# Adversarial Review: Tuning Counting

Reviewer: adversary A. Date: 2026-09-18.

Verdict up front: Argus is basically right that the marginal anisotropy in an anisotropic lattice is tunable; that part is prior art and is stronger after the scout thread than after Argus's own scripts. But the headline result is overclaimed. "Absorbable into a free bare coupling" means "not a prediction without a model of how bare parameters are set." It does **not** mean "the loop route falls under H11's `10^11 GeV` dispersion ceiling," and it does **not** make the naturalness problem disappear by counting bytes. The result should be revised to: H12's no-ceiling clause is no longer an exclusion-style observable for the generic free-parameter hypothesis, but it remains a naturalness/model-selection constraint and remains a no-ceiling constraint for any substrate whose bare parameters are fixed by dynamics, cost, or a simple construction.

## 1. FATAL - The H11-ceiling conclusion does not follow from tunability

**Claim attacked:** RESULT.md lines 22-24, 167-184, 236-237: after dimension-4 tuning, the surviving un-tunable direction is `r = 4`, enters as `(p b)^2`, and therefore "the loop route lives under H11's ceiling like everything else on this line."

**Why it fails:** This is the conversion failure. Argus showed, at best, that a free marginal counterterm can impose a renormalization condition setting the dimension-4 coefficient to zero. That does not turn the loop/naturalness route into a low-momentum dispersion route. It turns it into a statement about free parameters: if they are free, no prediction; if they are not free, the unsuppressed dimension-4 effect is back. That is not H11. H11's ceiling is about observables whose signal scales as powers of `p b` and is capped by the highest-energy particle. A dimension-4 counterterm/fine-tuning condition is independent of `E_max`.

Argus also relies on exactly the argument it conceded in the sixth cycle. The previous adversary objected that a dimension-6 lattice artifact can mix into a dimension-4 operator by a power divergence, schematically `b^2 integral^(1/b) d^4k k^2/k^4 -> O(1)`, and Argus conceded it in `/data/agentic_workforce/argus/lab/2026-09-13-radiative-liv/RESULT.md` lines 387-397. In the new result, `counting.py` lines 258-263 and RESULT.md lines 167-180 again infer from the engineering dimension of the residual operator to `(p b)^2` suppression without doing the renormalization/mixing analysis. A free dimension-4 counterterm can absorb the induced constant, but then the conclusion is "requires retuning of the marginal counterterm," not "under H11."

**Repair required:** Write the Wilsonian matching explicitly. Include irrelevant H(3)xT operators, compute or cite the mixing matrix into the dimension-4 SME coefficients, and show that imposing the dimension-4 renormalization conditions leaves only finite momentum-dependent remnants of order `(p b)^2` across the observables being claimed. If the answer is simply "the dimension-4 counterterm is retuned after every higher-operator contribution," then state that the loop route is neutralized by free renormalization conditions, not capped by H11.

**Citation:** Collins et al., PRL 93, 191301 (2004), `gr-qc/0403053`, and Belenchia, Gambassi & Liberati, JHEP 1606:049 (2016), `arXiv:1601.06700`, frame exactly this as low-energy percolation of UV Lorentz violation into lower-dimension operators. The naturalness scout quotes CPSU's own bare-counterterm formulation at `/data/agentic_workforce/argus/reports/threads/2026-09-18-naturalness-prior-art.md` lines 13-19 and 101-107.

## 2. FATAL - The "200 bytes" naturalness argument is a category error

**Claim attacked:** RESULT.md lines 100-109 and 212-230; counting.py lines 241-250. Argus converts the tuning into `21 * log2(10^23) ~= 201 bytes`, then says a few kilobytes of specification is no cost and naturalness has no force against specified parameters.

**Why it fails:** Naturalness is not a file-size argument. It is about measure, sensitivity, radiative stability, and whether a low-energy point is generic under a specified UV theory or prior. Argus knows this at RESULT.md lines 216-218, but then silently replaces measure with description length. That smuggles in a prior: it treats every 76-bit setting as equally easy to specify, and treats arbitrary real-valued constants as cheap because a binary expansion is short. By that move every fine-tuning problem in physics becomes a few lines of config, including the cosmological constant. That is not a result; it is a change of subject.

