FUTURE — what to do next, and why in this order
Read CONTEXT.md and DECISIONS.md first. Ordering here is by value per unit of
effort, not by how interesting the work is.
1. Expand the evaluation set — do this before any more training
Why first. Every conclusion in this project is capped by having 2 shared briefs, where the same adapter scored 0.000 and 0.149 on re-run. No amount of additional training makes the stage-2-versus-stage-3 question answerable; only more briefs do. This costs no GPU at all.
What to do
- Raise the shared-brief count to a number where the judge's ~0.05 test-retest noise is small against the effects you want to detect. At the observed spread you need tens of briefs, not two.
- Make it engine-balanced. Adding Godot briefs simultaneously closes the
measurement gap on 870 of 1,679 plan rows. There is already a pool of 2,023 Godot
briefs (
instructions/) to draw from. - Keep the existing protocol — build gate, 3 judged passes, judge errors excluded rather than zeroed. Changing the protocol and the brief count at once makes the new numbers incomparable to the old.
This one step dissolves two of the five standing limitations.
2. Retrain stage 1 on the expanded corpus
The corpus is built and validated (s1_train_full.jsonl, ~514 rows and growing —
+25% on the original 409). A 4×H100 job config is prepared with MAXLEN=81920, U13
keys, and fresh output paths.
Open question the user must answer: where to draw the cutoff. The full 1,297-brief campaign is ~775 stream-hours and will not finish. Training on what exists, then retraining later if more accumulates, is cheap relative to the campaign and guarantees a checkpoint exists.
Constraints from the user, carry these forward: fresh output paths (never
overwrite existing adapters), no grad_accum, do not change the loss.
3. Rebuild stages 2 and 3 on the full plan corpus
Both were trained before the gap distillation added 317 verified rows. The rebuilt mixes already exist:
| rows | generation share | |
|---|---|---|
stage2_gen_plan_full_v2.jsonl |
2,156 | 22% |
stage2_gen_plan_balanced_v2.jsonl |
954 | 50% |
build_s2_variants_v2.py is re-runnable and picks up whatever generation rows exist
at the time, so run it again after the campaign stops.
4. Settle the generation-to-plan ratio
The original mix was 47% generation. A reweighting experiment existed because someone judged that too low. Pooling all plan data takes it to 22% — further in the direction that caused the concern. Holding generation at 50% means discarding plan rows.
Both corpora are built; neither is declared correct because the current evaluation cannot tell them apart. This is blocked on step 1, not on training. Note that growing the generation corpus (step 2) pushes the ratio back up on its own — that is part of what the campaign is for.
5. Keep the generation campaign running while access allows
1,297 briefs queued, ~10 rows/hour at 10 workers, archived continuously to a2f and HuggingFace every 30 minutes. The queue is ordered so whatever completes is a balanced corpus, not a biased fragment. It is self-healing and needs no attention.
If restarting on a new machine, see REBUILD.md §5 — and note the two families
(tower_defense, ui_heavy, 100 briefs each) that stage 1 has never seen.
Open questions worth investigating
Why does S3 v1 score highest (0.293) with no recorded training metrics? It beats every reweighted arm, but nothing is known about how it trained. Either recover its logs or re-run it — at present the best-scoring arm is the least understood.
Is the Presentation-versus-Mechanics gap a training-data property? Every arm scores 0.60–0.87 on Presentation & Art and 0.00–0.61 on Core Mechanics; even the untuned base makes something that looks like a game while implementing none of it. If the corpus over-represents visual scaffolding relative to mechanics, that is fixable in the data — but it needs an evaluation that can measure it.
Does build-gate survivorship hurt stage 1? Only games that compile enter the corpus, so the student never sees a failed build in stage 1. That is what stage-3 repair data is meant to supply — but whether the split works as intended is untested.
Does the Godot plan data help or hurt web performance? 870 Godot rows sit in a corpus evaluated only on web. Cross-engine transfer might be positive (structural planning generalises) or negative (vocabulary interference). Currently unmeasurable — another thing step 1 unlocks.
Things NOT to do
- Do not rank the tuned arms against each other using the current evaluation. Tuned-beats-base survives; arm-versus-arm does not.
- Do not add Godot data to the generation corpus to "improve Godot" — it would be trained on and never measured. Fix the evaluation instead.
- Do not add lock files or claim directories to the campaign supervision. See
DECISIONS.md. - Do not overwrite the existing adapters. Fresh paths, always.
- Do not assert a problem before measuring it.
DECISIONS.mdPart 2 lists six confident claims that turned out false, several of which drove real work first.