Gate claims in page copy

The claims gate asks whether each rendered section violates a policy you write, such as promising a capability your facts do not support. Answers are committed to the repository and replayed at build time without a model or credentials. This guide configures the gate and runs pagegraph claims check.

The gate needs rendered Markdown capture. Set that up first with Publish Markdown twins and llms.txt.

Declare rules

Declare probability rules against claimsInput from the build-only pagegraph/claims entry. Each rule asks whether the section violates your policy:

import { Decision } from "effect/ai";
import { claimsInput } from "pagegraph/claims";
import { pagegraph } from "pagegraph/tanstack-start";

const rules = Decision.make({
  input: claimsInput,
  decisions: {
    unsupportedPromise: Decision.probability({
      instructions: "Does this section promise a capability unsupported by its facts or evidence?",
    }),
  },
});

const graph = pagegraph({
  origin: "https://example.com",
  markdown: { origin: "https://example.com" },
  facts: "src/facts.ts",
  claims: { rules, model: "typesafe/jev", cutoff: 0.8 },
});

Pass graph.prerenderPages to Start as shown in the rendered Markdown setup. The build persists captured documents and graph heads, then replays committed answers without a model or credentials. Missing answers and violations fail the build and name the page, section and rule.

Ask and commit answers

Run pagegraph claims check after changing prose, facts, context or rules, then review and commit .pagegraph/decisions/claims/. The cache key includes the rule definition fingerprint, model and input hash. Cached answers are reused, and --refresh asks again and replaces them. The command reports cached and asked counts and exits nonzero on violations. Set OPENROUTER_API_KEY (or TYPESAFE_API_KEY) when answers need asking. pagegraph check also replays configured claims without model calls.

pagegraph claims check
pagegraph claims check --refresh
pagegraph claims check --dev http://localhost:3000

Supply context and set the cutoff

claims.context supplies policy text to every input. Inputs include the section's authored Markdown, messages, all code-owned facts, and other sections marked as evidence. Page metadata is judged as a head section, including graph pages outside rendered capture. excludeHeads accepts path globs to omit metadata while keeping captured sections in the gate.

The default cutoff is 0.8 only when omitted. Any probability at or above the configured cutoff is a violation, and values below it pass. Changing the cutoff reuses answers and recalculates verdicts.

Choose the model transport

Jev uses OPENROUTER_API_KEY when set, sending bearer-authenticated requests to https://openrouter.ai/api/v1/systemone. Otherwise it uses TYPESAFE_API_KEY and the TypeSafe endpoint. A failed OpenRouter request does not fall back to TypeSafe. Claims cache keys retain the configured logical model string, so changing credentials or transport does not invalidate committed answers. See Run agent workflows for the attribution headers.