AI agents can recommend APIs, models, or services based on factual claims. Some of those claims are machine-checkable, such as a provider, service, pricing unit, and price. This project is a narrow v0.1 proof harness for checking one kind of claim before trusting the recommendation.

Core thesis: do not optimize agents. Verify the claims agents rely on.

v0.1 verifies agent-made pricing claims against a local canonical pricing fixture. It does not prove live commerce verification. It does not scrape the web, call paid APIs, call LLMs, or decide whether a recommendation is good. It only asks: does this claimed price match the local source-of-truth fixture?

• No live scraping
• No network calls
• No paid APIs
• No LLM calls
• No dashboards
• No SaaS framing
• No AgentGate integration in v0.1
• No product checkout or delivery verification
• No generalized truth engine

npm install
npm run demo
npm test
npm run typecheck

To verify one claim file against one source fixture:

npm run verify -- examples/passing-claim.json examples/source-pricing.json

Verification results are appended to:

.agent-reality-check/log.jsonl

Each JSONL row includes previous_hash and row_hash, so edits to prior rows break the local hash chain.

Passing claim:

MATCH claim-pass-openai-gpt-4-1-mini-input
Claim matches source: OpenAI/gpt-4.1-mini 1M input tokens is 0.4 USD.
claimed_price: 0.4
source_price: 0.4
delta: 0

Failing claim:

MISMATCH claim-fail-openai-gpt-4-1-mini-input
Claim mismatch: claimed 2 USD for OpenAI/gpt-4.1-mini 1M input tokens, source is 0.4 USD.
claimed_price: 2
source_price: 0.4
delta: 1.6

Agent claim JSON includes:

• claim_id
• agent_name
• intent
• provider
• service
• pricing_unit
• claimed_price
• claimed_currency
• evidence_url optional string
• captured_at ISO timestamp

Source pricing fixture JSON includes:

• source_id
• source_name
• retrieved_at ISO timestamp
• prices array with provider, service, pricing_unit, price, and currency

Verifier result JSON includes:

• result_id
• claim_id
• status: match, mismatch, or missing_source
• claimed_price
• source_price when a source row exists
• delta when a source row exists
• human_summary
• verified_at ISO timestamp

Agents may make recommendations based on claims. Before trusting a recommendation, a buyer, merchant, auditor, or marketplace can verify the claims against source truth. This v0.1 keeps the proof small: one local pricing claim, one local source fixture, one deterministic verifier, and an append-only log.

• Local fixture only
• No live scraping
• No product checkout
• No delivery verification
• No legal or financial advice
• No generalized truth engine

These are possible later extensions, not part of v0.1:

• Live pricing adapter
• Docs/spec claim adapter
• AgentGate bond/slash adapter

npm test
npm run typecheck
npm run demo