# RESULT.md — Cubic (ell=4) anisotropy search in the Auger 10% open data

**Experiment:** `lab/2026-09-08-auger-anisotropy/`
**Designed:** 2026-09-08 (first night cycle). **Run:** 2026-09-09 (second night cycle).
**Hypothesis under test:** H2 — *if simulated, the substrate is a lattice with spacing below current cosmic-ray sensitivity* (credence 0.20).
**Run by:** Argus, on claude-opus-5.

**Data:** Pierre Auger Observatory 10% open data release, DOI 10.5281/zenodo.4487612, CC BY-SA 4.0. Release paper arXiv:2309.16294, EPJ C 85:70 (2025). 25,106 events with complete energy and direction (22,751 vertical SD1500 with theta < 60 deg, 2,355 inclined with 60 < theta < 80 deg); 1,568 SD1500 rows dropped for a blank `sd_energy`. Span 14.6 years of GPS time. Max energy 144.1 EeV.

**Code:** `auger_lib.py`, `cubic.py`, `run_control.py`, `run_cubic.py`. Rerunnable with `/opt/argus-venv/bin/python run_control.py && /opt/argus-venv/bin/python run_cubic.py`. Total runtime about 70 s.

---

## 0. Headline

*This section was rewritten after two independent adversarial reviews (§8). Read §8 before trusting anything here.*

The search is **null at every energy cut**, as expected. What survived review:

1. **A cubic-anisotropy statistic that is unbiased against any declination-only exposure**, verified by breaking the exposure on purpose. Both adversaries attacked the proof and neither broke it. It is **not novel** — see §8 — it is a natural extension of the standard exposure-independent techniques in the field, and I initially overstated it.
2. **The sensitivity calculation.** Detecting a 1% cubic modulation above 40 EeV would take of order **4 x 10^5 to 8 x 10^5 events**, roughly **6,000 to 14,000 years** of full Auger running depending on the lattice orientation. This test is statistics-starved by three orders of magnitude and no amount of data access fixes it. This is the most useful thing the night produced.
3. **A conditional bridge from Beane et al.'s threshold formula to an observable anisotropy** (§7), which the literature does not contain. It yields `b^-1 > 6.2 x 10^10 GeV` **only if** the events above 40 EeV are protons **and** the observed spectral suppression is GZK attenuation. Auger's own composition data make the first condition doubtful. **It is not a bound. It is a conditional.**
4. **Three errors of my own**, all caught and named: a sign error in a logarithm, an optimistic pre-registered estimate, and a mislabelled statistic.

---

## 1. Controls — the gate everything else stands on

### C1. Observatory latitude, recovered from the data itself

The released files give both local coordinates (`sd_theta`, `sd_phi`) and equatorial ones (`sd_ra`, `sd_dec`). Those are redundant, and the redundancy determines the site latitude through
`sin(dec) = sin(lat) cos(theta) + cos(lat) sin(theta) cos(A)`, with `A = 90 - sd_phi` (the released azimuth runs counter-clockwise from East). Least squares over all events:

| subset | N | fitted latitude |
|---|---|---|
| vertical | 22,751 | **-35.2106 deg** |
| inclined | 2,355 | **-35.2504 deg** |
| all | 25,106 | **-35.2136 deg** |

The nominal Malargue site latitude is -35.20 deg. Agreement to **0.014 deg**, with a residual scatter of 1.7e-3 in `sin(dec)`. The coordinate columns are internally consistent and my transforms are right. *(Evidence class: Established — a direct consistency check on published data.)*

### C2. The exposure model — the real positive control

Sommers (2001), arXiv:astro-ph/0004016, gives the directional exposure of a ground array in closed form. The derivation is the hour-angle integral of the projected array area:

```
cos(theta) = sin(lat) sin(dec) + cos(lat) cos(dec) cos(alpha_h)
omega(dec) ∝ INT_{-am}^{+am} cos(theta) d(alpha_h)
           = 2 [ am sin(lat) sin(dec) + cos(lat) cos(dec) sin(am) ]
am = arccos[ (cos(theta_max) - sin(lat) sin(dec)) / (cos(lat) cos(dec)) ]
```
with `am = 0` if the argument exceeds 1 and `am = pi` if it is below -1. For the inclined band I use the difference of the two integrals at 80 and 60 deg, which is exact because theta increases monotonically with hour angle at fixed declination.

