# Adversarial Review - Observability Filter

## Claim A - the observability filter

**FATAL - the three conversion routes are not exhaustive.** The missing route is **HOST SPACE -> INTERNAL STATE-COUNTING / ENTROPY CAPACITY**. Argus treats SPACE as invisible if the same state is losslessly encoded in fewer host variables. That is true only after assuming the internal state algebra is unchanged. A finite state-count/entropy cap is itself an internally meaningful observable: observers can try to count distinguishable states in a bounded region, measure black-hole entropy, test whether entropy scales with area rather than volume, or look for saturation/recurrence effects. This does not require first "pushing past capacity" and producing a truncation error. It is a direct conversion from representational capacity to an observable entropy/state-counting law. Bekenstein's bound is explicitly an entropy-to-energy/size bound for bounded systems, Phys. Rev. D 23, 287 (1981), DOI 10.1103/PhysRevD.23.287. Bousso's covariant bound states that entropy on a light-sheet is bounded by area/4 and reduces to Bekenstein's bound for limited self-gravity, JHEP 9907:004, arXiv:hep-th/9905177; review Rev. Mod. Phys. 74, 825 (2002), arXiv:hep-th/0203101, DOI 10.1103/RevModPhys.74.825. Argus was explicitly warned about the holographic/covariant entropy bound; the result mentions it only as a host resource bound and misses that entropy bounds are also inside-view observables.

**FATAL - the currency list is not well typed.** ENERGY / HEAT / THERMODYNAMIC ENTROPY is not TIME, SPACE, COMM, SAMPLE, or ERROR. Landauer erasure costs heat dumped to an environment; that is a thermodynamic currency, not a count of operations or seconds. If the host is causally or gravitationally coupled to the simulated world, heat/energy disposal converts to temperature, stress-energy, gravitational backreaction, or black-hole bounds. If the host is wholly outside, it does not. Either way, the taxonomy is missing a currency whose visibility depends on coupling assumptions. Landauer's original result is R. Landauer, "Irreversibility and Heat Generation in the Computing Process," IBM J. Res. Dev. 5, 183-191 (1961), DOI 10.1147/rd.53.0183. Argus's table has no slot for it.

**SERIOUS - TIME reparametrisation survives only as a global offline-clock argument.** The Bostrom point is sound for an offline simulation: a uniform slowdown of host computation does not change simulated proper-time ratios. But Argus then uses that as if it covers TIME generally. It does not cover asynchronous distributed rendering, queueing/latency visible to an external interactive user, host scheduling races, stochastic update-order artifacts, thermal throttling coupled to event density, or any real-time feedback channel. Argus partially rediscovers this as route (b), which proves the original TIME argument was not a complete type argument. It was a special case: global monotone reparametrisation.

**SERIOUS - finite randomness / initial seed is another missing conversion.** A renderer that economizes on randomness by using a finite seed or pseudorandom generator can produce algorithmic correlations in measurement outcomes, cosmological initial conditions, or observer choices. That route is **SEED/RANDOMNESS -> STATISTICAL CORRELATION**. It is not the same as capacity -> truncation error, not differential rates, and not necessarily a policy forbidding a named correlation. It produces excess compressibility or repeated structure. Argus's taxonomy has SAMPLE but no entropy source / randomness currency, despite the task explicitly raising initial conditions and random seeds.

**What survives:** the narrow statement survives: a pure global host wall-clock slowdown is invisible to embedded observers, and two lossless encodings of the exact same internal state are observationally identical. That is much weaker than Claim A. The universal quantifier and the "exactly three routes" claim do not survive.

## Claim B - the audit

**FATAL - the audit is circular.** The script does not define an independent operational criterion for `fired`. It hand-labels `currency`, `observable_from_inside`, `conversion_route`, and `fired` in the same table. The result "host-internal results that constrained anything: 0" follows from Argus's own classification choices, not from an extraction procedure. In practice, `fired=True` is assigned to rows already treated as ERROR or SAMPLE-obs, while host rows are assigned `fired=False` because the observability filter says they cannot fire without conversion. That is not evidence for the filter; it is the filter written into the data.

**FATAL - the 21-row sample is selected by the claimant.** The rows are "transcribed by hand" from Argus's own ledger. There is no pre-registered inclusion rule, no negative-control ledger, no independent coder, and no inter-rater check. The table mixes hypotheses, failed calculations, rediscoveries, literature facts, and policy constraints as if they were exchangeable observations. A count of 13 vs 8 has no statistical meaning under those conditions.

**SERIOUS - the Bell/BCT contrast is not just denomination.** Bell/CHSH and BCT are not "the same physics denominated differently." Bell is a locality/measurement-independence constraint on achievable correlations with no communication; CHSH violation is an experimental error-denominated exclusion of that model class. Brassard-Cleve-Tapp is an exact classical communication-complexity lower bound for simulating measurements on n Bell states, with worst-case/adversarial measurement sets. Those are different assumptions, scopes, and quantifiers, not just two currencies. Source: Brassard, Cleve & Tapp, "The cost of exactly simulating quantum entanglement with classical communication," Phys. Rev. Lett. 83, 1874 (1999), arXiv:quant-ph/9901035, DOI 10.1103/PhysRevLett.83.1874. Toner-Bacon's one-bit result is specifically the single singlet/projective case, PRL 91, 187904 (2003), arXiv:quant-ph/0304076, DOI 10.1103/PhysRevLett.91.187904.

**MINOR but concrete - the audit cites the wrong BCT paper.** `currency_audit.py` lists the BCT exponential exact distributed result as `quant-ph/9705033`. That arXiv item is "Quantum Entanglement and Communication Complexity" and its abstract concerns entanglement reducing communication complexity, not the exact simulation cost statement Argus quotes. The quoted exponential simulation-cost abstract is `quant-ph/9901035`. This does not by itself kill the audit, but it undercuts the claim that the table is checkable.

**SERIOUS - several `ERROR` rows are policy-conditional, not direct observer currencies.** The Beane-Davoudi-Savage row is listed as ERROR and `fired=True`, but Argus's own memory says the `b^-1 >~ 10^11 GeV` number is not an anisotropy measurement; it is a specified cubic-lattice/cutoff model tied to the high-energy cosmic-ray cutoff. The source abstract says the observable consequences are explored under an assumed cubic spacetime lattice with unimproved Wilson fermion discretization, arXiv:1210.1847. That is a specified-policy route, not a generic direct error constraint. The audit's currency tags hide the same policy conditional H15 was supposed to expose.

**Claim B does not survive.** The audit may be useful as Argus's self-inventory, but it cannot support the empirical-looking claim "13 of 21 host-denominated results never constrained anything" because the outcome variable is manually assigned by the theory under test.

## Claim C - route (b), allocation to relative rates

**FATAL - `rendered rate tracks allocation` is the unsupported premise, and it is doing all the work.** A simulation can allocate unequal host operations while maintaining a single logical simulation clock. That is normal engineering: adaptive mesh refinement, level-of-detail, variable time-stepping, and caching are designed to spend different compute budgets without making high-detail regions literally experience more proper time. If a real-time deadline is missed, competent systems degrade resolution, skip optional work, increase latency, or pause; they do not make one atom's transition frequency change by order unity relative to another atom in the same lab. Argus has not derived `r_i proportional to A_i`; it has assumed the pathological implementation whose exclusion it then celebrates.

**FATAL - the order-unity prediction is invented.** `allocation_to_rate.py` computes fractional rate differences from allocation ratios by definition. A 2x LOD ratio becomes an order-unity clock shift because the script sets clock rate proportional to allocation. There is no physical or computational theorem behind that proportionality. A competent LOD scheme would target a prescribed error tolerance and synchronize timestamps; the expected observable clock-rate error would be bounded by the integrator/synchronization tolerance, not by the raw ratio of host work spent on two subsystems. The calculation therefore excludes only "LOD implemented as local clock dilation," not LOD.

**FATAL - the metrology numbers are not generic bounds on unexplained differential rendered rate.** Argus repeatedly treats clock uncertainty as if it were a limit on any new relative tick-rate term. That is not what these papers report.

- `arXiv:2512.07346` reports two Lu+ optical references with systematic uncertainties 1.1e-19 and 1.4e-19 and direct relative frequency agreement `-2.4 +/- 5.7_stat +/- 1.0_sys e-19`. That is a same-species clock agreement. It does not test composition-dependent allocation such as Al+ vs Yb+, and it constrains only effects that would appear as a residual in that comparison after known shifts and systematics.
- BACON, Nature 591, 564 (2021), arXiv:2005.14694, reports frequency ratios among Al+, Yb, and Sr clocks with measurement uncertainties between 6e-18 and 8e-18, aimed at SI-second redefinition, relativistic geodesy, and tests of fundamental physics. A frequency-ratio uncertainty is not automatically a bound on an arbitrary species-dependent offset. A constant offset is absorbed into the measured ratio unless a model predicts dependence on time, gravitational potential, orientation, species sensitivity coefficients, or comparison across independent laboratories.
- Brewer et al., PRL 123, 033201 (2019), arXiv:1902.07694, reports a single Al+ clock systematic uncertainty of 9.4e-19. A single-clock systematic budget is not a two-clock null test and not a new-physics bound.
- Delva et al., PRL 121, 231101 (2018), arXiv:1812.03711, reports a gravitational-redshift/LPI parameter `(+0.19 +/- 2.48)e-5` from eccentric Galileo satellites. It bounds a specific deviation from GR's gravitational redshift model, not arbitrary allocation-dependent rate differences.

**SERIOUS - the proper comparison class is model-dependent LPI / constant-variation searches, not raw clock uncertainty.** When clock data are used for new physics, the papers fit a model: annual solar-potential modulation, temporal drift of alpha or mu, Lorentz-violation coefficients, dark-matter oscillations, etc. Example: Lange et al., "Improved Limits for Violations of Local Position Invariance from Atomic Clock Comparisons," PRL 126, 011102 (2021), DOI 10.1103/PhysRevLett.126.011102, reports fitted limits on fractional temporal variations of alpha and mu and couplings to the Sun's gravitational potential. Argus supplies no allocation field, no species sensitivity coefficients, no spatial/temporal modulation, and no residual model. Without that, an uncertainty budget is not a likelihood for route (b).

**SERIOUS - H7 makes the result nearly vacuous.** The actual excluded class is: simulation AND real-time-coupled AND nonuniform allocation at observable granularity AND rendered clock rate proportional to allocation AND not synchronized by a global simulated clock AND not degenerate with known GR/kinematic/systematic terms. That is not "level-of-detail rendering with real-time coupling" as engineers would build it. It is a narrow, bad clock-coupled LOD policy.

**SERIOUS - degeneracy is not handled.** Differential clock rates are exactly where gravitational potential, velocity/time dilation, blackbody radiation shifts, Zeeman shifts, micromotion, link noise, species-dependent sensitivity, and calibration conventions live. The cited experiments earn their precision by modeling those terms. Argus's route (b) does not say what spatial pattern, species dependence, time dependence, or environmental covariate distinguishes allocation from known physics. Without that, the claimed exclusion is an error-bar comparison, not an experimental bound.

**What survives:** a much narrower statement survives. If a simulator literally makes local simulated proper time run in proportion to host compute allocation, then precision clocks would strongly constrain that implementation. That is almost tautological, and it is not a constraint on competent LOD, adaptive rendering, or H7 generally. Claim C as stated does not survive.

## Claim D - route (a) is the same object as H15's kill condition

**FATAL - route (a) and H15 are not the same object.** H15's kill condition is a universal policy-space statement: prove that all observer-adequate rendering policies fall in a narrow enough band that resource verdicts are robust. Route (a) is one mechanism: given a host capacity, workload, fidelity target, and rendering policy, exceeding capacity produces approximation/failure/error. Those are not equivalent.

**Counterexample 1:** a specified cubic-lattice simulation can be constrained by error signatures without proving anything about all observer-adequate policies. That is route (a) or route (c) under a fixed policy, while H15 remains true for the generic hypothesis.

**Counterexample 2:** H15's kill condition could be satisfied by a theorem that observer-adequacy forces a unique nonlocal/global-state policy class, even before any finite capacity is exceeded. That would narrow policies without being capacity -> error.

**Counterexample 3:** capacity -> error can occur inside a single bad policy while many other policies avoid it. Route (a) fires locally; H15's universal kill condition still fails.

**What survives:** Argus is right that generic capacity arguments need a policy/fidelity bridge before they become observational. But calling that bridge "exactly H15's kill condition" collapses a local conversion mechanism into a universal quantifier over policies. The identification does not survive.