Requesting Research Review

Dispatch eureka:research-reviewer subagent to catch scientific issues before they propagate. The reviewer gets precisely crafted context for evaluation — never your session's history. This keeps the reviewer focused on the evidence, not your thought process, and preserves your own context for continued work.

Core principle: Review before you conclude.

When to Request Review

Mandatory:

• After each major experiment phase (baselines, ablations, main experiments)
• Before writing any Results or Conclusion section
• Before submitting to any journal, conference, or preprint server
• Before reporting results to collaborators or supervisors

Optional but valuable:

• When results are surprising (real signal or artifact?)
• After revising analysis in response to feedback
• Before committing to a new research direction based on results
• When stuck interpreting ambiguous results

How to Request

1. Identify what you're reviewing:

Determine which research phase just completed and what the reviewer needs to evaluate.

2. Gather the review context:

RESEARCH_QUESTION   = The exact hypothesis being tested
DESIGN_DOC_PATH     = Path to the approved research design document
RESULTS_PATHS       = Paths to ALL result files (not just summaries)
PHASE_DESCRIPTION   = Which phase just completed (e.g., "baseline experiments")
DOMAIN_CONTEXT      = Field-specific context, tools, typical effect sizes
TARGET_VENUE        = Target journal or conference (sets the bar)
PASS_THRESHOLD      = Score threshold (85 for mid-project, 95 for pre-submission)

3. Dispatch the research-reviewer subagent:

Use the Agent tool with eureka:research-reviewer type. Fill the template at requesting-research-review/research-reviewer-prompt.md with the gathered context.

4. Act on feedback:

• CRITICAL issues: Stop work. Do not report results. Do not proceed to the next phase. Fix the issue and request re-review.
• MAJOR issues: Fix before next phase. Do not build on flawed foundations.
• MINOR issues: Track and address before publication.
• Push back if the reviewer is wrong — with specific evidence from your data.

Choosing the Right Threshold

Example

[Just completed baseline experiments comparing three models on a held-out test set]

You: Let me request a research review before interpreting these results.

RESEARCH_QUESTION   = "Does Model C outperform Model A and Model B baselines
                       on outcome Y prediction?"
DESIGN_DOC_PATH     = docs/eureka/designs/2026-04-10-model-baselines-design.md
RESULTS_PATHS       = results/run_20260411_171136/cv_results.json,
                      results/baseline_modelA/, results/baseline_modelB/
PHASE_DESCRIPTION   = Baseline experiment completion (Model A, Model B, Model C)
DOMAIN_CONTEXT      = Supervised prediction on cohort data,
                      Pearson r and R-squared are standard metrics
TARGET_VENUE        = [target journal]
PASS_THRESHOLD      = 85

[Dispatch eureka:research-reviewer subagent]

[Subagent returns]:
  D1: 88  D2: 82  D3: 75  D4: 70  D5: 72  D6: 78  D7: 80
  Overall: FAIL (D3, D4 below threshold)
  Critical: Single seed results reported — need ≥5 seeds
  Major: No confidence intervals on Pearson r comparisons

You: [Fix: re-run with 10 seeds, add bootstrap CIs]
[Request re-review after fixes]

Integration

• Called by: eureka:using-eureka (when experiment phase completion detected)
• Dispatches: eureka:research-reviewer agent (via Agent tool)
• Pairs with: eureka:receiving-research-review (for handling the feedback)

Red Flags

Never:

• Skip review because "the results are obvious"
• Ignore Critical issues
• Proceed to writing with unresolved Major issues
• Dismiss reviewer feedback without checking against your actual data

If reviewer is wrong:

• Push back with specific evidence (file paths, computed values)
• Show the data that contradicts the reviewer's assessment
• Request clarification on unclear feedback

See template at: requesting-research-review/research-reviewer-prompt.md