Against the observed declination distribution:

| sample | N | KS p | chi2/dof (30 equal-probability bins) | chi2 p |
|---|---|---|---|---|
| vertical, theta < 60 | 22,751 | **0.50** | 1.37 | 0.086 |
| inclined, 60 < theta < 80 | 2,355 | **0.29** | 1.56 | 0.028 |

The vertical set — where the statistical power is — is an excellent fit. This is the highest-power control available and it is the one that matters, because the exposure model is the component most able to fake an ell=4 signal. The inclined set shows mild tension (chi2 p = 0.028, one test among several, so not significant), plausibly because a purely geometric `cos(theta)` acceptance is a cruder approximation for very inclined showers.

The vertical:inclined event ratio is 9.66, against 3.41 from pure geometry, so the inclined sub-array's relative efficiency is **w_incl = 0.353**. Fitted, not assumed — and it never enters the final statistic.

### C3. Right-ascension uniformity

| band | N | first-harmonic r1 | 1 sigma | p |
|---|---|---|---|---|
| all energies | 25,106 | 1.17% | 1.26% | 0.42 |
| 1-3 EeV | 7,896 | 2.43% | 2.25% | 0.31 |
| 3-8 EeV | 13,827 | 2.48% | 1.70% | 0.12 |

Uniform, as a ground array integrating over many sidereal years must be. This is the one assumption the final statistic does rely on, and it is satisfied.

### C4. Positive control on the published dipole — **partially failed, honestly**

`PLAN.md` §7 set this gate: *"if the pipeline cannot reproduce a known published signal, no null it produces is worth anything."* Rayleigh first harmonic in right ascension:

| cut | N | r1 | 1 sigma | phase | p |
|---|---|---|---|---|---|
| E >= 4 EeV | 10,691 | 3.19% | 1.93% | 129.3 deg | 0.066 |
| **E >= 8 EeV** | **3,122** | **2.61%** | **3.58%** | **137.8 deg** | **0.59** |
| E >= 16 EeV | 796 | 8.56% | 7.09% | 180.5 deg | 0.23 |
| E >= 32 EeV | 156 | 26.93% | 16.01% | 186.9 deg | 0.059 |
| E >= 40 EeV | 87 | 40.88% | 21.44% | 170.9 deg | 0.026 |

Auger's published result above 8 EeV (Science 357:1266, arXiv:1709.07321, N = 32,187) is a first-harmonic amplitude of about 4.7% at phase about 100 deg, at 5.2 sigma.

My measurement is **consistent with it and cannot confirm it**. That is not a bug, it is arithmetic, and I should have computed it before designing around it. The amplitude resolution of a Rayleigh analysis is `sqrt(4/N)`, which at N = 3,122 is **3.58%**. A true 4.7% signal is therefore a **1.31 sigma** expectation in this sample. Simulating: the probability that this release yields p < 0.01 on the true dipole is **16.8%**, rising to only 38.9% if the true amplitude were the full 3D dipole value of 6.5%.

**So the intended positive control had roughly a one-in-six chance of firing, and did not fire.** The pipeline is therefore validated by C1 and C2, not by C4. I am recording that as a partial failure of the gate I set myself rather than quietly reclassifying it, because C2 happens to test the component that actually matters here.

The 40.88% amplitude at E >= 40 EeV (p = 0.026 before any trials penalty, five cuts examined) is a 1.9 sigma fluctuation on 87 events. It is not a result. Noted so that I do not rediscover it later and mistake it for one.

---

## 2. The statistic, and why it is not the obvious one

Under Beane, Davoudi & Savage (arXiv:1210.1847) the signature of a rigid cubic lattice is angular power at **ell = 4** in the cubic (Kubic) harmonic, with ell = 1, 2, 3 all killed by the octahedral group O_h. In Cartesian form, in the frame whose axes are the cube axes,

```
K4(n) ∝ (n_x^4 + n_y^4 + n_z^4) - 3/5
```
which is exactly the combination `Y40 + sqrt(5/14) (Y44 + Y4,-4)`, and which avoids every spherical-harmonic phase convention. Isotropic moments on the sphere give `E[K4] = 0` and `Var = 16/525`; I normalise by `sqrt(16/525)` so a fitted amplitude reads as a fractional modulation depth. Both moments were confirmed by Monte Carlo to four decimals.

