Case 04 / Gated publisher · safety-first
Nyx Persona
GPT-5.6 drafts replies for a real Threads business account — but only deterministic server code, triggered by one human clicking approve, can ever publish.
Demo
The rule
Approval is cryptographically bound to a sha256 of the final reply body — you cannot approve text A and ship text B.
Build notes
Publish as a lease
Publishing is a claimed lease with expiry and reclaim rules, not a retry loop — double-posting is structurally impossible, not just unlikely.
Unknown over guessed
When the provider's answer is ambiguous, the outcome is recorded as 'unknown' instead of guessed; a confirmed publish is trusted immediately, an absence must age before it's believed.
Wording-only zone
Untrusted social text can't pick evidence, capabilities, or a publish target — the demo ships with a prompt-injection attempt the system escalates instead of obeying.
Hand-built core
A 1,132-line workflow state machine, an 843-line Zod contract layer, a 1,536-line persona compiler. 480 unit tests run in ~10 seconds with no network.
More screens
