claim-gate

One evidence-gated ledger for public AI claims and signals — many domains, one CLI.

Public AI claims arrive faster than they can be verified: a model is "cheaper", an infra trend is "the next bottleneck", a bio result is "deployable". claim-gate keeps each claim in a local evidence ledger, checks it against domain-specific evidence rules, and emits a blunt recommendation — act / track / verify-first / reject — so hype never silently becomes a decision.

It is local-first and zero-dependency. Nothing leaves your machine.

Star This If

• You track AI/model claims and want them backed by local evidence, not hype.
• You want one CLI that gates each claim as act / track / verify-first / reject.
• You need an auditable, local-first ledger of what's verified, unproven, or rejected.

Why It Matters

AI teams lose time when hype turns directly into roadmap, dependency, or investment decisions. claim-gate inserts a small but strict step between "interesting claim" and "act on it":

1. record the claim with a source
2. classify the domain and risk
3. attach evidence or leave it blocked
4. export a decision-ready ledger

The default posture is conservative. A claim can be useful and still remain verify-first until independent evidence exists.

Quick Start

python -m pip install -e .
claim-gate --list-domains
claim-gate model add \
  --subject "ExampleModel" \
  --claim-type benchmark \
  --claim "Claims a new coding benchmark result." \
  --source-url "https://example.com/model-card"
claim-gate model validate

Expected behavior: the claim is tracked, but it does not become actionable unless the domain rules say the evidence is strong enough.

Domains

Each domain is a self-contained ledger with its own schema, vocabulary, validation rules, and scoring. They share one CLI, one package, and one shared engine.

Domain	claim-gate <domain>	What it gates
infra-cost	claim-gate infra-cost	Public AI infrastructure cost signals (GPU waste, cooling, energy, supply, cloud-margin) — scored by decision impact.
bio	claim-gate bio	Bio-AI claims behind an evidence and safety gate; refuses construction/wet-lab framing, requires public sources and reproduction.
model	claim-gate model	Model benchmark / cost / safety / adoption / medical claims, including physical-AI deployment gates and an event-graph export.

Install

pip install -e .

Use

# list domains
claim-gate --list-domains

# record an infra-cost signal, then validate and export the ledger
claim-gate infra-cost add \
  --title "GPU utilization waste becomes a major AI cost driver" \
  --source-type public_news \
  --source-url https://example.com/gpu-waste \
  --signal-type utilization-waste --cost-driver gpu \
  --summary "Public article reports deployed accelerators sit underutilized." \
  --impact high --status actionable
claim-gate infra-cost validate
claim-gate infra-cost export --output signals.md

# audit a model claim — verify-first until evidence is attached
claim-gate model add \
  --subject ExampleModel --claim-type benchmark \
  --claim "Independent benchmark result." \
  --source-url https://example.com/benchmark
claim-gate model validate

Each domain exposes add, list, validate, and export (model also has import-feed and an event-graph-csv export). Run claim-gate <domain> --help for the full command set.

Architecture

claim_gate/
├── engine.py          # shared primitives every domain sits on (timestamps, JSON ledger I/O)
├── cli.py             # `claim-gate <domain> ...` dispatcher
└── domains/
    ├── infra_cost/    # signal schema + cost vocabulary + impact scoring
    ├── bio/           # claim schema + hazard classes + safety refusal rules
    └── model/         # claim schema + physical-AI gates + feed import + event-graph export

Adding a domain means dropping a module under domains/ and registering its main(argv) entry — the CLI, packaging, and CI come for free.

License

MIT — see LICENSE.