The build/review loop

AI Factory^™

AI Factory is a build-and-review loop: a Prime Builder proposes a change, a Loyal Opposition reviews it, and a deterministic gate arbitrates the two.

What it isAn adversarial build-and-review loop

RolesPrime Builder · Loyal Opposition

Arbitrated byA deterministic, non-LLM gate

StageWorking prototype

The outcome

Before

Building a governed agent system means one model produces a change and a human reviews every one for correctness and safety, or, worse, nothing ships on more than a single model's say-so.

After

A Prime Builder's change is reviewed by a Loyal Opposition and arbitrated by a deterministic gate. Measured on 16 September 2026, cross-vendor review named 13 of 18 known defects (7 fully) and named something false in 14 of 18.

AI Factory is how the stack builds more of itself safely: governed systems generated and checked under the same identities and audit trail everything else uses.

How it works

The idea is that no change is trusted on one model's word. A builder proposes, an opponent tries to break it, and a deterministic gate decides.

ROLE 01

Prime Builder

Produces the change, the artefact under review, bound by a versioned role constitution that says what "good" means for this task.

ROLE 02

Loyal Opposition

An independent reviewer whose job is to find what's wrong: correctness, safety, and adherence to the same constitution, argued explicitly.

THE GATE

Deterministic arbitration

A pure, non-LLM gate arbitrates the two into one outcome, GO, REVISE, NO‑GO, or HUMAN‑ESCALATED. Same inputs, same decision, every time.

Why this shape The whole point of the stack is verifiable trust, so the tool that builds on it holds itself to the same bar. Every handoff is a VAID-signed artefact on a hash-linked chain: you can see which role produced what, and why the gate decided as it did. What that chain catches and what it does not is published in full, including one attack it does not yet address at all.

Proof

The gate is deterministic and pure, no model in the arbitration path, so the same builder output and the same review always produce the same decision. That's what makes the record trustworthy.

Recorded demo · the real build → review → gate loop

Autoplay

Watch a real build, review, gate cycle run

A recorded Prime Builder / Loyal Opposition campaign, arbitrated by the deterministic gate, start to finish.

~1.9 MB · loads on tap

Prefer a guided version? Walk through it in presenter mode →

Example · a gate decision

build → review → gate

prime_builder   → change: add rate-limit to /mint
loyal_opposition→ MAJOR: no test for the burst case
                  recommended_decision: REVISE
────────────────────────────────
gate (deterministic) → REVISE
  handoff: VAID-signed, on the audit chain

Two AI roles and one deterministic gate. An illustration of the loop, not a captured run.

The other half of the evidence A deterministic gate decides nothing if the record of what it decided can be edited afterwards. So we attacked our own audit chain in the production database of a deployed service and published all four results, and, more usefully, the finding that two of those attacks could not have been run the week before, because an earlier passing attack had certified the columns they target by association. Truncating entries from the tail is not tested in either direction and is not addressed. All of it is on the audit chain, attacked.

Demo

AI Factory demo, replay of a real run recorded 2026-10-07. Real terminal, real build submitted, real verdict.

SYNTHERA is the trust layer for multi-agent systems: every agent gets a verifiable identity, scoped authority and a tamper-evident record, so software from different teams, vendors and frameworks can act on each other’s behalf without custom glue between every pair.

Discuss a technical evaluation.

Discuss an evaluation