Product · Knowledge Layer
Nobody reads it.
The work asks it.
A working record of what your company knows, has decided and must not do, written by the AI as a side effect of work and read by the AI inside the next job. No page to open. No handover to write. No session that starts cold.
Three kinds of claim, kept apart
Your company runs on three kinds of claim. Most tools store them as one.
Facts
What happened. Moved by a new observation, and nothing else.
Rules
What must hold. Earned from the cases where they held, sunk by the cases where they didn't.
Decisions
What was chosen, by whom, on what date, replacing what. Moved only by another decision.
The Knowledge Layer stores each with its kind, its source, its name and its date. An AI that reads it knows which question to ask of every line: was this observed, was this applied, or was this chosen.
What it does
Know · Decide · Rehearse
Know
The work asks. The answer lands in the task, in the amount the task needs, and nothing is loaded in case. A claim surfaces only when it is asked for or when an outcome contradicts it.
Proof: Nobody Reads It. The Work Asks It. · A Hand-Written Map Can Be Worse Than No Map
Decide
Every decision has a name, a date, and what it replaced. Nothing is deleted; a replaced decision is marked superseded and stays readable. Ask what the company believed in June and it can answer. Ask why it stopped doing something and it can answer that too.
Rules can say no. A plan that breaks a current rule is refused before anyone does the work, and the refusal carries the rule and who decided it.
Proof: A Decision Is Not a Fact · Limitations, Perceived and Real · Chain of Thought You Don't Own
Rehearse
A plan is run against your rules before it runs for real. It fails on paper, with the rule it broke and the step where it broke, or it doesn't fail at all.
Proof: LLM-as-a-Judge: Perfect Scores 92% of the Time · Independent Convergence, the verifier-decides cluster
What it is not
- Not a wiki. Nobody goes to it.
- Not a chatbot. It answers the work, not the person.
- Not a model. It works with whichever model you already use.
- Not a search box over your documents. It returns claims you can act on, not passages you have to interpret.
How it learns
A claim that is merely present earns nothing. A rule earns from the jobs that followed it and scored above their class, and sinks when they scored below. Nobody sets a weight by hand; an operator's judgement enters as an observation and is counted, not imposed. The record interrupts once, on contradiction, with the evidence attached.
Proof: The Pigeon Was Not Superstitious. The Reader Was. · A Green Build Is Not Knowledge · A Wrong Theory in the Feedback Makes the Next Run Worse
Where it is short
The correction loop on our own record is still open: proposals to change our knowledge pages run at volume, and the part that measures whether a change helped has measured six. Our graph is young; of 444 links fed under the current rule, thirteen have earned a positive average. We publish both numbers because a record that hid them would be the kind of record this product exists to replace.
How it starts
It is seeded from what you already have: documents, tickets, code, past decisions. The AI distils; nobody re-types. From the first session, the work starts informed and leaves the record richer.
First customer: us. Five products, a small team, knowledge trapped in conversations and one person's head. It is deployed and in use.
The evidence
The learning rule series, in reading order.
- A Hand-Written Map Can Be Worse Than No Map
- The Pigeon Was Not Superstitious. The Reader Was.
- Nobody Types the Weights
- A Green Build Is Not Knowledge
- A Wrong Theory in the Feedback Makes the Next Run Worse
- Limitations, Perceived and Real
- LLM-as-a-Judge: Perfect Scores 92% of the Time
- Don't Build Your Safety on a Chain of Thought You Don't Own
- A Decision Is Not a Fact
- Nobody Reads It. The Work Asks It.
FAQ
Common questions
Is this a wiki with AI on top?
No. A wiki is read by people who go to it. The Knowledge Layer is asked by the work and answers inside the task. Nobody opens it.
Which AI does it need?
Any model that can make a request. The record is the product; the model is yours.
What happens when a rule is wrong?
A job that followed it scores below its class, the rule surfaces with that evidence, and someone re-decides it. A wrong rule in a wiki stays wrong until a person happens to read the page.
Can it refuse a plan?
Yes. A plan that breaks a current rule is refused before any work is done, with the rule and the name of whoever decided it.
Who owns the record?
You do. Self-hosted on your cloud, or hosted by us. Either way the claims, their history and their weights are yours.
Start with what you already have.
Tell us where your decisions live today. We'll show you the record built from them, on your own material.