Skip to content

docs(substantiate): driver signals — best/ over step counts; G2b unchained-eval rule #86

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 4, 2026
Merged

docs(substantiate): driver signals — best/ over step counts; G2b unchained-eval rule #86

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
23 changes: 14 additions & 9 deletions TODO.substantiate/CONTINUE.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,22 +20,27 @@ lands each step). Derive everything else from live state, not memory:
fewer than two ephemeral GPU apps are running (`modal app list`),
launch: `cd /Users/mulgogi/src/interscript/ml-qwen-feat && nohup
modal run --detach src/gpu/modal_distill.py::main --spec
ara-diac-small-2-gkd > /Users/mulgogi/gkd_distill.log 2>&1 &`. If the run dir
exists and training is past step 11073 but no `final_eval.json`,
run `::evaluate_der --spec-id ara-diac-small-2-gkd` the same way
(main does not chain it). When final_eval.json exists, verdict vs
ara-diac-small-2-gkd > /Users/mulgogi/gkd_distill.log 2>&1 &`. If the
run dir exists with a best/ checkpoint but no final_eval.json
(training complete, eval unchained), run
`::evaluate_der --spec-id ara-diac-small-2-gkd` the same way
(main does not chain it; never rely on step counts — they drift). When final_eval.json exists, verdict vs
the registered gate (adopt <= 4.5218; honest band to 4.8218;
prediction 4.30-4.65) into EXPERIMENTS.md E-GKD + PUBLICATION-
NOTES Paper-B lever table + TODO.substantiate statuses; mark item
03 COMPLETE either way. Never launch if the budget is full; just
report the blocker.
2. G2b watch (item 04): read `final_eval.json` under every
`run-*tashkeela*` / 48k run dir on the volume (tiny modal
volume-read, timeout 120). If a G2b verdict exists and is NOT yet
recorded in EXPERIMENTS.md, record it + the E6-swap vs G2b-add
pair in PUBLICATION-NOTES Paper-B framing, mark item 04 COMPLETE;
otherwise leave WATCHING and report progress (step count from
`modal app logs`).
volume-read, timeout 120). If G2b training completed (best/
checkpoint exists) but no final_eval.json and no app is running
it, run `::evaluate_der --spec-id ara-diac-small-2-6ep-tashkeela`
the same detach way (the other agents' rung; the eval is read-only
decode + scoring, safe to complete). If a G2b verdict exists and
is NOT yet recorded in EXPERIMENTS.md, record it + the E6-swap vs
G2b-add pair in PUBLICATION-NOTES Paper-B framing, mark item 04
COMPLETE; otherwise leave WATCHING and report progress (step
count from `modal app logs`).
3. If every item is COMPLETE: report the register closed and stop.
Do not invent new work; new items require the owner.

Expand Down