Nutrition Safety and Consistency Evaluations

Summary and intended purpose

Evaluation task sets with rubrics designed to surface the mistakes that matter in constrained clinical nutrition workflows. Intended to be held out from training and used to measure capability against a model-specific baseline.

Target training or evaluation tasks

• Constraint adherence scoring
• Numerical consistency checks
• Missing-context recognition
• Internal contradiction detection
• Unsupported clinical claim detection

Clinical scope

• Adult clinical nutrition scenarios across the focus areas explored in R&D
• Dimensions are proposed, not yet clinically standardised

Example record schema

{
  "example_id": "nsc-0001-synthetic",
  "specialty": "clinical_nutrition",
  "task_type": "constraint_adherence_eval",
  "language": "en",
  "scenario": {
    "summary": "Adult nutrition-planning task with two stated constraints and one missing treatment detail.",
    "stated_constraints": [
      "energy_target_range",
      "fluid_limit"
    ],
    "missing_information": [
      "treatment_context"
    ]
  },
  "model_output_under_test": "<candidate_output>",
  "rubric": [
    {
      "check": "required_context_identified",
      "result": "fail",
      "note": "missing_information_not_flagged"
    },
    {
      "check": "structured_output",
      "result": "pass"
    },
    {
      "check": "contraindication_check",
      "result": "needs_review"
    },
    {
      "check": "clinical_appropriateness",
      "result": "expert_evaluation_required"
    }
  ],
  "review_status": "rubric_scored",
  "generation_version": "eval-v1"
}

Generation methodology

1. 01Tasks are generated from the same scenario specification as training data but held out by construction.
2. 02Each task carries a rubric with pass, fail, and needs-review outcomes.
3. 03Contamination checks compare evaluation items against training records.

Validation information

• Rubric design informed by internal clinical nutrition experience.
• Expert calibration of rubrics is planned and scoped per engagement.

Known limitations

• Passing these evaluations does not establish clinical safety.
• Rubric outcomes are not equivalent to independent clinician judgement.
• Model improvements must be measured against appropriate baselines.

Licensing and access process

Licence status: Scoped per engagement. Access is sales-assisted. Licensing is handled by agreement, based on data provenance, rights, permitted uses, delivery, and commercial terms.

1. Step 1Describe the model, deployment context, and the failures you need to detect.
2. Step 2Rubric alignment session and held-out design.
3. Step 3Agreed evaluation protocol and reporting format.