Methods, Verification, and Validation

Defensible sequence

Confirm the current inventory and executable next action in Current Project Evidence; do not infer them from this procedure.
VERIFIED REVIEW RECORD: Complete the approved broad D3/D5 primary-source review. Four rounds reached the two-successive-no-new-results stopping rule; the review records candidate families, constraints, applicability, disagreements, and evidence gaps.
APPROVED DECISION: Use the owner-approved bounded candidate set and comparison procedure in RESEARCH_SCOPE.md.
Execute only predefined Paper 1 preliminary calculations, beginning with the approved short-history centre Case at 7.5 MeV, and retain failed, negative, excluded, and inconclusive records.
Approve one fixed D3 construction and complete the D5 amendment.
Complete the D6 comparison of conditional JENDL-5, ENDF/B-VIII.1, and PHITS built-in modelling before D4 calculations.
UNRESOLVED: Approve D7 grid, uncertainty, reportability, repeated-run, convergence, completion, and reporting criteria using preliminary room calculations.
UNRESOLVED: Approve D8 verification geometries, references, uncertainties, and comparison criteria before acceptance comparisons.
UNRESOLVED: Identify applicable D9 measurements, uncertainties, and physical-validation criteria.
UNRESOLVED: Predefine the qualifying Study, Assessment procedure, D10 groups and partitions, and D11 metrics before reference-Dataset generation.
Generate immutable room Cases and PHITS Results; retain failures, interruptions, exclusions, and inconclusive outcomes.
Admit Results only through an Assessment of a wholly qualifying Study.
Train baselines and surrogate candidates using training data and select or tune using only model-selection data.
Finalize the Model before opening the untouched final test; evaluate that partition once under the predefined procedure.
Independently compare selected Predictions with new PHITS calculations and report global, local, subgroup, uncertainty, and underprediction behaviour.

Application workflow checks belong in ramal-ebx-application-verification. Paper 1 preliminary calculations remain in the Paper 1 Project. A qualifying Study creates fresh Study-member Jobs; earlier preliminary Results remain Case history and are not silently repurposed.

Status distinctions

Successful execution: the process ended with the recorded return code and artifacts.
Transport completion: PHITS reached its declared end without a disqualifying transport failure.
Convergence: predefined comparisons show numerical stability for declared regions and statistics.
Verification: implementation and numerical behaviour agree with specifications or accepted references.
Validation: the physical model agrees adequately with applicable measurements and uncertainties for its purpose.
Uncertainty: statistical and other declared uncertainty is quantified.
Acceptance: an owner-approved criterion was satisfied.
Assessment and Dataset inclusion: the Result satisfied every applicable criterion in a qualifying Study and was admitted through its Assessment.

No later status follows automatically from an earlier one.

Approved D3/D5 method

The approved identity is a finite planar Ta–water–SS304 candidate family with a circular Gaussian source and uniform normalized one-dimensional scan. The screening coordinates and procedure are approved; numerical thresholds, retained spot/span/margin, and final construction remain amendment-open pending preliminary evidence.

Use +Y upward, electrons toward -Y, scan along X, and Z as the remaining transverse direction. Rotate source, layers, scoring geometry, outputs, and room integration together. See ADR-0001.

The authoritative decisions and unresolved gates are maintained in Research Scope and Decisions; the detailed evidence and derivations are maintained in D3/D5 Source-Head Review. Do not duplicate their full procedure here.

The current executable next action and its inspection checklist are maintained only in Current Project Evidence.

Study inclusion criteria

PROPOSED METHOD: A Result should enter a Dataset only when:

Case, Job, Attempt, Result, immutable input, and required artifacts pass integrity checks;
the physical definition is inside the approved domain;
required transport completed without disqualifying diagnostics;
PHITS version, libraries or models, controls, seed, and variance reduction are declared;
quantity, unit, normalization, multiplier, particles, grid, nominal field, and uncertainty field match the Study;
approved D7 reportability, uncertainty, convergence, and repeated-run criteria are met;
required D8 verification and declared D9 validation status are recorded;
unscored values are not converted to zero or repaired by undocumented interpolation; and
an Assessment records the outcome and every exclusion reason.

Current Result availability and the live Study-criteria implementation mismatch are maintained only in Current Project Evidence. A Paper 1 Study must not qualify until the approved method can be represented and assessed through the GUI.

Dataset partitioning

PROPOSED METHOD: D10 must predefine grouping keys, paired conditions, split proportions, seed, stratification, interpolation tests, and replica handling. Training, model-selection, and untouched final-test partitions are separated by complete physical configuration. All fields, voxels, profiles, replicas, and defined paired variants from one configuration stay together. If interpolation across 5-7.5 MeV is claimed, hold out at least one justified intermediate-energy configuration.

Retain Case and Result traceability, features and units, target and uncertainty fields, reportable masks, coordinates, material representation, particle pages, partition membership, and governing inclusion criteria.

Current Dataset-builder capability is maintained only in Current Project Evidence. D10 grouping keys, paired conditions, proportions, seed, and stratification remain unresolved.

Evaluation

PROPOSED METHOD: Compare nearest-configuration, interpolation or fitted attenuation, regularized regression, and compact nonlinear baselines before a spatial neural model. Select transformations, architecture, loss, hyperparameters, and stopping rules using only training and model-selection data.

Predefine signed error, bias, absolute error, relative error only for reportable PHITS values, logarithmic field error, underprediction and overprediction frequency and magnitude, high-quantile and maximum error, spatial maps and profiles, comparison with PHITS uncertainty, interval coverage, PHITS and inference time, and end-to-end speedup. Stratify by energy, geometry, barrier role, region, field level, particle, interpolation condition, and domain boundary. D11 formulas and tolerances require approval before final-test access.

Current Model, final-test, Prediction, and Prediction Verification capability is maintained only in Current Project Evidence. Capability does not establish model suitability or scientific acceptance.

Prediction verification

PROPOSED METHOD: Independently recalculate selected Predictions, including the largest underpredictions and overpredictions, high-uncertainty Predictions, representative low, middle and high energies, domain boundaries, and conditions near any approved design criterion. Statistical independence requires a new stochastic calculation. A Prediction remains a surrogate record and cannot be promoted to a PHITS Result.

Terminology

Design: editable authoring record.
Calculation: canonical domain term for one requested execution and its frozen input, lifecycle, diagnostics, files, and outcome.
Case / Job / Attempt: narrower persisted record kinds in the current application; use them only when identifying exact provenance within a Calculation.
Result: immutable publication from one successful Calculation; not automatically accepted evidence.
Study / Assessment: scientific question, membership, criteria, and evaluated outcome.
Dataset: Results admitted through the approved Assessment and partition procedure.
Model: trained surrogate finalized before final-test evaluation.
Prediction Verification: independent PHITS comparison for a selected Prediction.
Reportable voxel: scored voxel meeting approved D7 criteria.
Unscored voxel: missing Monte Carlo information, not zero.
Range of applicability: evidence-supported domain, not programme horizon.