**The obvious statistic is the wrong one.** Weighting each event by `1/omega(dec)` to undo the instrument makes the result only as good as the exposure model — and ell = 4 is precisely where a declination-band exposure residual lands.

**The better statistic.** A single-site ground array observing for many sidereal years has an exposure that depends on declination *alone*: it is azimuthally symmetric about the celestial pole. So subtract, at each event's declination, the right-ascension average of the cubic harmonic:

```
g_R(n) = K4(R n) - < K4(R n) >_RA at fixed dec
```

Then for **any** declination distribution P(dec) whatsoever,

```
E[ mean_i g_R(n_i) ] = INT P(dec) [ INT dRA/2pi (K4 - K4bar) ] = INT P(dec) * 0 = 0
```

The expectation is exactly zero and **contains no exposure model at all**. The only surviving assumption is right-ascension uniformity, which is a property of Earth's rotation rather than of anyone's model, and which C3 verifies directly on the data.

The cost is real and is measured, not assumed: immunity discards the m = 0 content, so sensitivity to lattice orientations aligned with the celestial pole is lost. Measured per-event noise falls from sd 1.000 to 0.998 — a negligible price, because the m = 0 part is a small fraction of the 9-dimensional ell = 4 space.

*(Evidence class: Inference (Argus), for the application to the cubic case. The underlying idea — that a ground array's RA analysis is exposure-independent — is standard in the field and is why Auger quotes first-harmonic-in-RA results. I have not yet found the ell=4 generalisation written down; prior-art search is incomplete. See §5.)*

### The immunity, demonstrated by breaking the exposure on purpose

200,000 genuinely isotropic directions drawn through a deliberately wrong declination distribution (the true Auger exposure times an extra `1 + 1.5 cos^2(2 dec)` tilt):

| exposure used to draw | naive weighted max abs S | **immune max abs S** |
|---|---|---|
| correct | 0.0466 | **0.0038** |
| deliberately distorted | **0.2583** | **0.0044** |

The naive statistic reports a **25.8% cubic anisotropy out of a perfectly isotropic sky**. The immune statistic does not move. *(Evidence class: Established, by construction and by simulation.)*

This is the result I would keep if I could keep only one thing from tonight.

---

## 3. The search — null at every cut

Null distribution by **right-ascension scrambling**: keep every observed declination exactly as measured and redraw RA uniformly, 20,000 realisations. This preserves the true exposure in declination by construction and requires no exposure model. Scan: 3,000 uniformly sampled cube orientations, statistic `max |S(R)|`.

| cut | N | observed max abs S | null median | sensitivity (null 95th pct) | global p | **true 95% UL** | naive/immune |
|---|---|---|---|---|---|---|---|
| E >= 8 EeV | 3,122 | 0.0468 | 0.0409 | 6.0% | 0.30 | **6.4%** | 2.12 |
| E >= 16 EeV | 796 | 0.0760 | 0.0806 | 11.9% | 0.58 | **10.7%** | 1.61 |
| E >= 32 EeV | 156 | 0.1208 | 0.1804 | 26.7% | 0.91 | **11.7%** | 2.95 |
| E >= 40 EeV | 87 | 0.2158 | 0.2414 | 35.8% | 0.66 | **30.8%** | 1.85 |

**No excess anywhere.** The smallest global p-value is 0.30.

**On the two limit columns.** My first draft of this file called the null 95th percentile a "95% upper limit". It is not one; it is the sensitivity, the limit one expects to set on average. Both adversarial reviews flagged the mislabelling independently, agreeing with my own re-derivation. The **true 95% upper limit** is the amplitude at which the probability of observing a scan statistic as small as the one I actually saw is 5%, computed in `limits.py` by injecting a cubic signal at the best-fit orientation and redrawing right ascension while holding every observed declination fixed.

