# Reactor diagnostic stress test: reviewed evidence notes

Prepared 27 October 2026 for the Laplace investigation page. This page compares recorded synthetic outputs; it does not run a solver in the browser.

## Displayed files

- `replay.json`: eight curated cases with indicated sensor observations, hidden physical states, wait/probe counterfactuals and stored diagnostic scores.
- `diagnostic-outcome-atlas.png`: unmodified source `atlas_fig6_atlas_map.png`.
- `conclusion-transfer-survival.png`: unmodified source `atlas_fig5_survival.png`.

The source report is `stage2_final_report.md`, with supporting `atlas_tables.zip`, `demo_transcript.md`, `final_test_log.txt`, `MODEL_A_V2.md`, `MODEL_B.md`, `SAMPLING_SCHEME.md`, `EVALUATION_CRITERIA.md` and `PREREGISTRATION.md`. The final report and its final corrections take precedence over earlier pilot or development reports.

## Separate counts and populations

The full campaign reports 99,790 scored diagnostic case definitions across 12,499 synthetic plant configurations. Its 64,241 non-excluded cases underpin the atlas. A case definition is not an ODE trajectory and a noise draw is not a physical integration.

The reported 5,718,064 physical integrations comprise 215,522 model-A settling integrations, 4,814,642 recording integrations and 687,900 partner-magnitude-solve integrations. The exported campaign records also report 898,064,000 Monte Carlo noise realizations and 1,531,877 candidate intervention evaluations. Neither quantity is an additional physical-trajectory count.

The atlas class counts, independently summed from the supplied CSV tables, are:

| Source class | Count |
|---|---:|
| INTERVENE | 14,617 |
| PARTIAL_PROBE | 7,161 |
| SENSE | 2,249 |
| PARTIAL_SENSE | 3,028 |
| WAIT | 9,783 |
| IRREDUCIBLE | 27,403 |
| Total | 64,241 |

Grouping the first two classes gives intervention helps, 33.9%; grouping the next two gives sensing required, 8.2%. Waiting accounts for 15.2% and unresolved cases for 42.7%. IRREDUCIBLE is relative to the tested hypothesis, action, sensing and horizon menu.

Pilot, development, full-campaign and challenge counts describe different tiers or rescorings and must not be added as unique physical cases. This study is separate from the other reactor investigation on the website.

## Model and comparison scope

The reported model inventory comprises a 66-state, lumped four-loop PWR model A and a separately written 146-state reference model B. The sampling framework varies 41 plant priors, 33 eligible fault mechanisms, six fault-composition strata and five observation models. State counts describe model architecture, not demonstrated real-plant fidelity.

All accuracy and confidently-wrong values are oracle-Bayes results from the profiled-likelihood classifier on simulated noise draws. No learned AI system, customer model or human operator was evaluated.

In the matched challenge, 969 of 1,427 reference conclusions retain their class under parameter mismatch; 690 retain it under structural mismatch; 550 retain it under both. Structural survival is 690 / 1,427 = 48.35%, shown as 48.4%. Failed and excluded twins count as not surviving. Only 1,194 structural twins were scored, and model B fails to settle for roughly 13% of mapped configurations.

On the scored structural twins, the source reports a confidently-wrong draw rate of 38.5% before the goodness-of-fit gate and 1.3% afterwards, alongside rejection of 85.2% of draws. For held-out combinations of known mechanisms, the final reported confident misattribution is 64.3% raw and 54.8% gated; the gate rejects 9.9% of draws. These are different challenge populations. The gate is not a universal protection against a missing physical explanation.

## Curated replay interpretation

The original explorer contains 314 cases: 300 stratified campaign examples, 12 replay-pack cases and two generated demonstrations. This page selects eight illustrative cases. Their category shares are not population estimates.

Source authors regenerated these time series on model A because the full campaign stored no traces. Indicated sensors show one noise realization. Hypothesis counterfactuals are noise-free. Source output spacing is 150 seconds; the displayed data retain at most 80 original samples per case without interpolation. Hidden thermal power is divided by that case's rated power to express a fraction of rated power.

Mismatch regime scores are stored scalar comparisons. Selecting those scores does not change the displayed model-A series into model-B traces. The SENSE case contains source sensing scores but no separate added-sensing time series in the curated file. ABSTAIN is a website label for source IRREDUCIBLE, not a tested abstention policy.

The accepted pump example has source oracle accuracy 0.821 after waiting and 0.998 after the designed intervention, with diagnostic separation gain 3.243. This is a classifier result within that synthetic case, not AI performance. Another demonstration is deliberately retained as EXCLUDED because its frozen metric-consistency check failed. It must not be shown as an accepted exam case.

## Verification boundary

The supplied final test log records 106 passing tests. Twelve source replay cases retained decision class and best design, with maximum relative separation difference about 0.0012 and maximum accuracy difference about 0.027 at 300 Monte Carlo draws. The stricter 1e-4 cross-platform target was not demonstrated. Those logs were reviewed as supplied evidence; the page preparation did not rerun the physical models.

This study does not establish reactor operating safety, real-world diagnostic accuracy or transfer to a real plant.