soft3/tru/eval/CT-0 end-to-end on a real graph — space-pussy.md

2026-09-11/12. the full run of the SHIPPED pipeline on real data: pussy_links.jsonl -> temporal 90/10 split -> passes 1-5 as compiled (index -> dialect -> arch -> embed -> attn, shipped hop2_mix and out_gain) -> glia-semantics llama forward. rs/examples/eval_pussy.rs (shipped Fx path), rs/examples/eval_pussy_mix.rs (f64 mirror on the scipy spectrum, eval/mirror_svd.py).

compiled architecture

d*=64, h*=1 (single token denomination), L*=36, diam=9, sigma_1/sigma_k = 218.8 (probe) -> out_gain = 8.77. vocab 65,121 particles (32,560 link endpoints + 32,561 axons — every cyberlink is itself a particle, §3).

task

predict the target of a held-out (temporal test) cyberlink from a train-graph walk prefix ending at its source: [f], [b,f], [a,b,f] with train edges a->b, b->f, held-out f->t. n = 1181 queries. metric: MRR over the full particle vocab.

results (n=1181)

shipped pipeline, 2026-09-12 revision: banded field SVD (full-rank tail), row-normalized E, quiet attention gain (tru#5).

zero-layer == embed-floor exactly: the plumbing is sound. the spectral ratio is now sigma1/sigma_k = 1.58e8 — the banded spectrum recovers the full 64-component tail (the collapsed 14-component version read 218.8). NOTE on the history below: the 0.6362 floor of the earlier revision rode on the collapsed-tail artifact (tru#7) — the honest full-rank field number is 0.4886, still above the exact scipy mirror at the same gauge (0.392 cosine). pussy remains the hard small graph; the bostrom ladder is where the floor reaches the counts baseline (0.89 at ~1M particles).

the attention verdict (measured, multi-mechanism)

the compiled attention stack is net-negative in EVERY configuration tested on real walks: shipped Fx weights, f64 mirror, walk kernels (P_dir, forward/backward), Gram kernels (G, G^2), gains 0.25–8.8, role-blend theta 0–45deg, spectral exponents p 0–0.5, rank truncations 4–64. three distinct failure mechanisms, all measured:

1. information ceiling. on a directed walk the answer t is never in the prefix; the only successor signal is E_f itself (the residual carries it for free). attention can inject functions of the prefix only, and on this graph they all hurt.
2. walk kernels dilute. P_dir retrieves predecessors; injecting gain-scaled predecessor embeddings moves the tied-head ranking away from the successor cone.
3. Gram kernels self-collapse. with near-orthonormal embeddings, G ~= the projector: scores become self-dominant (Cauchy-Schwarz), ctx comes out parallel to E_f, and the output is invariant to the gain (observed: identical MRR across a 35x gain sweep).

plus a scale pathology that made every gain meaningless: compiled embedding ROW norms span ~1e-4 (long tail) to O(1) (hubs), so with W_V = I any gain >= 1 swamps the residual 10^4x. rmsnorm inside the block hides it; the residual add does not.

design consequence (tru#5): attention ships as a training-ready substrate, quiet at init — the MLP's existing pattern (LayerScale gamma = 1e-5). keep the structural Q/K (a training prior), keep OUT_GAIN for post-training calibration, set the block's init gain to ~1e-2..1e-5. the open decisive experiment is a fine-tuning ablation: structural init vs random init on walk continuation. the claim "attention works" defers to that result; init-time MRR on pure structural walks is information-capped and cannot demonstrate it.

the embedding-floor anomaly (open, load-bearing — tru#7)

cross-checking the shipped Fx embedding against the exact scipy spectrum of the same operator (mirror_svd.py):

• the Fx SVD (fixed-point subspace iteration, k=64, 120 iters) resolves only ~14 of 64 components; the rest collapse to zero. the 14 columns are orthonormal; the middle ones are rotated 10-40 deg against the exact singular vectors (|cos| 0.76-1.0) and the spectrum is distorted (true sigma_3 = 0.48 rides at 0.16% of top; directions 5..20 carry ~10x their true weight).
• floor MRR: shipped Fx E = 0.64; exact scipy E at ANY truncation (k = 4..64), ANY spectral flattening (clamp, power law), ANY band pass = 0.38-0.41 — BELOW bigram (0.57).

so the current "compiled geometry beats counts" result on this task rides on an uncontrolled fixed-point artifact. nothing in the exact spectral subspace reproduces it. open: (a) reproduce from a clean build; (b) identify the mechanism (candidates: fixed-point noise as quasi-random features; Ritz-vector bias); (c) either control it deliberately or narrow the claim. NOTE the earlier LP-eval results (AUC-based, pairwise) are a different metric and remain valid on their own terms.

infra fixes this session

• tru#5 (fixed): attn.rs stored Q/K transposed ([in][out]) vs the .model [out,in] convention (glia matmul y = x @ W^T). a compiled model computed sqrt(S)(U^T V)sqrt(S) instead of P. W_V/W_O are identity- scaled and immune. found because mirror scores collapsed to ~1e-5.
• eval-side SVD used the symmetric-only Rayleigh shortcut sqrt(v^T M v); on directed M the correct sigma = ||M v|| (rank-2 garbage otherwise). fixed in eval_ct0/eval_attn/eval_pussy_mix.
• eval-side subspace iteration collapses the tail (exact zeros past component 6 on n=65k) — mirror now loads scipy svds.
• arch probe (k=1024, iters=120) was hours at fixed-point Gram-Schmidt cost; now (64, 30). Gram-Schmidt caches loop-invariant denominators.
• TRU_SVD_DEBUG=1 / MIX_DEBUG=1 give per-iter / per-layer telemetry.

honest status

compile machinery: works end-to-end, deterministic, glia-loadable. embedding floor: strong BUT the margin over counts is currently an uncontrolled artifact (tru#7). attention: net-negative at init in every configuration; ships quiet as a training substrate (tru#6); the fine-tuning ablation is the next real experiment.