AI-optimisation record

Type: Proposal
Status: Active candidate evaluation overlay v0.1
Canonical for: The proposed record shape for capability-scoped agent evaluation results
Depends on: Candidate package, canonical AI-optimised definition, and future evaluation protocols

This is a versioned evaluation overlay for the Design and Development Contracts. It is not part of the candidate AI-readiness acceptance profile and is not yet a full evaluation framework.

Required evaluation identity

Field	Requirement
evaluationId	Stable evaluation record ID.
conceptId	Component or pattern evaluated.
contractVersion	Exact Design/Development Contract version.
artifactVersions	Figma, package, Storybook, rules and validator versions.
model	Exact provider/model identifier and date.
agentHarness	System prompt, tools, retrieval configuration and structured-output mode.
taskSetVersion	Frozen scenario/gold-policy version.
runConfiguration	Temperature, samples, retries, context limits and ordering policy.

Capability results

Record separate results for every capability tested:

• discovery;
• component selection;
• no-match/clarification;
• property interpretation;
• enum/concept selection;
• legal configuration;
• composition;
• implementation binding;
• behavior/state correctness;
• accessibility correctness;
• content/token compliance;
• repair after validation.

For each, record task count, scoring rule, accuracy, uncertainty and failure distribution. Do not collapse schema validity, semantic correctness and visual similarity into one score.

Robustness results

Where relevant, record sensitivity to:

• prompt paraphrase;
• candidate and property order;
• aliases and renaming;
• opaque or conflicting labels;
• missing or truncated guidance;
• stale/conflicting artifacts;
• model family and model version;
• repeated runs.

Change record

An optimisation claim MUST identify:

• observed failure;
• contract/retrieval/validator change;
• hypothesized mechanism;
• pre/post comparison;
• regressions in other capabilities;
• human usability or maintenance costs;
• expiry or rerun trigger.

Changes to the public API are not assumed to be the appropriate response. An observed naming failure may be addressed by naming, descriptions, retrieval, metadata, constraints or another mechanism and must be tested against alternatives.

Optimisation status

Suggested record states:

• not-evaluated;
• evaluated-baseline;
• improved-under-tested-conditions;
• regression-detected;
• expired;
• not-generalizable.

Avoid a permanent, unqualified AI-optimised: true. Results are scoped to the recorded contract, artifacts, tasks, harness and model set.