{"id":"37ae4568-e31c-41c5-beb6-b3a2ec9d5a43","arxiv_id":"2607.08865","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A source-state-admitted multi-root boot entropy path for ESP32 IoT combines SRAM startup, public radio-burst local sampling, and asymmetric capsules, with measurements showing RF-disabled WDEV passes the same statistical screens as RF-active modes.","lead":"ESP32 sensors can emit statistically random-looking bytes from a documented pseudorandom RNG state when RF is off, so output tests alone fail. The paper designs and measures a multi-root boot path that credits SRAM, local radio-burst samples, and signed capsules only under explicit source-state and protocol checks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The radio-burst root's credit rests on an untested locality claim: a full observer of the public burst may still predict WDEV samples if the RF-active path is a well-seeded PRNG rather than fresh analog noise.","rationale":"The Reader correctly isolates the load-bearing empirical claim (full-observer unpredictability of WDEV under public burst) and the single-board limitation. My concern is the same claim, sharpened: the paper never measures whether RF-active WDEV is private to the node versus a shared or predictable process once the public stimulus and RF state are known. Table 1 and the Rabbit flags already show that output statistics cannot certify privacy; only a co-located second-receiver experiment can. Because the paper already withholds production credit pending multi-board work and is honest in §8, the verdict remains CONDITIONAL rather than REJECT. Capsules and the source-state admission thesis stand independently; only the quantitative radio-burst credit is at risk. The concrete co-located capture test settles the issue directly.","tokens_in":12555,"tokens_out":688,"duration_ms":8204,"concrete_test":"Place a second identical ESP32 co-located with the DUT, both associated and sampling WDEV during the same fixed public 64-packet burst (identical START nonce, payloads, schedule). Capture source-separated WDEV streams from both boards over ≥100 trials. Compute conditional min-entropy of DUT WDEV given the second board's WDEV (and the public Bi). If H∞(W_DUT | W_co-located, Bi) is near zero or << 7.2 bits/B, the locality claim fails and radio-burst credit must be withdrawn.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's strongest claim is that source-state admission (not output tests) is required, and that under that policy a multi-root combiner yields a seed as strong as the single best credited root. The radio-burst root is the only root that both (a) is presented as providing local unpredictability against a full observer of the public burst and (b) is given quantitative min-entropy credit (~59 kbit raw from SP 800-90B under RF-active state). Section 4.2 and 7.2 credit WDEV solely because Wi-Fi enables the documented analog path; Section 8 states the central assumption that \"a full observer of the public burst cannot predict the node's WDEV samples.\" That assumption is not tested. Table 1 shows RF-disabled (documented PRNG) and RF-active streams are statistically indistinguishable; SP 800-90B cannot distinguish a physical noise source from a well-seeded PRNG. If the RF-active WDEV path is effectively a PRNG whose state is set by a deterministic or low-entropy process known or controllable by a co-located observer (or is highly correlated across nearby receivers), then the public-burst stimulus produces no private entropy and the radio-burst root receives zero genuine credit. The one-board, fixed-stimulus experiment never measures cross-receiver predictability or residual entropy after conditioning on a second co-located capture of the same public train. Capsules remain a valid fallback, but the defense-in-depth claim for the radio-burst root is then unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper argues that ESP32-class IoT nodes cannot safely credit boot entropy from output statistical tests alone, because RF-disabled WDEV output is pseudorandom by specification yet passes the same batteries as RF-active modes (Table 1). It proposes a defense-in-depth boot path that mixes three roots under explicit source-state admission: SRAM startup material, radio-burst local sampling during a public packet train from a trusted node, and asymmetric entropy capsules (ML-KEM/ML-DSA) for cold start. A combiner hashes available roots and assigns the seed the credit of the single best supported root, never their sum. Measurements cover WDEV under controlled RF states, a fixed 64-packet burst window (~59 kbit conservative SP 800-90B raw min-entropy