01 / How it works

Reasoning, with a structure around it.

A model proposes. Evidence supports or contradicts it. The runtime decides what can proceed—and what can be claimed.

A request enters Merlin's plan, analyze, execute and verify workflow. Evidence and context, model routing, and tools and workspace connect to the orchestrator. Model routing connects to a GPU inference fleet and optional cloud fallback.
Conceptual architecture · not every request follows every stage. GPU symbols are representative.View full size ↗
01

Plan the work.

Name the steps, their dependencies and the evidence they require. Simple requests can take a shorter path; the architecture describes the system, not a mandatory journey for every question.

Spec 001 · Runtime and planning ↗
02

Gather the evidence.

Read the relevant files and sources. Record what was observed so later claims can refer to it. A previous conclusion is context; it does not replace a fresh observation.

AC-005-001-03 ↗
03

Challenge the finding.

Analysis separates a candidate finding from an established one. Rebuttal records whether it survived, was refuted, or remains inconclusive.

AC-006-005-01 ↗
04

Verify before committing.

Changes are staged. Reads and verification operate on the staged content where applicable; the workspace changes at commit.

AC-003-002-03 · AC-003-002-04 ↗
05

Authorize the answer.

The final decision reads task and operation summaries, the request contract, and the proposed answer. Deterministic guards can veto a success claim even after model authorization.

AC-004-004-13 ↗

The architecture

Each layer has its own responsibility.

Models provide reasoning. The runtime coordinates evidence, tools and decisions. Configured providers supply inference; they do not own permission to change the workspace or claim completion.

See the behavior

Read an actual decision.