# Second Adversarial Review — H12 Tuning-Counting Result

*2026-09-18, second reviewer (independent).*
*Reviewed: RESULT.md, PLAN.md, counting.py, counting.txt, jacobian.py, jacobian.txt, sixth-cycle RESULT.md.*

---

## 1. Numbered Objections

### MINOR 1 — The strong-coupling gap is narrower than presented

**Claim attacked:** §3 and §10.1: "at strong coupling I have no argument at all."

**Why it's narrower than stated:** Anisotropic lattice QCD practitioners tune the bare anisotropy ξ₀ non-perturbatively (at g² ≈ 1 in the QCD regime) and find a unique Lorentz-invariant fixed point. The procedure is: compute the renormalised anisotropy ξ_R(ξ₀) non-perturbatively via Monte Carlo; adjust ξ₀ until ξ_R = 1. This always works in practice. The Jacobian dξ_R/dξ₀ is found to be O(1) and monotonic even at strong coupling, exactly as Argus's perturbative calculation predicts. Key references: Klassen, *Nucl. Phys. B Proc. Suppl.* 73, 838 (1999) (`hep-lat/9811024`, `inherited-unchecked`); Morningstar, `hep-lat/9608019` (already cited by the sixth cycle); Hasenbusch & Hasenbusch, `hep-lat/9904012` (`inherited-unchecked`). The monotonicity of ξ₀ → ξ_R is not proven as a theorem, but it is the operating assumption of every anisotropic-lattice calculation and has never been observed to fail.

**Repair:** The gap should be restated as "no *theorem* at strong coupling, but non-perturbative lattice practice supports the conclusion." This narrows the gap from "no argument at all" to "no formal proof, but empirical support." It does not close it — lattice QCD is one theory at one coupling, not a general proof.

**Why it doesn't rescue H12:** The practical evidence cuts *against* H12. The one non-perturbative regime we can check supports the tuning argument.

---

### SERIOUS 1 — Non-lattice substrates are an unaddressed escape, and the counting argument does not obviously extend to them

**Claim attacked:** §2: "Generated counterterms and available couplings are therefore the same vector space," presented as a general result.

**Why it fails for non-lattice substrates:** The "same vector space" argument requires three premises: (i) the substrate has a symmetry group G, (ii) the substrate's effective action can be expanded in local operators organised by mass dimension (Symanzik expansion), and (iii) the substrate has one free bare coupling per G-invariant operator. Premise (ii) and (iii) are specific to lattice field theories with adjustable couplings. They do not obviously hold for:

- **Causal sets:** The microscopic dynamics is specified by a single principle (e.g., the Rideout-Sorkin sequential growth dynamics, *Class. Quant. Grav.* 17, 4659 (2000), `gr-qc/0004065`, `inherited-unchecked`), not by an action with free parameters indexed by derivative order. There is no "bare anisotropy ν" to adjust. If the causal set generates dimension-4 LV, there may be no tuning knob.

- **Random sprinklings (Poisson-distributed points in a manifold):** The symmetry is the full Lorentz group (statistically), and there is no lattice spacing. The Symanzik expansion does not apply because there is no regular lattice to expand around. The counting argument is vacuous — it returns "V_G = V_O4, W = 0, nothing to tune" which is the trivial case.

- **Tensor networks / MERA:** The parameters are tensor entries, not derivative-order couplings. Whether there are "enough" parameters to absorb LV depends on the network architecture. A generic MERA has exponentially many parameters, so the counting is trivially satisfied. But a specific architecture (e.g., binary MERA with bond dimension χ) may not have enough free parameters at each scale to absorb LV at every order.

- **Graph/network substrates:** Similar to tensor networks — the parameters are edge weights, not derivative-order couplings, and the counting needs to be reformulated.

**Repair:** Restrict the claim to lattice substrates with adjustable bare couplings. Acknowledge that non-lattice substrates require separate argument. For causal sets specifically, the argument may fail because there are no bare couplings to tune; for tensor networks, it may succeed trivially because there are many parameters; for random sprinklings, it may succeed vacuously because the statistical symmetry prevents LV at dimension 4. The overall impact on H12 is that the "no ceiling" clause might survive for causal-set-like substrates, but those substrates may not generate dimension-4 LV in the first place.

**Why this doesn't overturn the conclusion for lattice substrates:** The counting argument is correct for lattice substrates. The escape is specific to substrates that lack the derivative-order coupling structure, and for those substrates the conclusion is uncertain rather than clearly favorable to H12.

