research-loop

/research-loop — Research Derivation Loop

C4: Core research verification loop — drives the derivation forest from evolution results to deep motivation discovery.

Purpose

Starting from completed evolution results (code changes + performance data), iteratively build a derivation forest to discover why the changes work, find deep motivations, and grade contributions for paper writing.

Usage

/research-loop [evo_session_id]

If no session ID given, uses the current/most recent evolution session.

Behavior

Phase 1: Initialization

1. Call evo_get_status to get evolution results summary
2. Call evo_get_lineage for the best branch to understand the full change history
3. Call research_init_forest to create a new derivation forest
4. For each significant code change in the lineage:
   • Call research_add_node(type="change") to register it as a root node

Phase 2: Iterative Exploration (main loop)

Each iteration performs 5 steps:

Step 1 — Cut into code changes

• Read the current active change nodes
• Use code_qa to understand what each change does
• May merge or split change nodes via research_merge_nodes
• Add refined change nodes as needed

Step 2 — Reverse reasoning: why does it work?

• For each active change node, hypothesize why it improves performance
• Call research_add_node(type="hypothesis", parent_ids=[change_id])
• Consider: what unsolved domain problem does this address?

Step 3 — Literature search

• For each new hypothesis, call /ask-lit with the hypothesis as query
• Call research_add_node(type="evidence", literature_refs=[...]) for each finding
• Update hypothesis nodes with literature context

Step 4 — Experimental verification & visualization

• Design targeted experiments (ablation / control) to test hypotheses
• Call bench_adapt + bench_run to execute experiments
• Call bench_validate to check result reasonableness
• Call viz_generate to generate ablation curves, score distributions, and contribution heatmaps for each confirmed hypothesis. Save outputs to research/figures/.
• Call viz_polish to polish the figures for publication quality.
• Record figure paths in evidence nodes so write-experiment can reference them later.
• Supported: research_update_node(status="pruned") for rejected hypotheses
• Supported: continue deepening for confirmed hypotheses