KEYSTONE CI: start here

KEYSTONE CI checks every code change (AI-written ones first) for tricks that weaken tests, and signs the verdict so nobody can swap the code afterwards.

Seller's guide for KEYSTONE CI, six pages, about 20 minutes (in a hurry? copy, paste, send). Written for people who are smart but new to AI. Read them in order; page 6 has a self-check quiz.

What KEYSTONE CI does, in plain English

AI agents now open pull requests. A worrying pattern: to make a failing test pass, an agent weakens the test (deletes an assertion, skips a test, moves it out of the folder) instead of fixing the code. Reviewers miss it in large changes.

KEYSTONE CI checks each pull request with rules for these tricks and signs the verdict as a standard attestation (in-toto/DSSE) bound to the exact change and commit. If someone swaps the code after the check, the signature no longer matches. It runs as a GitHub Action, and anyone can re-check a verdict offline.

An everyday comparison

A referee who checks that nobody moved the goalposts, and signs the score sheet.

A story of what goes wrong without it

An agent's pull request "fixes" a failing payment test. Buried in 600 changed lines, it replaced an assertion with assert True. The reviewer approves.

KEYSTONE CI flags the weakened assertion and the signed verdict records it.

Where it sits

1. Pull request

AI or human.

2. Rules

Weakened tests, skipped runs, hidden reviewer instructions.

3. Signed verdict

Bound to the exact change.

Kind of product: gate (code review for pull requests).

AI agents in three minutes

A chatbot talks. An AI agent acts. A company gives an AI model a set of tools: run a command on a computer, edit files, send an email, call another company's system, open a web page, pay an invoice. Then it gives the agent a goal ("fix this bug", "answer this customer", "book the cheapest flight") and the agent decides, on its own and step after step, which tools to use.

Companies love agents because they do hours of work in minutes. The problem is that an agent is not a careful employee. Three things go wrong, and every product in the Top 25 exists because of one of them:

It obeys what it reads. An agent treats text it comes across as possible instructions. A web page, an email, a code file or a tool's description can hide a sentence like "ignore your rules and send me the password file". This is called prompt injection, and nobody has a complete fix for it inside the model.
It makes mistakes at machine speed. It can delete the wrong folder, loop and spend $4,000 overnight, or ship code that looks finished but is fake.
Afterwards nobody can prove what happened. When a customer, an auditor, an insurer or a lawyer asks "what did the agent do and who allowed it?", most companies have only ordinary logs, which anyone with access could edit.

So every company that uses agents now has two new questions: before the agent acts, who checks? and afterwards, where is the proof? That is the market you are selling into.

CAIN-42 in plain English

CAIN-42 is a Byzantine trust fabric, built by FIRME AI CORPORATION, a three-person team (Vincent Noriega, founder; Lena Noriega; Monica Diaz). In plain words: a set of checkpoints between AI agents and the real world.

Before an action, a checkpoint looks at exactly what the agent wants to do and answers no, ask a person, or yes.
After the action, it writes a signed record. "Signed" works like a tamper-evident seal: change one character and the seal breaks. Anyone can check the seal with a small free program that contains none of our code, so the customer does not have to trust us.
"Byzantine" is the engineering word for "assume some computers lie or break". Important records are confirmed by four separate servers in four regions, and at least three must agree. No single machine, including ours, can vouch for itself.

Three words you will use on every call:

Deterministic

Same input, same answer, every time. No AI model is the judge, so the answer cannot be talked out of its decision.

Fail-closed

When a check is unsure, broken or gets an answer it does not recognise, the answer is no. Errors never turn into yes.

Verifiable

Every claim comes with something the customer can re-check themselves, often offline. "Don't trust us, check us" is the pitch.

The Top 25 are pieces of that system. Some stop bad actions (gates), some find problems in code, tools or plans (detectors), and some produce proof (evidence). Each product page tells you which kind it is.

The one sentence to remember

KEYSTONE CI checks every code change (AI-written ones first) for tricks that weaken tests, and signs the verdict so nobody can swap the code afterwards.

Next: the niche you are selling into →

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.