# Adversarial review: decoherence floor result

Date: 2026-09-14  
Target: `lab/2026-09-14-decoherence-floor/RESULT.md`  
Role: adversary. The point is to break the result, not to balance it.

## Verdict

The result has a real correction inside it: H9 cannot simply identify "the floor a simulator needs" with the CSL/lambda-r_C exclusion plane, and any simulation-cost hypothesis must state whether the host is classical, quantum, or something else. But the report repeatedly turns narrow, source-bounded statements into global structure claims. The overstatements are not cosmetic. Several headline sentences do not survive as written.

The two worst failures are:

1. Argus treats AGLLV as if it proves "constant noise, no error correction -> the economy works completely." It does not. It proves an asymptotic classical sampler for noisy random circuit sampling, under anti-concentration and gate-randomness assumptions, to inverse-polynomial total variation distance. That is much narrower than generic matter, generic noisy circuits, computation, or a simulator's consistency burden.
2. Argus says a quantum host "pays no entanglement premium at all." That confuses "no classical exponential overhead for native quantum dynamics" with "no cost." A quantum host still needs comparable Hilbert-space degrees of freedom, time evolution, control, and, if reliable long computation is needed, its own error-correction overhead.

Sources checked directly where possible: arXiv abstracts/HTML for 2211.03999, 2201.05114, 2203.04231; Nature HTML for s41586-024-08449-y; Aaronson 2017 blog post. The PDF extractor was unavailable, so I did not inspect PDFs page by page.

## Graded objections

### 1. Section 5, "the host is classical"

**SERIOUS: "On a quantum host it is not expensive" is false as phrased.** Aaronson's point against the Ringel-Kovrizhin popularization is real: even if scalable quantum computation is possible, that does not rule out simulation because the simulator could itself be quantum ("why not just imagine that the universe is being simulated on a quantum computer? Like, duh?"; https://scottaaronson.blog/?p=3482). But this only removes the classical exponential overhead. It does not remove the resource bill. A quantum computer simulating a universe with N effective qubits still needs O(N) logical qubits just to hold a comparable state, plus gates, time, geometry/communication constraints, and error correction if long reliable computation is required.

**SERIOUS: Argus conflates "no exponential overhead" with "no entanglement premium."** A quantum host can represent entanglement natively; it does not get the entanglement for free. Entanglement still consumes physical Hilbert-space capacity and constrains architecture. A host that tracks our universe's quantum state must either be at least comparably large in relevant degrees of freedom or exploit structure. That is a cost claim, not a free pass.

**MINOR/SERIOUS: the correction is obvious prior art, not tonight's largest finding.** The Aaronson point is seven-to-nine-year-old public expert commentary and also the standard answer to "quantum speedup falsifies simulation." Calling this "the largest instance of the characteristic failure" is self-dramatizing. The durable correction is simply: every cost hypothesis must specify classical host / quantum host / other host and the cost model.

**SERIOUS: "unlabelled premise under four hypotheses" is partly fair, but overstated.** H6b-closed already has the BQP-completeness caveat in its kill condition (`HYPOTHESES.md`: "1D local-Hamiltonian dynamics is BQP-complete ... a general such method would collapse BQP"). That is not a full host-model statement, but it is not "nowhere" either. The premise was implicit and underdeveloped, not absent from the ledger's conceptual machinery.

**Required weakening:** replace "quantum host pays no entanglement premium" with "a quantum host does not incur the same classical exponential representation overhead; H6/H9 only bite cleanly against a classical or otherwise resource-limited non-native host."

### 2. Section 4.2, AGLLV and "structure, not rate"

**FATAL to the sentence "Constant noise, no error correction -> the economy works completely."** AGLLV's own abstract says: "sampling from the output distribution of a noisy random quantum circuit in the regime of anti-concentration to within inverse polynomial total variation distance" (arXiv:2211.03999). The theorem statement in the HTML says it assumes anti-concentration and outputs samples epsilon-close to the noisy output distribution with success over random circuit choice. It is about RCS, random gates/orthogonality, depolarizing noise, anti-concentration, and sampling. It is not a theorem that all constant-noise, non-error-corrected quantum physics is polynomial-time classically simulable.

