The ladder that lost: quantum filtering as a containment hierarchy over classical estimators, and a measured negative result

Working title: The ladder that lost: quantum filtering as a containment hierarchy over classical estimators, and a measured negative result

Author: [withheld in the repository copy]. Contact: [withheld]. Status: DRAFT — not submitted. Luke submits by hand, and only if he chooses to. Revision: 2026-11-02 second sweep (related work, venues). Issue: none filed; slate entry P2.

Tiers: PINNED (CI gate on every change), MEASURED (run once on real data the owner holds), DESIGNED (built, evaluation owed), DENIED (measured and rejected, kept with numbers), DARK (no measurement; no number typed).

Venues

Venue	Deadline (time zone)	URL	Fee / barrier	Checked
arXiv quant-ph primary; cross-list stat.ML, eess.SY	Rolling. Cutoff 14:00 US Eastern; announced 20:00 ET, Sunday–Thursday	https://info.arxiv.org/help/endorsement.html ; https://blog.arxiv.org/2026/01/21/attention-authors-updated-endorsement-policy	Free. One quant-ph endorsement; stat is its own domain; eess is a seventh domain the venues note did not enumerate (UNVERIFIED whether a cross-list triggers it). The note names this paper as the least ambiguous first quant-ph upload. Policy since 2026-01-21: automatic endorsement needs an academic email and prior arXiv authorship in the domain; every other new author takes the personal-endorsement route, which "arXiv staff cannot waive"; no numeric endorser threshold is published	READ (policy post READ 2026-11-02)
QIP 2027 poster (Singapore, 2027-02-20 → 26)	2026-12-04, 23:59 AoE	https://qipconference.org/2027/call/	HotCRP; not double-blind; title, abstract and technical PDF, no arXiv required; non-archival; presenter must be first author and register; fees TBA, early bird 2026-12-20; travel aid needs a supervisor letter, so none for an independent. Re-checked 2026-11-02: the registration page still says "Registration will open soon"; no fees posted	READ (re-checked 2026-11-02)
TMLR	Rolling; holiday pause about Dec 2 → Jan 5 (pattern; no 2026–27 pause posted as of 2026-11-02)	https://jmlr.org/tmlr/acceptance-criteria.html	No fees; double-blind on OpenReview (profile moderation up to 2 weeks on a public email); first criterion is evidence for claims; novelty "not a necessary criterion"	READ (re-read 2026-11-02)
Physical Review E (alternative)	Rolling	https://journals.aps.org/pre/about	Hybrid; subscription route lists no charge; OA APC $2,910, no waiver for a high-income independent; single-anonymized; ORCID required; empty affiliation field UNVERIFIED	READ
IEEE CDC / ACC, Automatica brief (slate)	—	—	Named in the slate; not checked in the venues note; Automatica "not sampled" by the filtering note. Deadline, fee and page limit UNVERIFIED	not checked
Quantum (viable since 2026-11-02, with the bar caveat)	Rolling; the quant-ph arXiv version must exist first	https://quantum-journal.org/editorial-policies/ ; https://quantum-journal.org/payment/	Single-blind; no fee at submission; publication fee on acceptance €600 by default, €100 reduced, or a complete waiver, "There is no need to justify your need for a waiver" (payment page READ 2026-11-02; the payment page dates the €600 default to 2024-01-01, the news post to January 2025; the amount is the same). The bar is unchanged and still the barrier: "a very significant technical or conceptual contribution"; "Correct research that incrementally improves a limited technique is below threshold". Fits only if the filtering theory (the ladder, Wonham as the diagonal reduction of the qubit SME) is the headline, not the market loss	READ (fee READ 2026-11-02)
Quantum Science and Technology, "Focus on Quantum Coherent Feedback and Reservoir Engineering" (IOP)	Ongoing; invited only; propose by email	https://iopscience.iop.org/journal/2058-9565/page/Focus-issues	One of 24 ongoing invited focus issues; no deadline; IOP hybrid with green OA free; only if the paper is framed as filtering theory, otherwise unchanged plan (QIP poster, TMLR)	READ 2026-11-02

Abstract (draft)

