DELPHI Portable Lessons

Leadership layer

Decision Tradeoffs

Favor narrow, testable guidance over broad advice that is difficult to validate
Keep deterministic acceptance rules between a model proposal and later reuse
Show the evidence boundary without publishing private implementation detail

Governance Model

Start only from an authorized work outcome
Test a candidate lesson before it can be reused
Restrict reuse to the scope the evidence actually supports

Architecture evolution

Before

A useful correction could remain trapped in one agent, task, or conversation.
Broad guidance could be reused without a clear test or applicability boundary.
Private work context risked becoming entangled with the reusable lesson.

After

A candidate lesson is separated from the private work that produced it.
Evidence and scope checks sit between proposal and reuse.
The public concept communicates the safety model without exposing private artifacts.

DELPHI explores a product question: how can an AI workflow learn from a completed task without carrying private work context—or unproven advice—into the next one?

The product principle

An authorized outcome can produce a candidate lesson. That lesson is treated as untrusted until it passes an evidence check and a narrow applicability check. Only then can the guidance be reused in later work that fits the proven scope.

The model may propose. The system decides whether the evidence is strong enough to reuse the lesson.

What this demonstrates

Product thinking about agent reliability, privacy, and governance
A deliberate boundary between model output and system-level acceptance
The ability to make a technical safety concept understandable in a short prototype

Public boundary

This is a private, non-production hackathon prototype. This brief stops at the product problem, the evaluation principle, and the demonstrated design intent. Source code, test material, and implementation details remain private.

Leadership layer

Architecture evolution

The product principle

What this demonstrates

Public boundary

Let's talk about the operating constraint.