under the RF-active model), deterministic capsule client latency, and preliminary SRAM reset dumps. Open Zephyr artifacts are provided.","tokens_in":12973,"tokens_out":1484,"duration_ms":77782,"significance":"Boot entropy starvation on headless wireless sensors is a real, historically validated failure mode. The paper’s strongest contribution is empirical and architectural: Table 1 cleanly shows that black-box statistical screens cannot distinguish the documented pseudorandom WDEV state from RF-active states, so admission must be source-state policy. The capsule path is a practical cold-start fallback with measured deterministic client cost, and the combiner’s “best single root” credit discipline is appropriately modest. Reproducible firmware, collector scripts, and a fixed-stimulus burst design are genuine strengths. If the radio-burst locality claim were better supported, the multi-root path would be a useful systems result for constrained IoT; even without that, the admission-policy lesson and capsule design are publishable contributions for the embedded-security community.","major_comments":[{"comment":"§4.2, §7.2, and §8: The radio-burst root is the only root given quantitative min-entropy credit (~59.0 kbit raw from SP 800-90B under RF-active state for an 8192 B window) while also being presented as local unpredictability against a full observer of the public burst. Section 8 correctly states the central assumption—that such an observer cannot predict the node’s WDEV samples—but the evaluation never tests it. There is no co-located second receiver, no residual-entropy estimate after conditioning on a second capture of the same public train, and only one board. Table 1 already shows RF-disabled (documented PRNG) and RF-active streams are statistically similar; SP 800-90B cannot separate physical noise from a well-seeded PRNG. Without a locality or cross-receiver experiment, the defense-in-depth claim for this root is unsupported, even though the RF-active gating policy itself is well m","section":"§4.2, §7.2, §8"},{"comment":"§7.2 and Table 1: Credit for the burst window is taken as the lower of two non-IID SP 800-90B rates (7.348 and 7.198 bits/B) under the RF-active source-state model. That model is documentation-based (high-speed ADC enabled with radio), not an empirical demonstration that the samples remain unpredictable given the public stimulus and a strong local adversary (§3.2). The paper should either (a) reframe radio burst strictly as an actuator that places WDEV into an eligible state plus audit/liveness side material, without claiming private min-entropy against a full observer, or (b) add a concrete predictability experiment (second receiver, or conditioning on public schedule and RF environment). As written, the abstract and §7.2 over-claim relative to the evidence.","section":"§7.2, Table 1"},{"comment":"§4.1 and §7.4: SRAM is mixed into the combiner as root P, but the manuscript correctly notes that EN/RTS resets do not fully discharge cells and that cold-power multi-board evidence is missing. Given that, the architecture section still lists “device intrinsic startup state” as a first-class root of unpredictability. Either demote SRAM fully to uncredited auxiliary material in the architecture narrative (matching the evaluation), or supply the cold-power study the paper itself says is required before credit. Leaving it as a named root while denying credit creates an inconsistency between the threat-model framing and the admission policy.","section":"§4.1, §7.4"}],"minor_comments":[{"comment":"Figure 2: The MCV min-entropy curve is labeled “diagnostic, not credited entropy,” which is good, but the caption and body should state more explicitly that timing Ji receives zero credit under the full-observer model, so readers do not treat the residual plot as an entropy claim.","section":"Figure 2"},{"comment":"Table 2: Reset-to-seed times (~10 s median) are dominated by boot/network setup and a fixed 1 s post-DHCP delay. Consider reporting a breakdown that isolates stack setup from crypto so the capsule path is not misread as cryptographically slow.","section":"Table 2, §7.3"},{"comment":"§4.5: The seed equation uses H for SHA-256 and then HKDF; a short note that domain separation tags and length prefixes are applied to each root (mentioned later in the paragraph) would help if placed next to the equation.","section":"§4.5"},{"comment":"Related work: Physical-layer key generation is correctly distinguished from the public-burst model; a one-sentence comparison to RFC 8937’s local hedging (already cited) versus capsule remote delivery would sharpen the