# Whole-pipeline power and calibration of the 15-context synthesis (3351)

29 September 2026. Documented rule `qa/portfolio_power_20260929/3351_rule.json` (sha256 `e6a9c99fef9d25295908843c882c2da91e493958170bf953dc49dc3ce5b5eeb1`), written before any simulation replicate; every cached replicate file and this report carry that hash and the report refuses to run if the file changes. Run time: 4.3 h wall-clock (simulation, checks and aggregation, at most 8 worker processes with 4 BLAS threads each, on a machine shared with other jobs) and 34.0 worker-hours of task time.

## What was simulated

Seventeen data-generating scenarios, each applied to every pooled context on its own cells and exposure path: no harm (R = 1 everywhere); a common true individual RR R in {1.05, 1.10, 1.25, 1.50} in every context (including the U.S. rotavirus-2006 path); and harm concentrated in one product (modern rotavirus, PCV or MMR) at the same R with every other context at 1.00. Each replicate is the block's noise-free null expectation plus the coverage-proportional mixture effect plus the block's own parametric noise (the harm-injection gate's bootstrap: residual serial and cohort dependence, observed nuisance trends). Every block's likelihood curve is refitted as the pipeline fits it, then the full primary hierarchy (15 contexts, 14 likelihoods, the calibrated bias SDs, U.S. rotavirus 2006 untied) and the common-effect model are recomputed. 400 pipeline replicates per scenario in two independent seed streams. Replicates per block and R: Gaussian blocks 400, Japan 400, U.S. joint likelihood 200, Scottish counts 100, Czech counts 50 (count refits cost 40-190 s each; smaller blocks are recycled across pipeline replicates and the Monte Carlo SEs are cluster-robust for it).

## Documented statements

**Detection of a common harm (>= 80% power).** The primary heterogeneous shared mean detects no common harm up to RR 1.50 with at least 80% power. Detection rates (P(shared mean > 1) > 0.975): RR 1.05: 0.0% (MC SE 0.0), RR 1.10: 0.0% (MC SE 0.0), RR 1.25: 0.0% (MC SE 0.0), RR 1.50: 0.8% (MC SE 0.4). The common-effect estimand detects no common harm up to RR 1.50 with at least 80% power: RR 1.05: 0.0% (MC SE 0.0), RR 1.10: 0.0% (MC SE 0.0), RR 1.25: 0.0% (MC SE 0.0), RR 1.50: 34.2% (MC SE 2.3).

**Exclusion when there is no harm.** Under R = 1 everywhere, the probability that the posterior 97.5th percentile falls below c is, for the shared mean: c = 1.05: 0.0% (MC SE 0.0), c = 1.10: 0.0% (MC SE 0.0), c = 1.25: 0.0% (MC SE 0.0), c = 1.50: 16.2% (MC SE 1.8); for the common effect: c = 1.05: 0.0% (MC SE 0.0), c = 1.10: 0.0% (MC SE 0.0), c = 1.25: 0.0% (MC SE 0.0), c = 1.50: 62.0% (MC SE 2.3). Reliably excluded (>= 80%) under no harm: shared mean none; common effect none.

**Calibration.** False-positive rate under no harm (nominal one-sided 2.5%): shared mean 0.0% (MC SE 0.0), common effect 0.0% (MC SE 0.0); product-level posteriors modern_rotavirus 0.0%, PCV 0.0%, MMR 0.0%. 95% interval coverage of the true common RR: shared mean R 1.00: 100.0%, R 1.05: 100.0%, R 1.10: 100.0%, R 1.25: 100.0%, R 1.50: 100.0%; common effect R 1.00: 100.0%, R 1.05: 100.0%, R 1.10: 100.0%, R 1.25: 100.0%, R 1.50: 100.0%.

**Shared mean versus common effect.** At a common RR of 1.25 the common effect detects 0% of replicates vs 0% for the shared mean, and at 1.50 34% vs 1%. Under no harm its median posterior 97.5th percentile is 1.471 vs 1.570, and it excludes RR 1.50 in 62% of null replicates vs 16%. The common-effect model (one theta, no product or context heterogeneity) is the more informative summary in both directions: it detects at least as often at every common RR and excludes at least as often under no harm. The shared mean additionally integrates the product- and context-heterogeneity components (posterior mean product SD 0.19 under no harm) and lets each product carry its own effect; both summaries are net of the shared-bias allowance (structural ceiling below).

