ChainCheck Action

Detect hallucinations in AI-generated PR descriptions and commit messages.

Uses ChainCheck to decompose text into atomic claims, verify each one, and post a per-claim verdict as a PR comment. Fails the check if the hallucination score exceeds a configurable threshold.

---

Example output

❌ ChainCheck — score 0.71 (HIGH)

Claim	Label	Conf	Evidence
This PR reduces API latency by 40%	⚠️ unsupported	0.89	no benchmark data provided
Adds async batching to the NLI pipeline	✅ supported	0.94	—
Fixes the rate-limit regression from #142	⚠️ unsupported	0.81	issue #142 not referenced

❌ Failing — score 0.71 exceeds threshold 0.7. Review the flagged claims above before merging.

---

Usage

# .github/workflows/chaincheck.yml
name: ChainCheck

on:
  pull_request:
    types: [opened, edited, synchronize]

jobs:
  hallucination-check:
    runs-on: ubuntu-latest
    permissions:
      pull-requests: write   # needed to post the comment

    steps:
      - uses: pauti04/chaincheck-action@v1.4
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}  # enables diff context
        with:
          openai-api-key: ${{ secrets.OPENAI_API_KEY }}

No actions/checkout step needed — the action fetches the PR diff directly from the GitHub API.

---

Inputs

Input	Required	Default	Description
openai-api-key	✦	—	OpenAI API key. Required unless anthropic-api-key or ollama-base-url is set.
anthropic-api-key	✦	—	Anthropic API key (alternative to OpenAI).
ollama-base-url	✦	—	Ollama base URL for local inference, e.g. http://localhost:11434.
model		auto	Judge model override, e.g. gpt-4o-mini, claude-haiku-4-5-20251001, ollama:llama3.
check		pr-description	What to check: pr-description, commit-messages, or any custom string.
threshold		0.7	Fail if hallucination score ≥ this value (0.0–1.0).
post-comment		true	Post results as a PR comment.
methods		judge	Detection methods: nli, judge, or nli,judge.
context		""	Optional reference doc to check claims against.

✦ At least one of openai-api-key, anthropic-api-key, or ollama-base-url is required.

Outputs

Output	Description
score	Aggregate hallucination score (0.0–1.0)
risk-level	low, medium, or high

---

Examples

Check PR description (default):

- uses: pauti04/chaincheck-action@v1.4
  env:
    GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  with:
    openai-api-key: ${{ secrets.OPENAI_API_KEY }}
    threshold: '0.7'
    post-comment: 'true'

Check commit messages (requires actions/checkout first):

- uses: actions/checkout@v4   # required for commit-messages mode
  with:
    fetch-depth: 10

- uses: pauti04/chaincheck-action@v1.4
  with:
    openai-api-key: ${{ secrets.OPENAI_API_KEY }}
    check: commit-messages
    threshold: '0.8'

NLI + judge ensemble (higher accuracy, slower):

- uses: pauti04/chaincheck-action@v1.4
  env:
    GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  with:
    openai-api-key: ${{ secrets.OPENAI_API_KEY }}
    methods: 'nli,judge'
    threshold: '0.65'

Use score in downstream steps:

- id: chaincheck
  uses: pauti04/chaincheck-action@v1.4
  env:
    GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  with:
    openai-api-key: ${{ secrets.OPENAI_API_KEY }}

- name: Print score
  run: echo "Score=${{ steps.chaincheck.outputs.score }} Risk=${{ steps.chaincheck.outputs.risk-level }}"

Use Anthropic Claude instead of OpenAI:

- uses: pauti04/chaincheck-action@v1.4
  env:
    GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  with:
    anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}
    model: 'claude-haiku-4-5-20251001'

Use a local Ollama model (self-hosted, no API costs):

- uses: pauti04/chaincheck-action@v1.4
  env:
    GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  with:
    ollama-base-url: 'http://your-ollama-host:11434'
    model: 'ollama:llama3'

Soft mode — comment only, never fail:

- uses: pauti04/chaincheck-action@v1.4
  env:
    GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  with:
    openai-api-key: ${{ secrets.OPENAI_API_KEY }}
    threshold: '1.1'   # effectively disables failing
    post-comment: 'true'

---

How it works

1. Reads the PR description (or commit messages) from the GitHub event payload
2. Fetches the PR diff via the GitHub API and uses it as grounding context for claim verification — no actions/checkout required
3. Decomposes text into atomic claims using gpt-4o-mini
4. Verifies each claim with the selected method (NLI cross-encoder and/or LLM judge)
5. Upserts a single PR comment — edits the previous ChainCheck comment on re-runs instead of posting a new one each push
6. Exits non-zero if aggregate_score ≥ threshold

Powered by ChainCheck — achieves 79% F1 / 94% precision on HaluEval-QA.

---

Cost

Each run makes 1–3 OpenAI API calls (claim decomposition + judge). A typical PR description (200 words) costs ~$0.001 with gpt-4o-mini.

---

License

MIT