The Machine

A reference architecture for reliable long-running applied intelligence. The harness, not the model. The prompt is the job ticket; The Machine is the operating system. The model is replaceable; the harness is not. Continuity lives in the artifact, not the session.

This repo is the canonical home of the standard and the tool that scores conformance to it.

• THE-MACHINE.md — the Conformance Spec (vNext, ratified 2026-06-14 by an independent multi-model council + operator signature). The spec is the source of truth; the poster is a derived view.
• kit/ — the conformance kit: an executable scoreboard. Conformance is run, not asserted.
• CONFORMANCE-MATRIX.md — the test matrix, a derived view of kit/obligations.py (single source of truth).

The loop, in one line

Durable state · a dumb deterministic driver · fresh workers · verify against reality (verifier ≠ subject) · autonomy earned at a non-bypassable gate. Six boxes (0–5), each with one falsifiable obligation. See THE-MACHINE.md.

Use the kit

python -m kit score <path-to-deployment-repo>   # dated, evidence-cited static candidate packet
python -m kit score <path-to-deployment-repo> --format json   # deterministic SDK-friendly JSON
python -m kit matrix                             # render the test matrix

Stdlib only. v0 runs static/structural checks and cites the file:line for every result; chaos/replay checks are declared NOT-RUN, never faked. Evidence comes from code and config, never from a deployment's own prose claims. The kit reports a static structural candidate, not a full conformance certificate; live chaos/replay evidence still has to be executed and attached before claiming production conformance. Shape is part of the stamp: Machine deployments score Machine-L*; model-in-loop orchestrators score Orchestrator-L*. If a repo has mixed signals, the kit requires an explicit --shape or executable-config declaration (conformance_shape / deployment_shape) instead of choosing the flattering label.

Score deployments, not the standard. The kit scores a wiring of The Machine against the six boxes. Pointing it at a library (e.g. the AAR signer), a web-protocol spec (e.g. AVL), or this repo itself returns NOT A MACHINE DEPLOYMENT — those aren't deployments. This repo therefore carries no CONFORMANCE.md: The Machine is the standard, not a deployment, and earns no level of its own. (The kit detects non-deployments by the Box-2 driver-loop signature and declines to score them, rather than floor-bumping to a meaningless L1.)

JSON output for SDKs

python -m kit score <repo> --format json emits schema the-machine.conformance.score.v1. It is the same canonical score as the markdown packet, normalized for orchestration: obligation rows include stable obligation metadata, status, and cited evidence; blockers, not_run, and vnext repeat the same row shape. The JSON intentionally has no timestamp, so repeated runs over unchanged source serialize deterministically. It also carries full_conformance_claimed: false and live_checks_executed: false; this remains a static structural candidate, not a live/full conformance claim. SDKs can use the same payload in-process via from kit import score_repo, score_to_json_data, render_score_json.

Conformance levels

L0 Look-Alike · L1 Declared · L2 Instrumented · L3 Enforced · L4 Receipted · L5 Trusted Autonomy. A deployment earns a level via the kit; it does not claim one. No deployment is grandfathered.

Part of Frontier Infra

Standards: AVL (view) · AAR / Agent Control Plane (proof) · Claude Layers (discipline). Deployments are wirings of The Machine — e.g. machine-driver (the code driver) and Conductor (a Hermes orchestration template for ops triage). None are the Machine itself.

License

MIT — see LICENSE.