wassname 2b020c95c0 fix: route2 Arm A flags per-rollout not per-token (external review) ...

The hook gate is necessarily per-token ([G*s, r], nn.Linear flattens the
batch). _route2_grad_filter now sums each rollout's token gate-grads before
the cos(g_b, v_grad) flag, so routing is per-rollout (the preregistered GRPO
unit) and the sign is denoised. Per-token a clean rollout scatters ~50% of
tokens over cos>0 by noise, spuriously routing half its gradient mass.

Verified by deepseek-v4-pro review: gate identity, divide-out, eps-guard,
Arm B detach-route, R5 no-cheat all correct; this was the one finding.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

2026-05-31 11:25:13 +00:00

..

blog

docs: fix review findings (global noise-floor, route one-sided, G3 xref)

2026-05-31 00:41:12 +00:00

brainstorm

ready

2026-05-23 14:19:41 +08:00

figs

docs: review outputs + figs; drop stale Qwen3.5-0.8B svd cache

2026-05-31 00:00:40 +00:00

grad_routing

docs: routing-v2 spec, related-work scorecard, paper fetches, journal

2026-05-31 10:16:21 +00:00

lab

lab report v3: TL;DR, three-line concept, PASS_RATE column, G_hack rename

2026-05-29 03:18:22 +00:00

papers

wip

2026-05-30 04:33:33 +00:00

personas

fix smoke.

2026-05-23 11:26:39 +08:00

reviews

tidy

2026-05-30 04:38:41 +00:00

spec

fix: route2 Arm A flags per-rollout not per-token (external review)

2026-05-31 11:25:13 +00:00

vendor

concepts

2026-05-29 06:29:20 +00:00

extract_vhack_grad-vec.md

Doc cleanup: mark susp gate as REMOVED in design doc

2026-05-27 09:08:34 +00:00

grpo_hyperparams.md

fix smoke.

2026-05-23 11:26:39 +08:00

human_journal.md

init

2026-05-23 10:22:54 +08:00

results.md

results: add Q10 (pairset mechanism>framing>placebo) + Q11 (60-step convergence gap closes)

2026-05-30 02:34:22 +00:00

spec.md

docs: fix review findings (global noise-floor, route one-sided, G3 xref)

2026-05-31 00:41:12 +00:00

table_proto.md

setup

2026-05-23 10:40:02 +08:00