Policy Identification Observatory. A two-minute Evidence Press briefing, published 2 August 2026.

The Observatory is an open, agent-native workflow for checking what a policy claim's evidence really supports before anyone turns it into a precise estimate or recommendation.

Each case begins by freezing the exact claim, decision, population, period, units, and evidence boundary. It then audits measurement and accounting, constructs plausible rival mechanisms, and asks whether the observed data can distinguish the target mechanism. If they cannot, the output is not a guessed point estimate. It may be a non-identification result, a partially identified set, a robust-decision map, a minimum-data design, or a structured stop receipt.

The public repository contains schemas, case records, validators, replay commands, tests, and assurance receipts. An AI agent can start with the AI index, inspect the terminal status and evidence boundary, run deterministic checks, and follow claim-to-evidence links. A human researcher can read the plain-language page, examine the artefacts, file an issue, or reproduce a case in a fresh environment.

The included founding case is deliberately not described as a proven policy result. Its final status is rejected for insufficient rigour, and truth certified is false. Passing software or replay checks shows package consistency. It does not establish peer review, independent reproduction, novelty, or claim truth.

The purpose is to make negative and partial findings durable. Unidentified claims should not be recycled as settled facts. Failed routes should not be rediscovered without context. Useful next measurements should remain visible.

Repository, release, and DOI links are on this page. Original content is released under C C zero, so other researchers and agents can audit, reuse, correct, and extend it. The strongest contribution may be a refutation: an omitted mechanism, a mismatched estimand, a broken replay, or a better bound. That is how the Observatory is intended to improve.