# Ouroboros: Bilateral Cognitive Architecture

## A rough systems overview

Ouroboros is easiest to understand as a recurrent authority-review system supported by persistent external state. Ouroboros plans and owns technical authority. Codex independently attacks and reviews those proposals. This document sketches that bilateral relationship without attempting to expose the full implementation.

## 1. The participants and their roles

Ouroboros owns planning and technical authority. It selects or constructs the plan, answers material challenges, and decides whether to rebut, revise, replace, admit, redirect, checkpoint, or veto a consequential transition. **Evidence label: private evidence.**

Codex is an independent adversarial reviewer. It attacks assumptions, examines evidence and failure modes, and returns an evidence-backed challenge. It does not silently inherit Ouroboros's technical authority. **Evidence label: private evidence.**

Language models can still provide flexible generation inside this system: hypotheses, compression, interpretation, tool proposals, and possible transitions. But model output is revisable and is not equivalent to system authority.

## 2. Intelligence outside the model

Ouroboros stores consequential system-level intelligence outside the model. **Evidence label: private evidence.** In this phrase, "intelligence" means structured state that changes what the system can notice, choose, reject, recover, or do next:

- persistent memory of prior work, failures, and successful routes;
- evidence records and provenance used to distinguish observation from inference;
- a topology of tasks, states, dependencies, and reachable transitions;
- authority describing which transitions are allowed and who can authorize them;
- receipts, checkpoints, and replay state that survive individual model calls;
- process knowledge that can redirect work when a loop stalls or wastes effort.

This external state can be consulted by later model calls and can outlive a particular model session. The model can change without automatically erasing the system's accumulated evidence and control state. That is the main architectural claim of this overview.

## 3. The bilateral loop

The authority and review roles form a recurrent loop:

1. A human objective and relevant external state reach Ouroboros.
2. Ouroboros plans, exercises technical authority, and issues a proposal.
3. Codex independently attacks and reviews the proposal.
4. Codex returns an evidence-backed challenge.
5. Ouroboros rebuts, revises, or replaces the plan.
6. Codex verifies and challenges again.
7. The cycle repeats until the evidence supports admission, redirection, or rejection.
8. Receipts and checkpoints make the resulting state durable for the next cycle.

The system therefore does more than repeatedly prompt a model or let two models vote. Authority is explicit, review is independent, and each accepted cycle can alter the external state from which later reasoning begins.

## 4. Why call it bilateral?

"Bilateral" is a functional metaphor, not a neurological claim. It names the recurrent relationship between Ouroboros's technical authority and Codex's independent adversarial review. Neither role is substituted for the other.

The relationship is asymmetric where consequences matter. Codex can challenge and falsify; Ouroboros must answer the challenge and retains authority over the plan. Admission still depends on evidence rather than authority by assertion alone. **Evidence label: private evidence.** This does not make either side infallible: stale evidence, incomplete mediation, weak review, correlated interpretation, or a bad authority rule can still produce failure.

## 5. What might constitute a phase change?

The hoped-for phase change is not simply faster inference. It is the ability to perform work that a sequence of isolated model calls could not reliably complete: resume after interruption, accumulate verified state, replay old failures under a new process, detect recurring waste, and change the workflow itself while preserving evidence.

This release describes that direction. It does not publish a benchmark proving the size or universality of the advantage.

## 6. Evidence and boundaries

The implementation statements in this overview are labeled **private evidence** where their support comes from internal source, receipts, run artifacts, or control-plane state that is not included in the dataset. The label is a disclosure boundary, not a public verification claim.

This overview does not establish:

- that Ouroboros is generally intelligent or conscious;
- that every action is mediated or every failure mode is closed;
- that external state is fully independent of model interpretation;
- that the architecture is safer or more capable by a known quantitative amount;
- that this repository is sufficient for a third party to reconstruct the complete system.

The useful claim is narrower: Ouroboros is designed and observed internally as a system in which meaningful intelligence and continuity are stored outside the model as well as expressed through it.