forge-graded-verify

Advisory graded quality scoring for RAG/research/scraping/prediction output. Use alongside forge-verify for subjective answer quality — rubric, per-criterion score, source grounding.

SKILL.md

GitHub

forge-graded-verify — advisory rubric-scored verification (scout #7)

Fills the graded-quality axis Forge deliberately lacks: forge-verify checks structural bookkeeping ("did the agent close its tickets"); forge-evals is intentionally deterministic-binary so it can safely gate git reverts. Neither scores subjective ANSWER QUALITY. This skill adds that — advisory only. Method reimplemented from the DeepVerifier failure-taxonomy pattern (arXiv 2601.15808); no code/dataset from that repo is used.

When to use

High-stakes answer-quality work where "did it pass an assertion" isn't enough: RAG answers, research syntheses, scraped-data summaries, prediction rationales. NOT for code/build gating (that stays with forge-evals + integration-gate). Reserve for genuinely high-stakes tasks — it costs an LLM review pass; don't duplicate ultra-review.

How it works (advisory, bounded, never gates irreversible actions)

  1. Pick the rubric for the domain from .claude/config/rubrics/<domain>.json. Each criterion has: id, descriptor (4-level: 1=poor … 4=excellent, what each level means), and raise (what would move the score up).
  2. Dispatch review-boss as a GRADED verifier (not the binary reviewer): it reads the output + the source/context, and returns per-criterion {id, score 1-4, evidence, feedback} via verification-by-decomposition (judge each criterion separately, cite evidence).
  3. Optional dual review for high-stakes work: for high-stakes answers only (RAG, research, prediction), an optional dual-review pattern with two independent reviewers using the same rubric can provide stronger confidence. A split verdict (reviewers disagree) gets one rework round, never a loop. This doubles review usage, so it is advisory-only, not the default.
  4. Bounded single rework loop: if any required criterion scores < the threshold (default 3), emit ONE rework_task_created → rework_assigned to the owning Boss with the NL feedback, re-grade ONCE, respect the usage-guard. No open-ended loops.
  5. Log advisory only: record the graded result via the existing gate_evaluated / lead_review_completed events (do NOT invent event types). NEVER let a graded score gate an irreversible action (deploy, git revert, send) — deterministic forge-evals / owner approval stay authoritative there.

Honesty

LLM-graded scores are judgment, not ground truth (a judge can mislabel). Frame results as advisory quality signals; cite per-criterion evidence; the reviewer model should differ from the builder model (evaluator independence). Rubrics are owner-editable JSON — start from the shipped ones and ADAPT per project, don't drop-in.

Rubrics

Shipped starters in .claude/config/rubrics/: rag.json (source-grounding, reasoning soundness, coverage, answer-correctness, completeness), payments.json (no-raw-PAN, money-integrity, idempotency, webhook-security, test/live-safety), ecommerce.json (real-catalog, inventory-safety, order-idempotency, flow-integrity, connector-safety), electron.json (renderer-hardening, ipc-safety, csp/secrets, installer-reality, no-mock-data), voice.json (dutch-first-ux, turn-discipline, barge-in/latency, consent/disclosure, action-safety). Add more per-domain rubrics as needed (research, scraping, prediction) with the same shape.