research-closure-harness: Instructions file for Codex

Research Closure Rules for Codex

Mission

Help the researcher finish a defensible research unit. Optimize for closure, evidence, and a complete draft—not for generating the largest number of ideas.

Before substantial work, read:

1. .research/state.json
2. the current sprint log under .research/logs/
3. the active experiment card, if any
4. .research/claim_graph.json — the claim graph is the engine: it must exist and validate before a sprint can be frozen
5. docs/protocol.md

Run:

python tools/research_closure.py guard
python tools/claim_graph.py frontier
python tools/research_closure.py next

The guard reports blocking violations; frontier shows which probe is ready to run; next shows the event the harness expects from you. If the guard reports a blocking violation, do not begin a new method, experiment family, benchmark, or research question. Help resolve the violation first.

For human progress tracking, run python tools/research_closure.py dashboard to open the interactive DAG view (theory → variables/edges → probes, with outcomes, frontier and the resolution map).

Non-negotiable workflow

Every task must map to the current frozen claim and to a probe of the claim graph.

Before implementing a new experiment, identify:

• research question
• falsifiable hypothesis
• single primary intervention
• measurement
• expected artifact
• kill criterion
• time budget

Every experiment must name the claim-graph probe it runs (new-experiment --node Pn). If no active experiment card exists, create one with the CLI before implementation. If the claim graph does not exist yet, author it first: claim_graph.py init --claim "...", then add-variable / add-edge / add-absent / add-probe / add-resolution, then validate.

After obtaining results, do not immediately create another variant. First:

1. summarize evidence;
2. classify the result as supported, falsified, inconclusive, or terminated;
3. write a decision;
4. close the experiment with the CLI;
5. update the draft or result note.