We build a strict containment ladder of estimators for a sparse binary signal and walk it on real data. The rungs are L0, an α-blend ρ ← (1−α)ρ + α·ρ_target; L1, a scalar Kalman filter; L2, a phasor Kalman filter; L3, the Wonham filter; L4, a qubit Belavkin filter (the stochastic master equation) closed at cumulant order 2; and L5, the full-ρ Belavkin filter, not built. Each rung reduces to the one below in a limit: coherences → 0 gives Wonham, linearisation gives phasor Kalman, a frozen gain gives the blend. L4 implements the conditioned Kraus update in closed form, z′ = tanh(atanh(z) + s·y) with a coherence factor; its gain is proportional to 1−z², so the filter goes deaf near certainty, and a purity floor Z_CAP = 0.98 is load-bearing. We score next-session direction prequentially against persistence on five tickers over two disjoint years. In 2025-07 → 2026-06 (n = 375 cells) persistence scored AUC 0.4902, the blend 0.4932, the Belavkin rung 0.4595 and a blank-booted engine 0.5170. In 2024-07 → 2025-06 persistence scored 0.4492, Belavkin 0.4754 and the engine 0.5563. The top rung never beat the blank engine and in one year fell below persistence. A free-run rollout matched hold-the-last-belief to about 0.001 AUC at every horizon, and on a synthetic song bench the zero-shot engine was anti-correlated (0.162 against a 0.500 prior). The rung ships default OFF with the loss in the ledger. No peer-reviewed negative result for an SME filter on real classical data was found.

Contributions

1. The ladder with its limits. L0 ⊂ L1 ⊂ L2 ⊂ L3 ⊂ L4 ⊂ L5, each rung a distinct estimator with the reduction to the rung below stated (DESIGNED theory; distinctness PINNED by [private, withheld]). (DARK: needs public evidence (a public repro or measurement) before it can be claimed.)
2. The closed-form Kraus update and its pole pathology. The log-odds shift, the 1−z² gain, the cumulant cross-update with regression gain cov/v, the exact full-ρ Kraus map as test oracle, and the saturation that makes Z_CAP = 0.98 necessary (PINNED, synthetic).
3. Cumulant-order-2 closure with measured fidelity. 100.00% decision parity against full ρ on ZZ couplings; under exchange stress a maximum divergence of 0.02 at J = 0.6 with parity 99.97% (PINNED).
4. The market ladder tables, two disjoint years, n stated. Every rung, coupled and uncoupled, against persistence, with the sentiment feed shown to add about nothing (MEASURED).
5. Rollout ≡ hold. A free-run forecaster at horizons 1–5 is indistinguishable from holding the last belief, both periods (MEASURED, NEGATIVE).
6. The hook-manifold anti-correlated null. A zero-shot engine scores 0.162 against prior 0.500, persistence 0.975 and a linear model 0.994; the anti-correlation is recorded, not tuned away; the comparison is category-unfair by design (MEASURED, NEGATIVE, synthetic).

Outline (about 7,000 words; a 6-page brief keeps sections 2, 3, 5 and 7)

