---
kind: project
name: Long-Horizon Verification
summary: A model for keeping claims, evidence, freshness, and next-action authority explicit across work that outlasts a single agent turn.
maturity: exploratory
period:
  start: 2026-09-06
status: exploring
areas:
  - verification
  - evidence
  - long-horizon-agents
owner: tim
started: 2026-09-06
updated: 2026-09-06
current_question: How should a long-running task represent what is proven, what is stale, and what remains safe to do next?
next: Define a small claim-and-evidence state model without assuming a particular agent framework.
featured: true
on_bench: true
artifacts:
  - label: Verification guide
    href: /guides/concept/verification-is-state
    kind: guide
  - label: Runtime authority guide
    href: /guides/concept/runtime-as-authority-system
    kind: guide
claims: []
open_questions:
  - Which evidence types can authorize a next transition?
  - How should independent verification be represented alongside proposer-generated checks?
  - What freshness rules are meaningful for different claims?
build_log: []
verifiedAgainst: []
proves: Verification is part of the control state of long-horizon work, not merely the final step after an agent produces an answer.
domains:
  - evals
  - agent-systems
  - trust-boundary
runtimes:
  - typescript
  - python
maintainer: tim
---

## Why this exists

Long-running work changes its plan, environment, permissions, assumptions, and
risk. A final pass/fail checkpoint cannot express what is currently proven or
what evidence has gone stale.

## Current model

Treat verification as a graph of claims and evidence. A deployment succeeding,
a service being healthy, and the original user-visible problem being resolved
are separate claims. Each needs evidence with provenance: producer, subject
state, observation time, verifier, and supported claim.

## What is unproven

This is a conceptual investigation. It does not report a live experiment, a
completed implementation, or a verified production result.

## Next experiment

Define a small claim-and-evidence state model with freshness rules, then test
whether it can distinguish stale evidence from evidence attached to an
immutable revision.

## Related

- [Verification is not the last step](/guides/concept/verification-is-state)
- [A runtime for agents is an authority system](/guides/concept/runtime-as-authority-system)
- [Authority Compiler](/work/authority-compiler)