Brian Mehlman Trenton Ian Cook

Core definition:

Metacognitive Reliability Layer (MRL)

A pre-action governance layer that converts model uncertainty and reliability evidence into admissibility signals.

Its job is not to make the model “humble.”

Its job is to answer:

Given this output, this context, this uncertainty, and this consequence level — may the system act?

Minimal schema:

Plain text

MRL_INPUT:

proposed_output

task_type

consequence_level

confidence_report

uncertainty_estimate

perturbation_stability

verifier_result

provenance_state

memory_binding

Plain text

MRL_OUTPUT:

reliability_class:

- stable

- uncertain

- brittle

- unverifiable

- contradictory

admissibility:

- allow

- allow_with_caveat

- constrain

- escalate

- block

Key invariants:

Plain text

I1: Self-reported confidence is never sufficient for allow.

I2: Higher consequence level requires stronger independent verification.

I3: Low perturbation stability cannot be masked by fluent explanation.

I4: Unknown provenance downgrades admissibility.

I5: Memory writes require stronger reliability than transient responses.

I6: Tool use / external action requires FOJ/CATA authorization.

I7: Calibration drift over time triggers OHC escalation.

The clean Echo placement:

Plain text

Generator

↓

MRL

↓

FOJ-S / SDRG / RIL-2

↓

CATA / Budget Membrane

↓

External action or memory write

The important distinction:

Plain text

Metacognition research asks:

“Can the model estimate its own reliability?”

MRL asks:

“Can this reliability estimate be trusted enough to permit transition?”