RESEARCH OUTLOOK · STUDY PROTOCOL PREPARED · SIMULATION NOT PERFORMED · NOVEMBER 2026

Experiment B — From detection to decision

Does a regret-inspired risk indicator (RRI) add predictive value over a strong baseline model with the same raw information when predicting probe-induced falls of a simulated inverted pendulum?

Study status Experiment B

PlannedPlanned study, execution not authorized
  1. 1 Planned
  2. 2 In preparation
  3. 3 In progress
  4. 4 Analyzed
  5. 5 Independently validated

DESIGN APPROVAL PENDING · PUBLICATION HOLD · DATA-GENERATION HOLD · EXECUTION_ENABLED = false

Duck Lab: Research project in preparation

Basis of classification: Classified according to the author’s instruction of November 15, 2026: study protocol prepared as a draft, design freeze not approved, simulation not performed, no empirical results. Approval of the design is separate from approval to run the simulation. On November 16, 2026, the author took the design decisions AD-01 to AD-18; the design freeze has not been granted and the simulation has not been approved. [S7, S9]

Section 1 of 7

Research question and hypothesis H2

Does a regret-inspired risk indicator (RRI) add predictive value over a strong baseline model with the same raw information when predicting probe-induced falls of a simulated inverted pendulum?

H2 (sole primary hypothesis, draft): adding the RRI to baseline model B9 increases sensitivity for probe-induced falls at an alarm load of 20 % by at least 5 percentage points, averaged with equal weights over nine pre-specified degradation types.

H2 concerns predictive performance, not a mechanism. Because the RRI is computed from the same raw data that B9 receives, an advantage would be a representation gain for the learner studied, not additional information.

Not approved H2, thresholds and decision rule are fixed and documented as design decisions. They become binding only with the design freeze, which the author has not yet granted.

Sources: [S1, S9]

Section 2 of 7

Theoretical background

Detecting a disturbance is not the same as correcting it in time.

The RRI builds on stochastic reach-avoid values: for the estimated state, it computes the probability of staying in the safe set over a horizon, once with the best available action and once with the command actually issued. The difference, averaged over a short window, is the RRI.

The RRI applies known principles (reach-avoid values, performance-difference lemma). No novelty claim is part of H2; the RRI does not measure the recoverability gap.

Sources: [S3, S1]

Section 3 of 7

Study design and twin protocol

Simulated system

Linearized inverted pendulum with 0.02 s sampling, exact zero-order-hold discretization, Kalman filter, event-triggered correction decision, saturated LQR controller and actuator with input limit. All values are specified in Technical Supplement S2 and fixed as design values; the design freeze is pending.

Nine degradation types

  • O – increased measurement noise
  • M_dyn – wrong dynamics model in filter and controller
  • M_obs – wrong measurement-noise model in the filter
  • F-suppression – corrections are suppressed with some probability
  • F-delay – corrections take effect late
  • F-threshold – higher decision threshold
  • A-α – reduced actuator effectiveness
  • A-ū – reduced input limit
  • A-d_a – actuator dead time

Twin protocol

  1. A 60 s mother run is simulated without a probe.
  2. At eleven pre-specified times the state is frozen.
  3. An intervention branch receives a fixed kick to angular velocity, a control branch does not; both use the same random numbers.
  4. Target event: the intervention branch falls within 3 s while the control branch does not.
  5. All monitors see only mother-run data up to 0.5 s before the probe.

No simulation has been run. Implementation and data generation are not approved.

Sources: [S8, S1, S6]

Section 4 of 7

Comparison, endpoints and decision rules

B9 versus B9 + RRI

Both models are gradient-boosting classifiers with the same search and the same data. B9 receives the raw window of estimated state and command over 1 s, filter covariance, decision counters and eight conventional monitor scores. B9 + RRI additionally receives only the RRI.

Primary endpoint

Difference in sensitivity for probe-induced falls at an alarm load of 20 %, equally weighted over the nine degradation types, evaluated in two variants: with pre-calibrated thresholds and at equal weighted alarm load.

  • Success: lower 95 % bound above 0.05 in both variants and better log score
  • Failure: upper 95 % bound below 0.05 in both variants
  • Inconclusive: all other cases
  • Invalid: a validity check fails

Thresholds and rules are fixed as design values. They become binding only with the design freeze, which has not yet been granted.

Sources: [S1, S9]

Section 5 of 7

Data and analyses

No data, no results

Experiment B has not been performed. There are no simulation data, no analyses and no indication of an RRI advantage.

After execution, all three possible outcomes (success, failure, inconclusive) as well as invalid tests will be reported with equal completeness.

Technical compatibility run (no study data)

On November 16, 2026, the author checked the fixed software environment on a MacBook Air with Apple M2 (macOS 27.2.1, Python 3.12.16, NumPy 2.5.3, SciPy 1.18.1, scikit-learn 1.9.1) using the prepared check script. All six mandatory checks passed. This run was a compatibility run on an earlier state of the package; it is not a reference run.

  • Installation of the fixed versions under Python 3.12
  • Environment check including chip identification
  • Known-answer tests for random numbers and checksums (46 of 46)
  • Recomputation of the documented reference numbers (47 of 47)
  • Logic test of check routine V5/V5b on a small test grid
  • Consistency of manuscript, study protocol and supplement

This is a software and environment check, not empirical research evidence and not a completed reference validation: it generates no study data, runs no simulation to test the hypothesis and tests neither the simulation model nor H2. Since author decision AD-18, the MacBook M2 is the sole binding reference platform; the Mac desktop M4 is the standby platform. The full reference run on the MacBook M2 is still pending.

Sources: [S7, S1, S10]

Section 6 of 7

Scientific limitations and pending approvals

A positive finding would hold only for this simulator, the nine degradation types at the intensities chosen in the pilot, the 20 % alarm load and the frozen model pair. It would not establish a recoverability mechanism or transfer to the robot Mali or to other systems.

The event and tables depend on the simulation model. Previous reviews of the protocol were AI-assisted and internal; no independent human review exists.

Not approved Open: the full technical reference run on the reference platform, the MacBook M2 (not yet performed), the design freeze, the one-time master seed and, separately, the approval of the simulation.

Sources: [S1, S9]

Section 7 of 7

Study protocol, manuscript and future results

Study protocol S1, Technical Supplement S2 and the methods and protocol manuscript exist as candidate version 1.0.0-candidate (status November 16, 2026). They are unpublished, not submitted and not publicly available; no DOI has been assigned.

No publication yet

A DOI and Zenodo link will be added here only after a publication approved by the author. Results will follow at the earliest after the simulation has been approved and performed.

Sources: [S8, S7]

← Back to Duck Lab