Jul 20, 2026
Scope addition: #463 (stage the release builder — the #406 pattern applied to
build_us_fiscal_refresh_release.py) joins as the dense-lane front item, replacing the closed #456. Tonight's full-scale dense measurements are in the issue: #461 fixed the leak class (RSS returns between families), and the remaining legitimate working set crests at ~429GB / ~536GB (eligibility high-water, SSI reconcile passes) — past physical RAM it swap-crawls (6h of zero log writes at ~1 effective core), which is a throughput wall, not a correctness one. Acceptance: dense full-pool on n2-highmem-32 with the #406-style bit-equivalence proof. The #406 checkpoint machinery is directly reusable; prior artcodex/jct-materialization-memory-20260617.Blocking Table 1.4 dollar-row fit + QRF tail-concentration gates; exp…
MaxGhenis commentedon Jul 21, 2026
Sequencing consequence of #464 (merged 9243675): the CGD route legs are now hard-required export columns with no reviewed exclusion — red on any build from the current base by design. That folds #462 fix (3) (the
puf_tax_detailQRF zero-inflation/donor-support guard + a base re-run through the staged #406 pipeline, frozen selection reused) into the N-lite scope: N-lite cannot certify without it. Revised N-lite contents: quartet release-side (tips target + protection, probes — #465 merged; v9 feed splice pending), CGD fix (3) + base re-run, CD target surface behind its flag for the timed A/B. The two Table 1.4 returns-count misses #464 observed on M (+495.9% / +328.1%, same #457 family) join the N gate-adjudication pass.MaxGhenis commentedon Jul 21, 2026
Timed A/B verdict: CD-on fails on quality, not wall-clock — per the decision rule, the flag stays
First CD-on run at full scale (Build N sparse attempt 2, commit fdbcf07, base 3573cdcd…, v9.2 feed, frozen rmloss100 selection, same machine as the M baseline, no contention):
Wall-clock and memory are non-issues (peak RSS fine, run faster than M's warm baseline thanks to the deleted reconcile loop). Quality is the cost: at naive equal footing, 27k CD rows drown the ~5.7k national/state rows — within-10% halves and a critical national aggregate misses by 3× its tolerance. The concept-budget grouping (
_fiscal_target_concept_budget_weights) as shipped does not protect the national surface at this scale.Decision per Max's pre-agreed rule ("if the CD-on delta is acceptable, remove the flag; if material, keep the flag and record the cost"): the delta is material → the flag stays, and Build N ships CD-off. The CD surface needs a loss-architecture rework before any future on-arm — e.g., CD families sharing a fixed fraction of their parent national row's budget rather than competing row-by-row — which is its own design piece with this run's diagnostics (32,820-row
calibration_diagnostics.jsonretained atpopulace-us-2024-buildn-sparse-rmloss100-fdbcf07-20260721T194425Z) as the baseline evidence.Operational follow-through: #484's one-surface parity coupling gets flipped back to the CD-off regime (with a fail-closed regime constant so the next flip is deliberate), and the sparse launcher drops
--include-congressional-district-targets.Parity CD regime: explicit constant, fail-closed both ways, surface b…
MaxGhenis commentedon Jul 22, 2026
Campaign index cross-post: the UK dense/local arm is now scoped and claimed as #495 — rowwise constituency-grain dataset by cloning the national staging pool (the UK analog of this build's dense/CD arm; the FRS pool is already 10× cloned, so K=1 is at US-dense per-area scale). First two increments (local-target census tool; rowwise fence-carriage + dry-run) are in progress on that issue.
MaxGhenis commentedon Jul 22, 2026
UK dense/local arm update: both scoping increments of #495 are now merged — #498 (local-target census: drift-gated inventory of the local surface with six reviewed binding fences) and #501 (rowwise fence carriage, exact mass conservation, pool-lineage layer, and a dry-run whose realized support runs the real sampler at the build seed). Both went through sol cross-family review with all confirmed findings fixed pre-merge. Remaining epic increments (ladder S/NI, solve doctrine, K=1 pilot, calibrated candidate) are open for claiming on #495.
