THEORETICAL BOUNDARY ANALYSIS · NO EXPERIMENTAL EVIDENCE · NOVEMBER 2026
Experiment C — Identifiability, input limits and recoverability
What can be inferred about identifiability and recoverability from finite observation windows when damping, input limits and model parameters are uncertain?
Section 1 of 7
Research question and motivation
What can be inferred about the identifiability and recoverability of an unstable, input-constrained system from a finite observation window – and when is the available information insufficient?
Experiment C began as a question about the incremental scientific value of BenchEWS, in particular through small, deliberate excitations of the system (Active Probing). The work of Phases 1 to 5.2 turned this question into a theoretical boundary analysis.
Original guiding question, types of benefit and hypotheses (Charter v0.1)
The author’s guiding question: “Can BenchEWS detect the loss of recoverability of a dynamical system earlier, more reliably or with a lower false-alarm burden than established early-warning and control methods?”
- Type of benefit 1 – earlier: lead time of detection
- Type of benefit 2 – more reliable: better detection at the same alarm load
- Type of benefit 3 – lower false-alarm burden at comparable detection
- The object of the test is a single, pre-registered candidate method, not BenchEWS as a benchmark infrastructure
- Proposed research EC-H0 (null hypothesis): no demonstrable additional information value. [C13]
- Proposed research EC-H1: better detection of critical changes of state (secondary, exploratory). [C13]
- Proposed research EC-H2: better prediction of a loss of recoverability (proposed as primary). [C13]
- Proposed research EC-H3: lower false-alarm burden at comparable detection (secondary). [C13]
- Proposed research EC-H4: the additional benefit persists under disturbances, observation errors and model uncertainty (secondary, robustness criterion). [C13]
Not approved Open scientific question: none of these hypotheses has been tested. The work up to Phase 5.2 led to a theoretical boundary finding; the confirmatory study is not authorized (NO-GO). The hypotheses are neither results nor the subject of an ongoing test.
Motivation: a recoverability statement depends on the input limit. If that limit cannot itself be determined from the data, scientifically honest decision support has to say so rather than produce a seemingly precise classification.
Experiment C is a purely theoretical investigation. No physical experiments, no study simulations and no data collection have been carried out. The confirmatory Active-Probing study is not authorized.
Section 2 of 7
Relationship to Experiments A and B, phase chronology
Experiment C uses the same reference model as the Experiment B planning (linearized inverted pendulum with an input limit) and the same nominal parameters. It replaces neither Experiment A nor Experiment B and presupposes none of their results; neither study has been carried out either.
Whereas Experiment A asks whether stability and recoverability come apart and Experiment B is to test a risk indicator, Experiment C asks which statements about recoverability are possible in principle from the available observations.
Phase chronology (all phases drafts only, not merged)
- v0.1 (Nov 16, 2026): design package, charter and draft protocol
- v0.2 (Nov 16, 2026): Active System Probing concept
- v0.3 (Nov 16, 2026): mathematical specification, probe safety concept
- v0.4 (Nov 17, 2026): safety and identifiability gate
- v0.5 (Nov 17, 2026): discrete-time, dead-time and probabilistic safety
- Phase 5.1 (Nov 17, 2026): boundary finding and recomputation of the Fisher information
- Phase 5.2 (Nov 17, 2026): consolidation and safety closure
- Phase 6 (Nov 17, 2026): theoretical question and protocol, proposal only
The documents are held in the project’s non-public research archive; the source list gives paths and commit identifiers.
Section 3 of 7
Mathematical model and assumptions
Reference model: linearized inverted pendulum θ̈ = ω0²θ − cθ̇ + b·sat(u) with ω0² = 19.62 s⁻², input gain b = 4.0, damping c = 0 (nominal) and input limit ū = 1.22625; exact discrete-time sampling at 0.02 s; Gaussian process and measurement noise.
Controller K0: linear-quadratic regulator with a stationary Kalman filter, computed at every sampling step. A probe is a small, known signal added to the control command.
Explicit assumptions
- ideal, symmetric input limit without rate limitation
- linear model; nonlinearities (e.g. sin θ) and friction not included
- zero dead time in the main statements; dead time examined separately
- nominal parameters; uncertainty in ω0² not covered
- Fisher computations asymptotic or local; no finite-sample statements proved
All statements hold only for this model. Transfer to physical robots or real-world systems has not been shown.
