Skip to content

docs: GKD verdict 6.0036 gate FAILED — register COMPLETE 5/5 #89

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ronaldtse merged 1 commit into from
Sep 5, 2026
Merged

docs: GKD verdict 6.0036 gate FAILED — register COMPLETE 5/5 #89

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion TODO.substantiate/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@ COMPLETE and closed: TODO.training-work, TODO.publish.
|---|---|---|---|---|
| [01](01-paper-currency.md) | Three latex papers current with the Sept results | P1 — highest value | COMPLETE (rababa #80) — venue formatting/submission is the owner's | nothing |
| [02](02-ce-curve-comparison.md) | E5/E6 CE-curve vs run-006 at matched steps | P1 — quick, mechanical | COMPLETE (ml #146) | nothing |
| [03](03-gkd-arming.md) | GKD rung: arm spec + registration (launch = owner) | P2 | IMPLEMENTED (ml #148) — launch queued on GPU slot; driver = [CONTINUE.md](CONTINUE.md) step 1 | GPU slot |
| [03](03-gkd-arming.md) | GKD rung: arm spec + registration (launch = owner) | P2 | COMPLETE (2026-09-05) — verdict 6.0036 GATE FAILED; implemented ml #148/#151, run-012 (nine preemptions absorbed), prediction missed; recorded ml EXPERIMENTS + PUBLICATION-NOTES; papers amended | nothing |
| [04](04-g2b-watch.md) | G2b verdict watch + cross-recording | P2 | COMPLETE (2026-09-04) — verdict 4.8231 recorded (ml PR pending); domain hypothesis negative both directions
| [05](05-project-closeouts.md) | Issue closures #38/#41 + legacy alert dismissals | P3 | COMPLETE — #38/#41 closed with cross-refs; all 56 Dependabot alerts auto-FIXED by #48+#78 on rescan (no dismissals needed) | nothing |

Expand Down
Binary file modified docs/paper-arabic/main.pdf
View file
Open in desktop
Binary file not shown.
30 changes: 19 additions & 11 deletions docs/paper-arabic/main.tex
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -88,10 +88,12 @@ \section{Introduction}
$\rightarrow$ \textbf{4.82\%} (42\% error reduction at
identical architecture) via a controlled lever ladder
(optimizer $-2.96$pp, fresher teacher labels $-0.47$pp,
longer training $-0.25$pp), plus two registered negatives
(multi-token-prediction auxiliary $+0.26$pp;
constant-budget register swap $+0.98$pp) showing neither
frontier-training techniques nor register mixing substitute
longer training $-0.25$pp), plus four registered negatives
(multi-token-prediction auxiliary $+0.26$pp; register swap
$+0.98$pp; register add flat-with-epochs-cancelled at 4.82;
on-policy GKD $+1.18$pp, the worst rung) showing that neither
frontier-training techniques, register mixing in either
direction, nor on-policy exposure substitute
for supervision quality.
\item A three-way RL negative result (RAFT, sequence-GRPO on Persian,
entropy-weighted GTPO-GRPO on Arabic): all flat or negative at
Expand Down Expand Up @@ -227,15 +229,21 @@ \subsection{The distilled client tier}
contributes $-2.96$pp (a controlled 2$\times$2 factorial closes
additively: optimizer $-2.96$, memory-layer capacity $-0.70$,
combined $-3.43$), fresher r7 teacher labels $-0.47$pp, and longer
training a further $-0.25$pp (4.57\%). Two registered negatives
training a further $-0.25$pp (4.57\%). Four registered negatives
bracket the ladder: multi-token-prediction as a training auxiliary
scored 5.09\% ($+0.26$pp vs control, with a disclosed preemption
confound), and swapping news-domain training units for classical
Tashkeela at constant budget scored 5.81\% ($+0.98$pp)---the causal
test of the domain-coverage hypothesis failing in the swap direction.
Together with the RL negatives, the pattern is one line: at SFT
convergence, supervision quality dominates policy optimization,
frontier training techniques, and register diversification alike. The
confound); swapping news-domain training units for classical
Tashkeela at constant budget scored 5.81\% ($+0.98$pp); adding
18k classical units on top at matched epochs scored 4.82\%---flat
against the 3-epoch control and 0.25pp behind news-only training at
the same epochs, the add canceling the epoch gain entirely; and
on-policy generalized distillation (reverse-KL on student-sampled
sequences scored by the frozen teacher) scored 6.00\% ($+1.18$pp,
the worst rung measured)---the domain-coverage hypothesis failing
in every direction it was tested. Together with the RL negatives,
the pattern is one line: at SFT convergence, supervision quality
dominates policy optimization, frontier training techniques,
on-policy exposure, and register diversification alike. The
tier ships as checksummed, runtime-agnostic artifacts (IMF v1) whose
quantization fragility was itself diagnosed and fixed---the quantized
output head, not the body, caused confident flip errors (9.34\%
Expand Down