[Phase 2 write-up] Sphere-frame cumulant propagation with age-structured κ₃ sources, and exact symmetry as a diagnostic of a tuned closure (1.50e-9, submission #337205)

Z. Dogus Ozel’s Phase 2 algorithmic-contribution write-up. It corresponds to #337205 only.

  • Score. 1.50329e-9 adjusted (final-layer MSE 1.262e-8 at a FLOP fraction of 0.119).
  • Grader run. flopscope 0.12.3 grader, 0 failed MLPs, wall max 105.5 s, residual max 0.365 s.
  • Nomination. Nominated for the private re-evaluation together with its parent #337177 (1.50288e-9). #337177 is 0.03% lower on the 50 public MLPs; #337205 is 0.098% better on 200 local MLPs (SE 0.029).

Write-up (PDF, 29 pages: main text 16 plus a one-page LLM-use section):
WhiteBox_Paper_v4.1_337205_2026-10-24.pdf (938.1 KB)

What it is. Layer-by-layer cumulant propagation through all 16 layers, with no sampling, built on ARC’s factorised K=3 method (Wu et al. 2026). Per layer the chain carries:

  • the mean and the full covariance;
  • the third cumulant as low-rank “hub/leg” sources, one born per layer;
  • the fourth cumulant as one scalar plus per-neuron diagonals.

The ReLU step uses Hermite/Price pair terms and an Edgeworth mean.

What I claim, narrowly (each item is measured in the PDF)

  • Sphere frame (C1).
    • Positive homogeneity lets the chain run on a spherical input. The sphere’s fourth cumulant, −6/(n+2), is exactly representable in the chain’s own isotropic κ₄ form.
    • The result is an exact radial Rao–Blackwellisation of every layer: 2 lines of code, 0 FLOPs, MSE ×0.967 with the corrector off.
    • il.miscusi independently reported the spherical input on a lean chain (18248). Ours adds the exact κ₄ seed and the corrector-target rescaling.
  • Age-structured factored κ₃ (C2). One hub/leg source per layer and exact (2,1)-slice extraction. Sources that reach age 7 are joined into one rank-256 basis by an exact, column-weighted truncated SVD.
  • Zero-matmul closure diagrams (C3).
    • These are folded into existing factors, so they need no extra matrix products.
    • New in this version: excess-kurtosis rows of the (3,1) slice. The diagonal kurtosis the chain already carries enters κ₄(a,a,a,b) on the triple index.
    • With a knob retune and a fitted counter-term (QB3), these give −0.30% (SE 0.07) against our previous named file on the real chain.
  • Exact symmetry as a diagnostic (C7).
    • We measure the Ward defect of our ReLU layer map against the exact scale symmetry. It sits in K₂₁ and grows with depth, but on the support the chain actually carries it is only 3.8%.
    • None of the exact completions of the closure we screened passed its gate; most were worse, by ×1.05–×9.
    • Our hypothesis, labelled as such: the tuned closure is calibrated against its own defect.
  • Measured error anatomy of #337205 (C4). Measured against Monte-Carlo oracles with 8×10⁶ samples, noise-corrected. Each ratio is the final MSE when that object is made exact:
    • 96–99% of the covariance residual is the scale mode, which the chain removes on purpose;
    • an exact carried mean at layers 8–13 leaves ×0.39 of the final MSE;
    • the non-scale covariance leaves ×0.59;
    • an exact diagonal variance makes it worse (×1.06–1.15 per layer, ×1.30 for the whole state);
    • the only directional lever left, the non-smooth part of the (3,1) slice, is reborn every layer, and no cheap carried state reproduces it.

Engineering, reported separately so it can be discounted

  • Fast-matmul engine built from metered flopscope operations: Strassen–Winograd, the Karstadt–Schwartz basis and Smirnov ⟨3,3,6;40⟩. It gives ×0.50, with vs without (60/60 MLPs, measured on the previous named file).
  • Wall guard with a residual-time rule.
  • Learned additive mean corrector, refit end to end on-policy.
  • Billing disclosure: the 0.12.3 truncated-SVD repricing and our own guard bug.

Negative results. The appendix condenses about 200 closed lines, among them:

  • about 25 further (3,1) structures;
  • exact-mode completions;
  • κ₅/κ₆ at the readout (×1.005);
  • DPST ⟨4,4,4;48⟩;
  • hub retirement;
  • four rounds of frontier-model proposals (QB3, derived from one of them, is the exception).

LLM use (also §10 of the PDF)

  • Who did what. The author worked with Claude agents (Opus 5.5 coordinator, Fable 5.1 sub-agents). The agents wrote the code, ran the experiments and drafted the write-up under the author’s direction. The author set the direction, selected the ideas and approved every submission.
  • LLM-originated ideas. _XSH, QB3 and the C7 closed forms; §10 labels them. The sphere frame (C1) is the author’s idea; its derivation and first measurements came from a 09-25 design round led by a Fable agent.
  • Free-model swarm. Idea generation was also delegated, through opencode with all tools disabled, to a swarm of 13 free-tier OpenRouter/OpenCode-Zen models.
    • It made 7,027 calls on 10-07…14.
    • All but 569 stayed inside the providers’ free quotas. The 569 were paid DeepSeek-flash triage calls on 10-07, made until a small prepaid balance ran out.
    • A guard process stopped the swarm if any session streamed from a model that was neither free nor on the allowlist of capped paid models. It never fired.
  • Paid models. Used only in short, capped rounds: GPT-6, Gemini 4, Muse, Grok 4.7, Kimi K3, Qwen 3.8, DeepSeek V4, Claude Opus 5.5 (critic and synthesiser) and Claude Fable 5.1.
  • Outcome, stated plainly. No idea from the free swarm survived code-level screening. About 40 frontier proposals were tested by 10-17, followed by four more rounds on 10-19/20; three proposals shipped in total.
  • Validation. Every idea, derivation and number was validated by measurement. Explanations we did not derive ourselves are marked as interpretations.

On the cost question from the “Good game well played” thread

  • Setup. This result came from the author working with Claude agents under one Claude Max subscription, plus a free-model swarm and a few capped paid rounds for idea generation.
  • Out-of-pocket cost. Phase 2 (about one month) cost about USD 290:
    • one month of Claude Max (20x), USD 200. It ran the Claude agents, and its included API credits covered the Anthropic API calls;
    • about USD 90 of API usage, from the provider balances: OpenRouter USD 82, DeepSeek USD 8.
  • My view. A cheap, tiered setup is enough for idea generation, as long as a measurement gate decides what survives: one coordinator writing briefs, free models generating and critiquing, and paid models used only for targeted questions.

Credits

  • il.miscusi’s coherent scale-mode extraction (18248): adopted; the recipe is theirs.
  • @Artyom_Mazur (Artyom Kuznetsov), young-κ₄ tree (18249): tested, not shipped.
  • Nurlan’s (3,1) law (18252): motivated our row term.
  • hyojun (18255): motivated two checks.
  • 504aldo (18218) and LordProcciuto (e30_a): prior chains we compare against.
  • ARC’s paper and code.

My repository is private. Every number in the PDF carries a source tag that maps to a file in it. It is available to ARC on request, and I will release it with the code if a prize requires it.

Thanks to ARC and AIcrowd for running the challenge.

1 Like

Congratulations on 1.50e-9, and thanks for the error anatomy. Your finding that 96–99 % of the covariance residual is the scale mode has a simple quantitative reason, which may be useful for C7.

For a random ReLU network, a generic error made at layer ℓ keeps S(ℓ) = ∏_{k>ℓ} (1 − arccos(c_k)/π) of its energy at the output: 0.21 from layer 6 at depth 16. An error along the incoming mean is correlated with the gates, and its per-layer gain is χ + (2c/π)·√((1−c)/(1+c)), 0.96 at layer 6 and 0.99 at layer 15. On fresh n = 1024 networks, a mean-direction error made at layer 6 keeps 0.86 of its energy against 0.22 for a generic one. So whatever the closure gets wrong in other directions fades, while the scale component piles up across layers. That makes it the dominant residual whatever its per-layer size, which also fits a Ward defect that grows with depth.

This is §2.3 of our write-up (18282).

Thank you @megan_park , and congratulations on your write-up. Your survival law matches what I measured independently. In my chain, a mean error injected at layers 4, 7, 10 or 13 loses its energy by ×0.85–0.87 per later layer, i.e. about 0.24 from layer 6 to the output, close to your S(6) ≈ 0.21. Your law also gives a clean reason why the scale component piles up while other directions fade.

One nuance: by 1-homogeneity the rank-one scale part (ε m mᵀ) cannot move any later mean. So it dominates the covariance residual but costs no final-layer MSE. That is why my oracles keep it, and why the chain removes only a small fraction of it on purpose. The MSE-relevant part is the non-scale remainder (oracle ×0.59). On the Ward defect: its growth with depth is mostly a representation gap. About 98% of the K₂₁ defect sits where the chain carries no T₂₁, and on the carried support it is 3.8%. So I would attribute only part of that growth to the accumulation mechanism you describe.

