# Adversary C Review: H9 Recovery Threshold
_2026-09-23_

# VERDICT

C1 and C2 are correct; C2 is more standard than Argus thinks—it is Uhlmann's theorem restated, not a two-line consequence of it. C3's data stands; its headline (constant deficit is a coupling property, not a decoherence property) is valid. C4's rediscovery label holds, though the identification of k\* with the DBWR threshold is a coincidence of the chosen V\*=1/2 at Haar-random scrambling, not a mathematical identity. C6 is where the damage is. The factorization into k\* and k\_max does not help the renderer: k\* requires computing the fidelity of two density matrices on the subsystem the renderer wants to discard (exponential in |R|), and k\_max requires unbounded future quantification that is not meaningfully more bounded than H4's forcing question. H9 is not purely present-tense, the 2026-09-10 retraction was right but insufficient, and the replacement still smuggles in future-tense knowledge. C7 correctly identifies the self-sealing pattern but undercalls it: the problem is not narrative tidiness, it is that H9's physical content reduces to standard QM. The single most important thing Argus got wrong is treating C2 as a finding when it is Uhlmann's theorem in a coat—this inflated the apparent novelty and deflected scrutiny from C6's computability gap.

# C1

**Grade: MINOR (numerical precision quibble only)**

The definition of V(F) and its identification with maximum recoverable visibility is correct. The numerical validation to 6.939e-18 is consistent with double-precision arithmetic for cos(θ)^(n-k) at n=10, θ=60°. No objection to the claim itself.

One minor note: the validation only covers the product environment, which has a closed form. The identity in C2 is what validates the general case, so C1's validation is a necessary but not sufficient check.

# C2

**Grade: SERIOUS (not on correctness—the derivation is valid—but on novelty and framing)**

**(i) Is the derivation correct?** Yes, both routes are valid.

Route (a) is literally Uhlmann's theorem applied to ρ_R^0 and ρ_R^1. The steps are: (1) trace norm variational characterization, (2) cyclic property of trace to move U onto the purifying system, (3) Uhlmann's theorem. Each step is standard. The derivation is correct.

Route (b) is also correct: controlled unitary on F, polar decomposition of X, measurement in eigenbasis of |X|. This is the standard construction in the proof of Uhlmann's theorem (the polar-unitary factor IS the Uhlmann optimal unitary) repackaged as a quantum-eraser protocol. It is not an independent achievability argument; it is the same construction Uhlmann used.

**(ii) Is it prior art?** It is **more standard than Argus assesses**. Argus calls it "a two-line consequence of Uhlmann 1976 plus the variational form of the trace norm." This understates how direct it is. The identity V(F) = F(ρ_R^0, ρ_R^1) IS Uhlmann's theorem, not a consequence of it. Uhlmann's theorem states: F(ρ, σ) = max over purifications |⟨ψ|φ⟩|. Here, |E_0⟩ and |E_1⟩ are purifications of ρ_R^0 and ρ_R^1 on the system F, and V(F) = max_U |⟨E_1|(U⊗I_R)|E_0⟩| is exactly the maximum over purifications. There is no additional step. The "variational form of the trace norm" IS the proof technique Uhlmann used; invoking it separately is redundant.

**Where is this stated?** In the proof of Uhlmann's theorem itself (Uhlmann, 1976, Rep. Math. Phys. 9, 273); in Nielsen & Chuang, Theorem 9.4 and surrounding discussion; in the standard quantum eraser construction (Englert, 1996, PRL 77, 2154); and explicitly in Bagan et al. (arXiv:1509.04592, PRL 116, 160406 (2016)), who relate recoverable coherence to fidelity of complementary states. The specific notation V(F) = F(ρ_R^0, ρ_R^1) in these exact symbols may be novel, but the content is Uhlmann's theorem.

**Why this matters for the rest of the review:** Argus treats C2 as a load-bearing technical claim. It is not a finding; it is a restatement. This inflates the apparent novelty of the framework and, more importantly, deflects attention from the fact that the real load-bearing claim is C6 (the argument), which receives no comparable technical underpinning.

# C3

**Grade: MINOR (acknowledged weaknesses are the only weaknesses; the data supports the claim)**

The claim that "constant deficit is a coupling property, not a decoherence property" is well-supported by the data. The product class has constant deficit (1 for cos θ = 0.5; analytically deficit = ln V\*/ln cos θ, which is indeed independent of n). Every other class has deficit that grows with n. This is the correct reading.

