Experiment Plan

Write

• projects/<slug>/plans/<idea>/experiment-plan.md
• projects/<slug>/experiment-memory.md
• projects/<slug>/STATE.md
• projects/<slug>/decision-tree.md

Must Not Write

• projects/<slug>/plans/<idea>/anchor.md
• projects/<slug>/results.tsv
• projects/<slug>/review-state.json
• memory/*
• projects/<slug>/workspace/dashboard-data.json

Required Input Checks

Before proceeding, verify:

• the target idea_id is known
• the planning mode is unambiguous: initial line or branch continuation
• the project brief constraints are readable
• the inputs are strong enough to populate all required experiment-plan.md sections

If the target line cannot be identified safely, stop and report the exact ambiguity.

Protocol

1. Determine Planning Mode

If there is no active experiment line yet:

• use bootstrap inputs from novelty-check.md

If judge already recommended branching:

• read the latest judge-report.json
• confirm the current idea-scoped plan slot is safe to reuse under the existing path contract

2. Establish Line Identity

Define or confirm:

• idea_id
• branch_id
• experiment_id
• parent_branch_id when the line is a branch continuation

Use stable repo-safe IDs that remain short and human-readable.

3. Author The Plan

Write projects/<slug>/plans/<idea>/experiment-plan.md using the documented input contract.

The plan must include:

• falsifiable hypothesis
• quantitative success criteria
• experimental design
• mutable versus immutable variables
• red lines and disconfirming signals
• evidence plan
• open risks

Do not leave placeholder-level vagueness for anchor-wrapper to guess around later.

4. Optional Cross-Model Review

If Codex MCP is available, you may request an advisory second opinion on:

• falsifiability of the hypothesis
• adequacy of baselines, metrics, and thresholds
• whether mutable versus immutable boundaries are drawn clearly
• confounders or execution risks that would make the plan hard to interpret
• whether the success criteria are strong enough to justify later anchor lock

Rules:

• treat the external review as advisory only
• revise the plan conservatively before anchor lock when the critique exposes a real planning gap
• if disagreement is material, keep the more conservative plan
• if no safe plan remains after review, stop without instantiating the active line

Record any adopted critique in experiment-plan.md.

5. Instantiate Or Refresh Active Experiment State

Update experiment-memory.md:

• set experiment_id
• set idea_id
• set branch_id
• set parent_branch_id
• set status: planned
• clear anchor_path and anchor_version
• reset latest result, analysis, drift, and judge refs for the new active line
• set success_criteria_status: unknown
• set archive_recommended: false
• set human_review_required: false
• set next_experiment_action: wait-human
• refresh the active snapshot, anchor summary, latest evidence summary, and last_updated

Preserve prior history tables and branch outcomes for older lines.

6. Update Branch Governance

Update decision-tree.md:

• set Governance Snapshot.active_idea_id
• set Governance Snapshot.primary_branch_id
• add or refresh the planned line in Active Branch Register
• append a Decision Log row when the plan came from a judged branch recommendation

Do not claim the anchor exists yet. This step establishes a planned active line, not a locked one.

7. Update Project Steering State

Update STATE.md:

• keep or set phase: phase1
• set project_status: running
• set active_idea_id
• set active_branch_id
• set next_action: lock the anchor for the planned line
• set decision_mode: auto-report
• set human_attention: none
• clear decision_type and decision_options_ref
• set last_completed_skill: experiment-plan
• refresh last_updated

Failure Handling

Stop without writing if:

• no selected idea or safe branch target can be resolved
• required plan sections cannot be authored from the available evidence
• the request would require rewriting a locked anchor or silently reusing an unsafe same-idea slot
• external critique reveals an unresolved planning gap that makes anchor lock unsafe

When stopping, report the exact blocking contract issue.

Self-Check

• Wrote a canonical experiment-plan.md
• Created or refreshed the active line in experiment-memory.md
• Updated decision-tree.md only to reflect a planned active branch, not a locked anchor
• Left anchor.md, results.tsv, review-state.json, and dashboard data untouched
• Preserved old history while making the new line explicit
• Treated any cross-model review as advisory only