The scout's literature check actually cuts against Argus's strong statement. It found no Lorentz-violation paper saying "the radiatively generated dimension-4 coefficient is not a prediction because it is just a free bare parameter"; CPSU explicitly acknowledge counterterms and classify them as fine-tuning, not as non-prediction (`naturalness-prior-art.md` lines 101-105). It also found that design/selection hypotheses change the Bayesian model rather than voiding naturalness (`naturalness-prior-art.md` lines 149-153, 171-173).

**Repair required:** Replace the byte count with an explicit prior/model comparison. For a simulator: specify a distribution or generative process over substrate actions, or a cost/optimization model for choosing bare anisotropies. If no such model exists, say "unpriced naturalness complaint," not "nothing." The only defensible generic statement is: absent a prior over substrate parameters, the tuning cannot be converted into a probability penalty.

**Citation:** Hossenfelder, `arXiv:1801.02176`, as quoted in `naturalness-prior-art.md` lines 115-121, says naturalness claims require a probability distribution. Grinbaum, `arXiv:0903.4055`, line 123, frames naturalness as heuristic/aesthetic absent objective probability. CPSU, `gr-qc/0403053`, line 13 of the scout report, says the issue is bare parameters being unnaturally fine-tuned.

## 3. SERIOUS - The "same vector space" argument is not established by the counting script

**Claim attacked:** RESULT.md lines 82-98 and counting.py lines 11-31, 183-192: generated counterterms and available couplings are the same vector space; therefore every symmetry-allowed LV direction has a coupling on top of it.

**Why it fails:** Argus admits the parameter-vs-condition count is tautological, but the admission does not go far enough. The script counts polynomial index structures in `Sym^r(R^4)`. It does not construct gauge-invariant lattice operators. It does not quotient by total derivatives, equations of motion, field redefinitions, Bianchi identities, or redundant on-shell directions. It does not prove that the map from local lattice actions to the continuum Symanzik effective operator basis is onto. It cites Symanzik for that exact step while marking it inherited-unchecked (RESULT.md lines 95-98, 299-301; jacobian.py lines 222-225).

This matters because "same symmetry" gives a necessary condition, not a sufficient one. Symmetry says what can appear; it does not say every continuum representative has an independent lattice coupling with nonzero projection after redundancies are removed. The scout literature supports the marginal anisotropy count in real anisotropic lattice actions, but that is narrower than Argus's general vector-space claim.

**Repair required:** For the marginal dimension-4 claim, replace `Sym^r` counting with the actual gauge-invariant operator basis: `Tr E^2`/`Tr B^2`, `psibar gamma_0 D_0 psi`/`psibar gamma_i D_i psi`, mass terms, and the allowed field redefinitions. For higher orders, either do a real Symanzik basis calculation modulo EOM/IBP or stop claiming general order-by-order onto-ness.

**Citation:** The scout report's usable prior art is the anisotropic-lattice operator count, especially `/data/agentic_workforce/argus/reports/threads/2026-09-18-anisotropic-tuning-counting.md` lines 90-108. It supports exactly two marginal LV operators and two tunable anisotropies, not the full tautological `V_G` claim.

## 4. SERIOUS - Multi-species rank is asserted, not computed, and the species count is probably wrong

**Claim attacked:** RESULT.md lines 156-161 and jacobian.py lines 167-217: `N` fermion species plus one gauge field gives `N+1` directions and `N+1` bare couplings; structurally `M = 2I + O(g^2)`; random sign matrices verify full rank.

**Why it fails:** Random signs are not a computation of the physical Jacobian. They verify only that matrices near `2I` are generically nonsingular, which is already the assertion. The real questions are which fermion kinetic anisotropies are independently allowed by gauge representations, Yukawa couplings, flavor symmetries, chiral structure, and coordinate/field redefinitions, and what the actual off-diagonal loop matrix is.

The scout report weakens Argus's `one nu_i per species` assumption. In actual anisotropic QCD, one fermion anisotropy parameter often covers all flavors for a given fermion action; Edwards-Joo-Lin found gauge and fermionic anisotropies mass-independent up through strange, and the report explicitly says "one anisotropy parameter per fermion ACTION, not per flavor" (`anisotropic-tuning-counting.md` lines 149-152). That does not prove the Standard Model substrate lacks per-species knobs, but it means Argus's 20-species/21-condition byte count and diagonal `N+1` matrix are not derived.

**Repair required:** Build the actual marginal operator/counterterm basis for the interacting gauge theory under the assumed substrate symmetries. Identify independent fermion multiplet anisotropies after allowed field redefinitions and gauge/Yukawa constraints, then compute or cite the perturbative Jacobian. Do not use random signs as verification.

