Skip to content

docs: G2b verdict — domain-residual attribution rejected - #170

Merged
ronaldtse merged 1 commit into
mainfrom
g2b-verdict
Sep 4, 2026
Merged

docs: G2b verdict — domain-residual attribution rejected#170
ronaldtse merged 1 commit into
mainfrom
g2b-verdict

Conversation

@ronaldtse

Copy link
Copy Markdown
Contributor

Result

G2b (G2a recipe + full cleaned Tashkeela corpus, 5x classical
coverage) full-set gate: student 4.8231, teacher 2.289, paired
bootstrap delta 2.3717 [2.194, 2.554], p=0.

  • Gate (>= 0.5pp over G2a's 4.5701): FAILED
  • Point estimate 0.25pp WORSE, CIs overlapping with G2a's [1.91, 2.35]

Interpretation

With E6 (register swap at constant budget: -0.98pp), both directions
of the classical-corpus manipulation fail. The student-tier residual
is not a classical-domain coverage deficit. Reframed as a
teacher-student interaction the corpus cannot reach; on-policy
distillation (GKD, in flight) is the live lever.

Provenance: labels sha256 b59e2f56 (235.0MB). Paper B's frontier cell
follows.

5x classical corpus left the rung statistically flat vs G2a (4.8231
[2.194,2.554] vs 4.5701 [1.91,2.35]); with E6's swap-negative, both
directions of the classical-corpus lever fail. The residual is not a
domain-coverage deficit; G2a stands as best student.
@ronaldtse
ronaldtse merged commit 77d0353 into main Sep 4, 2026
7 of 10 checks passed
@ronaldtse
ronaldtse deleted the g2b-verdict branch September 4, 2026 11:43
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant