run-phase

name	run-phase
description	Execute a single phase of the analysis pipeline by phase identifier
user-invocable	true

/run-phase -- Execute a Single Analysis Phase

Run a single phase of the analysis pipeline.

Arguments: $ARGUMENTS

The argument is a phase identifier: 1, 2, 3, 4a, 4b, 4c, or 5.

Step 1: Read Current State

1. Find the analysis directory (look for STATE.md in the current directory or immediate subdirectories, including under analyses/).
2. Read STATE.md to confirm the analysis is at the correct phase. If the requested phase does not match the current phase in STATE.md, report the mismatch and stop unless the user explicitly confirms they want to run this phase out of order.
3. Read analysis_config.yaml for configuration (model tier, channels, cost controls, pixi settings).

Step 2: Read Phase Methodology

Read src/methodology/03-phases.md and locate the section for the requested phase. Also read:

• orchestration/agents.md -- agent definitions
• orchestration/sessions.md -- session naming and isolation
• CLAUDE.md -- orchestration model

Read the conventions/ symlink in the analysis directory to identify applicable coding, plotting, and naming conventions. All spawned agents must follow these conventions.

Step 3: Identify Upstream Artifacts

Each phase depends on artifacts from prior phases. Locate the latest version of each required artifact by finding the most recent file matching the artifact pattern (sorted by timestamp in the filename).

Phase	Required upstream artifacts
1	prompt.md only
2	prompt.md, phase1_strategy/exec/STRATEGY*.md (latest)
3	prompt.md, STRATEGY*.md, phase2_exploration/exec/EXPLORATION*.md
4a	prompt.md, STRATEGY*.md, SELECTION*.md (per channel or combined)
4b	prompt.md, STRATEGY*.md, SELECTION*.md, phase4_inference/4a_expected/exec/INFERENCE_EXPECTED*.md
4c	All of 4b's inputs plus phase4_inference/4b_partial/exec/INFERENCE_PARTIAL*.md, and confirmed approved_for_verification: true in config
5	All prior phase artifacts

Step 4: Execute the Phase

Update STATE.md: status: executing, current phase: {phase}, timestamp.

All agents use pixi run for script execution and must read applicable conventions/ files.

Phase 1: Strategy

Spawn lead-analyst via SendMessage:

• Inputs: prompt.md, methodology (Phase 1 section), analysis_config.yaml
• Must read: applicable conventions/ files for naming and coding standards
• Working directory: phase1_strategy/
• Expected output: exec/STRATEGY.md, updates to experiment_log.md
• The agent should query the data corpus, identify signal/backgrounds, propose selection, define verification strategy, outline systematics
• All scripts use pixi run for execution

Phase 2: Exploration

Spawn three agents in parallel via SendMessage:

• data-explorer: sample inventory, data quality checks, baseline yields
• domain-specialist: variable definitions, data validation
• domain-scout: expected relationships, domain predictions, prior results

All read: prompt.md, latest STRATEGY.md, methodology (Phase 2 section) All read: applicable conventions/ files All write to: phase2_exploration/ All use pixi run for script execution

After all complete, spawn lead-analyst to consolidate into exec/EXPLORATION.md.

Phase 3: Selection and Background Modeling

Check analysis_config.yaml for channels. For each channel (or the single analysis):

• Spawn signal-lead and background-estimator in parallel
• Inputs: prompt.md, STRATEGY.md, EXPLORATION.md, methodology (Phase 3 section)
• Must read: applicable conventions/ files
• Output: exec/SELECTION.md (or per-channel SELECTION_{CHANNEL}.md)
• All scripts use pixi run for execution

If multi-channel, after all channels complete, spawn lead-analyst to produce SELECTION_COMBINED.md.

Phase 4a: Expected Results

1. Identify systematic sources from the strategy and selection artifacts
2. Spawn systematic-source-evaluator agents in parallel (one per source)
   • Must read: applicable conventions/ files
   • All scripts use pixi run for execution
3. After all complete, spawn systematics-fitter:
   • Builds statistical model, runs Asimov fits, signal injection tests
   • Output: exec/INFERENCE_EXPECTED.md

Phase 5: Partial Verification

1. Spawn systematics-fitter:
   • Runs fit on 10% SR data subsample (fixed random seed)
   • Output: exec/INFERENCE_PARTIAL.md
   • Uses pixi run for execution
2. Spawn note-writer:
   • Produces exec/ANALYSIS_NOTE_DRAFT.md and exec/VERIFICATION_CHECKLIST.md
   • Must read: applicable conventions/ files for document formatting

Phase 6: Full Verification

1. Verify analysis_config.yaml has approved_for_verification: true. If not, STOP and instruct the user to run /approve-verification first.
2. Spawn systematics-fitter:
   • Runs full fit on complete dataset
   • Output: exec/INFERENCE_OBSERVED.md
   • Uses pixi run for execution
3. Spawn cross-checker:
   • Validates consistency with partial and expected results

Phase 5: Documentation

Spawn note-writer:

• Reads all phase artifacts including ANALYSIS_NOTE_DRAFT.md and INFERENCE_OBSERVED.md
• Must read: applicable conventions/ files for document and figure formatting
• Output: exec/ANALYSIS_NOTE.md

Step 5: Run Review

After execution completes, update STATE.md: status: reviewing.

Invoke the review by running /review-phase with the current phase number, OR run the review inline:

Phase	Review tier
1	4-bot (physics + critical + constructive in parallel, then arbiter) with plot-validator
2	Self-review (no separate review -- already done by executor)
3	1-bot per channel (critical reviewer) with plot-validator
4a	4-bot with plot-validator
4b	4-bot, then PAUSE for human gate, with plot-validator
4c	1-bot with plot-validator
5	5-bot (physics + critical + constructive + rendering in parallel, then arbiter) with plot-validator

Step 6: Update State and Report

On review PASS:

1. Update STATE.md: status: passed for this phase, add row to Phase History table
2. For Phase 4b specifically: set status: human_gate and PAUSE
3. Report: which phase completed, artifact location, review outcome, iteration count

On review ITERATE:

• The review command handles re-spawning the executor. Wait for it to complete.

On review ESCALATE:

• Update STATE.md: status: blocked
• Report the escalation reason to the user

Check for regression triggers in review output. If found, handle per the regression protocol in CLAUDE.md.