Datasets:

dCoder30
/

RSIGame-sft-data

Read CONTEXT.md first. This file records the reasoning behind choices that look arbitrary from the code, and — more usefully — the beliefs that were tested and found false, so they are not re-litigated.

Part 1 — Decisions and their reasoning

Generation is web-only; planning is both engines

Two candidate brief pools existed for expanding S1:

the 2,023 tasks/ briefs — every one Godot;
largescale_1500 — the campaign that produced the original 409, confirmed by 403 of the 409 task IDs matching exactly, leaving 1,097 never generated.

The second was chosen, plus td_100 + uiheavy_100 (two families S1 has never seen) = 1,297 untrained web briefs.

Reasoning. Every S1 row and every eval brief is web. Godot generation data would be trained on and never scored — invisible at best, and at worst trading measured web ability for unmeasured Godot ability. Plans are different: they need no recorded session, so the second engine is nearly free there, which is why S2 is deliberately Godot-majority. The asymmetry is intentional, not an oversight.

The queue is balanced so any prefix is usable

1,297 games × ~36 min ≈ 775 stream-hours — it cannot finish in the available window. So the question is not "when does it finish" but "what do we have when the plug is pulled". order_queue.py round-robins family and difficulty, so the first 60 or 600 games have the same mix as the whole rather than 441 platformers.

No claim directories, lock files or `.done` sentinels — anywhere

Any first-writer-wins guard survives a SIGKILL with nobody left to release it, and the blocked re-run then exits 0, so the work silently never happens. Static shards have no state to strand. If you add supervision, keep this property.

The converter keys on the artifact, not the success flag

traj_to_sft_v2.py keeps the last session that made tool calls and requires dist/index.html. Measured on the live corpus: 5 games had a built dist/ and no session marked successful. Keying on the flag would have discarded them. This is also why campaign "failures" still yield rows — real yield was 78%, not the 66% the tick/cross counts suggested.

Engine detected per game, not per directory

The original distiller labelled engine per source directory. The later snapshots mix engines inside one directory — 42 Godot games sat in folders labelled web. A Phaser prompt aimed at a Godot game produces plans full of TypeScript files that do not exist. Engine now comes from project.godot vs package.json.

Plans are read off the real file tree, never a narration

The distiller receives the actual file listing and source excerpts. Otherwise a README claiming a physics system the game lacks would have that claim copied into the training target. Every backticked path is then checked to exist: 36,554 paths checked, 16 missing (0.04%), 23 rows dropped. Some drops are false positives of a strict checker (src/ is a directory, res://Main.tscn is valid Godot) — at ~2% of rows, dropping is cheaper than adjudicating.

The system prompt must stay byte-identical

Three plan corpora built at different times pool into one training set only because their system prompt matches exactly. One stray character gives the model two tasks that look the same. validate_plans.py checks this; it is the single most load-bearing check in the pipeline.

10 workers, not 12

At 12, rss sat at 26–28 GiB against a 32 GiB cgroup with page cache exhausted, and the kernel OOM-killed agents in bursts of 2–4. A burst defeats retry — the retry dies in the same pressure — and yield fell to ~71%. At 10 the kills are isolated singles that retry absorbs; yield returned to ~100%. A ~17% throughput cut bought back more than it cost.

429 and `exited with code 1` were made retryable

agent_run.ts's retry predicate matched neither, so both ended sessions outright. Measured: 2 of 5 early failures were capacity 429s and 6 of 7 CLI deaths produced no dist/, together discarding 6.6 hours of sessions that were otherwise progressing, with 60–348 KB trajectories already accumulated. A retry resumes in the same directory keeping prior work, so retrying costs a wait and risks nothing. Capacity faults get a 90 s backoff (the 15 s default is tuned to socket reuse and is far too short to outlast a shortage).

Part 2 — Beliefs that were tested and found FALSE

This section matters more than Part 1. Each of these was asserted, acted on, and then disproved. Do not re-adopt them.

✗ "My concurrency is causing the 429s"

Test: 12 heavy concurrent requests returned 12/12 in 3.7 s while the campaign was running 15 agents. The endpoint was not saturated. The 429s are intermittent org-wide capacity events. Backing off workers would have cost throughput and fixed nothing. (Workers were later reduced for a different, real reason — memory.)

✗ "Memory is critically high, intervene now"

Raised the alarm at rss 28/32 GiB. Reading the cgroup properly showed the kernel reclaiming page cache (18.9 → 1.9 GiB) rather than OOM-killing, and rss fell to 22 GiB on its own as games completed. free and uptime are misleading on this box; read /sys/fs/cgroup/memory/memory.stat and separate rss from cache.

✗ "new-S3 is clearly worse than old-S3"

Withdrawn after a second run. At n=2 briefs the same adapter scored 0.000 and 0.149. Run-to-run variance exceeds every between-arm gap.

✗ "122 rows were silently binned by `truncation_strategy=delete`"

Falsified by the source and by 409/409 step counts — after it had already justified a 17-hour GPU job. delete matches no branch, so rows pass through untouched; nothing was dropped. The real implication is the opposite of the alarm: set MAXLEN high enough rather than assuming truncation happened.

✗ "The duplicate-scene bug is specific to the reweighted run"

Falsified — old-S2 (v1) crashed on both briefs too.

✗ "There are 369 trajectory files" / "the HF corpus has new trajectories"

Both wrong. A bad grep produced 369; provenance showed 162+161+86 = 409. And the HF corpus contains no new trajectories — md5-verified identical. The ~900 extra finished games have no trajectory files at all, which is precisely why they were usable for plans and useless for generation.

Pattern worth internalising

Every one of these was a confident claim made from a partial reading, and several drove real work before being checked. The user noticed and asked why it kept happening. Measure first, then assert — especially before spending GPU hours or telling the user something is broken.

Part 3 — Things settled, so don't re-investigate

S1 needs no retrain for data-loss reasons — nothing was lost. (It may be retrained to use more data, which is different.)
The HF gamecraft-1700-corpus holds no new trajectories — md5-identical to what was already in hand.
godot_55high is excluded by user instruction; verified absent everywhere.
cloudnative-my / H20 clusters are unreachable for our jobs.
bi_base_model group is not accessible — explicit refusal, settled.
The trajectory recorder IS on GitHub (branch worktree-negative-evidence-scope contains class Trajectory) — but the retry patch on top is uncommitted, so a clone alone rebuilds a harness that loses ~30% of its games. harness/ in the handoff carries the patched files.
An a2f archive is self-sufficient — verified by extracting three at random where the originals were unreachable; all three regenerated valid training rows.

DECISIONS — why things are the way they are