Section 4 of 7
Theoretical findings and their status
- Theoretical result documented An input limit that is never reached in the observation window cannot be identified from saturation-free observations alone: the likelihood is the same for all limit values above the largest observed command. [C3]
- Theoretical result documented Confirming the nominal input limit with the same data whose safety it is meant to justify is circular. Active Probing does not resolve this limitation; an externally justified lower bound would be required. [C3]
- Result conditional on assumptions For known damping, a certifiable probe improves the estimation of the gain b only to a limited extent (e.g. window from about 9.3 s to 7.9 s at amplitude 0.10 ū). [C2]
- Result conditional on assumptions For unknown damping, formal asymptotic calculations indicate that certain required observation windows may decrease from approximately 65 s to approximately 26–38 s. The figures are model-specific and do not demonstrate usefulness for decisions within a few seconds. [C2]
- Theoretical result documented Saturation-free operation, controlled invariance, probabilistic safety and recoverability are distinct statements. None follows from another without its own argument; for example, the command can saturate immediately although the state lies inside the invariant region. [C2]
- Theoretical result documented Under Gaussian noise there is no deterministic safety guarantee, only probability bounds for defined events and windows. [C2]
- Withdrawn Withdrawn: an earlier draft statement that passive observation is risk-minimal under unknown damping relied on an overly pessimistic bound (finding R52-01) and is no longer maintained. Upper bounds do not rank risks. [C10]
The findings were checked only with AI assistance, most recently by a separate instance of the same AI model. This is not an independent review and not peer review.
Section 5 of 7
Data, evidence and reproducibility
No experimental data
Experiment C has produced no physical experimental data and no simulation data. Experimental validation is unavailable.
- Experimental validation unavailable No physical experimental evidence; no study simulation; no robot control. [C11]
- Theoretical result documented Numerical values come from deterministic matrix and spectral calculations (no random numbers, no trajectories), partly recomputed by two routes. [C2]
Reproducibility register
- Phase 5.1: Fisher recomputation by two routes (frequency domain and exact finite-N likelihood)
- Phase 5.2: static checks on identifiability, boundary parameters, conditional bounds and the usefulness criterion
- Phase 6: static check of the algebraic identities behind the recoverability boundary
- The scripts are held in the research archive and are not yet public
Section 6 of 7
Counterexamples, limitations, GO/NO-GO and Phase 6
Counterexamples
- Two systems that differ only in their input limit generate identical saturation-free observations but have different recoverability regions.
- Soft actuator limits can make the limit indirectly visible; the non-identifiability holds only for the ideal saturation model.
- At the boundary of the admissible range (damping exactly zero) the Cramér–Rao bound is not applicable.
- The choice of bound can reverse a usefulness statement (finding R52-01).
Unresolved conditions
- Not verified External bounds for input limit, gain and natural frequency are not available. [C3]
- Not verified Dead time combined with saturation is not covered. [C4]
- Not verified Model uncertainty, in particular in ω0², is not covered. [C4]
- Not verified Finite-sample validity has not been proved. [C4]
- Not verified Decision thresholds (error probability, minimum gain, decision window) have not been set. [C5]
- Not verified Independent source verification is pending. [C4]
- Experimental validation unavailable Experimental evidence is missing. [C11]
GO/NO-GO (as of November 17, 2026)
- Closing Phases 3 to 5.2 as a boundary finding: GO (recommendation)
- Confirmatory Active-Probing study: NO-GO, not authorized
- Lifting the probing lock, design freeze, data generation, simulation: NO-GO
Phase 6: theoretical research questions
- Proposed research Which statements about identifiability and recoverability are possible from a finite window under uncertain damping, unknown input limits and model mismatch, without assuming unidentifiable quantities to be known? [C8]
- Proposed research Separating structural, practical and decision-relevant identifiability; counterexamples with identical observations but different recovery capability. [C9]
- Proposed research An abstention criterion: without an external lower bound “recoverable” cannot be certified, and without an upper bound “not recoverable” never can. Being unable to certify is not proof of impossibility. [C9]
Not approved The Phase 6 question is a proposal. The decision register does not record an approval by the author as a research phase.
Section 7 of 7
Publication, DOI, versions and next steps
DOI pending — manuscript not published
An English manuscript (“Limits of Recoverability Inference”) is in preparation. It is not published, not peer reviewed and has no DOI. A DOI will be shown here only once it actually exists.
- Publication pending Publication pending; PUBLICATION HOLD active. [C11]
Version of this page
As of November 17, 2026, based on Phase 5.2 (commit 944a497) and the Phase 6 draft (commits 17550d0 and e7f0994).
Next steps (only after the author’s decision)
- Decision on the Phase 6 question
- External bounds for the input limit, if the branch is continued
- Setting the decision thresholds
- Review and later archiving of the manuscript