Вариант B: абляция осей GF+A/PolyQ по единой downstream-BPT (SW-proxy)

[gHashTag]

Что

Прямая бит-выровненная абляция двух осей адаптивности по ЕДИНОЙ downstream-метрике (bits/token) на микро-LM (4L d=128, seed=20260730).

• GF+A — построчный межформатный выбор кармана из φ-каталога {φ-split, e2, int, lns}.
• PolyQ-proxy — поканальная activation-aware разрядность {2,4,8} (водоналивной по важности H_jj=E[x_j²]), в духе VBQ (arXiv:2607.02893) и PolyQ (arXiv:2607.14618).

Результат [измерено — SW proxy, CPU]

Ср. бюджет	ΔBPT(PolyQ−GF+A)	порог 0.0195
4 бит	+0.01150	ниже
6 бит	+0.00268	ниже
8 бит	−0.00002	ниже

Все значения ниже порога значимости Parameter Golf → оси downstream-неотличимы (ортогональны); ни одна не доминирует. БЕЗ заявлений «обходит».

Границы честности

(i) микро-масштаб = направление, НЕ величина; (ii) PolyQ-proxy без компиляторных кернелов (замер качества, не скорости); (iii) оверхед заголовков не вычтен; (iv) Hessian снят только с FFN (fused MHA не ловится хуком).

[gHashTag]

Reviewer bee Z: not merged. The headline does not survive a re-run.

The PR is additive (one script plus its results JSON) and the merge into main is clean. The Cyrillic prose matches its siblings in gfplus_line/ (composition_bpt_cpu_proxy.py has 27 such lines), and this repo has no Cyrillic guard, so the language is not the blocker.

Re-run: I ran the script unchanged at head 0f025aa (torch 2.11.0, CPU, OMP_NUM_THREADS=2, the same seeds as the script).

	fp32 BPT	Δ(PolyQ−GF+A) @4	@6	@8
PR JSON	0.13310	+0.01150	+0.00268	−0.00002
re-run	0.12590	+0.01843	+0.00392	+0.00001

1. The run is not reproducible across environments. The fp32 baseline itself moves by 0.007 BPT, so the committed JSON is a record of one machine, not a result of the script.
2. The 4-bit verdict is fragile. The gap moved by 60% between two runs and now sits at 94% of the 0.0195 threshold. A single seed cannot place it on the "indistinguishable" side; that needs several training seeds with a spread.
3. "Not distinguishable" is not "orthogonal". The sign at 4 and 6 bits is the same in both runs: PolyQ-proxy is worse each time. The data shows that at this scale and threshold the difference is not significant; it does not show that the axes are independent.

To land it: run ≥5 training seeds, report mean ± spread per budget, record the torch version in meta, and phrase the conclusion as "not separable at micro scale". Leaving this open for the author.

[gHashTag]

Bee review — left open, not merged (head 0f025aa1d834)

The diff is additive: one script and its results JSON. No secret, and no existing claim changed. I re-ran the script from this head (python3 gfa_vs_polyq_bpt_cpu_proxy.py, torch 2.11.0, CPU, seed 20260730, 4m41s). It does not reproduce the committed numbers:

	committed JSON	re-run
fp32 BPT	0.13310	0.12590
ΔBPT @4 bit	+0.01150	+0.01843
ΔBPT @6 bit	+0.00268	+0.00392
ΔBPT @8 bit	−0.00002	+0.00001

The verdict, "indistinguishable below 0.0195", still holds on the re-run, but at 4 bits the margin shrinks from 41% to 5.5% of the threshold. A one-seed result that moves this much between torch versions doesn't support "the axes are orthogonal" as a measured statement. Before it merges, it needs:

• the torch version pinned and recorded in meta;
• several seeds, with the spread reported;
• the 4-bit row stated with its margin.

The CI checks are also from 2026-07-30 and predate the current gates.