jkf87/gnomon

Latest recorded source activity: May 9, 2026 at 22:25

SkillsMP catalog refreshed: November 9, 2026 at 14:22

skills collected: 1

GitHub stars: 3
GitHub forks: 2

Install with an AI assistant

Copy this prompt into the AI assistant you're using.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/jkf87/gnomon.

Repository README

Repository README source: README.md

Read the source README

Expand full README

# gnomon <p align="center"> <img src="assets/nomon-hero.jpg" alt="gnomon — a sundial casting an evaluation shadow" width="560"> </p> **Gnomon** — the needle of a sundial. When sunlight hits, the shadow becomes the answer. Rubric-first evaluation harness for AI agent loops — OpenClaw integration included, but the ideas and the gate engine work with any agent (Claude Code, Codex, or your own loop). Define what "good" looks like before writing a single line. --- ## What is Gnomon? **gnomon** is the pin on a sundial — the first quantitative verifier in human history. It casts a shadow against a fixed reference (the sun), not against itself. Gnomon applies that principle to AI evaluation: write the rubric first, then let the agent run. The harness measures output against ground truth — no LLM self-grading allowed past the threshold. | | Gnomon | |---|---| | Philosophy | rubric-first (define success before writing) | | Entry point | rubric.yaml | | Loop | verifier → writer (repeat until PASS) | | Measurement | quantitative / persona-LLM / human | | Guard | blocks entry if LLM ratio > 70% | --- ## Quick Start ```bash uvx --from git+https://github.com/jkf87/gnomon gnomon install ``` Then in OpenClaw: `/gnomon:setup` --- ## Usage ### Install & verify ```bash # Install into OpenClaw (idempotent) uvx --from git+https://github.com/jkf87/gnomon gnomon install # Check installation health gnomon doctor # Uninstall gnomon uninstall ``` ### Build a rubric ```bash # Generate a rubric.yaml scaffold gnomon rubric new # Validate before running (enforces label ratios) gnomon rubric check rubric.yaml ``` Rules enforced at `rubric check`: - `quantitative` items must be ≥ 30% — blocked otherwise - `persona-llm` items must be ≤ 70% — blocked otherwise ### Run evaluations ```bash # Run a task — spawns agents, writes report to .gnomon/runs/ gnomon run task.yaml # Side-by-side comparison view gnomon compare task.yaml # Dry-run (no agent spawn, placeholder output) gnomon run task.yaml --dry-run # Print a stored report gnomon report <run-id> ``` ### Workflow ``` gnomon rubric new # scaffold rubric.yaml ↓ gnomon rubric check # validate label ratios ↓ (PASS only) gnomon run task.yaml # spawn agents → collect outputs ↓ writer → verifier loop # iterate until PASS ↓ gnomon report <run-id> # review final result ``` ### task.yaml example ```yaml task: "Generate 10 card-news slides" goal_persona: role: "Office worker, 25–35" success_signal: "Core point summarisable in 5 seconds" items: - id: r1 label: quantitative description: "Font size >= 24 pt" pass_condition: "All text >= 24pt" - id: r2 label: persona-llm description: "Persona can summarise in 5 s" pass_condition: "PASS" - id: r3 label: human description: "Image–content tone match" pass_condition: "Reviewer approval" taste_gate: spearman_threshold: 0.7 ``` ### Available gates | Gate | Domain | Key signals | |------|---------|-------------| | `code_gate` | Software | test pass rate, lint score, complexity | | `blog_gate` | Writing | readability, structure, originality | | `translation_gate` | Translation | BLEU, terminology consistency | | `ui_gate` | Design | contrast ratio, spacing, accessibility | | `video_gate` | YouTube / Shorts | subtitle WER, audio LUFS, aspect ratio, chapter density | --- ## Gate Engine (TypeScript) — `packages/gates` A standalone