MaxGhenis commentedon Jul 22, 2026
Build O dense battery (evidence arm) —
populace-us-2024-buildo-dense-22bd902-20260722T175010Z, rc=1Chain: base-O built in 21.5 min at main
22bd902f(23 stages incl. #500'sadult_care_post_clone+ SE-health attribution; sha356cf501…; new columns verified: 14,462 persons incapable-of-self-care, 579 care-expense carriers with sane levels). Dense battery on feed v9.3 (eab45a2b…), 1h15m.What passed:
Terminal gate failures (same two classes as N dense, now with tighter evidence):
net_capital_gainsnational: −30.18% ($887.35B vs $1,270.86B) — byte-close to N dense's −30.20%, confirming the miss is structural to the 5× stretch bound on this base, not run noise. Owner: PUF donor uprating overstates mean-reverting distribution series: non_sch_d at 2.39x the SOI concept #487 (donor uprating / base CG mass). max_weight_ratio anchors differently per arm: 5x vs design (dense) but ~25x effective (sparse refit re-anchors) — declare one stretch contract #493's declared-stretch-contract question stands.First production past-cap census (#494, new in this run): 5,698 targets — init-past 484, final-past 459, escaped 31, frozen 453, pushed out 6 (two EITC-by-AGI cells 0.35→2.60 and 0.63→2.14; OK/NV small interest bins; SC real-estate taxes). Contrast with sparse N's hand census (929/377/569/17): the dense arm's clamped weights barely rescue dead rows (31 escapes vs 569) — a clean quantitative statement of what the 5× anchor trades away, and the dense baseline #492's experiment wanted.
Quality: within-10% 86.52% on 5,698 (N dense 86.59% — flat, as the bound predicts).
Disposition: this artifact is the intended first evidence-tier publish once #506's machinery lands (H5 + manifest carrying these exact failures with owners). Sparse-O holds for the #492 loss-shape verdict. Evidence on disk:
_buildo-runtime/out/buildo-run/dense/releases/…175010Z/.MaxGhenis commentedon Jul 23, 2026
Sparse-O attempt 1: rc=0 — GREEN ON THE FIRST ATTEMPT (Build N took six).
populace-us-2024-buildo-sparse-rmloss100-22bd902-20260722T232627Z, published immutable (latest.json deliberately untouched — still N; flip/cert = maintainer, and the recommendation is to hold for attempt 2, below).Scorecard vs certified N: within-10% 89.56% (N 89.49%), final loss 0.02813 (N 0.02824), 5,672 targets; CGD −0.03%, net CG −2.40%, SS benefits −0.26%, tips dollars −0.03%, medical +20.85% (inside the #490 0.25 bound); census 930/378/569/361/17 (same flat-cap equilibrium as N, as expected pre-#492 adoption). Coverage 163/163 including the #500 adult-care leaves — the quartet input stages are now on a gate-green sparse export.
Carried defects, both owned: the #507 SSI aged-band collapse (inherited by construction — this is why attempt 1 is NOT the certification candidate) and the tips return-count carrier deficit (−50.25%).
Why first-attempt green matters as system evidence: the six N attempts were spent discovering gate/register misalignments (#462 arc) — with #491's shared register, the measured tail register, the truthful surface, and the adjudicated bounds, the pipeline went straight through. The alignment work is what bought this.
Next: #508's SSI fix (universe adjudication posted — bands bind, national = Σ bands) → attempt 2 = the certification candidate. Equal-budget dense-3000 running now for the #493 bound-vs-convergence confound.