1. Why a negative result (500 words). The quantum-filtering literature approximates the quantum filter classically; nobody runs the quantum filter as the most general classical estimator and reports that it loses. We do, and we say why the loss is expected on sparse one-sided data. Simple baselines beating elaborate models is a repeated finding in forecasting (Makridakis et al. 2018; Zeng et al. 2023).
2. The ladder (1,200 words). Each rung as an update on a Bloch coordinate z ∈ [−1, 1] for a binary hidden state. The blend is a convex combination of the identity and a replacement channel. The scalar Kalman filter is the Gaussian assumed-density closure of the Bayes filter (Kushner 1967). The Wonham filter is the exact two-state filter; in the z coordinate its diffusion coefficient is proportional to 1−z² = 4p(1−p). The qubit SME (Jacobs and Steck 2006, eq. 91) adds a coherence and an η-weighted innovation; zero coherence recovers Wonham. This derivation is presented, not cited: no source writes Wonham as the diagonal reduction of the qubit SME.
3. The L4 implementation and the pole pathology (900 words). z′ = tanh(atanh(z) + s·y) plus a coherence factor; correlated peers move through a regression gain cov/v; the exact full-ρ Kraus map is the oracle. Repeated one-sided evidence saturates z, the 1−z² gain goes deaf, and the cross-update gain cov/(1−z²) explodes; Z_CAP = 0.98 holds the filter off the pole. This is the saturating, sinh-shaped evidence discounting of Veliz-Cuba, Kilpatrick and Josić (2016) seen in another coordinate.
4. Cumulant closure and decision parity (600 words). Order 2 keeps first and second cumulants and gives O(n²) equations (Plankensteiner, Hotter and Ritsch 2022). Decision parity against exact ρ is 100.00% on ZZ couplings and 99.97% under exchange stress. Kerber, Ritsch and Ostermann (2025/26) note that no general validity criterion for cumulant expansions exists; decision parity is proposed as an engineering one.
5. Method: prequential scoring with persistence as the bar (700 words). Next-step scoring on what was observed (Dawid 1984; Dawid and Vovk 1999), proper scores (Gneiting and Raftery 2007), stream conventions (Gama, Sebastião and Rodrigues 2013). Persistence is the honest baseline. The rung set, periods and metric were fixed before the second year was scored; the pre-registration record is the owner's ledger (Nosek et al. 2018).
6. Results (1,200 words). Both market tables in full, with coupled and uncoupled rungs, L4η, the blank engine, the sentiment ablation (0.5170 → 0.5169) and the membrane-cadence control (burst 0.5154, hold 0.5170, starvation 0.4764). Rollout ≡ hold at h = 1..5 (2025 h = 1: rollout 0.5167, hold 0.5168, persistence 0.495). The hook-manifold bench and its coherence-ablation seam. Cross-period: the engine is best-of-pack both years; the Belavkin rung never beats it and in 2025 sits below persistence.
7. Why it ships OFF (500 words). The engine's ledger records the Belavkin rung as DENIED and ships it default OFF ([private, withheld]). The real-data verdict that first produced that row came from an earlier deployment and is not claimable here; this paper rests the default-OFF decision on the ledger entry and on the market results above. "Persistence is genuinely hard to beat, and n is small." (DARK: needs public evidence (a public repro or measurement) before it can be claimed.)
8. Related work and the negative space (900 words). As below.
9. Threats, DARK, what this paper is not (500 words). As below.

Related work and the negative space

Quantum filtering. Belavkin (1987), "Non-Demolition Measurement and Control in Quantum Dynamical Systems", in Blaquière, Diner and Lochak (eds.), Information Complexity and Control in Quantum Physics, CISM Courses and Lectures 294, Springer, Wien, pp. 311–329, DOI 10.1007/978-3-7091-2971-5_19 (verified 2026-11-02 from the Crossref chapter record; a book chapter, not a journal paper; Belavkin's own list gives the book title as "...in Quantum Systems", a slip): an early "quantum Kalman filter" (priority as the first: DARK, not checked). Belavkin (1988), "Nondemolition measurement and nonlinear filtering of quantum stochastic processes", Lecture Notes in Control and Information Sciences 121, Springer, pp. 245–266. Belavkin (1992), J. Multivariate Analysis: the general quantum filtering equation. Bouten, van Handel and James (2007), SIAM J. Control Optim.: quantum filtering from the conditional expectation as a least-squares estimate; 2009 Best SICON Paper Prize. Bouten and van Handel (2005), arXiv:math-ph/0508006: the Belavkin–Zakai equation as the noncommutative Zakai equation. Wiseman and Milburn (1993), PRA, and (2009), CUP: the homodyne SME with detection efficiency. Jacobs and Steck (2006), Contemporary Physics: eq. 32 (the SME), eqs. 72–75 and 91 (inefficient detection; the √η innovation), Sec. 9 (back-action independent of η).

Quantum → classical reductions. Emzir, Woolley and Petersen (2017), J. Phys. A: a quantum extended Kalman filter by commutative approximation and linearisation. van Handel and Mabuchi (2005), J. Opt. B: the quantum projection filter. Amini, Pellegrini and Rouchon (2014), Russian J. Math. Phys.: filter stability under imperfect measurement. Ralph, Maskell, Ransom and Ulbricht (2021), IEEE FUSION: a particle filter replaces the SME for cost.

Classical rungs. Wonham (1964), J. SIAM Control: the finite-state filter. Kushner (1967), IEEE TAC: the Gaussian assumed-density closure. Särkkä (2013), CUP: the Kalman → EKF/UKF → particle ladder. Baxendale, Chigansky and Liptser (2004), SIAM J. Control Optim.: Wonham-filter stability. Sinopoli et al. (2004), IEEE TAC: Kalman filtering with intermittent observations.

