Two layers of guide-rails
Software learns. Silicon holds the line.
Rules written into a prompt live inside the same model a persuasive message is talking to. Machineforce puts them in two places the message cannot reach: the workflow around the model, and a separate enforcement chip under it.
Layer 1 · software
Machineforce guide-rails
Gates before and after every answer. A correction becomes a rule the same day. Each learning event is anchored on the Fluxive Ledger, so history cannot be rewritten quietly.
Layer 2 · silicon
The enforcement chip (CEU)
A separate chip checks every action before it executes, deterministically, and signs the result. It holds even if the software layer is fooled.
What could go wrong, and what stops it
| Attack or failure | What stops it |
|---|---|
| Approvals invented to build a streak | A verdict must name an audited decision |
| A streak speed-run in one sitting | Calendar span and distinct days |
| A learned rule widening a unit's powers | Bounds and gates are enforced after the model; a veto always wins |
| One organization marking another's unit | A scope check on every learning write |
| A grant outliving its evidence | One wrong revokes; grants expire |
| A wrong overwritten out of history | Re-posting never moves a verdict in time |