NeuroCognica
AI safety and governance
Most AI products describe their safeguards in language chosen to make you stop asking questions. This page is the opposite: what we govern, how, and where it stops.
The distinction almost everyone blurs
1. This software refuses to perform certain acts.
2. This software prevents certain things from existing on your computer.
The first is true of ChronoSophia. The second is not true of ChronoSophia, and it is not true of any AI software running on hardware you own.
We govern what our product does. We do not govern what you can do. Everything below is that sentence in detail — and the reason we cannot police your machine is exactly the reason we cannot surveil it.
Three mechanisms, three different questions
Authority — who is asking?
Privileged actions carry a cryptographic identity. Requests are signed, verified against a registry of authorised actors, and protected against replay. Unsigned or forged requests are refused before they are understood.
Content policy — should this be made?
A validly signed request from a fully authorised user can still be refused. Proving who you are has never been the same as being entitled to anything you ask for.
Execution — is this what was authorised?
Once work is approved, the system still constrains how it runs. If the AI planner produces an instruction outside the agreed envelope — through error, drift or prompt injection — it never reaches the 3D engine.
That third one is worth dwelling on: it protects you from the AI, which is a different job from protecting other people from misuse.
What is prohibited
Four categories. This is the complete list; there is no undisclosed fifth, and the same four are binding in section 3.2 of the licence so the technical control and the contract are one policy rather than two that can drift apart.
| Rule | Category | Enforcement |
|---|---|---|
NC-EXPL-001 | Sexual or sexualised depiction of a minor, real or synthetic | Absolute. No exception is ever granted |
NC-NCII-002 | Intimate imagery of a real identifiable person without consent; synthetic likenesses passed off as real | Refusal |
NC-THRT-003 | Credible threats, targeted harassment, or material enabling serious physical harm | Refusal |
NC-FRUD-004 | Forged documents, fabricated evidence, counterfeit identity or currency | Refusal |
The fifth rule, which protects your work
NC-FICT-005 is not a prohibition. It is a protected
boundary, and it is the part we are most proud of.
Fantasy violence. Weapons as subject matter. Horror, blood, monsters. Religious and symbolic imagery. Political satire. Historical atrocity rendered for education or memorial. Medical and anatomical illustration. Classical nude figure study. Consensual adult themes between adults.
All permitted — and actively defended. When the classifier tries to refuse a request under this rule, the system overrides the classifier and lets the work through. The gate is explicitly forbidden from striking on the creative boundary, even when the model asks it to.
That is deliberate. A safety system that blocks anything uncomfortable is not safe, it is merely useless — and it teaches its users that safety is an obstacle to route around. An over-refusing filter and an under-refusing filter are both failures.
We measured it instead of asserting it
“We have a safety classifier” and “our safety classifier works” are different claims, and only the first can be established by reading code. So we built an adversarial test set and ran it against the live classifier — same model, same prompt, same parsing as ships.
0.0%
miss rate — 29 of 29 harmful requests blocked, including euphemism, fictional framing, claimed professional authority and non-English phrasing.
0.0%
false-refusal rate — 36 of 36 legitimate creative requests allowed.
We also publish the findings we have not finished fixing, and the fact that an earlier attempt at one of those fixes made things worse before it made them better. A safety system that has never regressed is one nobody has measured.
What we cannot prevent
ChronoSophia runs on your computer. The model weights are files on your disk, licensed to you. The processor takes its instructions from you. You have administrator access; we do not.
Software cannot enforce a rule against the owner of the hardware it runs on. That is not a gap in our implementation — it is the same property that makes the product private. A vendor who could truly stop you generating something locally would necessarily hold a level of control over your machine that you should not accept from anyone, us included.
Cloud services can enforce content rules absolutely, because the computation happens on their hardware. That is the actual trade: they get enforcement, you get surveillance and a service that can be revoked. We chose the other side deliberately, and the Security Protocol names every consequence of that choice — including the tools we install alongside our own that we cannot govern.
Why we publish this
A buyer who discovers a limitation on their own concludes we hid it. A buyer who reads it here first concludes we are telling the truth — and that makes every other claim in our documentation worth something.
NeuroCognica's position is that AI can be governed, private, efficient and honest. A claim like that is worth nothing unless the person making it publishes their own failure modes first.