MaxGhenis commentedon Jul 23, 2026
Dense-3000 battery complete (
buildo-dense3000-22bd902-20260723T020906Z, rc=1, same two gate classes): the equal-budget arm for the #493 confound — netCG −29.59% (vs −30.18% at 1500 epochs; sparse reaches −2.40%), within-10% 87.10%, census 484/455/34/450/5. Verdict on #493: dense's CG gap is structural (optimization geometry, not budget). Build O battery day complete: base ✅, sparse-1 rc=0 published, dense-1500 ✅ evidence, dense-3000 ✅ evidence. Next certification motion = attempt 2 on the #508 SSI fix.MaxGhenis commentedon Jul 23, 2026
US local-artifact lane claimed → #512. The only local-area product in circulation (
populace-us-2024-buildl-acs-local-36de5d9a-20260712T104640Z) predates the entire M/N/O target-surface repair campaign — pre-truthful CG/interest chains, old SSI machinery, no adult-care/SE-health inputs — and pe.py'sdataset_overlays.us.populace_us_2024_acs_localstill pins it. Rebuilding it on the certified O lineage: base-O (sha356cf501…) donor + the O-1 surface vintage (feed v9.3), Build L calibration doctrine verbatim (no new knobs, #492), donor rows keeping base-O's certified block-ladder geography. Publish will implement #398's dataset-role contract rather than routing around the tracked path again; immutable tag, latest.json untouched; pe.py repin PR to follow. The #507 SSI aged defect rides this artifact by construction — declared in the release, with the committed one-command chain making the O-2 refresh mechanical once #508 certifies. Receipts on #512.MaxGhenis commentedon Jul 23, 2026
Feed v9.4 cut (owner):
b3c0835631a446eb…, 37,405 rows = v9.3 with thejct.tax_expenditures.cy2024family superseded 5→11 (ledger#118; values of the incumbent five byte-identical, six new #514 anchors added with concept caveats). Preflight vs base-O green-in-substance (standing 4-column advisory). Build O attempt 2 (post-#508) should launch on v9_4; the sparse launcher's FACTS pin updates accordingly. Pre-existing duplicate-srid wart filed as ledger#119.MaxGhenis commentedon Jul 24, 2026
#518 merged — Build O attempt 2 gains the truthful interest surface for free: the certified equilibrium already sits at $340.4B, so the ~470 past-cap interest rows flip near-truth at recompile without the solve moving. Attempt-2 launch config: base-O + feed v9_4 (
b3c08356…; ledger#117's Table 1.4 interest facts NOT in this cut — #518's lattice auto-prefers them whenever a future feed carries them, test-pinned) + the #508 fix when its lane delivers. Expect a material within-10% jump on attempt 2 from the interest family alone.MaxGhenis commentedon Jul 24, 2026
CD-on decomposition (owner, answering 'how did CD targets affect the non-CD surface') — attempt 2 (
fdbcf07, CD-on, 32,820 targets) split by the shared CD classifier:vs ~90% CD-off same era ⇒ the true national-surface cost was ~6–7pp, not the pooled 36pp — but the verdict-driver was critical-row breakage, which decomposition confirms: SS benefits −14.72% (hard fail of the 5% critical gate; CD-off runs: −0.25% to −2.9%), itemized total −11.24%, while income tax held at −1.09%.
Two caveats for any future re-test: (1) attempt 2 ran the PRE-repair surface (its own CGD row sits at +490%; the ~470 stale interest rows were then-impossible) — post-#486/#488/#518 a CD-on rerun should cost less; (2) per #492, the flat cap made small-area dumping free, which flattered CD-on's containment — a live-tail loss changes the dynamics both ways. The honest magnitude estimate therefore needs a post-repair A/B if national CD-on is ever wanted again — which the SLD/local lane may obviate (district credibility incl. CD in the local product, where small-area targets structurally cannot touch the national surface).
MaxGhenis commentedon Jul 24, 2026
Corollary to today's decomposition, per the owner's architecture ruling: #449's verdict re-reads as support starvation (27,148 CD rows on a 57,240-record support = ~2.1 hh/row; base has NO ACS spine — two channels only, asec+puf_tax_detail 168,852 each), not small-area-target toxicity. The retest that would actually answer the CD question: CD (and SLD) targets on, at ACS-spine record count (#512's expansion), post-#518 surface — one dataset, one target system, record count as the only variable.
