run

name	run
description	Visible Harness run entry. Use for WF10 experiment iteration: model exploration, hyperparameter changes, ablations, visualization, and quantitative runs.

Run

Use $run as the human-facing alias for WF10 iteration work. This is not a separate Skill Contract.

Read and follow:

• ../../../.agents/skills/iterate/SKILL.md
• ../../../.agents/skills/auto-iterate-goal/SKILL.md when starting auto-iterate
• ../../../AGENTS.md
• ../../../CLAUDE.md

Route the actual work through $iterate next by default. $iterate next reads the active iteration's action_state.next_action and then runs exactly one action from the WF10 action library: plan, code, run_screening, run_full, eval, debug, compare, ablate, register, promote, discard, or stop. Direct subcommands remain available when the operator names one explicitly.

Run-time code construction has three weights:

• config_only: create or edit only run-local configs and record no code manifest unless useful.
• run_local_code: create bounded run-local scripts under runs/wf10/<iter>/ and record runs/wf10/<iter>/code_manifest.json.
• stable_candidate or delegated_build: use heavier build/code-debug discipline, record the manifest, and require a promotion plan before merging back into stable src/, scripts/, configs/, tests, or project_map.json.

Meaningful train/eval work is commit-driven:

• before run_screening or run_full, create or verify pre_train_commit covering stable code, eval logic used by the command, durable configs, and run-local code/configs under runs/wf10/<iter>/
• before eval, create or verify pre_eval_commit, or record pre_eval_commit_NOT_CHANGED
• record those hashes in the run manifest or Gate ledger so the run can be reproduced or reverted
• do not ask for human approval during WF10 iteration when the action stays inside the active Automation Policy; use Gate ledgers and Claim Delta Evidence for traceability

When a run-local script or candidate implementation becomes reusable, use $iterate promote or route through $change classify before merging it into stable code. Promotion must read the run code manifest, write a promotion plan, run acceptance commands or report NOT_RUN, and update stable maps when public interfaces or responsibilities change.

When auto_paper_output/*/run_request_register.{json,md} has pending requests, fold the highest-priority unclosed request into the next $iterate plan unless the operator asks to ignore paper-driven experiments. After $iterate eval or $analyze updates completed run evidence, refresh the default light layer .evidence/light/index.json with tooling/evidence/build_light_evidence_index.py. Refresh the paper-facing docs/30_evidence/Experiment_Evidence_Index.{json,md} with tooling/evidence/build_experiment_evidence_index.py only when claim or writing evidence needs the detailed layer, or report NOT_RUN.

When eval creates follow-up questions, assurance gaps, or paper-driven missing evidence, append or refresh docs/40_iterations/Experiment_Queue.md. When a finding should remain searchable outside a single iteration report, append or refresh docs/45_discoveries/Research_Wiki.md. These files are source artifacts for later light evidence indexing, not approval records.