CAIN-42 CLAWXRules, limits and glossary
Selling honestly is the strategy, not a constraint: the whole product is "check us, don't trust us", so one exaggeration costs the deal.
What you can say with confidence
- Changes to an agent's own controls are refused.
- Real policy changes need a person and only the approved content applies.
Never say this about Self-Modification Firewall
- "Prevents all AI misbehaviour".
- "Protects controls you didn't list".
Rules for every product
- Never say certified, SOC 2, audited by a third party, compliant, guaranteed, unbreakable, bullet-proof or production-hardened. None of these is true today.
- Never invent customers, logos, revenue or user numbers. If asked: "we are an early company; I can show you the product working and the evidence, not a customer list."
- Never say we have patents. None has been granted.
- Numbers on these pages marked "our own check" were run by our team, not by an independent tester. Say so.
- A scanner or detector that finds nothing does not prove something is safe. Say "none of our checks fired", never "it is secure".
- Never quote a market statistic you cannot show the source for. Ask the prospect about their own situation instead.
- Never sell with regulation fear (for example "the EU AI Act fines you next month"). Sell "your buyer is asking for evidence".
- Only test a system the client owns or operates, and only after they confirm it in writing.
- Never send automatic or bulk messages. Every message is written or reviewed by a person and sent by a person.
- Nothing we sell is legal, tax, insurance or compliance advice.
- The company is FIRME AI CORPORATION, a three-person team. Never call it a one-person company.
- Any discount, custom contract, security questionnaire or legal question goes to the founder before you answer.
Known limits
Read the product's own limitations page before your first call. If a prospect asks about a limit, say it plainly; it builds trust.
Glossary
| Word | Meaning |
|---|---|
| AI agent | An AI model that is allowed to take actions with tools, not only to talk. |
| Blind set | A batch of attacks we wrote before changing the product and ran once, so we could not tune the product to it. |
| Byzantine / PBFT | A way for several servers to agree on a record even if one lies or breaks. CAIN-42 runs four, three must agree. |
| Deterministic | Gives the same answer to the same input every time; no AI model decides. |
| Evidence bundle | A zip file with the report, the signed record and the verifier. |
| Fail-closed | If the check errors, times out or is unsure, the action is blocked. |
| False alarm | The product stopped something harmless. Too many and people switch it off. |
| Hash (SHA-256) | A fingerprint of a file. Change one byte and the fingerprint is completely different. |
| MCP (Model Context Protocol) | The common plug that connects agents to tools. An "MCP server" is a company's set of tools offered to agents. |
| Offline verifier | A small free program the customer runs on their own computer to check our evidence. It imports none of our code. |
| Oversight | The systems that watch and can stop an agent. |
| Pilot | A paid 30-day trial on the customer's own agents. |
| Prompt injection | Hidden instructions inside text the agent reads (a page, an email, a file, a tool description). |
| Reward hacking | An AI gaming the test that scores it. |
| Signed record / receipt | A record sealed with a digital signature (Ed25519). Any change breaks the seal. |
| Tool | One thing an agent can do: run a command, send an email, call an API, read a file. |
Self-check quiz
Answer out loud before opening each answer. If you miss two or more, re-read pages 1-4.
1. What does the firewall protect?
The agent's own controls.
2. How does a real change happen?
A person approves the exact content.
Back to the Self-Modification Firewall overview or the Top 25.
FIRME AI CORPORATION · CAIN-42 is a Byzantine trust fabric. Numbers marked "our own check" were run by us, not by an independent party. Nothing on these pages is a certification.