MaxGhenis commentedon Jul 24, 2026
US local-area lane (#512): PUBLISHED.
populace-us-2024-buildo-acs-local-77e2061-20260724T110908Z— the ACS local artifact rebuilt on the certified O lineage (donor = the certified O-1 sparse release, sha-verified; base-O geography preserved). 1.589M households, 4,461-target state surface, within-10% 92.83% (buildl 87.87%), ESS 3× buildl, all gates green, consumer-loadable bytes, immutable tag withlatest.jsonuntouched, published through the tracked path under the #398 role contract. The #403 weight-composition skew is essentially resolved (persons/hh 2.72 vs 2.66; was 1.50 vs 2.93). Full receipts + known boundaries on #512; pe.py repin PR to follow (owner-gated).MaxGhenis commentedon Jul 24, 2026
The adequate-support CD retest (Max: 'lets try') — measured.
buildo-densecd-f70bb4b-20260724T115701Z: full 338k pool, v9.4, post-#518 surface, 3000 epochs, CD-on via a provenance-committed evidence branch (regime + feed pin + regenerated manifest, f70bb4b). 32,846 targets; rc=1 (3 critical fails).Verdict: support starvation was real but not the whole story. At 6× support the CD rows gain 17pp and the national cost shrinks from ~6-7pp to ~3.5pp — but criticals still break, and the census shows HOW: 414 in-cap rows dumped past the flat cap (vs 5-6 CD-off) — the #492 dumping dynamic is the mechanism by which the solve pays for CD rows. Notable inversion: CD-on FIXES medical (−1.7% vs +20.8%) — the district income-bin rows discipline exactly the weight mass the six-state problem lived in.
Composition hypothesis for the definitive A/B, after #492's sequence (donor tails → adoption): CD-on × rational tail — the rational arm independently killed the dumping (pushed-out→6) and fixed medical; combined with adequate support it is the plausible configuration where one dataset carries all geography levels. Until then the certified default stays CD-off. Evidence branch cd-retest-449-evidence (f70bb4b, not for merge); artifacts under _buildo-runtime/out/buildo-run/dense/releases/.
MaxGhenis commentedon Jul 25, 2026
CD-on × rational-tail composition run: verdict
Run:
populace-us-2024-buildo-densecdrat-61980d8-20260724T223254Z(full 338k pool, v9.4 feed, CD-on parity declaration,POPULACE_LOSS_SHAPE=rationalc=3, 3000 epochs, rc=1 on gates). Compared against the two flat-cap arms.Important caveat discovered during decomposition: this evidence branch forked before #518 (1c568c0) landed, so it still carries the 519 stale HT2 taxable-interest rows (register value $149.1B national vs the Pub 1304 actual $313.8B). That turns out to matter more than the composition question itself.
† symmetric exclusion of the 519
historic_table_2.*.taxable_interest_amountrows in all three runs (9 CD-side interest rows also excluded from the CD column; CD w10 moves ≤0.02pp).Findings
The apparent rational-tail recovery is mostly an artifact of the stale interest register. Naively, rational recovers +2.44pp of the 3.54pp non-CD composition cost (83.56 → 86.00 toward 87.10). Excluding the stale rows symmetrically, the recovery is +0.36pp (90.33 → 90.69, vs the 93.86 CD-off baseline). 120 additional stale rows moved inside 10% under rational — by hitting wrong targets.
The rational tail makes register truth load-bearing; the flat cap was accidentally shielding the frame from bad targets. National HT2 taxable interest (stale target $149.1B): both cap runs sit past-cap at truth (estimates $351.1B / $371.3B — error >100%, gradient dead, mass preserved). The rational run kept gradient alive and ground the frame to $158.7B — destroying ~$200B of real interest mass to satisfy a wrong target. The input-mass parity gate caught it: export $158.7B vs the Build G reference $320.2B (−50.4%, beyond ±50%, terminal fail). This is the empirical case for the Loss shape: flat cap at 1.0 makes past-cap rows free dumping grounds (17 rows pushed out in Build N); bounded-tail experiment + past-cap census diagnostic #492 adoption sequence: register re-verification (Rebase stale HT2 taxable-interest rows onto the Pub 1304 national actual (#489 adjudication) #518-class) strictly precedes any loss-shape change. With cap, a wrong register row costs you one abandoned target; with rational, it costs you the frame.
The CD composition cost on the clean surface is ~3.5pp regardless of tail shape (93.86 → 90.3–90.7), and the CD rows themselves are indifferent to it (65.1 vs 64.3). The loss tail is not the CD lever; support is where the remaining CD misses live — consistent with the calibration doctrine (misses are generally support).
