Adaptation Under Intermittent Power: A Pre-Registered, On-Silicon Demonstration that an On-Device Drift Adaptation of a Frozen Decision Survives Repeated Power Failure Bit-Exact by Progress-Free Re-Derivation of a Bounded Sufficient Statistic
Adaptation Under Intermittent Power: A Pre-Registered, On-Silicon Demonstration that an On-Device Drift Adaptation of a Frozen Decision Survives Repeated Power Failure Bit-Exact by Progress-Free Re-Derivation of a Bounded Sufficient Statistic Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: 10.5281/zenodo.24404565 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign · a pre-registered, on-silicon characterization of an on-device drift adaptation of a frozen, previously published decision encoder, together with a power-failure-survival mechanism for that adaptation (progress-free re-derivation of a bounded sufficient statistic, with an atomic-persistence fallback and a gated selection) that is the subject of a U.S. provisional patent application the authors filed before this deposit; the per-deployment selection procedure is retained as trade secret — see the notice below. Abstract A batteryless sensor draws its energy from light, radio, heat, or motion, buffers it in a small capacitor, and is therefore powered on and off repeatedly and unpredictably. A companion study established that a frozen, stateless, closed-form single-token decision survives such power failure bit-exact by progress-free re-execution: because the decision keeps no state between windows, a power failure costs it no progress, and it re-derives the identical token from a persisted input window across an arbitrary number of interruptions. But a long-lived sensor also drifts — a sensor ages, a baseline wanders, a mean shifts — and the one sibling of the frozen token that tracks drift is an adaptive computation: a geometry-gated drift compensator that carries, between windows, a running estimate of the feature-space location and subtracts it before the decision. That running estimate is exactly the kind of evolving, derived state a power failure destroys and the stateless survival argument does not cover. This paper asks, under pre-registration and on the hardware the method targets, whether an on-device adaptation can keep the frozen decision's bit-exact, deterministic survival guarantees under intermittent power — and finds a sharp, measured answer. The naive approach, updating the running estimate in place and resuming across the failure, is non-deterministic: the update reads and then rewrites the same non-volatile location, so a failure between the read and the write re-applies the update on the next power-up (a write-after-read double-apply), and we measured the adapted decision diverging on 81 of 360 windows under frequent failure and producing 32 distinct decision streams across 32 cold boots with boot-varying cut schedules. Atomically persisting the small (≤152-byte) estimate recovers determinism — we hammered it with 19,386 torn-write injections and observed 0 corruptions and 0 cross-boot decision flips — but it pays a recurring derived-state write and the burden of making that write atomic. The result the paper is about is the third strategy: progress-free re-derivation. We choose the adaptation estimate to be a bounded sufficient statistic of the committed input — a sliding-window mean over the last W committed input windows — so that the estimate is a pure function of windows the device already retains, carries no information not in those windows, and can therefore be re-derived, exactly, from the committed ring on each power-up, persisting no derived adaptation state at all. The derived-adaptation checkpoint is zero bytes. We measured this progress-free adapter re-deriving the adapted decision bit-exact across 32 of 32 cold boots (0 cross-boot flips, 0 fixed-versus-float disagreement) and across 12.95 million determinism comparisons with zero disagreements (Rule-of-Three 95% upper bound 2.3×10⁻⁷), and — the gold standard — across 24 real device reboots and 5 true external removals of the supply — the latter wholly erasing the device's volatile memory — on a Nordic nRF54L15 (Cortex-M33) metered with a Power Profiler Kit II; the adapted decision re-derived bit-exact every time, identically on a second core (nRF52840, Cortex-M4F) and matching the host reference. Recovery is not sacrificed for survival: the progress-free sliding-window adapter recovered accuracy 0.944 against the persisted exponential-moving-average compensator's 0.786 on the same drift battery (no-adapt 0.133, clean ceiling 0.983), beating it on 200 of 200 seeds. The survival, the determinism, and the bit-exactness are universal across the estimator family — any bounded-window pure reduction re-derives bit-exact — while the recovery of a given drift is species-and-regime-bounded: a mean compensation recovers a translational (level-shift) drift and does not recover a gain or shape drift, and the compensator's own label-free gate confines the adaptation to exactly the translational regime the re-derivation is valid in (it opened on additive drift, magnitude 0.747, and declined multiplicative drift, magnitude 0.000, recovering 