positioning.","section":"§2"},{"comment":"Minor polish: “In–fixed-nonce mode” (§6) has a stray en-dash; “Productionsystemsshouldinject” (§4.5) needs spacing; arXiv line-break artifacts in URLs should be cleaned for the camera-ready version.","section":"§4.5, §6"}],"recommendation":"major_revision","confidential_remarks":"The source-state admission result (Table 1) and the capsule design are solid enough for a systems/security venue after revision. The main risk is that the title and abstract lead with “radio burst extraction” as if it were a validated local entropy root; reviewers who fixate on that will reject unless the authors either add a locality experiment or demote the claim. I would accept a revised version that keeps the admission-policy lesson and capsules as the centerpiece and treats the burst primarily as a controlled RF-state actuator with honest non-claims about observer unpredictability."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is Table 1 plus the admission policy: RF-disabled WDEV is documented pseudorandom yet passes the same batteries as RF-active modes, so credit has to come from firmware-checked source state, not ENT/PractRand/SP 800-90B screens. That demonstration is clean, platform-specific, and useful for anyone still treating black-box batteries as gatekeepers on ESP32-class silicon.\n\nWhat is actually new is the packaging: fully public burst as actuator only (no shared-key claim), client-local WDEV + timing response, PQ capsules that put the randomized work on the trusted node, and an explicit per-root admission policy that never sums credit. They ship open Zephyr artifacts, fixed-stimulus burst measurements, and a deterministic capsule client path with real timings. The security argument is deliberately modest—seed strength equals the single best credited root that remains unknown—and the limitations section is unusually straight about one-board data, non-cold SRAM, and unexplained Rabbit flags.\n\nThe soft spot is real but already flagged by the authors. Radio-burst credit rests on the assumption that a full observer of the public train cannot predict the node’s WDEV samples. SP 800-90B cannot distinguish fresh analog noise from a well-seeded PRNG, and they never ran a co-located second receiver or residual-entropy check after conditioning on the public stimulus. If that locality claim fails, the burst root gets zero genuine credit; capsules and (eventually) cold SRAM remain. That does not sink the paper; it just means the defense-in-depth claim for the radio root is still provisional.\n\nMath and citation pattern look fine—HKDF under standard extractor assumptions, honest prior-art coverage of SRAM PUF, EaaS, and physical-layer keying. Free parameters are experimental knobs and the conservative min-entropy rate they already treat as such.\n\nThis is for people who build or review boot entropy on constrained wireless nodes and for standards folks who already care about source models. It deserves a serious referee. I would engage with the work, cite the Table 1 result and the admission framing, and treat the radio-burst numbers as provisional until multi-board locality data appears.","headline":"Solid ESP32 engineering that nails the source-state vs. output-test point; radio-burst credit is still an untested locality claim, but the paper is honest about it and the capsule path stands alone.","tokens_in":13588,"tokens_out":551,"would_cite":true,"duration_ms":6213,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"On ESP32-class sensors, RNG output can look random even when its entropy source is off, so boot seeds must be credited only by source-state admission across SRAM, radio bursts, and signed capsules.","keywords":["boot entropy","embedded systems","defense-in-depth","radio burst entropy","SRAM PUF","ESP32","source-state admission","entropy as a service"],"falsifier":"If a co-located second receiver, or an offline model of the fixed public burst schedule, can predict the target ESP32’s WDEV byte stream during the 64-packet measurement window with non-negligible advantage, the radio-burst root loses its entropy credit.","tokens_in":13493,"feed_emoji":"📡","tokens_out":943,"duration_ms":27620,"temperature":0.7,"pith_summary":"Inexpensive wireless sensors often boot under highly deterministic conditions and still ask an opaque RNG for the first cryptographic values. This paper shows that on ESP32-class boards the wireless RNG register keeps producing statistically clean bytes when its documented analog inputs are disabled—exactly the pseudorandom state the hardware specification already names—so output batteries cannot decide whether the source is live. The authors