Binding gates for a CD-on dense are unchanged: SS benefits and itemized. SS improves under rational (−8.52% → −7.11%) but still fails the 5% gate; itemized 1.2/2.1 sits at the 15% boundary in both CD-on runs. Medical stays healed under CD-on in both shapes (−1.7% / −2.2%). Net CG worsens under rational (−34.45% vs −28.23%, fails the SOI 1.4 25% gate) — plausibly capacity reallocating toward large-error rows the cap had abandoned, but that attribution is unverified.
Census: rational thins the boundary population. Final past-cap 1,940 → 953; frozen 1,526 → 561; pushed-out 414 → 392. Direction favorable, but partly confounded by the interest chase pulling those carriers off the boundary.
The six QRF donor-tail concentration failures (alimony_expense, alimony_income, bond_assets, estate_income, qualified_bdc_income, unrecaptured_section_1250_gain) fired identically on the flat-cap CD run — a dense-arm imputation property (the Build M live default: non_sch_d_capital_gains lands at $74.6B vs its $10.2B SOI target (+635%); Schedule-D CGD route is empty #462 signature), not a rational-tail effect. Already on the Loss shape: flat cap at 1.0 makes past-cap rows free dumping grounds (17 rows pushed out in Build N); bounded-tail experiment + past-cap census diagnostic #492 pre-adoption list.
Next
diagnose_us_target_support(c192b63) on SS benefits and itemized 1.2 under the CD-on frame to classify the two binding misses as support-starvation vs capacity competition.MaxGhenis commentedon Jul 25, 2026
CD targets on the SHIP arm, clean register: the composition answer
populace-us-2024-buildo-sparsecd-rmloss100-a8ac960-20260725T132040Z— sparse (the certified arm's configuration), CD surface ON, post-#518 register, post-#548 code, 3,000 epochs, 60 min, rc=1. This run exists because #548 landed: the identical configuration this morning failed the same SSI gate and left onlyus_ssi_take_up.json, destroying the surface it was built to measure. This one keptcalibration_diagnostics.json,input_coverage.json, andinput_mass_parity.jsonthrough a four-group gate failure.Composition cost
The headline non-CD comparison (89.56 → 89.20) is misleading and should not be quoted: O-1's population still contained 509 stale HT2 interest rows it was missing, which drag its own number down. On the comparable surface the cost of turning CD on is ≈5.5pp (95.95 → 90.40).
Criticals: SS benefits −14.1% (fails the 5% gate — the same break seen at −14.7% sparse and −8.5% dense under CD-on), itemized 1.2 −13.0% (inside the 15% bar), medical +2.8% (healed from O-1's +20.9%), net CG +1.6% (from −2.4%). The medical and CG improvements are confounded — #518's register rebase and the CD surface both changed here — so do not attribute them to CD.
The finding that matters: sparse cannot carry the CD surface
CD rows land at 47.4% within 10% here, against 65.1% on the full 338k-record dense pool with the same surface. The distribution is the story: median |error| 11.8%, p75 41.9%, p90 97.0%; only 36.7% inside 5%, and 21.4% of CD rows are off by more than half.
That is a support result, not a loss-architecture result. Same targets, same loss, same code — the only variable is record count, and the sparse frame does not have enough records per district to carry 27,148 small-area rows. It is consistent with the one-dataset architecture rather than an argument against it: CD rows bind where support suffices, and 57k records is not where it suffices. The path to CD-on shipping is the record count (full pool, and ultimately the ACS spine), not a CD-specific loss treatment.
Also visible: the pushed-out census goes 17 → 737 under CD-on, the flat cap paying for CD rows exactly as the dense arms showed.
Register recompile, quantified on the ship arm
Interest-amount rows within 10%: 25.0% → 44.2% (128/511 → 445/1,007). #518's rebase is delivering on the sparse arm, and the previously stale national HT2 row is gone from the miss list.
Residual found by the run
QRF tail concentration failed: exclusion register carries entries the checked surface did not use: ['partnership_self_employment_net_earnings']— the per-run tail register is stale under the CD-on surface. Fail-closed working as designed; the register needs a per-surface refresh before any CD-on ship candidate.