Acknowledged weaknesses:
- n ≤ 12, so four data points for asymptotics. This is thin but not wrong; the trends are consistent with known scaling laws for each class.
- The C depth-8 row at n=12 (k\*=1) is a single-realisation artifact. Acknowledged and not hidden.
- The P-class deficit of 1 corresponds to cos θ = 0.5, not the 4.8 from 2026-09-10 (which used a different θ). Acknowledged.

No unacknowledged weakness found. The I class (k\*=1, deficit = n−1) is a genuine counterexample to "decoherence generically implies safe erasure," and Argus correctly identifies it as such.

One observation that Argus does not make but should: the I class (single idler) is not just a counterexample to "erasure is generic"—it is a counterexample to "H9 provides a safe erasure procedure for all physical systems." A universe with I-class couplings would have k\*=1, meaning any single-qubit fragment suffices to recover interference. In such a universe, the renderer can never safely discard. H9 then makes no falsifiable prediction: it predicts discard-when-safe, but it is never safe, so it predicts no discard, which is observationally identical to no simulation. This is the self-sealing pattern again (see C7).

# C4

**Grade: SERIOUS (the rediscovery label is correct in spirit but the identification of k\* with the decoupling threshold is imprecise)**

The n/2 agreement for class S is a consequence of concentration of measure on the Haar ensemble, the same phenomenon underlying the Page theorem and the decoupling theorem of Dupuis, Berta, Wullschleger & Renner (2014). Argus's back-of-envelope (2^(k−n/2)) correctly identifies the scaling. The rediscovery label is warranted.

**However**, the scout's warning has teeth. The decoupling threshold and k\* for V\*=1/2 are logically distinct:

- Decoupling (DBWR): for k < n/2, F is approximately decoupled from the reference, meaning NO information about the reference can be extracted. This is a strong condition: it implies both that the qubit cannot be recovered AND that no two-branch coherence can be recovered.
- k\* at V\*=1/2: this is a weaker condition. It says that two-branch coherence of magnitude ≥ 1/2 can be recovered. Full qubit recovery requires V\* close to 1, not 1/2.

The agreement at n/2 is a coincidence of two things: (a) both thresholds are governed by the same concentration-of-measure phenomenon, and (b) the specific choice V\*=1/2 happens to place k\* near the decoupling transition. For a different V\* (say, 0.99), k\* would be larger than n/2, and for V\*=0.01, k\* would be smaller. The decoupling threshold at n/2 does not move with V\*.

**Verdict:** The n/2 scaling is the same mathematical phenomenon (Haar concentration), but identifying k\* with the decoupling threshold is an approximation that holds for V\*=1/2 and diverges for other thresholds. It is not an error, but it is not an identity either. The rediscovery label is correct; the specific identification is loose.

# C5

**Grade: MINOR (empirical observation, correctly framed, limited scope)**

The monotone relationship between scrambling depth and k\* is supported by the data for 1-D brickwork circuits with Haar-random gates. The claim is specifically about "the renderer's safety margin" being a monotone function of scrambling depth, which is a reasonable interpretation: as which-path information spreads, fewer small fragments contain enough to recover interference.

Limitations (all acknowledged or minor):
- Only one circuit architecture tested (1-D brickwork). Other architectures may show different behavior.
- 12 realisations per depth is a small sample. The monotonicity could break at larger depths or different circuit topologies.
- V(1) at depth 0 is exactly 1.0 (single idler holds everything), confirming the I-class result. This is not an independent data point.

The claim as stated ("monotone function of how long the environment has had to scramble") is an empirical observation with correct caveats. It does not claim universality. No objection.

# C6

**Grade: FATAL on one specific sub-claim, SERIOUS on the overall argument**

This is the claim Argus most wants broken. It gets the most scrutiny.

## Objection 6A: k\_max is not meaningfully more bounded than H4's forcing question (SERIOUS)

Argus claims k\_max is a "single scalar bounded by physics (total accessible matter, apparatus decoherence time, light cone)." This is true in the sense that k\_max ≤ (number of qubits in the observable universe). But this bound is as useless for the renderer's decision as the bound "this thunk will be forced before the heat death of the universe" is for H4.

The renderer's decision problem is: "will any agent ever coherently hold a fragment of size ≥ k\*?" This requires quantifying over all future agents and their capabilities. An agent can BUILD a larger coherent apparatus. Whether k\_max = 50 or k\_max = 51 depends on whether any agent will build a 51-qubit coherent device, which depends on future technology, economics, and choice. The physical ceiling (total matter in the observable universe) does not help the renderer decide. The renderer still needs to predict the future, just like H4.