Thanks, and good to see an independent chain land on ~0.24 against our S(6) ≈ 0.21.

Agreed on both points. On the scale mode, I only meant that it dominates the residual, not the MSE. Your 1-homogeneity argument is the clean way to say why it is harmless, and it fits the non-scale ×0.59 being the part that matters.

On the Ward defect, the 98% / 3.8% split is a better decomposition than mine. The survival argument can only speak to the carried-support part, so I would weaken “fits” to “is at most consistent with” that 3.8%.

Good luck in the private re-evaluation.

Your measurements locate the gap from the closure side: raw 1.262e-8 on 50 graded networks (#337205), with “99.9% of the squared defect” off the carried T21 support (Fig. 3). The §7 (3,1) remainder “is reborn every layer”; transport explains 1.5 to 4%, pointing toward 18282’s “A cheap κ4 memory” (§6.4).

Our full K=4 chain (topic 18254; logxio/fourth-cumulant) reads the gap from the carried side. At 1024 × 16 it reaches 1.72e-9 on official networks 0 to 4, outside the budget with zero samples, 7.35× below your raw on a different network set.

Handed the K=4 chain’s exact covariance and κ4 at every layer, augmented K=3 retains 0.29 of its original error; full K=4 retains 0.08 on the same scale (networks 0 to 3). The exact (3,1) slice in our K=3 chain recovers at most 10.0% of the K=4 log gain. Removing C·κ3 births from full K=4 costs 5.39×; the non-tree term contributes 1.56× in our portable K=4 set and about 1% in 18272’s K=3 chain. Over widths 256 to 1024, the two errors fall as n^−3.95 and n^−3.08. The width exponents match the fourth and third orders conjectured by Wu et al. (2026, Section 5), giving the same gap a separate measured scaling check.

A test with your §7 oracles: how much of the reborn (3,1) remainder do each layer’s C·κ3 births explain, beyond the contribution from transport?

Thanks for the question @Susupro . I ran it. I built the C·κ₃ tree term (one f’’ vertex, one covariance edge, one κ₃) of the post-ReLU κ₄ at layer l−1 from the exact Monte-Carlo κ₃ (8×10⁶ sphere-frame samples, independent of the oracle slices). I transported it by W to the (3,1) slice of layer l, with no n³ tensor needed: a few n×n Grams. Same metric as my §7 oracles: off-diagonal, coefficient on one half, explained energy on the other. MLPs 0–1, layers 9–13:

  • C·κ₃ births explain 31–47% of the exact (3,1) slice, with a coefficient of 1.2–1.65 (theory weight 1). That is the largest single structure I have found for this slice.
  • They explain 13–19% of the reborn remainder (the slice minus its smooth fit), against 1.5–4% for transport alone. Jointly 15–22%, so +14–19% beyond transport.
  • The same births built from my chain’s own carried legs explained only 7.5–7.8%. So part of the gap is the chain’s κ₃, not the birth mechanism.

More than 80% of the remainder is still unexplained by births plus transport at this order, which supports your reading that the gap sits in what a K=3 chain carries. Caveats: two networks, Gaussian Wick coefficients, coincident-index terms not removed, and the remainder is defined against my smooth fit, which already absorbs part of the birth (hence the lower coefficient on it).

Your 31 to 47% whole-slice result, and 13 to 19% of the reborn remainder versus 7.5 to 7.8% from carried legs, pinpoint the birth’s input state. Full K=4 closes this loop through (2,1,1) feedback into κ3 at each layer (§§2, 5).

For the remaining >80%, full K=4 carries two further mechanisms across layers: older sources b < l − 1 through cumulative transport Q_{b→l}, and non-tree C_jk κ3_ijk births. At 1024 × 16, cutting source lifetimes to three layers costs 67× final MSE on network 0 (§6). Removing non-tree births costs 1.95× in the chain starting at layer 4, and 1.56× in the portable set (§§3, 5). Together with C·κ3 births, they account for 96% of the counted budget from layer 4. Your test isolates the one-step tree birth; these ablations measure the state our full chain carries around it.

1 Like

Agreed. My test isolates the one-step tree birth, and your ablations measure the carried state around it, so the two are complementary. I can’t check the older-source and non-tree pieces in my chain without carrying them. Thanks for the exchange.