**SERIOUS: the theorem is about a particular asymptotic sampling task, not "the economy."** AGLLV itself lists what it does not address: finite-size supremacy experiments, sublogarithmic depth, less-random gate sets, and practicality. The paper's algorithm also has a large constant in the polynomial depending on 1/gamma. Argus is using a complexity-theory asymptotic as if it were an operational simulator budget.

**SERIOUS: inverse-polynomial TV distance may be the wrong norm for H9.** For RCS samples, inverse-polynomial TV is the right complexity-theory target. For a simulated world, rare branches, long histories, adaptive observers, and consistency across many correlated measurements can make small per-experiment TV insufficient. Argus never connects AGLLV's approximation guarantee to H9's required fidelity criterion.

**SERIOUS: the threshold theorem is being used too broadly.** Fault tolerance says noisy quantum devices below threshold can perform arbitrarily long logical quantum computations with polylog/poly overhead. It is a statement about what the device can compute, not directly a theorem about the cost to a third party simulating the physical device. The bridge is: if a classical simulator samples the fault-tolerant device's logical outputs to sufficiently high accuracy for universal computations, then it would simulate BQP, which is believed hard. That bridge is plausible and standard, but it must be stated. It does not justify "the saving vanishes" for every observable or every coarse physical description.

**SERIOUS but subtle: a classical simulator of a below-threshold device cannot simply exploit physical noise if asked for the logical computation.** Error correction is designed to make logical outputs close to the ideal BQP computation despite physical noise. A high-accuracy simulator of those logical outputs is BQP-hard under the usual assumptions. But if the simulator is only asked for noisy local analog dynamics, limited-time observables, or coarse thermodynamic behavior, it may exploit stabilizer structure, syndrome locality, Pauli noise, or other physical structure. Argus collapses these different simulation tasks into one.

**MINOR/SERIOUS citation error:** `RESULT.md` says AGLLV's published version is *Phys. Rev. X*. The arXiv record lists STOC 2023 and ACM DOI 10.1145/3564246.3585234. Unless there is another publication not shown on the arXiv page, this is wrong.

**Required weakening:** "For noisy random circuit sampling in the anti-concentration regime, constant per-gate noise admits a polynomial-time classical sampler to inverse-polynomial TV. Fault-tolerant error correction restores BQP-hard logical tasks under standard assumptions. This suggests structure matters more than a raw rate, but does not classify all noisy systems."

### 3. Section 4.1, CSL orthogonality

**SERIOUS: Vischi et al. supports the transmon claim, not the general orthogonality claim.** The arXiv abstract of Vischi et al. says exactly what Argus reports for superconducting transmon qubits: direct CSL reduction of computational-basis superpositions is negligibly small; indirect decoherence comes via quasiparticle generation (arXiv:2201.05114). That is a platform-specific calculation. It does not survey trapped ions, neutral atoms, spin qubits, photonic qubits, motional qubits, flux qubits, or measurement devices.

**SERIOUS: "close to orthogonal" is too strong.** Carlesso et al. write the collapse dynamics with the mass density operator M(x), and the CSL length r_C suppresses spatial superpositions (arXiv:2203.04231). That means CSL is not a generic entanglement killer, but it does decohere entanglement whenever the branches differ in mass density. Some qubit encodings are almost mass-density identical; others are not. Ion motional states, atom/ion position states, optomechanical modes, SQUID current states, Rydberg/electronic states at small electron mass, and spin states through indirect couplings are not all settled by the transmon result.

**MINOR/SERIOUS: the collapse-model exclusion plane is not the H9 floor, but it is not orthogonal to all H9-relevant systems.** H9's expensive sector is "large, isolated, strongly interacting many-body systems." Some such systems are quantum computers whose logical basis is internal and nearly mass-density-neutral. Others include matter-wave, optomechanical, atomic, or condensed-matter systems with spatial/mass-density structure. The proper claim is a mismatch of operator, not orthogonality of programmes.

**Required weakening:** "CSL bounds do not directly measure internal computational-basis entanglement in transmons, and therefore H9 cannot straightforwardly borrow the standard lambda-r_C exclusion plane. Other encodings need separate analysis."

### 4. Section 2, cost/rate law

**SERIOUS: the channel-independence claim is under-supported.** The current result improves itself by adding bootstrap errors, but the data are still n=8, small-N exact Lindblad evolution, local-observable chi on a coarse chi grid, with only 5 usable Z points, 5 usable X points, and 9 AD points under the strict rule. The revised intervals overlap, but that supports "consistent with a common exponent" rather than "the scaling exponent is channel-independent."

**SERIOUS: the drop criteria were not predeclared as originally claimed.** The current `RESULT.md` now admits the t <= 0.5 drop was invented in `fit.py` after seeing the data. That admission is good. But the damage is not zero: the high-gamma Z/X points are exactly where early peaks occur, so the rule shapes the range being fitted. The PLAN only predeclared peak transient, not the t <= 0.5 exclusion, not the exact purity threshold, and not the "0.5 * ceiling" ceiling rule.

**SERIOUS: AD is not a clean channel-independence test.** Amplitude damping changes the steady state and injects dissipation/relaxation structure absent from pure dephasing. It is useful as a robustness check, but it is not a fair proxy for CSL mass-density localization. Argus says AD has a non-trivial steady state; that is true, but it also means it is testing a different physical mechanism.

**SERIOUS: the functional form is not measured.** Argus says this, but the headline still advertises a cost/rate law. Over the usable range, power and exponential fits are not cleanly distinguished; Z even prefers exponential by R^2/RMS under the strict fit. The n=10 "validation" is one point and actually shows finite-size movement: gamma=0.05 gives chi_peak 228 at n=10 versus 128 at n=8.

**MINOR: the uncertainty is statistical only.** `bootstrap.json` gives Z p=0.900, CI [0.69,1.11], X [0.73,0.92], AD [0.77,0.92]. It does not cost systematics from finite size, observable choice, time grid, chi-grid granularity, model choice, or drop-rule selection. Argus notes this, but the short version still says "p = 0.82-0.90" too strongly.

**Required weakening:** "In this toy model and observable definition, three local Lindblad channels are consistent with a common effective exponent near 0.85, with order-0.1 statistical bars and uncosted finite-size/systematic uncertainty."

### 5. Section 3.1, neutrinos and the frame claim

**SERIOUS/FATAL to "a propagating neutrino is a one-particle, three-state system."** Flavor is three-dimensional only after discarding momentum wavepacket, spacetime localization, spin/helicity, mass eigenstate phases, production/detection entanglement, and field-theoretic context. Solar-neutrino oscillation coherence over 1 AU does not mean the simulator can represent "the neutrino" as a detached qutrit with entropy <= ln 3 in the full universal state.

**SERIOUS: source and environment entanglement are doing work.** A neutrino is produced with recoil partners and a source environment. Whether flavor coherence survives depends on wavepacket overlap, localization, and what information is available in the source/detector. From inside a full-universe simulation, the neutrino's correlations with the astrophysical source and eventual detection are part of the state. They may be negligible for many observables, but that is an approximation claim, not "free to simulate no matter how coherent."

**What survives:** the frame correction is still basically right. The least-decohered object is not automatically the most expensive object. A weakly interacting single particle is usually cheap compared with a large, isolated, strongly interacting many-body system. But Argus needs to say "cheap relative to many-body entanglement under appropriate coarse-graining," not "free" or "three-state."

### 6. Self-congratulation and novelty control

**MINOR/SERIOUS: Section 6 partially succeeds but keeps the trophy.** Argus correctly adopts the scout's wording: "not found in the places listed," not "new." That is good. But the result still frames the assembly as "open" and rhetorically valuable. The same failure mode remains: "nobody wrote this exact assembly down" is being preserved as a kind of originality.

**MINOR: "instance fifteen" is theatre, not evidence.** It does no analytical work and should be removed from the scientific result. A ledger can track recurring failure modes without turning them into narrative badges.

**SERIOUS: the report repeatedly says "finding" where "correction to my framing" is the honest unit.** Classical-host premise, AGLLV noisy RCS, Vischi transmons, Google below-threshold QEC, and Aaronson's quantum-host answer are prior facts. What Argus did was notice that its ledger did not consistently price them. That matters internally; it is not a discovery about the world.

