Technical

Technical notes for engineers and auditors.

This page describes SEED’s evaluation artifacts and data shapes at a high level. It is not a formal API specification and may evolve.

Back to the overview

Terminology

Consistency checks (regression pinning)

Variance and repetition are treated as signals, not errors, to surface shifts and instability across releases.

EvaluationRecord

A structured artifact containing distributed risk signals, audience interpretations, and review context. Not a single-score result.

Rewrite re-check

Rewrites are evaluated as mitigation hypotheses, not guarantees of risk reduction.

API endpoints

Endpoints shown here are illustrative. Field presence, structure, and verbosity may vary by domain, configuration, and review context.

POST /evaluations

GET /evaluations/:id

GET /evaluations/:id/report

PATCH /evaluations/:id/review

View example request

POST /evaluations
{
  "input_text": "Text to evaluate",
  "client_id": "optional",
  "engagement_id": "optional",
  "domain": "healthcare",
  "include_rewrites": true,
  "run_rewrite_eval": true
}

View example response

Example response (representative, non-exhaustive)

{
  "evaluation_id": "uuid",
  "timestamp": "2026-01-01T00:00:00Z",
  "overall_risk_level": "high",
  "confidence": "moderate",
  "top_concerns": [
    { "title": "Loss of agency", "category": "coercion" },
    { "title": "Catastrophic framing", "category": "fear amplification" }
  ],
  "risk_distribution": {
    "authority_pressure": "high",
    "emotional_safety": "elevated",
    "legal_claims": "low"
  },
  "audience_reactions": [
    {
      "audience": "literal interpreter",
      "interpretation": "Reads as a direct instruction with consequences.",
      "should_you_care": "Yes. Can override user judgment."
    },
    {
      "audience": "authority-sensitive reader",
      "interpretation": "Perceives a non-negotiable mandate.",
      "should_you_care": "Yes. Increases dependency risk."
    }
  ],
  "rewrite": {
    "status": "optional",
    "rationale": "Rewrites are evaluated as mitigation hypotheses."
  },
  "flags": [
    {
      "concern": "Loss of agency",
      "evidence": "must comply",
      "severity": "high"
    },
    {
      "concern": "Loss of agency",
      "evidence": "or you will be in danger",
      "severity": "high"
    }
  ],
  "recommendation": "block"
}