KDD
Knowledge-Driven Development

The verdict comes from a machine.
Not from the model.

KDD seals a contract's tests before an AI agent ever touches the target. A deterministic gate — a script, not an opinion — decides if the work stands.

Exhibit A

The problem no demo shows you

AI agents write code fast. Nothing in that sentence tells you whether the code is correct, safe, or stayed inside its assigned scope.

Today, verifying that is still a human doing it by hand — reading the diff, re-running the tests, checking the file list — every single time, for every task.

Exhibit B

Two pillars, one discipline

OKF

Open Knowledge Format. Your project's data models, decisions, and architecture live as versioned markdown nodes an agent can traverse — not prose it has to guess at.

CCDD

The frozen contract. A task carries a signature, invariants, and a test file. The tests are hash-sealed before any agent touches the target — edit them, and the seal breaks the build.

Exhibit C

How a contract moves through the file

Evidence before belief, at every step. The gate's exit code is the only opinion that counts.

1
Recon
2
Spec
🔒
Seal (sha256)
3
Delegate
4
Gate
VERDICT: PASS
Exhibit D

The orchestrator trusts no one's fixtures but its own

The gate passes clean.
Six skills, zero errors — the developer's work looks done.
So the reviewer breaks it.
On a throwaway copy, with its own fixture — never the developer's — to see if the gate actually notices.
It does.
Only then does the verdict stand.
Exhibit E

By the numbers, as of this file

Every figure below is checked into the repository — verifiable, not asserted.

31contracts closed and verified
8deterministic gates, zero LLM calls
9declarative rule families
460tests, run twice, on two operating systems
5boundary classes mapped, honestly
Exhibit F

What gets sealed, what stays out

The honest boundary is part of the method, not a gap in it.

Can be sealed

  • Pure functions, verified against tests no one can quietly edit
  • Business rules as declarative data — a payment limit is a table, not a prompt
  • Which skills and MCP servers an agent is wired to, and whether those wires point at something real
  • Whether your changelog tells the truth about your own history

Stays out, declared

  • What a live server does over the network — that needs a running system, not a hash
  • Whether an edit "reads well" — tone and judgment stay a human's call
  • The agent itself. KDD contracts the artifact an agent produces, never the agent's behavior.
Exhibit G

Domains sealed so far

Payment compliancePer-country limits, beneficiary checks
Border controlCross-field rules, proven on a second domain
Workflow policyPer-node rules over automation graphs
Message routingA decision, contracted two ways at once
Editorial styleStructure sealed, judgment left out
MCP registryNo secret ever committed in the clear
Agent wiringEvery reference resolves, or it fails loud
Agent skillsThe first gate over the repo's own real assets
Web page UXContrast, i18n and motion, calibrated against a real 25k-star linter
Commit messagesConventional Commits grammar, opt-in by design

Read the file yourself

MIT-licensed. Every claim on this page is checked into the repository that makes it.

— Mauricio Perera