The Frozen Token Under Intermittent Power: A Pre-Registered, On-Silicon Demonstration that a Frozen Single-Token Decision Survives Repeated Power Failure Bit-Exact by Progress-Free Atomic Re-Execution
Year: 2026
Type: preprint
Abstract: The Frozen Token Under Intermittent Power: A Pre-Registered, On-Silicon Demonstration that a Frozen Single-Token Decision Survives Repeated Power Failure Bit-Exact by Progress-Free Atomic Re-Execution Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: 10.5281/zenodo.24374592 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign · a pre-registered on-silicon characterization of a frozen, previously published decision encoder, together with a power-failure-survival mechanism (progress-free atomic re-execution and a window-persistence protocol) 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 — often many times per second. Every power failure erases volatile memory, so any computation longer than a single charge must be preserved across the failure or it never completes. The intermittent-computing literature preserves it: it checkpoints activations, or persists a loop index and resumes from the point of failure, or commits idempotent tasks — machinery whose premise is that the computation is long relative to a charge. This paper measures what happens when the computation is not long. The method under test is a frozen, deterministic, class-discriminant single-token decision encoder — a fixed projection and a nearest-of-≤256-centroid lookup that reduces a window of sensor observations to one 8-bit token — whose entire decision is a bounded (~1.57 ms) pure function of a single input window and a fixed model, carrying no state between windows. We show, under two pre-registered campaigns with frozen Go/No-Go bands, that such a decision survives power failure not by any progress-preservation machinery but by progress-free atomic re-execution: on power-up the device restores no intermediate state and tracks no progress index, and simply re-executes the entire fixed computation afresh over a persisted copy of the input window, re-deriving a decision that is bit-exact identical across an arbitrary number of power interruptions, because it keeps no progress to lose. The persisted state is the input window alone; the activation checkpoint is zero bytes, against the 33 KB of genuine computed activations a comparison int8 convolutional network must checkpoint and restore on every power cycle. We measured the decision re-deriving bit-exact across 15,136 cold-start decision comparisons (from independent reboot processes) with zero disagreements (Rule-of-Three 95 % upper bound < 2.0×10⁻⁴), across 24 of 24 mid-computation hardware-reset injections, and — the gold standard — across real external removals of the supply on a Nordic nRF54L15 (Cortex-M33) instrumented with a Power Profiler Kit II, in which the device's volatile memory was wholly lost and it rebooted from an unpowered state, the same token re-deriving bit-exact every time; the result holds identically at 1.8 V and on a second core (nRF52840, Cortex-M4F). The one genuine piece of state — the input window — is persisted by a power-failure-atomic protocol (double buffer + cyclic-redundancy check + commit indicator), which we hammered with 4,000 torn-write injections (3,188 interrupted mid-commit) and observed zero corrupt reads. The resume overhead is not an activation checkpoint but the rebuilding of volatile working tables, which we measured (≈2.1 ms / 16.4 µJ for a transform front-end with a RAM-cached model) and reduced to approximately zero with a stateless ordinal front-end and a flash-resident model. We settle the standard objection — that re-executing an interrupted computation wastes work — with a measured energy law: the re-execution waste scales with the ratio of the decision time to the mean time between failures and therefore vanishes for a bounded decision (8.1 % at a 10 ms mean on-time, 0.08 % at 1 s), so progress-free re-execution is the energy-optimal strategy precisely in the regime this token occupies, where a 756 ms int8 network is instead checkpoint-mandatory. From the measured energies we derive the forward-progress threshold (the token completes from a ~4–10 µF capacitor where the int8 network needs ~2 mF, a 206× capacitor-size advantage) and the non-volatile-memory endurance (a byte-writable FRAM/RRAM with wear-leveled double-buffering lasts tens of thousands of years at one decision per second, where a block-erase flash is exhausted in days). We establish that the survival property is architectural, not an artifact of one encoder or dataset, by re-deriving the decision bit-exact across 3,872 cold boots spanning six distinct decision computations and four signal modalities (transform, ordinal, recurrence, topological, and a four-head product-of-posteriors token ladder). We