Closures. Plankensteiner, Hotter and Ritsch (2022), Quantum: QuantumCumulants.jl; order 2 gives O(N²) equations. Kerber, Ritsch and Ostermann (2025/26), arXiv:2511.20115: "no general criterion for its applicability or convergence has been found". The four closure papers re-read in full on 2026-11-02, each with its recorded second-order loss: Kirton and Keeling (2018), New J. Phys. 20, 015009, DOI 10.1088/1367-2630/aaa11d: order 2 "only describes the behaviour of approximately Gaussian states", so it is less accurate in the lasing than the superradiant phase, the Fano factor and the full photon distribution "require evaluating moments beyond second order", and the cumulant equations show initial-condition-dependent bistability that mean-field does not. Robicheaux and Suresh (2021), Phys. Rev. A 104, 023702: "MF2" sets the third-order cumulant to zero; two-time correlations show "a complete breakdown of the approximation even for very low light intensities", g²(0) comes out "too large by a factor of ≃2.3" and "negative ... which is unphysical", while one-time values are right except for the most subradiant state (5.0% error against 0.18% at third order). Sánchez-Barquilla, Silva and Feist (2020), J. Chem. Phys. 152, 034108: at second order the emitter population reaches "nonphysical values" (negative) even at weak coupling, weak-driving oscillations are "not accurately described, neither in shape nor in amplitude", and "the second-order approximation fails to describe the dynamics in several cases". Schnoerr, Sanguinetti and Grima (2015), J. Chem. Phys. 143, 185101: the classical counterpart; second-order closures produce "negative mean values or variances, divergent time trajectories, unphysical oscillations and unphysical bistability", "MAs are an ad-hoc approximation and there is no straightforward way to predict their accuracy", with the rule that a closure is valid only if its trajectories show damped or no oscillations. Validity-criterion neighbours: Fowler-Wright, Arnardóttir, Kirton, Lovett and Keeling (2023), Phys. Rev. Research 5, 033148, DOI 10.1103/PhysRevResearch.5.033148: against exact results to N ≈ 150 the cumulant error "does not decrease monotonically" with N or with order, a higher order can be worse than mean-field, and whether mean-field captures N → ∞ depends on how parameters scale with N. Egenlauf, Březinová, Andergassen and Klopotek (2026), Mach. Learn.: Sci. Technol. 7, 025062, DOI 10.1088/2632-2153/ae57f8: a time-local closure holds only while two- and three-particle cumulants stay strongly correlated; buildup of the three-particle cumulant predicts breakdown (a diagnostic, not a threshold; the closest in spirit to order 2 zeroing the three-body cumulant). Carollo (2026), arXiv:2607.25434: cumulant truncations "lack clear justification" and in the strong long-range regime are "at best, conceptually redundant", with one regime where low-order cumulants are exact. Chen et al. (2025), arXiv:2512.12010: convergence proved for weakly interacting fermions, no truncation-order criterion. All of these judge a closure by physical observables; none by whether downstream decisions change.

Saturating gains. Veliz-Cuba, Kilpatrick and Josić (2016), SIAM Review: dy = g dt − 2ε sinh(y) dt + ρ dW; certainty saturates; linear discounting "performs worse".

Quantum models on classical data, and their concessions. Monras, Beige and Wiesner (2011), AMCS: hidden quantum Markov models. Srinivasan, Gordon and Boots (2018), AISTATS: HQMMs "cannot model data any better than a sufficiently large HMM". Ning (2025), arXiv:2510.23237: HQMM learning "highly sensitive to data corruption"; all-synthetic data (N = 30 sequences, T = 100; log-likelihood), and in print: "For datasets generated by classical HMMs, the classical EM algorithm is preferred over the RILA". Ning (2026), arXiv:2608.06554: the same author now claims the opposite on synthetic HMM data, an HQMM beating EM-HMM by 38.5% on average, and on real Splice "only in higher latent-dimension regimes"; the HQMM-versus-HMM record is contested within one author's work. Fellner, Kreplin, Tovey and Holm (2026), Mach. Learn.: Sci. Technol. 7, 010501, arXiv:2504.12416: a peer-reviewed, hyperparameter-matched benchmark of variational quantum time-series models on three chaotic systems over 27 forecasting tasks; "in many cases, quantum models struggle to match the accuracy of simple classical counterparts". Sundar and Elliott (2026), Phys. Rev. A 113, 062461, arXiv:2601.16126: HMM → quantum compression with favourable memory–accuracy trade-offs against a state-merging baseline, the opposite direction to this ladder. Kappen (2026), arXiv:2608.21169: quantum measurement, filtering and smoothing as Bayesian inference on a joint quantum–classical process; the SME recovered from a Kushner–Stratonovich equation with correlated noise, so the quantum-filter ⊃ classical-filter relation is stated in continuous time, for a hybrid latent state, with no estimator ladder and no classical data. Arrazola, Delgado, Bardhan and Lloyd (2020), Quantum: quantum-inspired algorithms useful only for low rank and low condition number.