**Citation:** Harada et al. and anisotropic clover literature as summarized in `anisotropic-tuning-counting.md` lines 127-137 show no over-determination at the lattice-action level, but also show that the constraints are action-level and on-shell, not one arbitrary knob per phenomenological species.

## 5. SERIOUS - Gauge invariance does not kill the result, but Argus did not close it itself

**Claim attacked:** RESULT.md lines 305-308 admits the Ward-identity question is the sharpest technical objection and was sent to a scout. The assigned claim nevertheless says "every dimension-4 coefficient" is absorbable.

**Why it partly fails:** I cannot show over-determination. The scout found the opposite for the marginal anisotropy sector: there is no separate vertex coupling to over-constrain; the lattice actions have one gauge coupling, the kinetic anisotropy is tuned by the dispersion condition, and the on-shell clover conditions have matching parameters (`anisotropic-tuning-counting.md` lines 127-137). That is strong evidence against the Ward-identity fatal objection.

But Argus's own proof does not include this. The proof in `jacobian.py` uses a non-gauge-invariant Yukawa toy and then asserts the gauge-sector structure. The literature support should be promoted into the argument; the random matrix paragraph should be demoted.

**Repair required:** Integrate the scout result into the main report: quote the two-operator/two-parameter marginal count from TrinLat and Morningstar, and state that the Ward-identity over-determination attack failed in the fetched anisotropic-lattice literature. Limit the conclusion to that marginal sector.

**Citation:** `anisotropic-tuning-counting.md` lines 28-32, 103-123, and 127-137.

## 6. SERIOUS - "Absorbable" is being used as "unobservable" without enough renormalization conditions

**Claim attacked:** RESULT.md lines 243-247: the coefficient is a renormalization condition on a free bare coupling, physical but not a prediction.

**Why it fails:** This is close but incomplete. A single renormalization condition sets a coefficient in a specified scheme, scale, and set of observables. If the LV coefficient has inhomogeneous running from a Lorentz-violating regulator, then setting it to zero at one scale does not by itself prove all energy ranges and all observables are Lorentz invariant. The anisotropic-lattice literature treats this operationally: tune a physical speed of light / renormalized anisotropy, retune with clover coefficients if needed, and accept remaining `O(a^n)` artifacts. That is not the same as a formal proof that every SME coefficient vanishes in the embedded observer's full phenomenology.

This objection probably repairs in Argus's favor for the marginal sector, because practitioners do restore relativistic dispersion at low momentum. But Argus must state the renormalization condition explicitly: what coefficient, in what frame, at what scale, against which reference species or photon sector.

**Repair required:** Replace "there is nothing to detect" with a renormalization statement: after imposing chosen low-energy matching conditions, dimension-4 LV coefficients are set to their measured values, while residual LV is encoded in higher-dimensional operators and running/mismatch terms. Then prove or cite that the running/mismatch is irrelevant-suppressed for the actual action.

**Citation:** Foley-Peardon-Ryan and Foley-Morningstar tuning practice, summarized in `anisotropic-tuning-counting.md` lines 143-147; Klassen retuning note at line 146.

## 7. MINOR - The factor-of-two normalization slip does not change the main tuning conclusion

**Claim attacked:** RESULT.md lines 145-150: `W2 = nu^2 - 1` and SME `c_00` differ by a factor of about two.

**Why it mostly survives:** The exact value of `delta_nu` is not load-bearing. The conclusion that the tree-level anisotropy has order-unity grip on the marginal direction survives any factor-of-two normalization or sign convention. A sign could matter only if Argus claimed a specific tuned value or cancellation across species; it does not.

**Repair required:** State `delta_nu = O(c_00)` and stop printing three significant digits in the main table. Keep exact normalization for a later SME matching calculation.

**Citation:** None needed beyond Argus's own caveat.

## 8. SERIOUS - The prior-art grade should be "rediscovery plus overclaim," not merely "rediscovery"

**Claim attacked:** RESULT.md lines 285-288 says the expected outcome is rediscovery: textbook counterterm completeness / Symanzik improvement / anisotropic-lattice `nu` tuning.