I predicted in an earlier draft that the true limits would come out *weaker* than the sensitivity. **That prediction was wrong**, and my analytic estimate behind it was wrong because it ignored the noise floor of a scanned maximum: even with zero signal, max|S| over 3,000 orientations sits well above zero, so an injected signal pushes the statistic up faster than a single-orientation Gaussian argument suggests. Three of the four true limits are *tighter* than the sensitivity, and they are tighter exactly where the observed value fell below the null median. The one cut where the observed value fell *above* the null median (E >= 8 EeV) is the one cut where the true limit is looser. The pattern is right, which is the check that the Monte Carlo is doing what I think it is.

Two things worth noticing in the table:

- The **naive/immune ratio is 1.6 to 3.0 on real data at every cut**. The exposure-weighted statistic consistently reads higher than the immune one on the same events. That is what residual exposure leakage into ell = 4 looks like. Anyone running the naive version and quoting a limit is quoting a number partly made of their own instrument.
- Three of four observed values sit **below** the null median. With four correlated cuts that is unremarkable (p of order 0.3), and I mention it only to have said so.

---

## 4. Sensitivity — the actual deliverable

Signal injection: draw from a density proportional to `omega(dec) * (1 + A * K4(R0 n))` by rejection, and measure the response.

- response `<g>/A` = **0.9786** per unit amplitude
- per-event noise sd(g) = **1.0324**
- single-orientation significance = **A * 0.9479 * sqrt(N)**

Trials factor, from the scrambled nulls rather than a rule of thumb. The 95th percentile of the null max sits at 3.24 single-orientation sigma, which by the Sidak relation `P = 1 - (1 - p_local)^N_eff` gives **N_eff = 42** effective independent orientations out of 3,000 scanned. That number is a sanity check in itself: the ell = 4 subspace is 9-dimensional and the cube orientations trace a 3-parameter family inside it, so tens of effective trials is exactly what it should be, and thousands would have meant a bug. Global 5 sigma (p = 2.87e-7) then needs a **single-orientation z of 5.80**.

Full Auger accumulates about **60 events/yr above 40 EeV** (inferred: 87 events in a 10% release spanning 14.6 yr).

| cubic amplitude A | events needed (median orientation) | full-array years | range over orientations |
|---|---|---|---|
| 30% | 415 | 7.0 | 7 - 15 |
| 20% | 935 | 15.7 | 16 - 34 |
| 10% | 3,739 | 62.7 | 63 - 135 |
| 5% | 14,956 | 251 | 251 - 542 |
| 2% | 93,474 | 1,569 | 1,569 - 3,389 |
| **1%** | **373,895** | **6,275** | **6,275 - 13,555** |

The right-hand column is the correction demanded by the statistics adversary and it is a fair hit: the response coefficient depends on where the cube axes sit relative to Auger's declination coverage, ranging from 0.67 to 1.00 with a median of 0.959. My original single-orientation number understated the requirement for unfavourable lattice orientations by up to a factor of 2.2. The adversary also recomputed N_eff by simulating the covariance of the actual 3,000 templates with per-orientation normalisation and got **108** rather than my 42, which raises the threshold from 5.80 to 5.92 and the event counts by about 4%. Their method is better than mine; the effect is negligible next to the orientation spread, and neither changes the conclusion by an order of magnitude.

And what the full dataset in hand today would reach (10x this release, scaling as 1/sqrt(N)):

| cut | full N | projected 95% UL |
|---|---|---|
| E >= 8 EeV | ~31,200 | 1.9% |
| E >= 16 EeV | ~7,960 | 3.8% |
| E >= 32 EeV | ~1,560 | 8.5% |
| E >= 40 EeV | ~870 | **11.3%** |

**This is the number Travis should take away.** Full access to the Auger dataset buys a factor of about 3.2 over what I just computed from public data. It moves the limit above 40 EeV from 36% to about 11%. It does not approach 1%, and nothing built or planned will: 6,275 years of the largest cosmic-ray observatory ever constructed.

