18  Sub-Module 3.8-A

AI/ML Provenance Requirements and Acceptance Protocol

NoteNode Declaration — SM-3.8-A: AI/ML Provenance Requirements and Acceptance Protocol
Field Content
Tier Sub-Module
Status ○ Specified
Assumes §3.8
Contributes Complete specification of the five additional provenance fields, the confidence scoring protocol, the regime boundary detection and response protocol, and the human acceptance workflow
Skip condition Skip until surrogate modules are being deployed
Passes to Module 6 SM-6.6-E
Sub-Modules here None

18.1 SM-3.8-A: AI/ML Provenance Requirements and Acceptance Protocol

The five additional provenance fields for AI/ML-produced artefacts are specified as follows.

The model_type_version field carries a structured identifier of the form [architecture]:[version]:[commit], for example gradient_boosting_regressor:v2.1:a3f7c9d. It references an entry in the model registry, a governed table in the backbone that stores the full model specification, hyperparameters, and training configuration.

The training_data_id field carries the artefact identifier of the governed training dataset from which the model was trained. The training dataset is itself a governed artefact in the backbone with its own provenance, schema, and validation record.

The confidence_score field carries a float in [0,1] representing the surrogate’s self-assessed reliability for this specific prediction. A confidence score of 1.0 indicates the input is well within the interior of the training distribution; a confidence score below the declared threshold (default 0.7) triggers a truth model verification request.

The regime_flag field is a boolean indicating whether the input was detected as near a regime boundary. Regime boundaries are pre-identified from the full model’s output surface during training; inputs within a declared distance of a boundary in the input feature space receive a True regime flag regardless of their confidence score.

The human_acceptance_id field carries the identifier of the human acceptance record in the validation registry. This record is created when a qualified reviewer examines the artefact and confirms that its content is plausible and appropriate for the analytical context.

Confidence score protocol. The confidence score is computed as \(c = 1 - d_{\text{norm}}\), where \(d_{\text{norm}}\) is the distance from the input feature vector to the nearest training sample, normalised by the maximum nearest-neighbour distance in the training set. Equivalently, a Gaussian process surrogate can use the posterior standard deviation normalised by the prior standard deviation as the uncertainty measure. When \(c < 0.7\), the orchestration layer queues the input for truth model evaluation, runs the full PyPSA model (or equivalent full model) for that future, and replaces the surrogate prediction with the full model output before the artefact is admitted to the backbone. The full model output is recorded as a non-surrogate artefact with its own provenance.

Regime boundary protocol. During surrogate training, a preliminary grid scan identifies the input regions where the full model output changes most rapidly. These regions, typically near the transition from feasible-without-upgrade to exceedance conditions for the regional electricity module, are declared as regime boundaries and stored in the model registry. When a new input is evaluated, its distance to the nearest declared boundary is compared to a declared tolerance. Inputs within tolerance receive a True regime_flag and are automatically queued for truth model verification regardless of their confidence score.

Human acceptance workflow. A generated AI/ML artefact with a confidence score above the threshold and no regime flag is reviewed by a designated analyst before backbone admission. The review checks: (1) whether the output values are physically plausible given the input conditions; (2) whether the confidence score is consistent with the analyst’s domain assessment of the input’s difficulty; and (3) whether the artefact’s declared provenance is complete and consistent. The analyst records the review outcome in the validation registry. A positive review produces a human_acceptance_id that is embedded in the artefact’s provenance. A negative review returns the artefact to provisional status and flags it for investigation.

Table SM-3.8-A: Standard versus AI/ML artefact provenance fields

Provenance field Standard artefact AI/ML artefact
artefact_id Required Required
schema_version Required Required
artefact_family Required Required
run_id Required Required
future_id Required Required
produced_by Required (module name and version) Required (includes surrogate model reference)
validation_status Required Required
checksum Required Required
model_type_version Not applicable Required
training_data_id Not applicable Required
confidence_score Not applicable Required
regime_flag Not applicable Required
human_acceptance_id Not applicable Required