**Concentrated harm.** Detection rates when only one product is harmful: harm confined to modern rotavirus - RR 1.25: shared mean 0%, common effect 0%, modern rotavirus product posterior 0% (its median posterior median 1.135); RR 1.50: shared mean 0%, common effect 0%, modern rotavirus product posterior 1% (its median posterior median 1.277). harm confined to PCV - RR 1.25: shared mean 0%, common effect 0%, PCV product posterior 0% (its median posterior median 1.050); RR 1.50: shared mean 0%, common effect 0%, PCV product posterior 0% (its median posterior median 1.097). harm confined to MMR - RR 1.25: shared mean 0%, common effect 0%, MMR product posterior 0% (its median posterior median 0.991); RR 1.50: shared mean 0%, common effect 0%, MMR product posterior 0% (its median posterior median 1.005). At RR 1.05 and 1.10 the largest detection rate of any estimand under concentrated harm is 0% (no-harm rates: shared mean 0%, common effect 0%, products modern_rotavirus 0%, PCV 0%, MMR 0%). The product posteriors are those the pipeline reports (global_hierarchy.posterior_detail): each product effect is centred on the hypermean INCLUDING the shared-bias term, so, unlike the shared mean and the common effect, they are not net of the N(0, .15^2) shared bias and the structural ceiling below does not bind them. Exclusion under concentrated harm (dilution): with RR 1.50 confined to one product, the posterior 97.5th percentile nevertheless falls below 1.50 - i.e. a global summary would "rule out" RR 1.50 - in modern rotavirus: shared mean 0%, common effect 0%, the harmed product's own posterior 0%; PCV: shared mean 2%, common effect 33%, the harmed product's own posterior 0%; MMR: shared mean 11%, common effect 56%, the harmed product's own posterior 2% of replicates. A global exclusion therefore does not rule out a harm concentrated in one product.

**Structural ceiling (analytic).** With shared bias N(0, .15^2) and prior N(0, log 2), theta | phi has SD 0.1466 even with infinitely precise data (phi = theta + shared bias): the 97.5th percentile can never fall below RR exp(1.96 x 0.1466) = 1.333 when phi = 0, and P(theta > 1) > .975 needs phi > 0.301 (RR 1.35). In the no-harm replicates the smallest posterior 97.5th percentile was 1.359 for the shared mean and 1.301 for the common effect (0.0% and 1.8% of replicates below 1.333, possible only when the data place the hypermean below 1).

**Where the observed result falls.** The observed primary shared mean 0.957 (0.599-1.517; P(> 1) = 0.425) sits at the 28th percentile of the no-harm distribution of posterior medians (null 95% range 0.892-1.098) and at the 0th / 0th percentiles under a common RR of 1.25 / 1.50; its 97.5th percentile is at the 25th percentile of the null upper bounds. The observed common effect 0.977 (0.659-1.446; P(> 1) = 0.455) sits at the 34th null percentile and the 0th / 0th percentiles under RR 1.25 / 1.50. (The observed-data grid computation of the primary is 0.956 (0.598-1.516).)

**Conditionality.** Conditional on the model: the specified exposure paths (point estimates), the bias allowances (.35/.70 and the shared .15), additive coverage-proportional mixture effects and the nuisance structure of each block. True bias is zero in every scenario, so these are operating characteristics under an unbiased world, not guarantees against real bias. Approximations: the U.S. joint likelihood is its grid version (point NIS exposure draw instead of the 512 bootstrap draws; interaction and receipt dependence integrated by quadrature), which reproduces the observed-data primary within 0.0012 (tolerance 0.005; the primary's two Sobol scrambles differ by 0.002); the U.S. series-variance plug-in is held at its observed value; count-block noise extends the gate's bootstrap (Scotland: lognormal extra-Poisson variance 0.0105; Czechia: none, its counts are under-dispersed around the nuisance fit, df-corrected dispersion 0.69, so Poisson noise only). True joint receipt is independent with no PCV x RotaShield interaction.