disclose the invention's boundaries honestly and do not claim them surpassed: a forward-progress threshold shared by every non-sub-checkpointing classifier; an uncommitted-transient-input acquisition limit shared by every sensor; the one adaptation-carrying sibling (a geometry-gated drift compensator) whose ≤152-byte state either persists more cheaply than the window or, if lost, degrades gracefully to a defined commissioning-anchor decision and re-converges geometrically over a median 41 windows with no hard-failure mode; and a platform system-wake (≈833 ms full-OS boot, trimmed to ≈532 ms) that dominates the per-power-up time and is a separable platform cost, not a token property. The progress-free re-execution mechanism and the window-persistence protocol are the subject of a U.S. provisional patent application the authors filed before this deposit; the base encoder, its ladder, nearest-centroid monitor, and drift adapter are previously published and separately patent-pending; the per-deployment harvesting-aware selection procedure is retained as trade secret. Highlights · A frozen decision that survives power failure with a zero-byte checkpoint. On power-up the device restores no intermediate state and tracks no progress index; it re-executes the whole fixed computation afresh over a persisted input window, re-deriving a bit-exact decision across an arbitrary number of interruptions. The activation checkpoint is 0 bytes against the 33 KB an int8 CNN must save and restore every power cycle. · Measured on real silicon under real power removal. The token re-derives bit-exact across 15,136 cold-start comparisons (0 flips; 95 % upper bound < 2.0×10⁻⁴), 24/24 mid-computation reset injections, and genuine external supply removals on a Cortex-M33 — identical at 1.8 V and on a Cortex-M4F. The input-window persistence protocol survived 4,000 torn-write injections with 0 corrupt reads. · Re-execution is the energy-optimal strategy for a bounded decision, not a waste. The re-execution waste scales as (decision time)/(mean time between failures): 8.1 % at a 10 ms mean on-time, 0.08 % at 1 s. The 1.57 ms token wastes essentially nothing; a 756 ms int8 network cannot complete by re-execution at those rates and is checkpoint-mandatory — the standard objection, inverted by measurement. · Forward progress from a small capacitor; millennia of memory endurance. The token completes a decision from a ~4–10 µF capacitor where the int8 network needs ~2 mF (206×); a byte-writable FRAM/RRAM with wear-leveled double-buffering lasts ~63,000 years at one decision per second, where block-erase flash fails in days (so the design mandates byte-writable NVM). · Architectural, not one-encoder. The decision re-derives bit-exact across 3,872 cold boots over six distinct decision computations and four modalities — a transform front-end, a stateless ordinal front-end, recurrence and topological relational front-ends, and a four-head product-of-posteriors token ladder — so the survival property is a property of the stateless, bounded computation, not of any one front-end or dataset. · The boundaries, disclosed straight. A forward-progress threshold shared by all non-sub-checkpointing classifiers; an uncommitted-transient-input acquisition limit shared by all sensors; the one small-state drift-adapter sibling that degrades gracefully (fallback to a defined anchor, geometric re-convergence over a median 41 windows) rather than failing; and the platform system-wake (≈833→532 ms) that dominates the per-power-up budget and is a separable platform cost, not a token property. What must be persisted, and the survival evidence The structural fact: the activation checkpoint a stateful computation needs to resume mid-failure is, for the frozen token, zero bytes — it is re-run, not resumed — against ~33 KB of computed, not-re-acquirable activations for the comparison int8 network. The one object the token persists is the input window, which the sensor must acquire regardless, by the M2 power-failure-atomic protocol (double buffer + CRC + commit indicator). The disagreement rate is structurally zero (the frozen integer / fixed-point path is deterministic) and empirically bounded by the Rule of Three (the 95 % upper bound on a zero over N trials is ≈3/N): Evidence Flips / errors N 95 % upper bound Host f32 cold-boot determinism (pooled) 0 15,136 < 2.0×10⁻⁴ Int8 cold-boot determinism 0 10,240 < 2.9×10⁻⁴ M2 window-persistence integrity 0 30,720 < 9.8×10⁻⁵ Torn-write commit atomicity (3,188 interrupted) 0 4,000 0 corrupt reads On-silicon true power-cut + reset events 0 347 < 8.6×10⁻³ Mid-computation reset injection (bit-exact) 24/24 24 point of interruption irrelevant Cross-capability sweep (6 computations / 4 modalities) 0 3,872 survival is architectural The on-silicon tr
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236766715