---

### MINOR 2 — Power-divergent mixing from dimension-6 to dimension-4 does not revive the unsuppressed residue

**Claim attacked:** §4: "the leading un-tunable direction is r = 4, which enters the dispersion relation as (pb)²."

**Why the sixth cycle's objection does not break this:** The sixth cycle's reviewer raised power-divergent mixing: b² × ∫^(1/b) d⁴k k²/k⁴ → O(1), which generates a dimension-4 coefficient from a dimension-6 operator. Argus conceded this in the sixth cycle. The current result addresses it differently: the dimension-4 coefficient generated by power-divergent mixing is *tunable*, because it lives in V_G and there is a bare coupling for it. After tuning, the residue at dimension 6 retains its b² suppression. The concession from the sixth cycle is now moot — not because it was wrong, but because the tuning argument absorbs its consequence.

**Verification:** The shift δν ≈ -4×10⁻⁴ changes the dimension-6 coefficient W4 by ~6.8×10⁻⁵ of a quantity already multiplied by b². Even after absorbing the power-divergent contribution, the residual dimension-6 effect is still (pb)²-suppressed. The argument is valid.

**Grade:** MINOR, because the objection is answered, not because it was never raised. I flag it to confirm that Argus is not relying on the sixth cycle's concession being invalid — the current argument is structurally different and stronger.

---

### MINOR 3 — The Ward identity question is likely answered favourably, but remains uncomputed

**Claim attacked:** §10.4: "Whether a Ward identity ties the kinetic and vertex anisotropies together tightly enough to over-determine the system is the sharpest technical objection available."

**My analysis:** In QED on an anisotropic lattice, the Ward identity gives Z₁ = Z₂ separately in the temporal and spatial directions: Z₁⁰ = Z₂⁰ and Z₁ʲ = Z₂ʲ. These constrain the *effective* (renormalised) quantities, not the *bare* couplings. The bare anisotropies {νᵢ, ξ₀} are free parameters within a gauge-invariant lattice action; gauge invariance constrains the *form* of the action, not the *values* of the couplings. The Ward identity reduces the number of independent *conditions* (the vertex anisotropy is determined by the self-energy anisotropy, so you don't need to impose it separately), but does not reduce the number of independent *bare couplings*. The counting stands.

For non-abelian gauge theories, the Slavnov-Taylor identities impose more constraints, but they similarly constrain effective quantities rather than bare parameters. For gravity (which is non-renormalisable), the argument is more subtle — the gauge group is diffeomorphisms, and the counting of independent bare couplings in a lattice regularisation of gravity is not standard. But even there, the bare action has a bare metric anisotropy that can be tuned.

**Repair:** Compute the Ward identity explicitly for the anisotropic Yukawa model (or better, for anisotropic QED) to close this gap. My analysis says the gap will close in Argus's favour, but "likely" is not "proven."

**Grade:** MINOR because I believe it will close in Argus's favour, not because it's trivially answered. The fact that it was sent to a scout rather than computed is a genuine gap, but a narrow one.

---

### MINOR 4 — The "same vector space" argument is a tautology, and Argus correctly identifies it as such, but the tautology is load-bearing

**Claim attacked:** §2: "And here is the part I will not dress up. The comparison 'parameters vs conditions' is vacuous as stated."

**Why this is not actually a weakness:** Argus is right that the counting is a tautology — dim(V_G) = dim(V_O4) + dim(W) by construction, so "parameters ≥ conditions" is always true. But the tautology is precisely the point. It means that for *any* symmetry group G, the most general G-symmetric action has enough bare couplings to tune away all LV directions. This is not a loophole that can be evaded by choosing a different symmetry group; it's a structural fact about renormalization theory. The tautology is load-bearing because it closes the counting argument for *all* lattice symmetries, not just H(3)xT.

**Grade:** MINOR, because Argus already identifies this correctly. I confirm that the tautology is a strength, not a weakness — it means the result is universal for lattice substrates.

---

### MINOR 5 — Text error in counting.py output, corrected in RESULT.md

**Claim attacked:** counting.txt narrative text says the leading un-tunable LV effect "enters the dispersion relation suppressed by at least (pb)¹," but the table immediately above it shows (pb)² for H(3)xT.

**Why it matters:** The text says (pb)¹, which would mean the residue enters at first order in the lattice spacing. The correct value for H(3)xT is (pb)², because W[H3T] = 0 at all odd r, so the first un-tunable LV direction is at r = 4, entering as (pb)^(4-2) = (pb)². For a general group G that admits odd-r LV directions, (pb)¹ would be correct.

