auto-mdp-solver

Stages 1–2 of the MDP pipeline: generate the spec-conformant domain (_uncertainty/_scenarios/_mdp/_gym, adapter, tests, campaign docs) from a frozen IR, gated by conformance + laws + differential. Use when {name}.signoff.json exists beside the schema and the domain code does not, or must be regenerated. Full pipeline: mdp-solver.

Contribute a finished MDP case, its playbook riding in the folder, upstream to github.com/tong-wang/auto-mdp-solver — optionally re-skinned to hide business context; a case too sensitive even re-skinned is not contributed. Use on "contribute this case", "share this upstream", "submit to auto-mdp-solver". Spec/schema extensions go through mdp-propose. Nothing is sent without the user's approval of the final content.

L2+ rounds of the MDP pipeline (hp / gym / arch levers) after mdp-solve's L1 gate shows a gap with a healthy build — playbook-consulted, ledger-addressed, within the run plan's budget. Use when an L1 leaderboard row underperforms. Full pipeline: mdp-solver.

Phase A of the MDP pipeline: a guided interview (or faithful paper translation) that turns a verbal decision problem into a validated, human-signed-off MDP-IR. Use when no signed-off IR exists yet, or a model change re-opens the mdp block. Full pipeline: mdp-solver.

Stage 5 of the MDP pipeline: read the winning policy back into the domain's policy-structure vocabulary (spec §14) — probe, fitted rule scored paired, overlay figure — anchored to a reference when one exists. Use when a winning artifact exists and a declared confirm/discover stance owes the readback. Full pipeline: mdp-solver.

Stage 6 of the MDP pipeline: wrap the winning artifact as a deployable {domain}_policy.py (spec §12) and finish the campaign README in its §1.3 shape. Use when eval TSVs exist for the run plan's target and the winning artifact is chosen. Full pipeline: mdp-solver.

File an upstream extension proposal against the auto-mdp-solver spec, IR schema or pipeline process, as one GitHub issue, when a campaign hits a wall the current version cannot express. Use on "propose this upstream", "file a proposal", "this needs a spec/schema change", or when UPSTREAM_PROPOSAL_*.md drafts exist in scratch/. Nothing is sent without the user's approval of the final body.

Stages 0+3–4 of the MDP pipeline: confirm a run plan, run baselines, train PPO at L0 and L1, eval under the shared seed protocol, diagnose. Use when a built domain passes its gates and a leaderboard row is wanted, or to re-enter for another target. L2+ is mdp-escalate; full pipeline: mdp-solver.

Build a trained, deployable RL policy from a verbal description of a dynamic decision problem: formalize it into an MDP-IR, generate the domain, baselines, PPO training, a policy-structure readback and a {domain}_policy.py wrapper, with executable gates between stages. Use when the user describes a sequential decision problem ("build a domain for...", "train a policy for...", "formalize this problem"), names an existing {name}/{name}_schema.json, or asks for "the MDP pipeline" / "Phase A" / "Phase B".