ADR-0211: Conformal Falsification Bench

Status: Accepted Date: 2026-06-06 Domains: sensorium/environment/, evals/environment_falsification/ Depends on: ADR-0198, ADR-0208, ADR-0209

Context

ADR-0208 established ObservationFrame as a deterministic bundle of already compiled afferent units. ADR-0209 established sensorimotor feedback as afferent evidence. ADR-0198 established that real motor emission is fail-closed until a verdict-enforcing efferent gate exists.

The missing contract was falsification: a replayable way to say, "this expected environmental evidence was or was not observed," without turning observation frames into fusion, memory mutation, hardware control, or learned truth.

Decision

CORE will use the Conformal Falsification Bench as a Python-reference replay contract:

hypothesis / plan
  -> ExpectedObservationFrame
  -> actual ObservationFrame
  -> FalsificationRun
  -> replay-stable report

The v1 bench is exact and binary:

SUPPORTED when every expected slot is present with the expected merge key and no unexpected slot appears.
FALSIFIED when any expected slot is missing, changed, or accompanied by unexpected evidence.

Contract

ObservationFrame remains afferent-only evidence, not fusion and not a mutable world model.
ExpectedObservationFrame and FalsificationRun are replay artifacts, not learned truth, reviewed memory, or pack mutation proposals.
v1 verdicts are only SUPPORTED and FALSIFIED.
No probabilistic confidence, numeric tolerance, hardware-noise envelope, or learned latent is part of the v1 verdict.
No motor/efferent unit, actuator trace, raw pixel buffer, PCM buffer, event payload, decoded action payload, Vault mutation, or generate/* dependency is allowed in the bench.
Public API names are stable: ObservationUnitRef, ExpectedObservationFrame, FalsificationResidual, FalsificationRun, build_expected_observation_frame, and compare_expected_to_observation.

Consequences

The bench gives CORE a falsifiable environmental replay surface before any hardware, motor, native backend, or learned world model is admitted. Later event vision, witness-log import, tabletop lab, and motor governance work must feed or consume this contract rather than bypass it.

Proof Obligations

Expected frame hashes are order-invariant and duplicate-safe.
Raw payloads and efferent units are rejected before traces are built.
Missing, unexpected, or changed evidence yields FALSIFIED.
Exact matched evidence yields SUPPORTED.
The bench does not import generate, mutate Vault, or call ModalityRegistry.decode.
core eval environment-falsification --json is hash-pinned.

2.7 KiB Raw Blame History