---

## 5. Novelty gate

| step | status |
|---|---|
| 1. Prior art | **Done.** `reports/threads/2026-09-09-beane-scaling.md` (GPT-5.5) read Beane et al. in full and searched arXiv, INSPIRE, OpenAlex, Crossref and the Auger/TA methods literature. The statistics adversary independently searched the cosmic-ray multipole-estimation literature. |
| 2. Own check | **Passed.** Everything above is my own computation on public data, rerunnable from this directory. Machinery verified to 3.6e-15 against direct evaluation; analytic moments confirmed by Monte Carlo. |
| 3. Adversarial review | **Done, twice.** `2026-09-09-adversary-cubic.md` (statistics and method) and `2026-09-09-adversary-derivation.md` (the physics chain), both GPT-5.5, both told to break it. Outcome in §8. |

**Verdict: `rediscovery` for the method, `open` for the conditional bound, `null` for the search.**

**The method is not novel and I was wrong to flag it as a candidate.** The statistics adversary found the ancestry: the East-West method (Bonino et al., ApJ 738:67) is explicitly sold as an exposure-independent technique for large-scale anisotropy; Roulet & Mollerach (astro-ph/0504630) state outright that the simplest version of their method assumes exposure depends only on declination; Aublin & Parizot (astro-ph/0504575) and Billoir & Deligny (arXiv:0710.2290) treat multipole reconstruction under partial and non-uniform sky coverage. What I built is a cubic-template matched filter projected orthogonal to the space of declination-only exposure functions. That is standard in spirit. The specific application to the ell=4 cubic template appears not to have been done, but "nobody has applied a standard technique to this particular template" is a much smaller claim than the one I was drifting toward, and I am recording the correction rather than the drift.

**Prior art on the search itself: genuinely absent.** The thread found no published implementation of the oriented cubic-symmetry test on Auger or Telescope Array data, and no published reanalysis converting UHECR anisotropy nulls into a bound on `b`. The 2024 Auger large-scale anisotropy paper (arXiv:2408.05292) states that partial sky coverage prevents estimation of individual `a_lm` for `l > 2` — which is precisely the limitation a template-projection approach sidesteps. So the gap is real even though the tool is standard.

---

## 6. Errors I made, recorded

1. **The trials-factor arithmetic was wrong on the first run.** I inverted the Sidak relation with `ln(0.05)` where it should have been `ln(0.95)`, inflating N_eff by a factor of 58 (2,467 instead of 42) and the 5 sigma threshold from 5.80 to 6.44. Caught because 2,467 effective trials out of 3,000 random orientations is impossible for a statistic living in a 9-dimensional space — the physics caught the algebra. Corrected; all numbers above are post-fix. The 95% upper limits were never affected, since those come straight from the Monte Carlo quantile.
2. **My pre-registered upper-limit estimate was optimistic by 1.7x.** `PLAN.md` §8b predicted about 21% at E >= 40 EeV from the naive `1/sqrt(N)` floor. The true answer is 35.8%, because the scan over orientations costs a trials penalty of 3.24 sigma that the back-of-envelope ignored. Estimates that skip the trials factor are optimistic, and mine was.
3. **I set a positive-control gate without computing its power first.** See C4. Designing a validation test that had a 17% chance of passing under the true hypothesis is a planning error, not a data problem.
4. **I called a null critical value an upper limit.** See §3. Caught by me and by both adversaries independently. Then I compounded it by predicting, from a crude analytic argument, that the corrected limits would be weaker — and they are mostly tighter, because my analytic argument ignored the noise floor of a scanned maximum. Two errors in the same paragraph, in opposite directions.
5. **I measured the signal response at one cube orientation and quoted it as the sensitivity.** The response coefficient ranges from 0.67 to 1.00 across orientations (median 0.959); my single sample gave 0.948, slightly below median. Unfavourable orientations inflate the event requirement by up to a factor of 2.2. The sensitivity table in §4 now spans that range instead of quoting one number.

---

## 7. Effect on the ledger

**H2** (lattice spacing below current sensitivity), currently 0.20. This null does not lower it — a null with a 36% upper limit was never going to constrain anything, and I said so before running. What the night actually changed is the *reason* H2 sits where it does. It is not "the tests came back null"; it is now, with a number attached, **"the tests cannot be run"**. That is a different epistemic situation and a worse one for the hypothesis's testability, though not for its truth.

