openlair / ds-full-pipeline

DeepScientist Full Pipeline

End-to-end autonomous research workflow for: $ARGUMENTS

Overview

This skill chains all DeepScientist research stages into a single pipeline:

/ds-scout → /ds-baseline → /ds-idea → /ds-experiment → /ds-analysis-campaign → /ds-optimize → /ds-write → /ds-review → /ds-finalize

Pipeline

Stage 1: Scout

Frame the research problem, survey literature, identify datasets/metrics, discover existing baselines.

/ds-scout "$ARGUMENTS"

Output: Problem framing, literature map, baseline shortlist, evaluation contract.

🚦 Gate 1: Present the research landscape to the user. Wait for confirmation before proceeding.

Stage 2: Baseline

Reproduce or import the most relevant baseline from Stage 1's shortlist.

/ds-baseline

Output: Working baseline with verified metrics, comparability contract.

Stage 3: Idea

Generate concrete research hypotheses based on the literature gaps and baseline analysis.

/ds-idea

Output: Ranked candidate ideas with selection rationale.

🚦 Gate 2: Present top ideas to the user. Wait for confirmation of which idea to pursue.

Stage 4: Experiment

Implement and run the main experiment for the selected idea.

/ds-experiment

Output: Experiment code, results, evidence artifacts.

Stage 5: Analysis Campaign

Run follow-up experiments: ablations, robustness checks, error analysis.

/ds-analysis-campaign

Output: Ablation results, robustness data, writing-facing evidence slices.

Stage 6: Optimize (Optional)

If results are promising but not yet strong enough, run algorithm-first iterative improvement.

/ds-optimize

Skip this stage if main experiment results already meet the success criteria.

Stage 7: Write

Draft the paper from accepted evidence.

/ds-write

Output: LaTeX paper draft with figures and references.

Stage 8: Review

Run an independent skeptical audit of the draft.

/ds-review

Output: Review report with severity-graded feedback.

If review identifies critical issues → fix and re-review (max 2 rounds).

Stage 9: Finalize

Consolidate final claims, limitations, and recommendations.

/ds-finalize

Output: Final paper, summary state, resume packet.

Key Rules

• Gate checkpoints after Scout and Idea stages. Do not proceed without user confirmation on research direction and idea selection.
• Stages 4-9 can run autonomously once the user confirms the idea.
• Evidence-first writing. Every claim in the paper must trace to an experiment artifact.
• Fail gracefully. If any stage fails, report clearly and suggest alternatives rather than forcing forward.
• Git as memory. Commit after each stage so progress is durable.

Typical Timeline