Evaluation and negatives. Dawid (1984), JRSS A, and Dawid and Vovk (1999), Bernoulli: prequential probability. Gneiting and Raftery (2007), JASA: strictly proper scoring rules. Gama, Sebastião and Rodrigues (2013), Machine Learning: prequential error for streams. Makridakis, Spiliotis and Assimakopoulos (2018), PLOS ONE, and (2020), IJF: statistical methods dominated ML on M3 and M4. Zeng, Chen, Zhang and Xu (2023), AAAI: one-layer linear models beat Transformers on nine datasets. Nosek, Ebersole, DeHaven and Mellor (2018), PNAS: preregistration.

The negative space. The nearest neighbours are Bouten, van Handel and James (2007), who build quantum filtering as the generalisation of classical conditional expectation but present neither Kalman nor Wonham as explicit rungs and test nothing on classical data; Emzir, Woolley and Petersen (2017) and van Handel and Mabuchi (2005), who go the other way and approximate the quantum filter classically; Srinivasan et al. (2018), whose concession about HQMMs concerns expressiveness, not held-out prequential loss on a physical stream; Arrazola et al. (2020), the one published "quantum-inspired method failed" result, which concerns linear algebra and not filtering; Veliz-Cuba et al. (2016), who derive the saturating gain in neuroscience and never connect it to the qubit SME or to a filter losing to persistence; and Kappen (2026), who states the quantum ⊃ classical filter relation in continuous time for a hybrid latent state and never builds or scores a ladder. No paper writes the Wonham filter as the diagonal reduction of the qubit SME; none assembles a containment ladder topped by the SME and walks it on classical data; none reports a Belavkin-class filter losing a prequential contest to a blend and to persistence. Stated narrowly: no peer-reviewed negative result for a quantum-formalism filter on real classical data was found. The nearest published results are Fellner et al. (2026) and Ning (2025), both forecasting or sequence models on chaotic-system or synthetic data, neither an SME filter, neither scored prequentially against persistence, neither shipping the loser default OFF; the one self-reported loss of a density-matrix belief filter to a Gaussian Bayes filter is the quantum-mind library README (0.347 against 0.280 at 60 training sequences; synthetic data; a PyPI and GitHub self-description, not peer-reviewed). No occupancy or market study scores a filter against persistence with a proper score, and no Wonham or Bernoulli filter for presence sensing exists in anything retrieved: the web query returned only PIR and ultrasonic bandpass-filter patents, and the nearest paper is Esmaieeli-Sikaroudi et al. (2024), arXiv:2409.11743, a switching autoregressive process with Markov regime on indoor CO2 that "outperforms simple Hidden Markov approaches" with no proper score and no certainty trap named. The arXiv query "density matrix" AND belief AND (robot OR sensor OR IoT OR cyber-physical) returned zero results on 2026-11-02; "quantum filtering" AND Kalman returned two post-2018 physics-side items (Dehaghani 2024, arXiv:2403.07021; Vladimirov 2020, arXiv:2012.08212), neither against a classical estimator. The filtering note states these as "not found" under a search budget exhausted twice; phrase searches for quantum-like or density-matrix Kalman filters on classical data returned only quantum-system filters.