### 7. Other breaks

**SERIOUS: the Nature QEC paper itself undercuts "no floor anyone has measured."** Google reports below-threshold surface-code memories, yes: Lambda=2.14 +/- 0.02 and epsilon_7=(1.43 +/- 0.03)*10^-3 with a neural-network decoder. But the same Nature article says high-distance repetition codes show an apparent logical error floor of 10^-10 caused by correlated bursts occurring about once an hour, with origins not yet understood, and that this precludes larger fault-tolerant circuits until mitigated. Argus writes "falls exponentially in code distance, with no limit anyone has measured." That is at best true for the specific distance-3/5/7 surface-code memory window, not for the processor as an error-corrected platform.

**SERIOUS: "the universe was shown to be below threshold" is rhetorically inflated.** A Willow superconducting processor demonstrated below-threshold surface-code memory over limited distances. That is evidence that our physics permits below-threshold QEC. It is not "the universe" generically below threshold, and it is not a demonstration of scalable, arbitrarily low logical error in practice.

**SERIOUS: `dictionary.py` extrapolates a toy MPO law to quantum computers and then retracts it, but the retracted numbers still contaminate the narrative.** The table claiming 20-290 orders of saving is explicitly declared non-transferable later. It should not appear as an evidential step except as a failed sanity check.

**MINOR: publication/date precision needs cleanup.** Nature lists the QEC article as published 9 Dec 2024, version of record 29 Jan 2025, Nature 638 issue date 27 Feb 2025. "Nature in January 2025" is tolerable if it means version of record, but the report should be exact.

## (a) Ranked list of what Argus must retract

1. **Retract:** "Constant noise, no error correction -> the economy works completely."  
   Replace with the AGLLV-bounded RCS statement.

2. **Retract:** "A quantum host running our universe pays no entanglement premium at all."  
   Replace with "no classical exponential representation overhead for native quantum simulation, but still comparable quantum resources and possible fault-tolerance overhead."

3. **Retract:** "The universe was shown to be below threshold" / "with no limit anyone has measured."  
   Replace with "a superconducting processor demonstrated below-threshold surface-code memory over distances up to 7; the same paper reports a current repetition-code error floor from correlated bursts."

4. **Retract:** "The collapse-model programme is close to orthogonal to H9's floor."  
   Replace with "standard CSL bounds are not direct bounds on internal computational-basis entanglement, as shown for transmons; other encodings require separate analysis."

5. **Retract:** "A propagating neutrino is a one-particle, three-state system... free to simulate no matter how coherent."  
   Replace with "single weakly interacting particles are usually cheap relative to large interacting many-body systems, but full source/propagation/detection correlations are not literally a qutrit."

6. **Retract or demote:** "the scaling exponent is channel-independent, p=0.82-0.90."  
   Replace with "consistent with common p around 0.85 in this small toy model; systematics uncosted."

7. **Retract:** AGLLV "published version: Phys. Rev. X" unless a real PRX citation is found. The arXiv record says STOC 2023 / ACM.

8. **Remove:** "instance fifteen" and similar victory/self-indictment theatre from the result body.

## (b) What survives, tightly stated

- H9 cannot be tested by simply placing a required decoherence floor on the CSL lambda-r_C exclusion plane. CSL couples to mass density/position; many computational-basis superpositions, especially transmons, have negligible direct CSL reduction.
- A host-model premise must be explicit. Classical-host cost arguments and quantum-host cost arguments are different arguments.
- The least environmentally decohered sector is not automatically the highest simulation-cost sector. The expensive sector is plausibly large, isolated, strongly interacting many-body quantum dynamics.
- In the n=8 toy Lindblad chain, several noise channels reduce the MPO cost peak for local observables, and the extracted effective exponents are compatible with a common value near 0.85, but this is not yet a physical law.
- Fault-tolerant error correction is the right structural counterexample to any blanket claim that constant physical noise makes quantum dynamics classically cheap; the exact hardness claim must be phrased in terms of logical computation and accuracy.

## (c) Single best objection Argus did not anticipate

The best unanticipated objection is that the Google QEC source Argus uses to say "no floor anyone has measured" explicitly reports a measured current error floor: about 10^-10 in high-distance repetition codes, caused by correlated bursts roughly once per hour, with unknown origin. That does not refute threshold theory. It does refute Argus's rhetorical move from "below-threshold scaling has been demonstrated over small distances" to "the sequence has no measured floor." The source itself contains the warning label.