TypeScript engine that turns plain-string acceptance criteria into judgeable evaluation gates and decides loop closure deterministically — the Backward Design counterpart to the Python harness above. - Derives one gate per AC (content-hash keyed): observable conditions + evidence spec + calibration pair - Qualitative ACs → rubric LLM-judge gates (`ac_compliance` + `score ∈ [0,1]`); `uncertainty > 0.3` resolves via multi-judge consensus, never human review - Mixed closure policy: all `critical` gates pass AND non-critical pass ratio ≥ 0.8 → `close`, else `continue` - Merge precedence `override > derived > default`; editing AC text invalidates its stored gate ```bash cd packages/gates && npm install && npm test # 12 tests ``` Built through a full Ouroboros loop (interview → QA-refined seed → autonomous run → 3-stage evaluation, APPROVED). The origin seed ships in [`packages/gates/seeds/`](packages/gates/seeds/). See [`packages/gates/README.md`](packages/gates/README.md). --- ## OpenClaw Skills (after install) | Skill | Description | |-------|-------------| | `/gnomon:rubric` | Scaffold a rubric for a task | | `/gnomon:verify-check` | Validate rubric.yaml | | `/gnomon:run` | Run the full evaluation loop | | `/gnomon:calibrate` | Taste-gate calibration (sample runs) | --- ## Development ```bash git clone https://github.com/jkf87/gnomon cd gnomon pip install -e ".[dev]" pytest tests/ -v # 63 tests ``` --- ## 한국어 가이드 **노몬(gnomon)** — 해시계의 바늘. 태양 빛이 닿는 순간 그림자가 정답이 된다. 루브릭-퍼스트 방식의 AI 에이전트 평가 하네스. OpenClaw 연동이 기본 포함이지만, 개념과 게이트 엔진은 어떤 에이전트 루프(클로드 코드, 코덱스, 자작 루프)에서도 쓸 수 있다. 한 줄도 쓰기 전에 "좋다"의 기준을 먼저 정의한다. ### 노몬이 뭔가요? **노몬(gnomon)**은 해시계의 바늘이다. 태양을 기준으로 그림자를 드리워 시간을 측정하는 도구 — 인류 최초의 정량 verifier. LLM에게 묻지 않고 ground truth에 직접 대조한다. 노몬은 그 원리를 AI 하네스에 적용한다. `rubric.yaml`이 먼저, workflow가 나중이다. | | 노몬 | |---|---| | 철학 | rubric-first (평가 먼저, 작성 나중) | | 진입점 | rubric.yaml | | 루프 | verifier → writer (PASS할 때까지) | | 측정 방식 | 정량 / 페르소나-LLM / 사람 | | 차단 조건 | LLM 비율 70% 초과 시 진입 불가 | ### 빠른 시작 ```bash uvx --from git+https://github.com/jkf87/gnomon gnomon install ``` OpenClaw에서: `/gnomon:setup` ### 사용법 #### 설치 및 확인 ```bash gnomon install # OpenClaw에 설치 (멱등) gnomon doctor # 설치 상태 진단 gnomon uninstall # 제거 ``` #### 루브릭 작성 ```bash gnomon rubric new # rubric.yaml 스캐폴드 생성 gnomon rubric check rubric.yaml # 실행 전 검증 (라벨 비율 강제) ``` 검증 규칙: - `quantitative` 항목 최소 30% — 미달 시 진입 차단 - `persona-llm` 항목 최대 70% — 초과 시 차단 #### 평가 실행 ```bash gnomon run task.yaml # 에이전트 spawn → 리포트 작성 gnomon compare task.yaml # 나란히 비교 출력 gnomon run task.yaml --dry-run # 드라이런 (에이전트 미생성) gnomon report <run-id> # 저장된 리포트 출력 ``` #### 워크플로우 ``` gnomon rubric new # 루브릭 스캐폴드 ↓ gnomon rubric check # 라벨 비율 검증 ↓ (PASS만 통과) gnomon run task.yaml # 에이전트 → 결과 수집 ↓ writer → verifier 루프 # PASS까지 반복 ↓ gnomon report <run-id> # 최종 리포트 확인 ``` ### 평가 게이트 | 게이트 | 도메인 | 주요 신호 | |--------|--------|-----------| | `code_gate` | 코드 | 테스트 통과율, 린트, 복잡도 | | `blog_gate` | 글쓰기 | 가독성, 구조, 독창성 | | `translation_gate` | 번역 | BLEU, 용어 일관성 | | `ui_gate` | 디자인 | 대비비, 여백, 접근성 | | `video_gate` | 유튜브/쇼츠 | 자막 WER, 오디오 LUFS, 화면비, 챕터 밀도 | ### 게이트 엔진 (TypeScript) — `packages/gates` 문자열 AC에서 판정 가능한 평가 게이트를 도출하고 루프 종료(close/continue)를 결정론적으로 판정하는 독립 엔진. 정성 AC는 루브릭 LLM-judge 게이트로, 불확실성이 높으면 다중 판정자 합의로 — 사람 검수 없이. 위 Python 하네스의 Backward Design 짝이다. [한국어 문서](packages/gates/README.ko.md) 참고. --- ## License MIT — See `LICENSE` file.

Loading the repository README…

Skills in this repository

Computer Occupations, All Other

1 occupation category · 100% classified

skill

occupation

description

updated

nomon

Computer Occupations, All Other

Rubric-first evaluation harness. Define what good looks like before writing. verifier → writer, not the other way.

Source text: Korean

2026-05-09

Showing 1 of 1 collected skill.