Adaptive FIM next-experiment proposals¶

Adaptive FIM feedback turns leakage and exact-state-retention counts into conservative proposals for a future static lambda_fim batch. It is an offline experimental-design surface, not a realtime controller and not evidence that FIM feedback protects hardware coherence.

Decision path¶

Supply disjoint leakage and exact-state-retention event counts from one shot block with provenance.
Approve the complete paired control/candidate shot plan through the hardware-safe no-submit budget.
Compute a two-sided Wilson score interval for the configured witness.
Decrease lambda_fim only when the harmful-direction confidence bound clears the target and deadband. Ambiguous, missing-count, and underpowered evidence holds the current value.
Export an observer record and an unapplied scalar proposal to the co-design ports.

The product deliberately has no increase action. The committed repeated IBM follow-up found that lambda_fim = 4 increased leakage and reduced retention for the tested circuit family, so this policy does not reward larger feedback without a new, separately preregistered policy.

from scpn_quantum_control.analysis.adaptive_fim_feedback import (
    AdaptiveFIMConfig,
    FIMWitness,
    plan_adaptive_fim_schedule,
)

witness = FIMWitness.from_counts(
    leakage_events=60,
    retention_events=400,
    shots=512,
    depth=2,
    source="simulator",
    artifact_id="example-count-block",
)
plan = plan_adaptive_fim_schedule(
    4.0,
    (witness,),
    policy_id="ci_dry_run_only",
    shots_per_arm=128,
    config=AdaptiveFIMConfig(target_leakage=0.05, step_gain=4.0),
)

assert plan.allowed
assert plan.steps[0].decision == "decrease"
assert plan.budget.would_submit is False

shots_per_arm describes the future paired control/candidate batch. The hardware-safe execution policy gates that complete plan before any interval is evaluated or observer is created. An over-budget or hardware request returns a refused plan with no steps and no observer records.

Interfaces¶

Interface	Role
`FIMWitness.from_counts(...)`	Count-bound, provenance-carrying leakage and retention witness
`wilson_score_interval(...)`	Closed-form binomial interval used by the decision gate
`propose_count_aware_lambda(...)`	One conservative decrease-or-hold proposal
`adaptive_count_aware_schedule(...)`	Deterministically thread proposals through committed witnesses
`plan_adaptive_fim_schedule(...)`	Hardware-safe budget composition and co-design observer creation
`codesign.adaptive_fim_proposal_port(...)`	Convert a step into an unapplied one-parameter `ControllerProposal`
`propose_next_lambda(...)`	Legacy point-estimate compatibility helper; not a product/evidence route

Evidence¶

Run:

PYTHONPATH=src python scripts/run_adaptive_fim_evidence.py

The committed evidence contains three synthetic calibration controls and an offline replay of circuits 0, 3, and 7 from the already committed repeated FIM raw-count artefact. The source SHA256 and exact job identifier are bound into the payload. The synthetic actions are decrease -> hold -> hold; the selected historical adverse witnesses produce three deterministic decreases.

This establishes input validation, uncertainty gating, deterministic replay, budget refusal, and hardware refusal. It does not establish that executing a proposed batch would improve leakage or retention. Closed-loop efficacy remains untested.

Scientific basis and limits¶

Adaptive experimental design chooses later experiments using information from earlier outcomes. Ferrie, Granade, and Cory developed that structure for bounded Hamiltonian estimation (AIP Conference Proceedings 1443, 165); Hincks and colleagues evaluated online design against fixed sweeps at matched data volume (arXiv:1806.02427). For binomial uncertainty, Brown, Cai, and DasGupta document the poor coverage of the Wald interval and recommend Wilson or Jeffreys alternatives (Statistical Science 16, 101).

Those sources motivate sequential proposal semantics and interval-aware count handling. They do not validate this proportional update rule. This product claims no optimal policy, Bayesian posterior, realtime feedback, provider submission, hardware execution, FIM protection, control stability, or quantum advantage.