Second sweep (2026-11-02). Twenty-two web queries ran on the filtering note and twenty on the belief note before the shared budget closed; the remainder ran as direct fetches (arXiv API, ar5iv full texts, Crossref, Semantic Scholar, publisher pages), with four arXiv-API queries rate-limited on every retry and seven web queries never run (quantum particle filters for tracking; quantum-inspired state estimation against a Kalman baseline; Bernoulli or continuous-time HMM occupancy; innovation-based per-sensor reliability; Belavkin applied to finance or cognition; density-matrix belief in IoT; Lanza follow-ups). Verdicts: the ladder topped by the SME, tested on classical data, UNCHANGED (nearest new: Kappen 2026; Sundar–Elliott 2026); a published negative result for a quantum-formalism estimator, NARROWED by Fellner et al. 2026 and Ning 2025, with the specific claim (SME loses to the blend and to persistence on sparse real data, shipped OFF) still without a counterpart; η as observation confidence with η = 0 a no-op, UNCHANGED (nearest: Kappen 2026; Amoros-Binefa 2025, arXiv:2512.05265, an SME with imperfect detection in a simulated magnetometer); the p(1−p) certainty trap used to explain a loss to persistence, UNCHANGED; a decision-theoretic validity criterion for order-2 closures, NARROWED by the model-specific results of Fowler-Wright 2023, Chen 2025, Egenlauf 2026 and Carollo 2026, none decision-based; prequential scoring against persistence in presence sensing, UNCHANGED (the three review hubs, Chen–Jiang–Xie 2018 now placed in Energy and Buildings 169, 260–270, Saha et al. 2019 and Candanedo–Feldheim 2016, all use accuracy or F1); a prior-initialised per-source reliability web, UNCHANGED (the 2024–2026 innovation-based items adapt noise covariances or reject outliers; learned per-sensor trust appears only in patents); GKSL density matrices as an engineering belief for cyber-physical systems, UNCHANGED, with quantum-mind (synthetic, self-reported loss) displacing Lanza et al. 2020 (now known to be pure-state, simulation only) as the closest object. Corrections recorded in the notes and carried here: Belavkin 1987 is the CISM chapter above; quant-ph/0410080 is Bouten's 2004 thesis, not a James–Mabuchi paper. Still UNVERIFIED after search 2026-11-02: a BITS Pilani 2025 "quantum computing-accelerated Kalman filtering for satellite clusters" item claiming lower MAE than a classical KF (search snippet only, not fetched); Lanza and Mannone follow-ups 2021–2024 beyond the records above. Owner numbers in this file are unchanged by the sweep.

Evidence each claim rests on

Claim	Pointer	Tier
The ladder L0 ⊂ L1 ⊂ L2 ⊂ L3 ⊂ L4 ⊂ L5 with limits; rungs are distinct estimators	DARK: needs public evidence (a public repro or measurement) before it can be claimed	DARK
L4 closed-form Kraus update; Wonham gain ∝ 1−z²; cross-update gain cov/v; exact full-ρ Kraus map as oracle; [private, withheld] default	DARK: needs public evidence (a public repro or measurement) before it can be claimed	DARK
Pole pathology; Z_CAP = 0.98 load-bearing	DARK: needs public evidence (a public repro or measurement) before it can be claimed	DARK
Cumulant fidelity: 100.00% decision parity on ZZ; 0.02 max divergence at J = 0.6, parity 99.97%	DARK: needs public evidence (a public repro or measurement) before it can be claimed	DARK
Market 2025-07 → 2026-06, 5 tickers, 76 held-out sessions, n = 375 cells: persistence 0.4902; L0 0.4932; L4 0.4595; L4η 0.4694; coupled L0 0.4932, L4 0.4615, L4η 0.4677; engine 0.5170; sentiment 0.5170 → 0.5169	DARK: needs public evidence (a public repro or measurement) before it can be claimed	DARK
Market 2024-07 → 2025-06, 75 held-out: persistence 0.4492; L0 0.4672; L4 0.4754; L4η 0.4796; coupled L4 0.5001, L4η 0.4955; engine 0.5563 (5/5 tickers)	DARK: needs public evidence (a public repro or measurement) before it can be claimed	DARK
Cross-period: engine best-of-pack both years; L4 never beats the engine; below persistence in 2025	RESULTS.md + RESULTS_2024.md	MEASURED, two disjoint periods
Membrane cadence control: burst 0.5154 / hold 0.5170 / starvation 0.4764	DARK: needs public evidence (a public repro or measurement) before it can be claimed	DARK
Rollout ≡ hold at h = 1..5 to ~0.001 AUC, both periods (2025 h = 1: 0.5167 / 0.5168 / persistence 0.495)	RESULTS_S2.md E-2	MEASURED, NEGATIVE
Belavkin filter as default is DENIED (default OFF)	DARK: needs public evidence (a public repro or measurement) before it can be claimed	DARK
Hook-manifold bench, 6 synthetic songs, grouped split: prior 0.500, persistence 0.975, linear 0.994, U1 zero-shot 0.162, diagonal U0 0.113, reservoir 0.388	hook-manifold docs/experiments/HOOK_MANIFOLD.md	MEASURED, NEGATIVE, synthetic
Coherence ablation seam: diagonal-only ≡ coherent under pure z-dissipation and zero Hamiltonian; distinct canonical hashes on the full spec	hook-manifold tests	PINNED

