| name | auto-research |
| description | Use when preparing Ralphthon research specifications or 연구 명세, automated experiments, a Track 1 short paper or 논문, a Track 2 review agent or 리뷰, or an evidence-backed submission under the event deadline. |
Auto Research
Overview
Turn one testable idea into reproducible evidence, then produce a submission-ready artifact. When the task needs new autoresearch training evidence, run the official VESSL cookbook recipe on one approved A100 with a bounded safety and W&B observability overlay. Keep claims no stronger than the recorded results.
Read the event workflow and evidence contract before planning. For a Training campaign, also read the complete official VESSL autoresearch runbook. Copy the appropriate template from assets/ into the participant workspace.
Preflight
Confirm the selected Track, current time, 16:30 submission hard cut, permitted public data, available evidence, and whether the work creates new autoresearch training evidence. Do not load private participant, reviewer, messaging, or operations records.
Submission Contract
| Path | Required output |
|---|
| Track 1 | Agent workflow plus a 2-4 page workshop-style short paper and self-review |
| Track 2 | Review agent plus an ICML-style review result for a Track 1 paper |
| Both | Complete Track 1 evidence first, then review the frozen paper with Track 2 |
Treat the final paper/agent submission at 16:30 as a hard cut. Peer and self-review may follow. During the Ralph Loop, operate through the coding agent and preserve an audit trail of prompts, code, runs, and outputs.
Workflow
Select exactly one path below. Do not merge the compute and no-compute preflight requirements. Use the bundled references, assets, and recorder directly from this skill.
Select exactly one path
Training path (Track 1 or Both)
Choose this path only when generating new autoresearch training evidence. It includes all compute, cost, metric, and onboarding requirements below.
- Freeze a research spec containing one falsifiable hypothesis, a named baseline,
val_bpb as the evaluation metric, dataset, budget, and stop condition.
- REQUIRED SUB-SKILL: Use
wandb-onboarding, including its synthetic offline run. Before online sync, show entity, project, visibility, and the exact W&B allowlist and obtain explicit confirmation.
- REQUIRED SUB-SKILL: Use
vessl-cloud-onboarding. Run vesslctl resource-spec list --usable-only -o json; show the live exact single A100 spec, hourly price, credit, image, object volume, wall-clock cap, total estimate, and cleanup/timeout plan. Provision only after explicit confirmation.
- Use
vessl-ai/vessl-cloud-cookbook/autoresearch pinned at 97a0af14b0acae042162b1f70f17fbe2d570afa2 as the execution SOT. Use a participant-owned writable fork and the recipe's vesslctl job create flow. Override its default with the approved live A100 spec; do not fall back to another GPU, CPU, or a larger model, and never substitute H100.
- Keep the official benchmark unchanged: this is an unchanged benchmark, not a reimplementation. The baseline uses the pinned
prepare.py and train.py; candidates may modify only train.py, one hypothesis and one change at a time. Do not modify the evaluation harness, dependencies, benchmark reports, data, tokenizer, or batch-job scripts.
- Copy the campaign control file, VESSL A100 run card, experiment ledger template, and
scripts/record_experiment.py, then execute the complete official VESSL autoresearch runbook. The recorder consumes each fetched VESSL log locally, appends experiments.jsonl, and creates only an allowlisted W&B offline run until separately approved sync.
- Run one baseline, at most three candidate trials executed sequentially, and one winner confirmation. Do not use the cookbook's parallel fan-out or unbounded loop. Keep only a lower
val_bpb.
Do not fall back when the approved single A100 is unavailable; stop and report the blocker.
8. Separate evidence from interpretation, report failure modes and limitations, and produce Track 1 plus its self-review. For Both, freeze Track 1, copy the Track 2 Review Agent template to review-agent.md, freeze its version/input hashes, and run it with the Track 2 review result template. Submit both artifacts.
Training uses a W&B allowlist containing only run_tag, trial, Git and cookbook SHAs, remote branch/commit, VESSL Job slug/name/state, approved/actual resource spec, exact A100 model/count, canonical cache and evidence hashes, val_bpb, reported parameter/VRAM/time values, and status. Store runs locally with group=run_tag and job_type=autoresearch-trial. Exclude console logs, source code, Git patches, requirements, machine metrics/info, datasets, checkpoints, artifacts, secrets, and private participant data. W&B's offline format still includes internal SDK metadata (sanitized host, SDK/Python/platform versions, timestamps/runtime, and SDK telemetry records); disclose it in the action-time sync approval.