therefore build a defense-in-depth boot path that mixes three roots under an explicit admission policy: SRAM power-up material, a local response to a fully public radio packet burst from a trusted node, and pre-provisioned asymmetric entropy capsules that let the trusted node deliver a signed seed the client can verify and decapsulate before it has local entropy. Each root is credited only when firmware can check the required physical or protocol state; the combiner never adds credits across roots. If at least one credited root remains outside the adversary’s view, the first seed stays unpredictable.","feed_headline":"ESP32 RNG looks random even when entropy is off","feed_subtitle":"Boot seeds must credit roots by source state—SRAM, radio bursts, capsules—not by statistical screens alone","key_machinery":"Source-state admission policy: each root (SRAM startup, radio-burst local response, asymmetric capsule) receives entropy credit only when firmware-checkable physical or protocol conditions hold; the combiner derives the seed from all available material but assigns the credit of the single best supported root, never the sum.","core_discovery":"RF-disabled WDEV output on ESP32-class boards is pseudorandom by specification yet passes the same statistical screens as RF-active states, so output tests cannot replace source-state admission. A defense-in-depth boot path that independently admits SRAM startup material, radio-burst local responses, and asymmetric entropy capsules, then hashes them under a non-summing credit policy, produces a seed only as strong as the single best root that remains unknown to the adversary.","pith_inferences":["Other SoCs whose RNGs silently fall back to PRNG modes when analog inputs are disabled likely need the same source-state gating rather than black-box statistical admission.","A multi-board cold-power study of SRAM would convert the current auxiliary material into a second credited local root and reduce dependence on the network path.","If co-located receivers can partially reconstruct the target’s WDEV samples under identical public bursts, radio-burst credit would have to be reduced or withdrawn for that platform family."],"forward_implications":["Firmware must gate WDEV credit on verified RF-active state, not on byte statistics that a documented pseudorandom mode can also pass.","A short public packet burst from a local trusted node can open a measurement window that, under the RF-active model, yields tens of kilobits of conservative raw min-entropy before extraction into a 256-bit seed.","A cold-start device with no credited local entropy can still obtain a first seed by performing only deterministic verification, decapsulation, and hashing of a signed capsule.","The derived seed remains secure unless every credited root is predictable given the adversary’s view of traffic and influence.","Any root that fails its admission checks can still be mixed as auxiliary input without receiving entropy credit."],"fun_headline_variants":["ESP32 WDEV fools screens with RF and entropy off","RF-disabled ESP32 RNG passes the same statistical tests","Output tests cannot replace source-state admission on ESP32","ESP32 boot needs SRAM bursts capsules not screens alone","Pseudorandom ESP32 WDEV screens as random by design"],"cache_read_input_tokens":128,"weakest_assumption_plain":"A full observer of the public radio packet burst cannot predict the sensor’s own chip RNG samples collected during that short measurement window.","fun_headline_variants_meta":{"raw":{"variants":["ESP32 WDEV fools screens with RF and entropy off","RF-disabled ESP32 RNG passes the same statistical tests","Output tests cannot replace source-state admission on ESP32","ESP32 boot needs SRAM bursts capsules not screens alone","Pseudorandom ESP32 WDEV screens as random by design"]},"model":"grok-4.5","effort":"low","cost_usd":0.008552,"raw_usage":{"total_tokens":2046,"prompt_tokens":835,"num_sources_used":0,"completion_tokens":83,"cost_in_usd_ticks":85520000,"prompt_tokens_details":{"text_tokens":835,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1128,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":835,"tokens_out":83,"duration_ms":12468,"temperature":1.0,"reasoning_tokens":1128,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T06:09:25.993564+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"If a co-located second receiver, or an offline model of the fixed public burst schedule, can predict the target ESP32’s WDEV byte stream during the 64-packet measurement window with non-negligible advantage, the radio-burst root loses its entropy credit.","supporting_citations":[],"review_version":1}