DARK (do not claim a number)

• Coherence advantage (U1 > U0) on real music or real listeners: owed.
• Forecast skill on live data, for any rung: owed.
• The L5 full-ρ Belavkin filter: not built, so nothing is measured.
• Any result at larger n: the market tables are the only real-data walk.
• Not claimable, distinct from DARK: the real-data ladder verdict on the earlier home deployment. It exists in the engine's documentation and is why the DENIED row was written, but the rescission ceiling forbids naming or using it. This paper rests on the ledger entry and the market results only.

Threats to validity

n is small: 375 cells in 2025, 76 and 75 held-out sessions, five tickers. Every AUC sits near 0.5, and the owner's ledger calls the 2025 result "inside the noise band at this n"; the dossier records no interval or test, so one must be computed, not asserted. One owner tuned the rungs and chose the periods; the pre-registration is a ledger, not a third-party registry. The Belavkin rung tested is the cumulant-order-2 closure, so the loss is a result about L4, not the exact filter, and closure fidelity was pinned on a synthetic gridworld stream. The hook-manifold bench is synthetic and category-unfair by design, as its own document says. Next-session direction is a hard target on which persistence itself scored below 0.5 in both years.

What this paper is not

• Not a claim that quantum filtering is wrong for quantum systems; the SME is exact there.
• Not a report on, or a number from, the earlier home deployment the engine descends from.
• Not a theorem of quantum advantage or disadvantage; it is one measured walk.
• Not a trading or forecasting product; AUC near 0.5 is the finding, not a pitch.
• Not a claim that persistence is optimal, only that the top rung did not beat it.
• Not a human study.

Luke's steps, if he wants to submit

1. Register at arxiv.org with "Independent" as the organization; link an ORCID. Start a submission with primary quant-ph to trigger the endorsement email with its six-character code.
2. Find a quant-ph endorser through "Which authors of this paper are endorsers?" on a recent neighbour (Kerber, Ritsch and Ostermann, arXiv:2511.20115; Ning, arXiv:2510.23237 or arXiv:2608.06554; Mannone et al., arXiv:2509.08002; Kappen, arXiv:2608.21169). Send the code and the PDF to one person; endorsers need papers in the domain dated 3 months to 5 years ago. Since 2026-01-21 this personal route is the only one for an author without an academic email and a prior arXiv paper in the domain, and "arXiv staff cannot waive endorsement requirements" (READ 2026-11-02).
3. Add stat.ML as a cross-list (a stat endorsement may be requested). Treat eess.SY as a seventh domain; if it asks for endorsement, drop it rather than delay.
4. Pick the arXiv perpetual non-exclusive license if the journal is undecided; submit before 14:00 ET.
5. QIP 2027 poster: by 2026-12-04, 23:59 AoE upload title, abstract and the technical PDF on HotCRP (https://qipconference.org/2027/call/). Not double-blind; no arXiv needed. The presenter must be first author and register; fees are TBA (registration page still "will open soon" on 2026-11-02) and early bird closes 2026-12-20; the meeting is in Singapore 2027-02-20 → 26. Decide whether the trip is worth a poster.
6. TMLR: create the OpenReview profile now (homepage showing name and email; "Independent Researcher"), wait out moderation, use the TMLR LaTeX class, anonymize the PDF and ship code as an anonymized zip. The paper may sit on arXiv under Luke's name during review; the submission must not link to it.
7. Before any submission: compute an interval or permutation test for every AUC in both tables; re-run [private, withheld] and [private, withheld] and record the dates; confirm [private, withheld] is still the default. (DARK: needs public evidence (a public repro or measurement) before it can be claimed.)
8. CDC, ACC or Automatica: read their calls by hand; nothing about them is verified here.
9. Quantum, only if the filtering theory leads: the quant-ph arXiv version must exist first; on acceptance the fee is €600 by default, €100 reduced, or waived outright with no justification required (tell them for accounting; READ 2026-11-02). The "very significant" bar is unchanged, so do not send the market-loss framing there. QST's "Quantum Coherent Feedback and Reservoir Engineering" focus issue is invited-only: propose by email, same framing condition.
10. Log the submission in the authority record when there is a receipt.