## Operating characteristics

Detection = P(target > 1) > 0.975. Exclusion of c = posterior 97.5th percentile < c. Rates in percent with cluster-robust Monte Carlo SEs in parentheses.

| Scenario | Estimand | Detection | Excl. 1.05 | Excl. 1.10 | Excl. 1.25 | Excl. 1.50 | Median of medians | Coverage |
|---|---|---:|---:|---:|---:|---:|---:|---:|
| no_harm | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 16.2 | 0.991 | 100.0 |
| no_harm | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 62.0 | 0.996 | 100.0 |
| no_harm | PCV product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 12.8 | 0.995 | 100.0 |
| no_harm | MMR product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 9.2 | 0.979 | 100.0 |
| no_harm | Modern rotavirus product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 63.5 | 0.997 | 100.0 |
| common_R1.05 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 4.5 | 1.035 | 100.0 |
| common_R1.05 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 29.5 | 1.042 | 100.0 |
| common_R1.05 | PCV product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 2.5 | 1.041 | 100.0 |
| common_R1.05 | MMR product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 2.2 | 1.022 | 100.0 |
| common_R1.05 | Modern rotavirus product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 34.5 | 1.045 | 100.0 |
| common_R1.10 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 1.0 | 1.078 | 100.0 |
| common_R1.10 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 8.5 | 1.088 | 100.0 |
| common_R1.10 | PCV product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.2 | 1.086 | 100.0 |
| common_R1.10 | MMR product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.8 | 1.069 | 100.0 |
| common_R1.10 | Modern rotavirus product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 9.8 | 1.092 | 100.0 |
| common_R1.25 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.0 | 1.202 | 100.0 |
| common_R1.25 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.0 | 1.226 | 100.0 |
| common_R1.25 | PCV product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.0 | 1.222 | 100.0 |
| common_R1.25 | MMR product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.0 | 1.202 | 100.0 |
| common_R1.25 | Modern rotavirus product (primary hierarchy) | 0.5 (0.3) | 0.0 | 0.0 | 0.0 | 0.0 | 1.236 | 100.0 |
| common_R1.50 | Shared mean (primary hierarchy) | 0.8 (0.4) | 0.0 | 0.0 | 0.0 | 0.0 | 1.404 | 100.0 |
| common_R1.50 | Common effect | 34.2 (2.3) | 0.0 | 0.0 | 0.0 | 0.0 | 1.442 | 100.0 |
| common_R1.50 | PCV product (primary hierarchy) | 0.5 (0.3) | 0.0 | 0.0 | 0.0 | 0.0 | 1.439 | 100.0 |
| common_R1.50 | MMR product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.0 | 1.417 | 100.0 |
| common_R1.50 | Modern rotavirus product (primary hierarchy) | 61.0 (1.9) | 0.0 | 0.0 | 0.0 | 0.0 | 1.469 | 100.0 |
| concentrated_modern_rotavirus_R1.05 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 10.5 | 1.008 |  |
| concentrated_modern_rotavirus_R1.05 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 45.0 | 1.022 |  |
| concentrated_modern_rotavirus_R1.05 | Modern rotavirus product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 44.8 | 1.025 | 100.0 |
| concentrated_modern_rotavirus_R1.10 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 6.5 | 1.024 |  |
| concentrated_modern_rotavirus_R1.10 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 26.0 | 1.047 |  |
| concentrated_modern_rotavirus_R1.10 | Modern rotavirus product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 24.8 | 1.053 | 100.0 |
| concentrated_modern_rotavirus_R1.25 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 1.5 | 1.070 |  |
| concentrated_modern_rotavirus_R1.25 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 2.5 | 1.121 |  |
| concentrated_modern_rotavirus_R1.25 | Modern rotavirus product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 1.0 | 1.135 | 100.0 |
| concentrated_modern_rotavirus_R1.50 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.0 | 1.135 |  |
| concentrated_modern_rotavirus_R1.50 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.0 | 1.233 |  |
| concentrated_modern_rotavirus_R1.50 | Modern rotavirus product (primary hierarchy) | 0.8 (0.4) | 0.0 | 0.0 | 0.0 | 0.0 | 1.277 | 100.0 |
| concentrated_PCV_R1.05 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 13.2 | 0.998 |  |
| concentrated_PCV_R1.05 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 58.5 | 1.001 |  |
| concentrated_PCV_R1.05 | PCV product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 8.0 | 1.006 | 100.0 |
| concentrated_PCV_R1.10 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 11.2 | 1.004 |  |
| concentrated_PCV_R1.10 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 55.5 | 1.007 |  |
| concentrated_PCV_R1.10 | PCV product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 3.5 | 1.018 | 100.0 |
| concentrated_PCV_R1.25 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 6.2 | 1.023 |  |
| concentrated_PCV_R1.25 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 46.2 | 1.020 |  |
| concentrated_PCV_R1.25 | PCV product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.2 | 1.050 | 100.0 |
| concentrated_PCV_R1.50 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 2.5 | 1.045 |  |
| concentrated_PCV_R1.50 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 32.8 | 1.040 |  |
| concentrated_PCV_R1.50 | PCV product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 0.0 | 1.097 | 100.0 |
| concentrated_MMR_R1.05 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 15.8 | 0.992 |  |
| concentrated_MMR_R1.05 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 61.0 | 0.997 |  |
| concentrated_MMR_R1.05 | MMR product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 8.5 | 0.980 | 100.0 |
| concentrated_MMR_R1.10 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 15.5 | 0.993 |  |
| concentrated_MMR_R1.10 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 60.2 | 0.998 |  |
| concentrated_MMR_R1.10 | MMR product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 6.8 | 0.984 | 100.0 |
| concentrated_MMR_R1.25 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 13.5 | 0.997 |  |
| concentrated_MMR_R1.25 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 59.2 | 1.000 |  |
| concentrated_MMR_R1.25 | MMR product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 4.2 | 0.991 | 100.0 |
| concentrated_MMR_R1.50 | Shared mean (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 11.0 | 1.005 |  |
| concentrated_MMR_R1.50 | Common effect | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 56.2 | 1.006 |  |
| concentrated_MMR_R1.50 | MMR product (primary hierarchy) | 0.0 (0.0) | 0.0 | 0.0 | 0.0 | 2.2 | 1.005 | 97.8 |

