# Evidence notes: reactor diagnostic experiment exam

Source study dated 26 Oct 2026. This is a synthetic experiment generator based on a lumped reactor model, not a plant diagnostic or operating procedure. No world model, learning system or deployed detector was evaluated against the exam.

## Final source hierarchy

Use METHODS.md, final v2; certification_table.csv; verification_benchmarks.csv; and pwr_exam_pack.tar.gz, including sweep_results.json and the self-contained case reports and record CSVs. The final METHODS and certification CSV match their nested copies in the supplied pack.

## What was checked during curation

- certification_table.csv has 20 cases across four families. Sixteen have degen=True. FLUX-04, FLUX-05, THERM-04 and THERM-05 are retained boundary failures, correcting the earlier prose count.
- sweep_results.json has 1,440 intervention-design records: 72 per case. There are 20 case-record CSVs and 15 rows in the analytic benchmark table.
- The displayed THERM-03, FLUX-03 and REACT-03 sensor traces each sample 79 actual rows, including phase transitions. Each is one noisy generated record, not the clean response under both hypotheses.

These are supplied-file checks. Physics integration, analytic benchmark calculations, intervention search and Monte Carlo classification were not independently rerun. Numerical sweep time is not the complete research-session duration.

## Central comparison

The intervention control is the same total observation interval without a probe. The source reports THERM/SG separation gains from 14.24 to 15.10 times relative to that control. REACT waiting alone already gives reported Monte Carlo classification accuracy from 0.996 to 1.000; FLUX probes give a separation gain of 1.00. Thus the outputs distinguish acting, waiting and obtaining an independent calibration reference. A separation ratio is not an accuracy multiplier.

## Unresolved source consistency

METHODS.md claims closed-form separation and Monte Carlo checks agree across all cases. The passive REACT rows do not support that claim. REACT-03 has full-precision passive separation S=1.0605001468, predicted accuracy 0.7020307212, and reported Monte Carlo accuracy 0.52025. The other four passive REACT cases also disagree materially. The page keeps these metrics separate and does not present a universal validated Bayes bound. A numerical-method audit is required to resolve the discrepancy.

FLUX-04 displays S=1.20 in the rounded certification CSV, but its full-precision value exceeds the 1.2 threshold. Use the supplied degeneracy flag or full precision, rather than recomputing status from rounded values.

## Model assumptions and boundaries

The source describes a 20-state, four-loop lumped PWR with six delayed-neutron groups, thermal feedback, iodine/xenon, boron and rod control. It omits spatial neutronics and decay heat, permits one fault at a time, simplifies the secondary inventory, linearizes saturation, and omits a real-plant controller deadband. Instrument nuisance priors and modeled limits are engineering assumptions rather than calibration to a particular plant.

Reported Monte Carlo accuracy of 1.000 means a finite set of simulated classifications, not perfect real-world diagnosis. Constant boron dilution and constant rod-drive slip retain an ambiguity under all available actuators; resolving it requires information outside these channels. No observed learning-system exam performance is established.

## Figures

- certification-overview.png is the unaltered fig_certification.png. Its reported values must be read with the passive REACT consistency issue above.
- temperature-bias-experiment.png is the unaltered fig_therm_case.png for THERM-03. The reported 0.213 K difference and 0.170 K single-sample temperature-rise noise are synthetic model quantities, not plant-calibrated operational limits.