What genuinely moves is my confidence in the ledger entry that calls cosmic-ray anisotropy "the sharpest empirical constraint on the ledger." If reaching 1% cubic amplitude takes 6,275 years of Auger, then that entry is not a sharp constraint. It is a *weak* constraint that has been repeated until it sounded sharp. Whether Beane et al.'s `b^-1 >~ 10^11 GeV` survives that reframing depends entirely on what amplitude their lattice actually predicts, which is exactly the number the prior-art thread is fetching.

Held. Recorded in `HYPOTHESES.md`.

---

## 7. The missing link: from Beane's threshold to an observable — and what it is worth

`derive_bound.py`. **Evidence class: Inference (Argus).**

The prior-art thread established that Beane et al.'s `b^-1 >~ 10^11 GeV` is **not an anisotropy result at all**. It comes from equating the lattice's maximum representable energy `E_max ~ 1/b` to the observed GZK cutoff — one line of dimensional analysis, quoted directly: *"the cut off from the dispersion relation is E^max ~ 1/b. Equating this to the GKZ cut off corresponds to a lattice spacing of b ~ 10^-12 fm."* I have been repeating that number as "the sharpest empirical constraint on the ledger" without knowing what it was. It is a statement that we have not yet seen the lattice cutoff, not a measurement of lattice structure.

Their only angular result is eq. (18), a direction-dependent **threshold** for `gamma_CMB + N -> Delta`, which they never propagate to a flux. So I built the propagation:

1. **Normalisation.** Beane's harmonic combination `Y_4^0 + sqrt(5/14)(Y_4^+4 + Y_4^-4)` equals `(15/(4 sqrt(pi))) (x^4+y^4+z^4-3/5)`, i.e. **0.369349** times my unit-variance K4. Derived analytically, confirmed numerically to 3.7e-16, and independently re-derived by the adversary, who got the same 12 digits. Threshold modulation = `0.07274 (b|p|)^2` in K4 units.
2. **Wien amplification.** The threshold sits at `omega_0/kT = 16.7` at 40 EeV, deep in the Wien tail, so the available CMB photon density above threshold has `-dln n(>omega)/dln omega = 14.78` by exact blackbody integration.
3. **Attenuation.** `dF/F = +(D/lambda) x amplification x (d omega/omega)`. Sign confirmed by the adversary: a higher threshold means fewer photons, a lower rate, a longer attenuation length, and therefore *more* flux.
4. **`D/lambda` is read off the measured spectrum, not modelled.** Auger (arXiv:1002.1975 — the very paper Beane et al. cite as their input) reports the flux suppressed by a factor two at `log10(E/eV) = 19.61`, giving `D/lambda = ln 2` at 40.7 EeV. That lands on my 40 EeV cut by luck, not design.

Assembled: **`A = 0.745 (b|p|)^2`**. Inverting the true 95% upper limit of 30.8%:

| | limit on A | implied bound |
|---|---|---|
| this work, public 10% release | 30.8% | `b^-1 > 6.2 x 10^10 GeV` |
| full Auger today, projected | 9.7% | `b^-1 > 1.1 x 10^11 GeV` |
| Beane et al., cutoff route | — | `b^-1 >~ 1 x 10^11 GeV` |

**This is not a bound, and I will not present it as one.** See §8.

---

## 8. What the adversaries did to it

Two hostile reviews, both GPT-5.5, both instructed to break rather than assess. Full texts in `reports/threads/`.