Concentrated-harm rows show the harmed product's own posterior; the full table (every product posterior in every scenario, coverage of 1 and of R, seed-split halves) is in `3351_portfolio_power_calibration.csv`.

## The observed result against the simulated distributions

Observed posterior (the primary IS result for the shared mean; the grid computation on the observed data for the common effect and the product posteriors) and its percentile among the simulated posteriors under no harm and under a common RR of 1.25 and 1.50.

| Estimand | Observed median (95% CrI) | Observed P(> 1) | No-harm posterior medians, 2.5-50-97.5% | Percentile of observed median: no harm / 1.25 / 1.50 | Percentile of observed 97.5% bound, no harm |
|---|---:|---:|---:|---:|---:|
| Shared mean (primary hierarchy) | 0.957 (0.599-1.517) | 0.425 | 0.892 - 0.991 - 1.098 | 28 / 0 / 0 | 25 |
| Common effect | 0.977 (0.659-1.446) | 0.455 | 0.906 - 0.996 - 1.108 | 34 / 0 / 0 | 35 |
| PCV product (primary hierarchy) | 0.966 (0.608-1.553) | 0.436 | 0.898 - 0.995 - 1.096 | 27 / 0 / 0 | 32 |
| MMR product (primary hierarchy) | 0.951 (0.520-1.634) | 0.414 | 0.864 - 0.979 - 1.103 | 31 / 0 / 0 | 38 |
| Modern rotavirus product (primary hierarchy) | 0.979 (0.670-1.449) | 0.456 | 0.899 - 0.997 - 1.118 | 36 / 0 / 0 | 42 |

## Where the operating characteristics come from: block-level recovery

Posterior median of each block's own replicate likelihood under the gate prior N(0, log 2): median across replicates [10th-90th percentile], with no harm and with R = 1.5 on that block. A block whose median barely moves carries little information about the effect; a no-harm median away from 1 is that block's null lean.

