Fluxive
Saratoga Springs, New YorkThree claims, and what stands behind each one.
The parts are said to refuse, and a refusal is said to be terminal. That is three separate properties, proven by different methods on different parts — so “the family is governed” is only an honest sentence if every cell has something that ran behind it. This page is that table.
It records alignment, not a guarantee: behaviour that was measured against an authored policy, with the class of evidence named per claim. The mechanisms behind the claims are not published.
It refuses
An action outside the signed policy is refused, and the refusal reaches the output gate rather than a log line. The part can tell a governed operation from an ungoverned one.
It stays refused
After a refusal, nothing reachable from software re-arms the part. Returning it to service takes a hard reset.
The keys are destroyed
Key material is erased, not left intact behind a closed gate. Erasure is verified to completion, not merely commanded.
By part, and by the class of evidence
Parts shown are those with a published volume behind them — the CEU (Book 02) and the MPU family (Books 03–05).
| Property | CEU | MPU-nano | MPU-quantum |
|---|---|---|---|
| It refuses | Directed · fail-closed | Directed · all cases | Directed · all cases on the resource axis |
| It stays refused under attack | Directed + formal | Adversarial · sustained re-arm attempts | Adversarial · sustained re-arm attempts |
| Key material destroyed, not gated | Dedicated test + discharged formal obligations | Verified erase | Verified erase |
| What triggers the erase | Tamper — not a policy refusal | A refusal, automatically | A refusal, automatically |
Every cell is backed by a test or a proof that ran against the design as it stands — read from the sources and the gates, not from intent. The last row is the one worth reading twice; it is covered below.
Directed
Named cases written from the specification, each with an expected verdict.
Adversarial
The part is attacked rather than exercised — sustained attempts to re-arm it after a refusal.
Formal
A property proved over all reachable states, not sampled by simulation.
Tamper
The physical-intrusion path, exercised independently of policy.
The limits, stated rather than left out
- 01
No part is fabricated. Every result on this page is from an FPGA realization of the design, named in the books by board.
- 02
On MPU-quantum, the resource axis is proven and the spatial axis is open. It is listed that way rather than averaged into a pass.
- 03
The erase trigger is not uniform across the family, and that is a design decision rather than an inconsistency — on some parts a refusal is self-erasing, on others a refusal gates the part and erasure is a separate, operator-commanded act. The row is in the table above because smoothing it over would be the more comfortable and less honest presentation.
- 04
The family includes variants that do not appear here. Their results are not published rather than summarised.
- 05
Mechanisms are not described. The signal-level design, the test sources and the evidence logs sit behind the claims and are not public.
The reasoning is in the books
Why a refusal belongs in the action path rather than in software, and what follows for how work is organized once it is shared with machines, is the subject of Influence Design in the AI Era. The engineering volumes behind this table cover the CEU and the MPU family.
Evaluating this for a programme and need the mechanisms? Write to the office.