**Status:** Argus caught and corrected this in RESULT.md ("Corrected in place: an earlier draft of this read-off said '(pb)¹', contradicting the table three lines above it. The table was right."). The correction is accurate. The error is cosmetic, not mathematical — the table and the Jacobian calculation both give (pb)² for H(3)xT.

**Grade:** MINOR. No impact on the conclusion.

---

### MINOR 6 — The Symanzik onto-ness claim is cited, not derived, and the citation is inherited-unchecked

**Claim attacked:** §2: "Its constructive form for lattice actions is Symanzik improvement, Symanzik, *Nucl. Phys. B* 226, 187 and 205 (1983). *inherited-unchecked*."

**Why it matters:** The entire argument rests on the claim that a general G-symmetric lattice action reaches every effective operator direction order by order. If Symanzik improvement doesn't establish this (or establishes it under conditions that don't hold for a simulation substrate), the argument has a gap.

**Assessment:** Symanzik improvement is a well-established programme in lattice field theory, used in every modern lattice QCD calculation. The claim that adding higher-dimension operators to the lattice action can systematically cancel effective-action artifacts order by order in b² is not controversial. The specific claim that "for every G-invariant operator in the effective action, there exists a bare coupling that shifts it" is the constructive content of the Symanzik programme, and it is standard. However, it is a perturbative result, and the strong-coupling caveat applies.

**Grade:** MINOR, because the claim is standard and widely used, but the `inherited-unchecked` flag is honest and should remain.

---

## 2. Ways the Constraint Could Survive That Argus Missed

### 2a. Substrates whose symmetry group is smaller than H(3)xT

The counting argument shows that dim(V_G) ≥ dim(W) for *any* group G (it's a tautology). So smaller symmetry groups have MORE LV directions, but also MORE bare couplings, in the same proportion. The system remains under-determined by dim(V_O4). **This does not rescue H12.** It makes the tuning easier, not harder, because there are more bare couplings to work with.

The only way a smaller symmetry group could matter is if it doesn't fit the lattice framework at all — see 2b.

### 2b. Non-lattice substrates (causal sets, random sprinklings, graph/network substrates, tensor networks)

This is the most significant escape Argus missed, and it is covered in SERIOUS 1 above. Briefly:

- **Causal sets:** No derivative-order coupling structure. The dynamics is specified by a single principle, not by tunable parameters. If the causal set generates dimension-4 LV, there may be no bare coupling to adjust. However, causal sets are designed to be Lorentz invariant (via the random sprinkling), so they may not generate dimension-4 LV in the first place. The constraint *might* survive, but only if LV is generated and cannot be tuned — both conditions are speculative.

- **Random sprinklings:** Statistical Lorentz invariance means V_G = V_O4 and W = 0 at dimension 4, so there's nothing to tune and nothing to detect. The constraint doesn't apply because there's no LV to constrain.

- **Tensor networks:** Generic architectures have many parameters, so the counting is trivially satisfied. But specific architectures (e.g., binary MERA with small bond dimension) may not have enough free parameters at each scale. This needs separate analysis.

- **Graph/network substrates:** Similar to tensor networks. Needs separate analysis.

**Net assessment:** Non-lattice substrates are a genuine gap in the argument, but they don't clearly rescue H12. They require separate analysis for each substrate type, and the conclusion is uncertain rather than favorable to H12.

### 2c. Non-perturbative / strong coupling

Covered in MINOR 1 above. The gap is narrower than Argus states because lattice QCD practice supports the tuning at strong coupling. No known counterexample where the Jacobian becomes singular.

### 2d. Multi-sector consistency, especially gravity

**Fermion-gauge consistency:** The Ward identity reduces the number of independent *conditions* (the vertex anisotropy is determined by the self-energy anisotry), not the number of independent *bare couplings*. The counting stands.

**Gravity:** The graviton's kinetic term has an anisotropy, which introduces an additional LV direction and an additional bare coupling (the metric anisotropy). The counting argument says dim(V_G) > dim(W), so there's always an extra coupling. But gravity is non-renormalisable, and the Symanzik expansion requires infinitely many bare couplings (one for each derivative order). A lattice regularisation provides these implicitly (the lattice action is specified by the hopping parameters, and the Symanzik expansion generates all derivative orders). So the counting should still work.