| Context | R = 1.00 | R = 1.50 |
|---|---:|---:|
| US_PCV | 1.047 [0.937-1.162] | 1.554 [1.393-1.706] |
| US_RotaShield | 0.968 [0.765-1.177] | 1.445 [1.119-1.758] |
| JPN_MMR | 0.769 [0.559-1.103] | 0.895 [0.643-1.324] |
| SWE_PCV_corrected | 0.989 [0.850-1.134] | 1.468 [1.257-1.683] |
| DEU_PCV | 0.983 [0.851-1.127] | 0.990 [0.859-1.134] |
| HKG_PCV_autism | 0.976 [0.814-1.146] | 1.283 [1.144-1.461] |
| ENG_RV_backfill | 0.998 [0.949-1.048] | 1.471 [1.408-1.546] |
| SCT_RV | 1.537 [1.522-1.553] | 1.654 [1.639-1.670] |
| NIR_RV | 0.983 [0.917-1.073] | 1.476 [1.376-1.609] |
| CAN_noQC_RV_newly_identified_cnics | 1.035 [0.759-1.434] | 1.487 [1.175-2.038] |
| QC_RV_ASD_cnics | 0.890 [0.705-1.175] | 1.313 [1.053-1.710] |
| CZE_MMR_counts | 1.026 [0.859-1.295] | 1.088 [0.952-1.357] |
| BRA_MDTPA_census_autism | 1.036 [0.857-1.343] | 1.550 [1.279-1.925] |
| COL_VARICELLA_school_autism | 1.029 [0.851-1.268] | 1.512 [1.262-1.856] |
| TZA_SCHEDULE_autism | 0.962 [0.719-1.320] | 1.277 [0.961-1.653] |
| US_RV2006 | 0.919 [0.828-1.050] | 1.393 [1.241-1.564] |

Posterior medians of each block's own replicate curve under N(0, log 2); U.S. RotaShield at R = 1.5 is from the common-harm DGP (every U.S. path at 1.5), U.S. PCV and rotavirus 2006 from their concentrated DGPs.

## Numerical checks

| Check | What | Value | Tolerance | Pass |
|---|---|---:|---:|---|
| A | grid hierarchy on the observed data vs the primary headline (median, 2.5%, 97.5%) | 0.0012 | 0.005 | yes |
| B | foreign product functions vs global_hierarchy.product_inputs (max relative difference) | 0 | 1e-12 | yes |
| C | grid core integrals vs global_hierarchy.component on a tensor-grid core (max abs value difference) | 9.09e-13 | 1e-08 | yes |
| D1 | Gaussian refits at the observed data vs production curves (max abs difference) | 9.8e-09 | 1e-06 | yes |
| D2 | count refits at the observed counts vs production curves on the window (max abs difference) | 6.49e-06 | 0.0001 | yes |
| E1 | Asimov R = 1, 2 medians vs the recorded harm-injection gate (Gaussian blocks and Japan) | 5.33e-15 | 0.0001 | yes |
| E2 | Asimov R = 1, 1.5, 2 medians vs a fresh gaussian_block_gate call (Gaussian blocks) | 5.33e-15 | 1e-06 | yes |
| E3 | count-block Asimov R = 1, 2 medians vs the recorded gate | 0 | 0.0001 | yes |
| E4 | the gate's 353-cell U.S. engine re-run: R = 1, 2 medians vs the recorded gate | 0 | 0.0001 | yes |

Noise-free (Asimov) posterior medians at an injected R = 1.5 under the gate prior N(0, log 2): SWE_PCV_corrected 1.497; DEU_PCV 0.991; HKG_PCV_autism 1.317; ENG_RV_backfill 1.446; NIR_RV 1.522; CAN_noQC_RV_newly_identified_cnics 1.702; QC_RV_ASD_cnics 1.382; BRA_MDTPA_census_autism 1.558; COL_VARICELLA_school_autism 1.467; TZA_SCHEDULE_autism 1.270; JPN_MMR 1.163; SCT_RV 1.652; CZE_MMR_counts 1.093. U.S.: the gate's 353-cell engine gives PCV 1.506 and RotaShield 1.511; the 1,469-cell grid engine used here gives PCV 1.526, RotaShield 1.485, rotavirus 2006 1.432. Seed split: the largest |z| between the two independent halves over all 425 rates is 1.79 (0 scenario-estimand rows with |z| > 3). U.S. grid edge mass: max 1.0e-04 over all replicates. Count likelihood maximum at a window edge: SCT_RV 0% of replicate curves, CZE_MMR_counts 0% of replicate curves; replaced count draws: {'SCT_RV': 0, 'CZE_MMR_counts': 0}.

## Files

- `07_analysis_code/3351_portfolio_power_calibration.py` (this analysis), `07_analysis_code/test_3351_portfolio_power_calibration.py`
- `3351_portfolio_power_calibration.csv` (scenario x estimand operating characteristics), `.json` (everything, with checks and sha256s)
- `08_figures/3351_portfolio_power_calibration.png`
- `project_results/portfolio_power_3351/` (cache/ replicate curves, aggregate/ per-replicate posteriors, checks/, logs/)
- `qa/portfolio_power_20260929/3351_rule.json` and `3351_rule.sha256`