EPS
← All tasks·#457On hold for now

Pre-experiment planning gate: declare target plot / expected result before agents implement

kind: infra#needs-thomas
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Idea

Front-load the human (needs-thomas) thinking/reading/planning for experiments so the agent-ok implementation step is high-leverage instead of churning shallow runs. The point: more time planning up front, then agents implement well afterward.

Concrete mechanism from Thomas's note: a pre-experiment GATE that declares the target plot / expected result BEFORE the experiment is run (or before an agent picks it up). "What plot do we want to see at the end?" becomes a required field on a proposed experiment.

Why

  • Operationalizes Dan's recurring breadth-vs-depth feedback (fewer experiments nailed carefully > many shallow). See mentor_feedback_breadth_vs_depth.
  • Fits the agent-ok / needs-thomas split: target-plot declaration + planning is the needs-thomas step; once the gate passes, agents take the agent-ok implementation.
  • Guards against the uninterpreted-clean-results problem (EPS #434/#435): declaring the target plot up front defines what "done + interpreted" looks like before the run.

Open questions

  • Where does the gate live? A required target_plot: / expected_result: field in the task body frontmatter for kind=experiment, enforced by task.py at the proposed->approved transition?
  • Should an experiment be blocked from moving to running / agent-ok until that field is filled?
  • Granularity: one target plot per experiment, or a small set of expected figures?
  • Add a short "reading/thinking done" checkbox before greenlight?

Source

my-goat store todo: need-to-spend-more-time-thinking-reading-plannin-20260531-01 (Todoist 6gmMggVwxr27Gjpv)

Activity