Cortex Loop
A loop should not be able to finish by asserting that it finished. Here is one that cannot.
The problem: an AI agent can always say it finished. Saying it is free. That is why so much "progress" evaporates on inspection.
What this page proves: a loop that refuses to mark work complete until the evidence matches. Pick a tick, and watch it succeed on three items and refuse on two - because the evidence was not there.
No model call, no server: the whole state machine runs in your browser over a synthetic ledger.
Note: items edit-2044 and edit-2045 are designed to fail their check. The loop will refuse them. That refusal is the feature.
| item | intent | state |
|---|
Why it matters: once completion requires evidence, "done" stops being a claim and starts being a check result. That is the difference between an agent that reports progress and one that produces it.
The loop engine is open source: cortex-loop and the harness it drives: cortex-harness-open. Example targets come from a public storefront, fit finder and today's deal. All ledger rows here are synthetic. No customer records, contact details, hostnames, or internal identifiers are used.