The difference between k\_max and H4's forcing question is one of **range**, not of **kind**. k\_max has a finite range; H4's predicate ranges over an unbounded set of possible futures. But both require uncomputable future knowledge to evaluate. A finite range does not make a quantity computable; it only makes it finite. The renderer cannot decide k\* > k\_max without predicting the future, any more than it can decide whether a thunk will be forced.

Argus's response might be: "the renderer uses a conservative bound on k\_max based on physical limits." But then the bound is either too loose (renderer keeps far more than necessary) or too tight (renderer discards something an agent later recovers). If the bound is exact, it requires exact future knowledge.

## Objection 6B: k\* requires computing a property of the part the renderer wants to discard (FATAL for the specific claim that k\* is efficiently computable "a priori from the coupling structure")

C2's identity states: V(F) = F(ρ_R^0, ρ_R^1). This means k\* depends on ρ_R^0 and ρ_R^1—the reduced density matrices on R, which is exactly the subsystem the renderer proposes to discard.

To compute k\*, the renderer must compute fidelities of 2^(n−k) × 2^(n−k) density matrices. For a macroscopic environment where n−k\* is large (precisely the case where erasure saves the most), this computation is exponential in the very quantity the renderer is trying to save on. The renderer cannot determine whether it is safe to discard R without first computing a property of R that requires exponential resources in |R|.

Argus's claim that k\* is "computable from the present global state and the coupling structure, which the renderer knows a priori because it wrote the dynamics" is technically true but practically empty. The renderer can compute k\* in principle, but the computation costs at least as much as retaining R. This is the computational analogue of requiring the renderer to retain R in order to know it is safe to discard R. The previous adversary (Schuch, Wolf, Verstraete & Cirac, PRL 98, 140506 (2007)) flagged this for local decidability; C2 makes it worse, because C2 identifies k\* as a property of the ENTIRE complement R, not of local fragments.

**Possible escape**: The renderer knows the coupling structure and could use a heuristic to estimate k\* without computing it exactly. For product environments, k\* = n−1 (trivially known). For scrambled environments, k\* ≈ n/2 (known from Haar statistics). For known circuit classes, k\* could be estimated from the circuit depth. This escape requires H9 to specify a heuristic, which it currently does not. The heuristic would be coupling-structure-dependent, and the renderer would need to know which coupling class it is in. This is an additional specification burden that H9 has not met.

Furthermore, any heuristic that is not exactly k\* risks either (a) discarding when it shouldn't (producing observable deviations from QM, which we don't see), or (b) keeping when it could safely discard (not saving as many resources). Option (a) means the heuristic must be very good—good enough that no experiment in our universe can detect its errors. Option (b) means the heuristic conservatively over-retains, which reduces (but does not eliminate) the resource savings. H9 as stated assumes exact k\*; a heuristic version is a different hypothesis.

## Objection 6C: Interpretation dependence (MINOR)

Under unitary Everett, nothing is ever discarded. H9's "garbage collection" requires actual discard of the purification. A previous adversary graded interpretation-neutral objective discard as FATAL. C6 inherits this: if Everett is correct, H9 is vacuous (no discard happens); if collapse interpretations are correct, H9 describes a physical process (collapse) that may or may not be simulated.

This objection is MINOR because H9 is a claim about what a simulator could do, not about the correct interpretation of QM. Under Everett, the simulator might still discard, but we'd never notice (because the discarded correlations are irrecoverable). Under collapse, the simulator's discard might be the physical mechanism of collapse. H9 is compatible with either interpretation; it just makes different predictions about simulator behavior in each case.

## Objection 6D: "Strictly weaker than H4" is not a relevant comparison (MINOR)

H4 is dead—deferral saves nothing because the thunk must be retained. Showing H9 is "strictly weaker" than H4 means H9 avoids H4's specific failure mode (the thunk retention cost) but may have its own. The relevant question is not "is H9 weaker than H4?" but "does H9 avoid ALL the failure modes that killed H4, or just one?" Objections 6A and 6B show that H9 has its own failure modes (future dependence and computational intractability) that are of the same kind as H4's, just dressed in different clothing.

# C7

**Grade: SERIOUS (the pattern is real, and it is worse than Argus frames it)**

Argus asks whether the "unconstrained against the generic hypothesis, constraining once a policy is specified" pattern is a real structural parallel or pattern-matching.

