Skip to content

docs(substantiate): CONTINUE.md — standing idempotent continuation prompt #82

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 3, 2026
Merged

docs(substantiate): CONTINUE.md — standing idempotent continuation prompt #82

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
47 changes: 47 additions & 0 deletions TODO.substantiate/CONTINUE.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,47 @@
# CONTINUE — the standing, idempotent continuation prompt

This is the fixed continuation prompt. It is stateless: re-run it any
number of times, in any session — it derives state from this register,
the Modal volumes, and git; it never relies on conversation memory and
never changes per iteration. Copy the block below verbatim.

---

Work the register: `interscript/rababa/TODO.substantiate/README.md` is
the index of record. For every item whose Status is not COMPLETE, do
that item's next open step exactly as its file and its linked
registrations specify; skip anything already done (statuses in the
README are the source of truth — update them in the same PR that
lands each step). Derive everything else from live state, not memory:

1. GKD (item 03, spec `ara-diac-small-2-gkd`, implemented in
interscript-ml): if no run dir `rababa_arabic_distill_small/
run-011-r7-muon-gkd` exists on the `rababa-checkpoints` volume AND
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 > /tmp/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
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`).
3. If every item is COMPLETE: report the register closed and stop.
Do not invent new work; new items require the owner.

Standing rules: numbers only from `docs/RESULTS.md` (the ledger);
commit via PRs off fresh `origin/main` (rebase-merge; squash if the
branch carries merge commits); never commit to main; no AI
attribution in commits; `git add` explicit paths only; Modal always
`--detach`, <= 2 big GPU apps; version numbers and releases are the
owner's; the GPU worktree is `/Users/mulgogi/src/interscript/ml-qwen-feat`.
7 changes: 6 additions & 1 deletion TODO.substantiate/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,10 +11,15 @@ 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 | ARMED (ml #147) — launch on owner word | owner ordering |
| [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 |
| [04](04-g2b-watch.md) | G2b verdict watch + cross-recording | P2 | WATCHING — no final_eval yet as of 2026-09-03; rung census: 8.26/7.56/4.83/5.29/5.08/4.57/5.09/5.81/5.78/73.95 | other agents' run |
| [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 |

## Continuation

[CONTINUE.md](CONTINUE.md) is the standing idempotent prompt — stateless,
repeatable, unchanged across iterations. Point any future session at it.

## Working rules (unchanged)

- Protocol-matched numbers only; every paper number must trace to
Expand Down