Home › Specifications
Prove it · step 4 of the loop · open standards and research
Specifications
The interesting failures are almost never in the model or the engine. They are in what surrounds it: the context assembled wrongly, the retry that charged a customer twice, the file that arrived twice and doubled revenue. That part can be specified, built and tested by ordinary means. These specifications are published in the open, name no product, and are versioned and machine-readable.
Each family asks the same three questions: what must exist, what must be true, and what must the system face before it meets real people or real data. The working code built to these specs is under Assets.
Clean AI Engineering: for AI agents
Published and in use: one reference agent passes all four gates.
AI Assurance Catalog
AAC · what must be TRUE
Test obligations for AI applications, by architecture type, with crosswalks to external frameworks.
AI Harness Catalog
AHC · what must EXIST
Harness capabilities across sixteen layers and ten archetypes. Observability and evals are two of sixteen layers, and that is the point.
AgentTwin
what must be FACED
Keeps the agent real and simulates its world: customers, systems, data, approvers, faults and time. Then judges what changed, not what was said.
Reference Agent
all of it, built
A customer-support agent with no agent framework: one module per harness layer, exercised against AgentTwin worlds.
The agent guardrail catalog
57 guardrails
Must or should have, rule, classifier or model judge, where each sits, and the test that proves it.
The AgentTwin Lab
in progress
An agent runs for weeks inside a simulated business while an AI fixer repairs it. Does the fixer quietly weaken the checks that caught it?
Clean Data Engineering: for data platforms
The same three questions asked of a data platform. Designed; being written spec by spec, with the learning studio already built.
DataTwin
what must be FACED planned
Keeps the data platform real and simulates the business that feeds it, with planted incidents. Every simulated day, the dashboard is checked against the world's truth.
Data Platform Catalog
DPC · what must EXIST planned
The spine (sources to experiences), the 22 cross-cutting concerns and the port each is reached through. The sibling of the AI Harness Catalog.
Data Assurance Catalog
DAC · what must be TRUE planned
Checkable obligations such as 'every silver row traces back to a bronze row', citing DAMA and ODCS instead of restating them.
Code under Apache 2.0, specification prose under CC BY 4.0. Free to use, cite and disagree with.