QRF amount forests lose donor design weights after the n-of-n bootstrap: extreme-tail draws are a seed lottery

[MaxGhenis]

Summary

The QRF's design-weight handling for amount forests is a single n-of-n weighted bootstrap of the positive donors, after which the forest fits unweighted and within-leaf quantiles are occurrence-weighted over bootstrap multiplicities (qrf.py:187-216, :493-513; the module header documents the design). For heavy-tailed sparse dollar targets this amplifies the extreme tail by construction:

• a tiny-weight extreme donor's inclusion is a coin flip (the non_sch_d_capital_gains $594,483.64 row: 0.00755% of positive donor weight, expected bootstrap copies 0.47, present-at-least-once probability 37.5%);
• once included it is floored at 1/n of the training array (2.13× its weight share for that row), and empirically drawn at 7.11× its weight share on Build M (4 of 7,452 positive draws);
• the larger donors ($5.96M actual max, weight 1.1) were all omitted at seed 0 — so the shipped tail is a bootstrap lottery, not a bounded estimate, and a different seed could ship a far worse ceiling.

Deterministic replay evidence (seed-0 production target sequence reproduced the exact bootstrap composition and the $594,483.6449131259 ceiling) is in the #462 fix-3b diagnostic record.

Ask

Weight-aware within-leaf draws (or an equivalent estimator change that keeps original design weights available at draw time), so a donor's draw probability tracks its weight share instead of its bootstrap multiplicity. This is the root statistical fix behind #462's interim per-target tail bound and should come with cross-validation evidence on the affected stages.

Blast radius (from the diagnostic)

QRF is the public alias — every consumer redraws: ~25 US runtime constructors (puf_tax_detail chain, SCF, SIPP, ORG, housing, retirement, disability, voluntary filing, …) plus two UK SPI stages. Behavioral tests to adapt are enumerated in the diagnostic (weighted-fit contract, tail-share/interpolation, RNG determinism, chain bit-equality — behavioral constraints, not goldens).

Relates #462 (interim bound), #457 (same failure family in the frozen eCPS), #464 (tail-concentration gate = the backstop that detects this class).

[MaxGhenis]

Confirmed on the certified N default, in a worse form than filed: the bootstrap lottery also runs at release time, on a 5,000-donor sample, outside every tail guard

Forensics on populace-us-2024-buildn-sparse-rmloss100-c3e378a-20260722T010408Z (receipts on #462 and #451) adjudicated the two open support defects to owners. Verdict: this issue owns the dominant mechanism of the six-state medical blowout and a contributing mechanism of the tips carrier deficit. #487 is cleared for both (measured ASEC dollars and a SIPP concept gap — no uprating in either chain).

Medical (the #462 six-state family): a release-side QRF redraw broadcast one extreme measured donor ~40×

with_us_other_health_insurance_inputs — the last frame mutation before calibration — redraws health_insurance_premiums_without_medicare_part_b for the entire puf_tax_detail channel (100,320 persons) from a ≤5,000-person ASEC donor sample using populace.fit.QRF, i.e. exactly the estimator this issue describes, operating at its most fragile ratio yet: ~20 recipients per donor row, sparse positives, heavy tail.

• The ASEC donor pool contains a measured $680,000 individual premium (one MA record; 2 occurrences in the 865k-person base pool including its clone copy). The release redraw shipped 40 persons at exactly $680,000 (39 puf-channel) plus 80 at exactly $96,000 — a ~100× frequency enrichment of the extreme donor, the n-of-n bootstrap inclusion-then-overdraw pattern this issue documents on non_sch_d.
• Same-person receipts (base → published): $12,000→$680,000, $0→$680,000 (the AR couple), $1,900→$680,000 (DC, age 28), $1,200→$680,000 (NY, age 30), $2,600→$96,000 (NV), $6,000→$96,000 (UT). ASEC-channel persons 100% unchanged, per the stage contract.
• Downstream: single households at 3.9× an entire state's SOI medical target (AR), top-1 carrier shares of 63–85% of state medical mass, all six rows infeasible at design weights (+133% to +373% pre-solve), zero gradient past the loss cap. Revert counterfactual: the redraw explains ~100% of the AR/DC/NV/NY excess and injected ~$47B weighted deduction nationally (48.1% of puf-channel persons changed; unweighted channel premium mass +49%).

Tips: the same machinery halves an already-deficient donor concept

sipp_tips (also release-side, post-selection, populace.fit.QRF, ≤10,000 training rows of which ~57 are positive — with the in-code note that seed 0 "materially under-samples positive tip rows") transfers the SIPP donor's 2.20M weighted positive persons onto the frame at 1.14M design-weighted — the estimator loses another ~2× on the participation margin before calibration starts. Root cause of the −50.27% count is the donor's December-snapshot participation concept (receipts on #451), but the transfer loss is this issue's mechanism on the participation margin: positive-donor inclusion is the same coin flip whether the tail is dollars or participation. 95.6% of shipped tipped persons carry byte-exact donor values (44 distinct values across 549 persons).

What this adds to the ask

1. Weight-aware draws must cover the release-side consumers too — other_health_insurance_premiums and sipp_tips run after frozen-support selection inside build_us_fiscal_refresh_release.py; a base-only fix misses the worst instance.
2. The interim tail-bound machinery has a coverage hole: base-n puf_tax_detail_tail_bounds binds only non_sch_d_capital_gains, and the shipped qrf_tail_concentration.json surface contains none of health_insurance_premiums_without_medicare_part_b, other_medical_expenses, over_the_counter_health_expenses, tip_income in any list. Until weight-aware draws land, the interim per-target bound and the Blocking Table 1.4 dollar-row fit + QRF tail-concentration gates; export-guard both CGD route legs (#462) #464 gate surface should extend to every populace.fit.QRF output column, wherever the stage runs.
3. Sample caps compound the lottery: 5,000 of 66,001 ASEC persons (premiums) and 10,000 of 39,350 (tips) shrink the positive-donor pool before the bootstrap ever runs; the cap is part of the estimator's effective weight distortion.

All measurements read-only from the release evidence dir, the run's target_frame_checkpoint.h5 (compiled columns verified byte-exact against the diagnostics finals), and base-n (3573cdc…, confirmed the run's input base).