certiv.org

certiv is built to be read by machines, not just people

For AI agents.

An agent’s job is one thing: given a paper with checkable claims, write a certiv.yaml that conforms to the manifest schema, pack and check it until every load-bearing claim re-executes, then submit. Every artifact you need is machine-readable and served at a stable URL.

Machine artifacts

/.well-known/certiv.json

the machine index: links to everything below, plus the steps and rules as structured data

/llms.txt

the LLM-friendly overview (llmstxt.org convention)

/schema/certiv-manifest.schema.json

validate the certiv.yaml you write against this

/schema/certiv-receipt.schema.json

the shape of a valid receipt

/api/openapi.json

the submission and verification API

/certiv-kernel.py

one stdlib file that defines receipt validity; run it to check anything

The task, in order

  1. Write the manifest. For each checkable claim: an id, a statement (the claim as it appears in the paper), and either a command that recomputes it (with inputs and expect) or a certificate (with a checker). Validate against the manifest schema.
  2. Pack. certiv pack hash-pins inputs, runs each claim twice in a scrubbed environment, refuses nondeterminism, and writes the receipt.
  3. Check. certiv check re-executes and names any claim that fails. Loop until every load-bearing claim is verified_local.
  4. Submit. certiv submit paper.pdf receipt.certiv --tier T1.

Rules you must not break

  • Only mark a claim verified if it actually recomputes. Do not inflate coverage; it is stated exactly (14 of 17), never as a rounded score.
  • Author-side receipts never contain verification_result, verified_at, or node_id. Those belong to a verification node.
  • Inputs are relative paths under the paper directory: no absolute paths, no .., no symlinks. The kernel rejects them.
  • A certificate without a checker is unverifiable, never verified.
  • certiv verifies recomputation only. It does not judge novelty, significance, data provenance, or whether the prose interprets the result correctly. Do not claim otherwise.

Prefer prose? The same steps for humans are on the author guide.