Sooth

LLMs generate. Sooth verifies.

Claim-by-claim fact-checking for AI output, designed for CI. Sooth checks every sentence in a draft against your source material and returns PASS / FAIL / REVIEW with calibrated probabilities and the exact source line behind each verdict.

See it catch a hallucination

pip install sooth
sooth demo        # no API key needed

Real output — seven claims in an AI-written news summary, three numbers quietly wrong:

# Sooth

**PASS 2 · FAIL 3 · REVIEW 1 · UNCHECKABLE 1** — threshold 0.70

| # | Claim | Verdict | P | Why (P distribution) | Evidence |
|---|-------|---------|---|----------------------|----------|
| 3 | BNBR baru menuntaskan rights issue bernilai besar di harga Rp 33. | ❌ FAIL | ████████ 1.00 | contradicts 1.00 / not_found 0.00 / supports 0.00 · details 0.01 · missing #s: 33 | `examples/news-1.md:25` "Saham ini juga baru menyelesaikan rights issue dalam juml…" |
| 5 | BUMI hanya perlu turun sekitar 10% untuk menyentuh level Rp 50. | ❌ FAIL | ████████ 1.00 | contradicts 1.00 / not_found 0.00 / supports 0.00 · details 0.01 · missing #s: 10 | `examples/news-1.md:23` "Adapun PT Bumi Resources Tbk (BUMI) di sekitar Rp192 haru…" |
| 6 | BEI juga menetapkan batas atas harga saham Rp 5.000 per saham. | ⚠️ REVIEW | ████████ 1.00 | contradicts 0.00 / not_found 1.00 / supports 0.00 · details 0.02 · missing #s: 5.000 |  |
| 7 | Para investor ritel sangat senang dengan aturan baru ini. | ➖ UNCHECKABLE | ██░░░░░░ 0.24 | P(checkable)=0.24 |  |

Every verdict carries its probability distribution and the source span it was judged against. When Sooth is unsure, it says REVIEW instead of guessing.

Why

AI writes fast, nobody checks. Claims ship wrong — usually a number, a hedge, or a name that drifted. Asking another LLM "is this right?" just produces more prose.

Sooth takes a different bet: the model makes typed judgments, code makes the decision. Every rule that turns a probability into a verdict is a readable line of Python, not a prompt.

Install

pip install sooth
export TYPESAFE_API_KEY=...       # get one at console.typesafe.ai

Usage

sooth --source docs/policy.md --source tickets/t123.md --text draft-reply.md

# CI-friendly exit codes
sooth --source policy.md --text draft.md --format plain
#   0 = clean  ·  1 = any FAIL  ·  2 = any REVIEW  ·  3 = usage/config error

# Options
#   --confidence T   REVIEW below this confidence (default 0.7)
#   --format md|plain
#   -o FILE          write report to file
#   --log FILE       append full judgment trace (one JSONL line per run)

Wire it into CI as a quality gate for generated content:

- name: Verify AI output
  run: sooth --source docs/policy.md --text generated-reply.md
  env:
    TYPESAFE_API_KEY: ${{ secrets.TYPESAFE_API_KEY }}

How it works

1. Draft is split into claims (one sentence each).
2. Each claim gets four questions to Jev, fanned out in parallel batches: is this checkable?, does the source support it? (supports / contradicts / not_found), do all details match exactly?, and which source span is the evidence?
3. Verdicts are mapped in code: uncheckable → UNCHECKABLE; low confidence → REVIEW; then PASS / FAIL. Safeguards demote PASS to REVIEW when details drift or claim numbers are absent from the source (checked in plain code).
4. The report shows the full probability distribution per claim — not just a label.

The decision logic lives in src/sooth/verify.py in a dozen readable lines. Change thresholds and rules there, not in prompts.

Judgments come from Jev (TypeSafe System One), which returns typed distributions instead of generated prose.

Why not just use an LLM judge?

Not claiming Sooth is universally more accurate. Claiming it gives you architectural guarantees a prompt cannot:

Who it's for

Developers building systems that generate text from trusted material:

• RAG pipelines — retrieved the right document, still wrote the wrong number
• AI agents — before an agent's reply reaches a human
• Support / ops copilots — policy-backed answers, checked against the policy
• Report & summary generation — figures that must match the source
• Compliance workflows — evidence trail for every claim, not vibes

Known limits (alpha)

• Safeguards catch most drift and smuggling: claim numbers absent from the source demote PASS to REVIEW, and a detail-match gate flags altered hedges ("about 74%" → "over 74%"). Not perfect — read the Why column before trusting a verdict.
• Derived numbers (totals, values computed outside the source) look "missing" and land in REVIEW.
• Source + questions must fit ~64k tokens — split long documents yourself.

Development

pip install -e ".[dev]"
python3 tests/test_core.py     # pure checks, no network
pytest                         # same suite
bash tests/smoke.sh            # live smoke (needs TYPESAFE_API_KEY)
PYTHONPATH=src python3 tests/calibrate.py   # live calibration, 30 labeled claims

Docs: PRD · Design · Testing

Roadmap

• v0.2.2 shipped: sooth demo (offline, no API key)
• v0.2.1 shipped: source-span evidence, published on PyPI
• GitHub Action
• later: hosted web app — paste UI, history, team review queues

License

MIT © Naufal Hilmiaji. Powered by TypeSafe / Jev.