Track · critic

Critic track

Cross-provider taste auditor. Evaluates generated output (illustrations, layouts, prose, scene compositions) against authoritative reference sets with weighted rubrics. By design the provider that critiques is independent from whatever generated — so a Claude-authored rig gets judged by GPT, or vice versa.

Endpoints

Route	Auth	Purpose
`GET /api/v1/critic/rubrics`	`critic:v1:read`	List every rubric (title, id, weight schema).
`GET /api/v1/critic/rubrics/[id]`	`critic:v1:read`	One rubric's full definition.
`GET /api/v1/critic/references`	`critic:v1:read`	List reference sets (labels, image counts).
`GET /api/v1/critic/references/[id]`	`critic:v1:read`	One reference set's images + notes.
`POST /api/v1/critic/critique`	`critic:v1:use`	Submit an output for scoring. Billable.

:read is free (or plan-default). :use is metered.

Scopes

Scope	Grants
`critic:v1`	Umbrella. Every endpoint above.
`critic:v1:read`	Rubric + reference inspection. No critique calls.
`critic:v1:use`	`POST /critique`. Implicitly grants `:read`.

Quick critique

curl -X POST "https://urja.insightsbyomkar.com/api/v1/critic/critique" \
  -H "x-api-key: $CRITIC_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "rubric":    "great-dane-construction",
    "reference": "great-dane",
    "target":    { "svg": "<svg>…</svg>" },
    "meta":      { "iteration": 12, "hypothesis": "shorter muzzle" }
  }'

Response:

{
  "score":         7.2,
  "perCriterion":  [{ "name": "muzzle-skull-ratio", "score": 0.74, "weight": 0.15 }, ...],
  "text":          "Improved vs iter 11. Muzzle length closer to breed standard...",
  "provider":      "openai:gpt-4o",
  "referenceUsed": "great-dane"
}

Cross-provider independence

The rubric evaluator lives in lib/critic/providers/. Environment config (CRITIC_DEFAULT_PROVIDER=openai|anthropic) picks the judge; the default provider for Urja today is OpenAI. Rotating to Anthropic is a per-customer tweak we route through at mint time or at request time via a header.

Reference sets

Every reference set is a public-domain, pre-1929 authoritative corpus. No modern Solar Fire / Rudhyar / Hand translations — rubric language cites only works whose copyright has lapsed. The intent is legal-clean AI grounding rather than taste-via-vibes.

Examples of shipped sets:

great-dane — breed-standard photos + anatomical plates (public domain).
art-nouveau-ornaments — decorative motifs from c.1900–1920 prints.

Automatic runs (GitHub Actions)

The lucky-critique workflow in Urja runs scripts/critic/sweep-lucky.ts on every character-manifest change, posts a score table as a PR comment, and nightly-regresses against the great-dane reference set. See scripts/critic/README.md for the CI harness.

Response contract

All endpoints return JSON with cache-control no-store (per-call results + per-customer billable). Rate-limited per key — watch the X-RateLimit-* headers.