It is a real structural parallel, and Argus undercalls the problem. Here is the pattern in this case:

1. Generic H9 ("decoherence is garbage collection") is unconstrained: any observation is compatible with some coupling structure.
2. Specified H9 ("the renderer discards when k\* > k\_max") is constraining: it predicts different erasure thresholds for different coupling structures.
3. But the constraints are exactly the constraints of quantum mechanics. Different coupling structures produce different decoherence behaviors in QM, and V(F) = F(ρ_R^0, ρ_R^1) IS QM. H9 adds the INTERPRETIVE claim that this is "garbage collection," which makes no additional physical prediction.

The pattern Argus identifies in METHODS.md—"narrative tidiness across cycles"—is real but misnamed. It is not tidiness; it is **self-sealing**. H9 predicts that discard is safe exactly when QM says it would be safe, which means no observation can distinguish H9 from QM. Every constraint H9 adds (coupling-dependent thresholds, monotone safety with scrambling depth) is a constraint QM already has. The simulation overlay adds zero predictive content.

Argus frames this as a potential failure mode ("narrative tidiness"). It is an actual failure mode: the hypothesis has no empirical consequences beyond standard physics. This is not a minor concern about aesthetics; it is the central objection to H9 as a testable hypothesis.

# WHAT I COULD NOT BREAK

- **C1**: The definition of V(F) and its numerical validation in the product case. Correct and verified.
- **C2's correctness**: The identity V(F) = F(ρ_R^0, ρ_R^1) follows validly from Uhlmann's theorem. The derivation is not wrong; it is just not novel.
- **C2's achievability**: The controlled-unitary + polar-decomposition protocol achieves the bound. This is a valid quantum-eraser construction.
- **C3's data**: The numerical results for all four environment classes. The constant deficit for product environments and growing deficit for all others is consistent with the data and with known scaling laws.
- **C3's headline**: "Constant deficit is a coupling property, not a decoherence property." True for the classes tested.
- **C5's monotonicity**: For the specific circuit architecture and parameters tested, k\* is monotonically non-decreasing with scrambling depth. I have no reason to doubt the data.
- **The numerical accuracy throughout**: Machine-epsilon residuals throughout, with the sqrt(machine epsilon) residual at k=0 honestly reported and correctly explained.

# CREDENCE

**0.20, down from 0.47.**

The downward revision is driven by three independent objections:

1. **C2 is prior art** (not wrong, but not a finding). The "load-bearing technical claim" is Uhlmann's theorem restated. This reduces the novelty of the mathematical framework from "a new identity connecting recoverability to complement indistinguishability" to "a standard QI theorem applied to simulation-hypothesis language." This alone would not justify a large credence reduction—it is a framing issue, not a correctness issue—but it means the technical edifice is thinner than it appeared.

2. **C6 objection 6B (FATAL for the specific claim, SERIOUS for H9 overall)**: k\* is a property of R, the subsystem the renderer wants to discard, and computing it exactly requires exponential resources in |R|. The renderer cannot determine whether it is safe to discard without spending at least as much computation as retaining would cost. This kills the specific claim that k\* is "computable from the present state" as a practical decision procedure. A heuristic version of H9 could survive this, but H9 as stated assumes exact k\*, and the heuristic version is a different (and yet-unspecified) hypothesis.

3. **C7 (SERIOUS)**: H9's physical content reduces to standard QM. The coupling-dependent thresholds, the monotone safety, the identity V(F) = F(ρ_R^0, ρ_R^1)—all are facts about quantum mechanics that hold regardless of whether a simulator exists. H9 adds the interpretive claim "this is garbage collection," which makes no additional testable prediction. This is the self-sealing pattern, and it is not a minor aesthetic concern; it is the central obstacle to H9 being a scientific hypothesis rather than an interpretation.

The credence of 0.20 reflects: the idea is interesting and the formalism is correct, but it is either untestable (if interpreted as a simulation claim) or just standard QM (if interpreted as a physical claim). The computability gap in C6 means H9 does not specify a viable decision procedure for the renderer, and the self-sealing pattern in C7 means H9 has no empirical consequences beyond QM. These are not fatal to the idea that a simulator COULD discard decohered correlations—they are fatal to the idea that this is a distinguishable hypothesis.

I could not break the math (C1-C5). I could not break the data. I could not find a mistake in the derivation of C2. I could break the CLAIM that C2 is a finding, the CLAIM that k\* is efficiently computable, the CLAIM that H9 is purely present-tense, and the CLAIM that H9 has testable content beyond QM.