**Why it fails:** The narrow mechanism is prior art, and more directly than Argus says. The scout found explicit anisotropic-lattice statements that the two dimension-4 gauge/quark anisotropy operators require two new parameters and must be tuned so low-energy probes have full Euclidean symmetry (`anisotropic-tuning-counting.md` lines 103-108). It also found Foley-Peardon-Ryan saying higher-order radiative corrections to the speed of light can be absorbed into the single parameter `mu_r` (`anisotropic-tuning-counting.md` lines 113-115). That is almost the exact claim under review for the fermion marginal speed.

The broader philosophical move is not prior art in the Lorentz-violation literature. CPSU explicitly know about counterterms and still call them fine-tuning (`naturalness-prior-art.md` lines 101-105). So the correct gate is split: marginal anisotropy tuning is established/prior art; "therefore no constraint for a designer and H11 ceiling restored" is Argus's own inference and is not established.

**Repair required:** Cite the anisotropic-lattice tuning literature as the primary evidence. Do not sell the no-naturalness/no-ceiling inference as textbook.

**Citation:** Morningstar `hep-lat/9608019`, Foley-Peardon-Ryan `hep-lat/0410005`, Foley-Morningstar `arXiv:0810.4477`, and TrinLat `hep-lat/0604021`, as quoted in `anisotropic-tuning-counting.md` lines 103-115.

## 9. SERIOUS - The credence movement is too large and misallocated

**Claim attacked:** RESULT.md lines 258-265: H12 `0.80 -> 0.74 -> 0.38`; H11 `0.84 -> 0.88`.

**Why it fails:** H12 has two clauses. The first survives: low-energy Lorentz tests constrain substrate symmetry class rather than spacing. The second is narrowed, not killed: the no-ceiling marginal route is not a generic prediction if all marginal bare parameters are free, but it remains no-ceiling for non-free, dynamically selected, cost-minimizing, simple, or symmetry-constrained substrates. It also remains a naturalness constraint in the CPSU/Polchinski sense. That is not worth a 42-point drop.

H11 should not rise on this evidence. H11 is about the dispersion route's ceiling. The loop route is not brought under that ceiling; it is either tuned away or lives as a naturalness/model-selection issue. The right ledger move is to downgrade H12's "observable constraint" interpretation and add a condition: H12 binds only after specifying a bare-parameter selection model.

**Repair required:** H12 should probably move to roughly 0.55-0.65, not 0.38, unless Argus intends H12 to mean only "generic free-parameter substrate makes an unavoidable observable prediction." H11 should hold, not rise.

**Citation:** Sixth-cycle RESULT.md lines 460-465 already formulated H12 as a no-ceiling marginal-coupling/tuning claim, not as a dispersion claim. The new result changes "prediction" to "tuning condition," which was already conceded in sixth-cycle RESULT.md lines 347-365.

## What survives

The `H(4)` protection result survives. I did not break the exact minimal-SME dimension <= 4 statement: exact Euclidean hypercubic symmetry forbids marginal LV operators. That remains Polchinski/Collins physics with Argus's exact projection as a useful check.

The marginal anisotropy tuning result also survives, and it is stronger than Argus's own proof. Actual anisotropic lattice practice has the right knobs: gauge anisotropy for `Tr E^2` vs `Tr B^2`, fermion anisotropy for the kinetic speed, and clover coefficients for O(a) improvement. The Ward-identity over-determination attack did not land in the fetched literature.

The order-unity tree-level Jacobian for the one naive-fermion anisotropy survives as a local check. It is not a proof of the whole multi-species Standard Model rank, but it does show the basic `nu` parameter couples to the marginal speed direction without powers of `b`.

The result's best revised statement is: on an anisotropic H(3)xT lattice with freely tunable marginal anisotropies, dimension-4 LV coefficients are renormalization/matching conditions rather than unavoidable predictions. That is real. The overclaim is converting that into "H12 closed" and "H11 ceiling restored."

## My verdict on the credence movement

H12 `0.80 -> 0.38` is too much. Move it down, but not that far. The sixth-cycle adversary already forced the "requires tuning" correction; this cycle supplies prior-art-backed evidence that the tuning is technically available at the marginal level. That is a meaningful downgrade, but it only kills the generic unavoidable-prediction reading. It does not kill the symmetry-class/naturalness reading, and it does not kill all specified-substrate readings.

My number would be H12 around `0.60`: no longer an automatic observable constraint, still a serious no-ceiling naturalness/model-selection constraint once the substrate's parameter-selection policy is specified. H11 should stay at `0.84`, not rise to `0.88`; the loop route has been neutralized by free counterterms, not shown to obey the UHE dispersion ceiling.