Two theorems from running the network

[gHashTag]

Two theorems from trinity-fpga#588, which finally ran the network the earlier benchmarks kept refusing to claim anything about.

T43 — a proxy gate must yield to the task metric it stands in for. The 1%-flushed feasibility gate introduced in #587 was a hypothesis about perplexity. Perplexity disagreed: e3m4 zeroes 1.75% of GPT-2's weights on average and gives the lowest perplexity of four candidates, while e2m5 zeroes 12.65% and costs 16%. The gate was retracted to a reported number rather than kept beside the contradicting result — a repository holding two verdicts on one question holds none.

T44 — at 8 bits the golden-section split ties the deployed one.

split	gpt2	pythia-160m	opt-125m
rule e3m4	1.003×	1.021×	0.998×
e4m3 fields	1.009×	1.020×	1.005×
e5m2 fields	1.015×	1.081×	1.016×

First on two models, second by 0.1% on the third. That is a tie, not a win — and 0.998× is quantisation slightly beating the baseline, which is the noise floor these comparisons sit on. At 16 bits every candidate is lossless, so the rule is not distinguished; at 32 bits no task measurement exists and the proxies place it far from optimal.

T44's refusal is the one that matters: three models under 500M parameters is not a survey, and weights-only says nothing about the regime E5M2 exists for — activations carry outliers ~100× typical.

Ratchet: 179 errors across 26 files against baseline 184 across 27, no file gained. Language audits run locally before pushing: RU PASS (27 маршрутов), EN PASS (27 маршрутов).