ORRERY

The split

ORRERY and the theory are two systems on two clocks, coupled through exactly one thing: tool contracts (CLI flags + JSON output schema + exit-code semantics + a determinism clause, semver'd). Same parameters and seed ⇒ byte-identical declared output, frozen as a golden hash. Exit codes never lie: 0 pass · 1 a declared gate fired (a real negative result) · 2 error. Every tool ships a built-in --selftest and --golden; every build is verified by an independent cold-context two-pass agent that never reads the source.

The catalogue (11 tools built + verified; generated live from the registry)

Every golden is a frozen blake2b-256 of the tool's canonical declared output at fixed parameters — a rewrite must reproduce it bit-for-bit or it is rejected. The four CUDA tools were migrated onto a shared KAT-pinned core (liborrery) in July 2026 with zero golden changes — "the code is ephemeral," exercised for real.

Measured, not asserted

• The branching threshold reproduced in-silico: ratchet's GPU Monte-Carlo matches the Galton–Watson analytic law to a relative error of 0.06% at 4,000,000 trajectories — the law that contains the (1−p)ρ=p critical point.
• The critical point located blind: autotune swept ratchet across ρ against a pre-registered target and located the threshold at ρ_c = 0.2581 (analytic: 0.25).
• The entropy divergence checked against ground truth: algebra measures the critical free-boson central charge at c = 0.9963 (Calabrese–Cardy: c = 1), with the massive control flat at c ≈ 0 — and deliberately computes only the receipted half of the claim; a value the theory itself withdrew is excluded by contract.
• An honest tie, reported as a tie: the someone experiment (n = 24 de-confounded replicates per regime) found the self-model's fitness advantage not supported in any regime — earlier stronger claims were overturned by the instrument, and that overturn is the published result. Instruments that can only confirm are not instruments.
• The citation chain is machine-checkable: both surfaces (the MCP server and the job daemon) return every result with its declared-output hash; end-to-end, a run's hash equals the frozen golden it reproduces (I-12).
• Physics rendered on RT cores — and the tempting claim withheld: lens ray-traces the Schwarzschild photon-capture shadow on hardware ray-tracing cores and measures its cross-section at 27π M² (matched to the exact general-relativity oracle to 2.8×10⁻⁵); the hardware render and an independent analytic baseline agree to the pixel. The seductive claim — that RT cores accelerate the physics — is left as a pre-registered experiment with a kill criterion, not asserted (the RAYFORMER lesson, in-house).

The firewall

Every tool measures structure — whether a mechanism confers fitness, where a threshold sits, how an entropy scales — and says nothing about whether anything feels. Simulations prove structure, never acquaintance. That sentence ships verbatim in every tool's output notes, and no result on this page crosses it.

The repository

The full instrument — contracts, schemas, goldens, verification verdicts, and source — lives at github.com/bochen2029-pixel/ORRERY (goes live with publication). Every cited result carries a result.lock: tool version, binary hash, GPU architecture, CUDA version, seeds, parameters — enough to reconstruct the run from the repository at that commit.