### What survived

| claim | outcome |
|---|---|
| Harmonic normalisation (0.369349) | **Survived.** Independently re-derived to 12 digits. |
| Sign of the flux response | **Survived.** |
| Exposure-immunity proof, within its stated domain | **Survived both attacks.** "No declination-only exposure residual can bias the statistic. I could not break that proof." |
| Azimuth convention in the latitude fit | **Survived.** Adversary tried six alternative conventions; only the correct one and its exact degeneracy give -35.2136. |
| Sommers exposure implementation | **Survived**, including the inclined band-difference trick. |
| Monomial expansion, signs, offsets in `cubic.py` | **Survived.** No bug found. |
| Large-amplitude nonlinearity in the injection | **Survived.** Coefficient moves from 0.99556 to 0.99823 between A=0.01 and A=0.30. |
| The continuous-energy-loss objection I most feared | **Survived, surprisingly.** Photopion production degrades a proton by ~20% rather than removing it, so `lambda = 1/R` is the wrong label — but the *first-order logarithmic response* still carries a factor of `ln 2` provided the observed suppression comes from that same channel. The adversary's own words: "The derivation survives this specific attack more than I expected." |
| The scan-max comparison being conservative | **Survived.** For an unknown lattice orientation, comparing against the scanned maximum is the right conservative object. |

### What broke

1. **Composition — the fatal one.** I assumed every event above 40 EeV is a proton undergoing `gamma + p -> Delta`. Auger's own Xmax measurements (arXiv:1701.06812) favour a progressively heavier mix above 10 EeV, and their source fits (arXiv:2211.02857) reproduce the data with mixed composition and a low rigidity cutoff. For nuclei the relevant process is photodisintegration, the relevant Lorentz factor is `E/A`, and Beane's proton threshold formula is simply the wrong calculation. Optimistic nucleon scaling gives `b^-1` weaker by 4x for helium, 14x for nitrogen, 56x for iron. **For a plausible mixed composition the bound is somewhere between weaker by half and meaningless.**

2. **Causation.** `D/lambda = ln 2` assumes the measured spectral suppression *is* proton GZK attenuation. It may instead be the end of the source acceleration spectrum. Berezinsky et al.'s "disappointing model" (arXiv:0907.5194) is explicit: with average energy per nucleon below a few EeV, CMB pion photoproduction is absent and the GZK cutoff does not exist. If only a fraction `f_GZK` of the suppression is GZK, the bound scales as `sqrt(f_GZK)`: 4.1e10 GeV at f=0.5, 1.8e10 at f=0.1, **no bound at all at f=0**.

3. **"Exposure immune" was overstated prose.** The correct statement is *immune to declination-only exposure, conditional on RA uniformity after all energy and time cuts.* The adversary quantified the residual: a 15.5% fourth-harmonic RA exposure modulation could fake the E >= 8 EeV null level; at E >= 40 EeV it would take order-unity modulation. Auger's published RA-modulation systematic is at the sub-percent level over long integrations (arXiv:0706.2640), so **the present nulls are safe by a wide margin** — but a future 1% cubic target would need RA control at 2.6%, and a time-correlated 1% energy-scale drift could produce a fake `S` of 0.012-0.016. That is irrelevant now and disqualifying later.

4. **Magnetic deflection** damps ell=4 along the propagation path. Small for protons in quiet fields (6.3 deg RMS gives 3% damping) but severe for nuclei, where deflection scales with Z.

5. **The expansion parameter** is used at `b|p| ~ 0.4-0.7` where Beane's formula says `|p| << 1/b`. The adversary computed the next ell=4 term in the massless lattice dispersion at only 1.6% at `q=0.4` and 4.8% at `q=0.7` — so the *formal* expansion is not exploding. The real problem is that the proton/Delta threshold has not been rederived at that `q` at all.

### My verdict on my own result

The **null is solid**. The **method is sound but standard**, and I have downgraded the novelty claim accordingly. The **sensitivity calculation is the real deliverable** and it survived with a factor-2 orientation caveat.

The **bound is a conditional, not a bound.** The correct statement is: *if the highest-energy Auger events are protons, and if the spectral suppression above 10^19.6 eV is GZK attenuation, then the absence of cubic ell=4 anisotropy in the public data implies `b^-1 > 6.2 x 10^10 GeV`.* Both antecedents are contested by Auger's own measurements. I built the bridge the literature was missing, and the bridge turns out to rest on two piers that may not be there. That is worth knowing and worth writing down, and it is not a result I would defend as a constraint on anything.

The most useful thing to say about Beane et al.'s number, having done this: **their bound and mine are the same order of magnitude by two completely independent routes**, which is mildly reassuring about the order of magnitude and says nothing whatever about whether there is a lattice.