0.133→0.786 in the first and correctly refusing the second). The survival is moreover architectural, not confined to the one adapter: the frozen decision, its nearest-centroid competence monitor, the token ladder, the on-device safety interlock, and the adaptation re-derived bit-exact across 32 cold boots as one composed stack, and the adaptation composed constructively with the monitor, cutting its benign-drift false-out-of-distribution rate from 0.900 to 0.000 under power loss. This coincidence — power-failure-atomic and deterministic with bounded memory *if and only if* the adaptation statistic is a bounded sufficient statistic, which is exactly the regime the gate already accepts — is the unifying result, and it is a measured boundary, not a claim of generality: we report the honest negatives with equal prominence (the mean-based gate is gameable by a crafted near-zero-mean adversary, driving recovery 0.944→0.231 while the gate stays open; affine, median, and trimmed estimators do not extend the recoverable boundary to shape drift; and a real gas-sensor-array drift is predominantly non-translational, so the recovery is a niche while the survival is universal). The adaptation adds no measurable energy (un-adapted base active power 7.89 mW, drift-compensated 7.61 mW, an overhead of −3.58% within measurement variance) and does not raise the device's minimum storage capacitor. The progress-free re-derivation mechanism, with its atomic-persistence fallback and the gated selection among {re-derived adaptation, amplitude-invariant front end, recommission}, is the subject of a U.S. provisional patent application the authors filed before this deposit; the base encoder, its ladder, nearest-centroid monitor, and the drift adapter are previously published and separately patent-pending; the per-deployment harvesting-aware selection of strategy and window depth is retained as trade secret. Highlights · An on-device adaptation that survives power failure with a zero-byte derived-state checkpoint. By choosing the adaptation estimate to be a bounded sufficient statistic of the committed input — a sliding-window mean — the device re-derives it from windows it already retains on every power-up, persisting no running estimate, checkpoint, or progress index, and re-executes the frozen decision. The derived-adaptation checkpoint is 0 bytes; the decision is bit-exact across an arbitrary number of interruptions, because it keeps no adaptation progress to lose. · The naive update is non-deterministic; measured. Updating the estimate in place and resuming across the failure double-applies the interrupted update (a write-after-read hazard): the adapted decision diverged on 81/360 windows under frequent failure and produced 32 distinct streams across 32 cold boots. Atomic persistence of the ≤152-byte estimate recovers determinism (19,386 torn-write injections, 0 corruptions, 0 cross-boot flips) but pays a recurring derived-state write — disclosed as a fallback, not the point of novelty. · Recovery is not traded for survival. The progress-free sliding-window adapter recovered 0.944 against the persisted exponential-moving-average compensator's 0.786 (no-adapt 0.133, clean ceiling 0.983) on the same drift battery, beating it on 200/200 seeds — while the exponential average needs a deep window (on the order of 76–100 committed windows) to re-derive faithfully and must otherwise be checkpointed. · Measured on real silicon under real power removal. The adapted decision re-derived bit-exact across 24 real device reboots (0/24 cross-boot flips) and 5 true external supply removals on a Cortex-M33 metered with a Power Profiler Kit II, identically on a Cortex-M4F and matching the host reference; 12.95 million host determinism comparisons gave 0 flips (95% upper bound 2.3×10⁻⁷). The adaptation added no measurable energy (overhead −3.58%) and did not raise the minimum capacitor. · The boundary is a theorem the gate already enforces. Power-failure-atomic + deterministic + bounded-memory iff the adaptation statistic is a bounded sufficient statistic — which is exactly the translational-drift regime the compensator's label-free gate accepts (gate opened on additive drift 0.747, declined multiplicative 0.000; correlation of gate signal with recovery ρ = +1.000). The survival is universal across the estimator family; the recovery is species-bounded, and the gate routes gain/scale drift to an amplitude-invariant front end and abrupt drift to a recommission. · The boundaries, disclosed straight. The mean-based gate is a safety detector, not a security control — a crafted near-zero-mean drift keeps it open while defeating the compensation (recovery 0.944→0.231); affine/median/trimmed estimators do not extend recovery to shape drift; a real gas-sensor-array drift is predominantly non-translational (translational fraction ≈ 0.286), making the recovery a niche while the su
Authors
- Randolph James Ferlic (ORCID: https://orcid.org/0009-0006-2367-5205)
- Kimberly Kate Ferlic (ORCID: https://orcid.org/0009-0007-3012-3430)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24404564
- Citations
- 1
- Primary Topic
- Data Stream Mining Techniques
- Type
- preprint