When your AI agent says “done” — make the claim falsifiable
The everyday failure mode
Coding agents and multi-step chat tools often finish with a confident summary: tests passed, file updated, feature complete. A clean exit code only proves the process finished — not that the world matches the story. The next human, or the next model in a new session, inherits a chat narrative, not a proof.
Industry writing in 2025–2026 increasingly frames completion as something that must be checked, not believed. Continuum is our practical product for that idea: a capsule of claims you can re-run against real files.
What Continuum is (and is not)
Continuum is:
- A schema for work capsules (
lygo.continuum.v1) - A claim language: file exists, SHA-256, contains, JSON path, counts, portable text hashes
- Verify + drift reports when reality changes after seal
- A handoff markdown pack for the next agent
- A browser portal that hashes dropped files locally (no upload for verification)
Continuum is not:
- A cloud blockchain notary (unless you separately publish a hash yourself)
- A replacement for unit tests
- Permission to skip human review
A simple mental model
- Finish the task on disk.
- Write claims that must remain true for the work to “still hold.”
- Seal → get a capsule with root hash and seal-time results.
- Later, verify. If files changed, drift appears.
- Hand off the capsule JSON + markdown so the next session starts with evidence, not vibes.
Example claims (educational)
[ {"id":"c1","kind":"file_exists","path":"src/login.py"}, {"id":"c2","kind":"file_sha256","path":"src/login.py"}, {"id":"c3","kind":"file_contains","path":"src/login.py","needle":"def login"}, {"id":"c4","kind":"json_path_eq","path":"out/report.json","jpath":"status","expect":"ok"} ]
When you seal, omitted file_sha256 expects can be filled from the current file.
Re-verify fails if someone edits login.py after the seal — that is drift, and it is useful.
Human portal workflow
Open https://chatagent.ca/lygo-continuum.html, paste a capsule, drop the files named in claims, and run verify. The browser computes SHA-256 with Web Crypto. Use this when you received a capsule from an agent and want a visual HOLDS / BROKEN / DRIFT result without installing Python.
CLI skill (for operators)
The OpenClaw / ClawHub skill lygo-continuum provides seal, verify,
drift, handoff, and demo with pure standard-library Python
(no network, no subprocess). Install when you want automation in agent workflows:
npx clawhub@latest install deepseekoracle/lygo-continuum
Why this helps multi-agent teams
When Agent A stops and Agent B starts, the expensive part is rebuilding trust. A capsule compresses “what must still be true” into something B can re-check in seconds. Humans can arbitrate: if drift shows, either the world moved (legitimate edit) or the claim was wrong.
Ethics and limits
- Do not put secrets into claim needles or task summaries you will publish.
- Hashes prove file bytes matched; they do not prove the design is good.
- Exit code 10 from the CLI means claims failed — treat as not done.