# RESULTS v5e (AMENDMENTS_9160 A7) beside v5d (A4) and v5c with A3 applied

Written by `07_analysis_code/9998_v5e_results.py` from the 9505 numbers files (`qa/claude_release_v4/v5e_numbers.json`, `qa/claude_release_v4/v5d_numbers.json`, `qa/claude_summative_20261008/work_v5c_a3/v5c_numbers.json`). Full method, disclosures and sensitivities of the shared estimator: `RESULTS_v5c.md`.

**Headline rule** (A4, A7 and the author's choice of 14 October of which version is the result; nothing is published): v5e if it passes its gates, otherwise v5d if it passes, otherwise v5c with A3 applied. **Headline version: v5e (A7).**

| version | gates pass | per-child RR | random-effects 95% CI | shared-bias allowance (sensitivity) | placebo-calibrated (sensitivity) | K_eff | countries (weight > 1%) | stacks | G1 planted x1.25 / x1.5 (scaled) | G2 placebo | 
|---|---|---|---|---|---|---|---|---|---|---|
| v5e (A7) | yes | 0.973 | 0.836-1.131 | 0.699-1.354 | 0.855-1.107 | 26.04 | 40 (28) | 39,753 | -8.3% / -6.7% | 6/6 |
| v5d (A4) | yes | 0.976 | 0.838-1.135 | 0.701-1.358 | 0.857-1.111 | 25.96 | 39 (28) | 39,558 | -8.3% / -6.8% | 6/6 |
| v5c with A3 applied | yes | 0.973 | 0.834-1.136 | 0.698-1.357 | 0.853-1.111 | 24.90 | 37 (27) | 39,548 | -8.4% / -6.9% | 6/6 |

v5c as declared (A2): 0.972; version 4 (RULE_9159): 0.965.

## Tipping point on shared bias (headline version)

A shared (all-country) bias component with SD 0.102 would be needed before the upper end of the interval reached a 25% increase (0.207 for a 50% increase; 0.063 before the lower end reached a 20% decrease). The placebo upper bound on a shared component is 0.080; the ramp model estimates 0.185 (profile interval 0.085-0.410), which would widen the interval to 0.657-1.441.

## What v5e adds (factory v5e input set)

* 107 new parts admitted with a full test pass (A7 cutoff 2026-10-15 02:00 CDT); errors 0.
* Primary outcome cells 12,134,984 (52,126 new): USA_WI_DPI_GRADE 22,789, USA_AL_ALSDE_DISTRICT 13,851, USA_IDEA_PARTB_STATE_NC 8,060, USA_TX_TEA_DISTRICT 3,681, USA_PA_PENNDATA_LEA 1,910, USA_MA_DESE_DISTRICT 796, HKG_EDB_SEN 384, GBR_ENG_HES_CHILD_CANCER 368, USA_FL_FDOE_ESE_DISTRICT 146, USA_CO_CDE_AU 69, GBR_ENG_SEN_ASD_NATIONAL 33, BEL_FL_TYPE9_PROV 30, USA_OR_ODE_EIECSE 9.
* Roles (A7): 43 new outcome sources enter as sensitivity (max_incl only), 40 are descriptive; 19 _EXT parts were examined for automatic continuation of a held primary source (exact overlap agreement: 0 disagreements).
* Exposure: 534 new part rows and 12,934 sub-state rows; 34 new sources meet the A2 single-cross-section definition; only the ranked ones (not descriptive / sensitivity) are path_eligible False contrasts in the design, the rest never enter the exposure ranking.

## Maximal-inclusion supplement (A6, headline version)

Every held autism series that can be placed on birth cohorts, with the A2 restrictions dropped (minimum span, cross-section as contrast only, strict level choice) and the sensitivity-role parts entering with their own source fixed effect: per-child RR 0.961 (random-effects 95% CI 0.835-1.107; allowance 0.694-1.332; calibrated 0.856-1.079), 44 countries, K_eff 28.8, 48,939 stacks, 59 sensitivity parts entering. Gates: G1 -9.6% / -7.9%, G2 6/6, ramp validated 1. Supplement only; no headline role.

## Robustness battery for the per-child label (A5, v5e)

Full report: `qa/claude_summative_20261008/work_v5c_v5e/robustness_A5/ROBUSTNESS_A5.md` (forest chart beside it).

**The per-child label is NOT FULLY SUPPORTED: 6 of 7 criteria pass; failing: M1.**

| criterion | verdict | numbers |
|---|---|---|
| P1 measured-p and modeled-p stacks: allowance intervals overlap the headline | **PASS** | measured 0.934 (0.629 to 1.387); modeled 0.967 (0.684 to 1.366); headline allowance 0.699 to 1.354 |
| P2 excluding stacks on modeled pre-programme zeros (a and b): allowance intervals overlap | **PASS** | P2a 0.894 (0.613 to 1.303); P2b 0.851 (0.514 to 1.409) |
| P3 excluding national-broadcast-only countries: allowance interval overlaps | **PASS** | 0.993 (0.698 to 1.412) (18 countries) |
| P4 large-jump stacks (dp >= 0.30): allowance interval overlaps | **PASS** | 0.984 (0.706 to 1.371) |
| M1 intercept CI includes 0 (REML random country intercept) | **FAIL** | intercept +0.1802 (95% CI +0.0961 to +0.2644); slope -0.2032; implied r 0.797; country-cluster robust CI +0.0678 to +0.2927; nested country+event RE CI +0.0228 to +0.1744; random-slope CI -0.0153 to +0.2057 |
| M2 tertile estimates agree (Cochran Q p > 0.05) | **PASS** | T1 0.895, T2 0.910, T3 0.982; Q 0.40 on 2 df, p 0.819 |
| O3 no leave-out moves the RR outside the headline allowance interval | **PASS** | dropped BRA, USA, GBR, PER, KOR; RR range 0.970 to 0.977; largest move: drop USA -> 0.977 |

## Gate detail

* **v5e (A7):** G1 raw x1.25 / x1.5 -10.6% / -14.5%; placebo max |z| 1.20; country placebo coverage 0.80 of 187; attenuation (recovery) 0.925.
* **v5d (A4):** G1 raw x1.25 / x1.5 -17.2% / -18.6%; placebo max |z| 1.18; country placebo coverage 0.80 of 184; attenuation (recovery) 0.924.
* **v5c with A3 applied:** G1 raw x1.25 / x1.5 -14.5% / -16.0%; placebo max |z| 1.18; country placebo coverage 0.79 of 177; attenuation (recovery) 0.924.