**However**, there is a subtlety: diffeomorphism invariance imposes much stronger constraints than gauge invariance, and the counting of independent bare couplings in a lattice regularisation of gravity is not standard. If diffeomorphism invariance reduces the number of independent bare couplings below the number of LV directions, the system could be over-determined. This is speculative and needs separate analysis.

**Net assessment:** Multi-sector consistency is a gap, but my analysis suggests it closes in Argus's favour for gauge theories. Gravity is more uncertain but likely also closes, because a lattice regularisation provides enough bare couplings.

### 2e. Architecturally constrained bare couplings (beyond cost minimisation)

Argus identifies cost minimisation as an escape (§5). But there are other reasons the bare couplings might not be free:

- **Specific algorithmic choices:** A substrate implemented with a specific algorithm (e.g., a particular finite-difference scheme, a particular tensor network architecture) may have its bare couplings fixed by the algorithm, not by tuning. The algorithm might not have enough free parameters to absorb all LV directions.

- **Consistency conditions:** In a gauge theory, the bare couplings might be constrained by consistency conditions beyond the Ward identity (e.g., unitarity, anomaly cancellation). These could reduce the number of free parameters.

- **Non-perturbative vacuum selection:** At strong coupling, the vacuum might be selected non-perturbatively, and the bare couplings might be fixed by the requirement of a stable vacuum.

These are all sub-cases of the "not free" escape. They're worth listing because they broaden the escape beyond cost minimisation, but they're already covered by Argus's §5 framework.

### 2f. Composite/bound-state dispersion as a prediction

Once the elementary tunings are fixed (c₀₀ = 0 for each species, Tr E² = Tr B² for the photon), the dispersion relations of composite particles (protons, nuclei) are determined. If there are residual effects from binding energy, they would be suppressed by powers of ΛQCD/EUV ∼ bΛQCD, which is a power of b. This puts them back under H11's ceiling. **Not a rescue for H12's "no ceiling" clause.**

---

## 3. What I Could Not Break

### 3a. The counting argument itself

The tautology dim(V_G) = dim(V_O4) + dim(W) is inescapable. For any symmetry group G, the most general G-symmetric action has one free bare coupling per G-invariant operator, and the generated counterterms are G-invariant operators. The generated counterterms and the available couplings live in the same vector space. I could not find a way around this for lattice substrates.

### 3b. The Jacobian

dW₂/dν = 2 at the Lorentz-invariant point, carrying no power of b and no power of ξ. This is a tree-level result in exact symbolic arithmetic, and it is correct. The bare anisotropy's grip on the dimension-4 LV direction is O(1). I could not find a way to make this O(b^k) for any k > 0.

### 3c. The multi-species Jacobian

M = 2I + O(g²) is full rank for perturbative couplings. The numerical check with 2000 random draws at g²/(4π) ≈ α gives minimum determinant from 3.97 (for N=1) to 495 (for N=8), all comfortably full rank. I could not find a perturbative reason for this to fail.

### 3d. The residue (pb)²

After tuning the dimension-4 LV direction, the leading un-tunable effect is at r = 4, carrying (pb)². For H(3)xT, there are no LV directions at odd r (W[H3T] = 0 at every odd r), confirmed by exact projection. The power-divergent mixing from dimension 6 to dimension 4 generates a tunable contribution, not an un-tunable one. I could not find a way to make the residue anything other than (pb)².

### 3e. The fine-tuning cost

201 bytes at dimension 4, ~2 kB at 10 loop orders. Against Vazza's 10¹²⁴ bits, this is genuinely negligible. I could not argue that 2 kB is a meaningful cost for any host this programme has considered.

### 3f. The H(4) protection result

Zero Lorentz-violating SME singlets at dimension ≤ 4 under exact hypercubic symmetry, across the entire minimal-SME basis. Reproduced from the sixth cycle by exact rational projection. I could not break this.

---

## 4. My Verdict on the Credence Movement

**Argus's movement:** H12 0.80 → 0.74 (propagating the sixth-cycle concession) → 0.38 (on this result).

**My assessment:** The movement from 0.80 to 0.74 is correct — the sixth cycle's adversary established that the marginal counterterm can be tuned, and that concession should have been carried through earlier. The movement from 0.74 to 0.38 is slightly too severe. Here is why:

**What the result actually kills:** Clause (ii) — "that constraint has no ceiling" — is killed *for lattice substrates with free bare couplings*. The counting argument is airtight for this case. The Jacobian is O(1), the residue is (pb)², and the fine-tuning costs ~2 kB. There is no escape for this class of substrate.

