Transformation baseline published · August 4, 2026

Meridian Science

Measure the scientific path, not the eloquence of the answer.

A frozen instrument tests whether an AI model preserves one physical system while translating among physics, mathematics, computation, observation, uncertainty, and repair.

Open the latest evidence demo See the future comparison

Foundational baseline · MSIO-OSCILLATOR-v0.2

The earlier six-task interpretation baseline remains visible.

Every answer is schema-validated and scored by exact assertions constructed before the model runs.

Reconstruct

Physics → signed equation

Notation residual

Compute

Equation → stable contract

Strict pass

Infer

Observation → bounded inference

Strict pass

Preserve ambiguity

Unknown law → open branches

Strict pass

Diagnose

Planted sign defect → repair

Strict pass

Round trip

Physics → code → observation

Strict pass

Ground truth firstHidden assertions were sealed before paid inference.

Fail-closed scoringMalformed or incomplete structured output fails deterministically.

No answer selectionEvery authorized attempt and residual remains in the audit record.

Protocol note. An earlier pilot exposed a harness defect. Its records remain in the audit trail but are excluded from the model baseline; no task was retried.

The measurement path

Interpretation becomes an auditable transformation.

The model may reason freely inside a bounded envelope. It cannot see or redefine the oracle that scores it.

Physical system

A declared system, units, initial state, constraints, and planted uncertainty.

Structured interpretation

One typed response spanning equations, computation, diagnosis, and epistemic state.

Deterministic residual

Pre-registered assertions expose transport, collapse, invariant, and repair failures.

Claim boundary

Scientific interpretation—not scientific discovery.

This early baseline measures conformance to a declared scientific contract. It does not establish new physics, replace experiment, or substitute for expert review. The Python computation tool was available but not invoked, so this release does not claim autonomous numerical experimentation.

Longitudinal observatory

A place is reserved for what comes next.

The future column will use an official model name only after OpenAI releases it publicly.

EvaluationCurrent OpenAI baselineFuture official OpenAI model

Scientific execution5 / 6 strict passes Reserved — identical frozen run

Graph interpretationNext protocol being sealed Reserved — identical graph contract

Improve MeridianNext protocol being sealed Reserved — independent proposal

Same instrument. Same hidden oracle. New engine.

That comparison isolates model improvement. A separate evolved-Meridian track will measure improvement in the application itself.

Next research layer

From language to an addressed scientific graph.

The next Meridian iteration will construct typed scientific objects, retrieve already verified paths, preserve unresolved branches, and spend new model computation only on the unresolved frontier.

See Meridian beneath AIPI

Natural-language input

Addressed object graph

Verified cached path

Unresolved frontier

Benchmark seal5e37511bd1ed2b02b8779a976bf700a2904a81eae70e982f526f6497e609dcf1

Integrity auditca1e1ac128911cce045cd49f8b0e84e7e542e396baa5e5bcdc74fdc504bc0fea