The cookbook recipe is the sole execution profile; no custom micro preset is provided. Preserve its benchmark settings and record A100 results only against the same pinned cache/evaluation identity. The cookbook's polling timeout does not terminate a remote Job: inspect the Job and execute the approved termination action when it fires. Append every outcome to experiments.jsonl, and preserve unrelated files with file-scoped Git operations. DO NOT USE git reset --hard, git add -A, LOOP FOREVER, or NEVER STOP.
General Track 1 path (or Both)
Choose this path for a non-Karpathy research workflow or when credible evidence already exists. It does not automatically require W&B, VESSL, or A100. Do not force this path through the A100 runbook or require val_bpb when a different metric is appropriate.
- Copy the Track 1 submission template.
- Freeze a research spec with one falsifiable hypothesis, a fair baseline, primary evaluation metric and threshold, evidence inputs, budget, stop condition, and failure modes.
- Record the research agent workflow, versions, commands/prompts, human checkpoints, raw outputs, exclusions, and negative or inconclusive results.
- Recompute the headline result from the frozen evidence, then produce the 2-4 page paper and self-review. Never present planned work as an observed run.
- For Both, freeze the paper and evidence bundle before running the Track 2 Review Agent below.
Track 2-only frozen-paper path
Choose this mutually exclusive path only to review a frozen Track 1 paper and its existing evidence. It can skip training onboarding and the A100 runbook: it does not require val_bpb, does not require a GPU, and does not require W&B or VESSL, a cost card, account creation, or new experiments.
- Record the frozen paper version and the supplied evidence paths without changing them.
- Copy the Track 2 Review Agent template to
review-agent.md; freeze its name/version, input hashes, instruction, output contract, and evidence guardrails.
- Run that agent with the Track 2 review result template to assess soundness, evidence support, weaknesses, questions, and confidence.
- Trace review statements to the frozen artifact. Do not invent a missing run or demand Karpathy-specific metrics from unrelated research.
- Return both
review-agent.md and the ICML-style review result with blockers. If evidence is insufficient, say so; do not silently switch paths.
Integrity Gate
Do not fabricate results, citations, runs, reviewer evidence, or sponsor claims. Never backfill experiments to match prose. Label planned or expected results explicitly. If a run fails, report the failure and narrow the claim.
Stop and correct the artifact when:
- a number cannot be traced to a saved result;
- the paper and review use different experiment versions;
- a claimed improvement lacks the named baseline or metric;
- negative or inconclusive evidence was omitted;
- private participant, reviewer, or operations data appears in the artifact.
On the Training path only, also stop when the live GPU is not exactly one A100, an approved price/resource field changed, or a candidate changes the data, tokenizer, evaluation harness, or more than one experimental factor. These compute gates do not apply to Track 2-only.
Verification
- Recompute headline values from the frozen evidence and confirm the winner rerun before claiming improvement.
- Trace every number and central claim to a saved result; verify citations, the 2-4 page limit, baseline fairness, metric validity, negative results, and self-review.
- For Track 2, verify the paper/evidence hashes and that
review-agent.md produced the structured result without inventing missing evidence.
- For Training, inspect the append-only ledger, offline W&B run directories, VESSL Job correlation, exact experiment count, best confirmed
val_bpb, artifact paths, cost exposure, and cleanup state.
- Treat MFU as diagnostic only; use
val_bpb as the research metric. Do not compare A100 measurements directly with the official H100 benchmark.
Output
For Training, return the frozen research spec, resource/approval state, exact experiment count, best confirmed val_bpb, artifact paths, Track artifacts, cleanup, blockers, and next action. For General Track 1, return the research spec, agent workflow, frozen evidence, short paper, self-review, blockers, and next action. For Track 2-only, return the frozen paper version, supplied evidence paths, review-agent.md, ICML-style review result, confidence, blockers, and next review action; omit all compute, cost, W&B, GPU, and training-metric fields.
Next Steps
- Track 1 complete: freeze the paper and run self-review.
- Track 2 complete: attach both
review-agent.md and the structured review with evidence trace.
- Both complete: confirm the reviewer used the frozen Track 1 version.
- Training complete: obtain the documented Pause/Terminate cleanup decision and report any remaining storage exposure.
Common Mistakes
- Training: broad topic without a falsifiable hypothesis → reduce to one measurable comparison.
- Training: paper-first workflow → finish one real end-to-end run before polishing prose.
- Review-as-summary → score soundness, evidence, weaknesses, questions, and confidence.
- Deadline optimism → preserve a complete small result instead of an unfinished large experiment.
- Training: cheap-credit optimism → live hourly price, storage, duration, and cleanup still require explicit confirmation.
- Training: single lucky score → rerun the winner before making a Track 1 improvement claim.