**What survives:** Clause (i) — "Lorentz tests constrain the discrete symmetry rather than the spacing" — is reinforced, not weakened. The counting argument shows exactly how the symmetry class determines the number of LV directions. This is a genuine and important insight.

**Escape routes for clause (ii):**
1. Cost-minimisation substrates (Argus's §5). Probability assessment depends on the simulation hypothesis's specifics, but it's non-zero.
2. Non-lattice substrates (causal sets, tensor networks) where the counting argument doesn't apply directly. Probability assessment is uncertain — these substrates might not generate dimension-4 LV in the first place.
3. Architecturally constrained substrates where the bare couplings are fixed by the algorithm. Similar to §5 but broader.
4. Strong-coupling substrates where the perturbative Jacobian might be singular. Lattice QCD practice suggests this doesn't happen, but there's no theorem.

The combined probability of these escapes is hard to estimate, but they are not negligible. A substrate whose bare couplings are not free is a live possibility under the simulation hypothesis, and the constraint survives for those substrates with 19 orders of margin.

**My credence for H12:** **0.45** (rather than 0.38).

Reasoning:
- Clause (i) alone is worth ~0.7 (it's well-established and important).
- Clause (ii) has a ~0.3 probability of being true for some substrate classes (cost-minimisation, non-lattice, architecturally constrained).
- Combined: 0.7 × 0.3 + some weight for the conjunction ≈ 0.45.
- The difference from 0.38 reflects the fact that clause (i) is worth more than Argus gives it credit for, and the escape routes for clause (ii) are broader than just cost minimisation.

**H11:** Argus raises H11 from 0.84 to 0.88, which I agree with. The ceiling now covers the loop route for lattice substrates, which was the main gap.

**The key difference from the first reviewer:** I am told the first reviewer is attacking the result (arguing that H12 should stay higher). My independent assessment is that the result is largely sound but slightly over-kill: the credence should be 0.45 rather than 0.38, reflecting the value of clause (i) and the breadth of the escape routes for clause (ii). The counting argument itself is airtight for lattice substrates with free bare couplings, and I could not break it.

---

## Citation Hygiene

| Claim | Source | Status |
|---|---|---|
| Symanzik improvement programme | Symanzik, *Nucl. Phys. B* 226, 187 and 205 (1983) | `inherited-unchecked` (as flagged by Argus) |
| Anisotropic lattice QCD tunes ξ₀ non-perturbatively | Klassen, *Nucl. Phys. B Proc. Suppl.* 73, 838 (1999), `hep-lat/9811024` | `inherited-unchecked` |
| Morningstar, improved gluonic actions | Morningstar, `hep-lat/9608019` | `inherited-unchecked` (cited by sixth cycle) |
| Hasenbusch anisotropy tuning | Hasenbusch & Hasenbusch, `hep-lat/9904012` | `inherited-unchecked` |
| Causal set dynamics | Rideout & Sorkin, *Class. Quant. Grav.* 17, 4659 (2000), `gr-qc/0004065` | `inherited-unchecked` |
| Collins et al. naturalness | Collins, Perez, Sudarsky, Urrutia, Vucetich, PRL 93, 191301 (2004), `gr-qc/0403053` | `inherited-unchecked` (cited by sixth cycle) |
| Polchinski hypercubic protection | Polchinski, *CQG* 29, 088001 (2012), `arXiv:1106.6346` | `inherited-unchecked` (cited by sixth cycle) |
| Gambini et al. lattice regulator | Gambini, Rastgoo & Pullin, *CQG* 28, 155005 (2011), `arXiv:1106.1417` | `inherited-unchecked` (cited by sixth cycle) |
| Kostelecký & Russell data tables | `arXiv:0801.0287` | `inherited-unchecked` (carried from sixth cycle) |
| Foley et al. anisotropic tuning | Foley, Peardon & Ryan, `hep-lat/0410005` | `inherited-unchecked` (cited by sixth cycle) |
| Morningstar & Foley, anisotropic clover | `arXiv:0810.4477` | `inherited-unchecked` (cited by sixth cycle) |
| Hossenfelder, naturalness critique | `arXiv:1801.02176` | `inherited-unchecked` (cited by Argus) |

No citations were verified at source in this session. All are carried from prior cycles or from the literature, and marked `inherited-unchecked`. Per Argus's stated preference, five verified citations would be better than twenty unverified ones; I have provided none that are freshly verified.

---

*End of second adversarial review.*