Mvp0

AutoResearch Paper MVP0

Use Codex as the Host and one physically separate Claude Code session with the frozen MiniMax model as the bounded Worker. P1–P5 remain the only scientific and experiment truth. P6 observes and advances those stores one transition per Codex heartbeat; it never creates a second research truth.

Required reading

Read these resources completely before acting:

• mvp/README.md;
• mvp/prompts/codex-research-compiler.md and mvp/schemas/research-ir.schema.json for P1;
• the Worker task/result schemas and mvp/worker_adapter.py for P2;
• the Experiment Receipt schema and mvp/experiment_ledger.py for P3;
• the evaluator/decision schemas and mvp/evidence_gate.py for P4;
• mvp/prompts/codex-recompile-analyst.md, the P5 schemas, and mvp/recompile_loop.py for P5;
• mvp/prompts/codex-supervisor-heartbeat.md, the supervisor/runtime schemas, mvp/supervisory_controller.py, and mvp/runtime_assurance.py for P6.

Use examples/mvp0/fixed-wing-visual-guidance/ only as a shape example. Never copy its claims, paths, evidence, or approval.

Non-negotiable boundaries

• P1 must discuss the research contract with the owner before freezing the initial IR. Publish a proposal, stop for Human Critique, revise, stop for Human Approval, then freeze OWNER_REVIEWED and replay it.
• ENGINEERING_ACCEPTANCE is test-only. Never use it for live work.
• P6 may use DELEGATED_ENGINEERING_REVIEW only for a version N+1 delta that passes the deterministic execution-only policy and a fresh non-MiniMax Codex review. Problem, claim, baseline, metrics, thresholds, falsification, safety, evaluator semantics, and scientific hypothesis fields remain immutable.
• The compiler, reviewer, revision author, and approver must be distinct codex/<role> identities for delegated review. MiniMax is never a reviewer, approver, or final authority.
• Never treat Worker COMPLETED, P3 recording, or a Watchdog heartbeat as scientific success. Only the frozen P4 truth table can decide KEEP, PIVOT, STOP, or RECOMPILE.
• Never repeat a failed task contract, reset its rejected worktree, rotate the Worker model/session, or accept rejected files. A successor uses a clean new worktree, an exact P5 freeze, the predecessor terminal receipt, the same session UUID, and exact --resume.
• Unattended dispatch is forbidden until the full L0/L1/L2 activation receipt replays. A Codex automation file alone is not a complete Watchdog.
• One heartbeat performs at most one state transition. Stop after the committed tick and let the next heartbeat derive the next action.
• Pause on scientific/ambiguous changes, identity mismatch, unknown scheduler drift, exhausted deterministic recovery, or inconsistent lineage.
• Do not claim 24h, 7x24, production cutover, SOTA, or paper readiness from P6.

P1–P5 flow

1. Ground the brief in real local evidence and hash every cited local file.
2. Draft and validate one Research IR with a fair baseline/evaluator, falsification conditions, allowed search space, finite budget, and STOP plus RECOMPILE rules.
3. Publish the proposal and render a review card. End at AWAITING_HUMAN_CRITIQUE.
4. Translate owner feedback into a structured critique, publish explicit JSON Pointer changes, render the semantic diff, and end at AWAITING_HUMAN_APPROVAL.
5. After later explicit approval, freeze OWNER_REVIEWED and run verify-freeze --check-paths.
6. Only on execution authority, initialize P2 in a clean detached worktree, compile one task from one frozen experiment, and dispatch once.
7. Record every terminal P2 receipt into P3 before another Worker turn.
8. Run P4 once for the next P3 receipt. Preserve negative and failed evidence.
9. On PIVOT/RECOMPILE, publish a P5 failure analysis, request, and IR N+1 proposal. Do not silently broaden the scientific contract.

Exact commands and receipt semantics are in mvp/README.md.

P6 supervisor initialization

Initialize one plan-bound store under the research run:

python3 mvp/supervisory_controller.py init \
  --run-dir /absolute/research-run \
  --target-thread-id <exact-codex-task-id> \
  --adapter-dir /absolute/research-run/adapter \
  --ledger-dir /absolute/research-run/ledger \
  --gate-store /absolute/research-run/gate \
  --p5-store /absolute/research-run/p5-recompile

Run inspect first. Its action envelope binds the controller, exact Codex task, run, sequence, and one permitted action. If it returns NEEDS_ENGINEERING_REVIEW, inspect the full parent/child semantic diff and P4/P5 evidence, then supply the closed engineering-review input to one tick.

Render the L1 automation using render-automation. Register it through the Codex App scheduled-task capability when available. Otherwise write only its exact returned path, normalize it with codex-automation-registration, and verify it appears in Scheduled before claiming App registration.

Bootstrap the complete runtime closure only after L1 exists:

python3 mvp/supervisory_controller.py bootstrap-assurance \
  --store-dir /absolute/research-run/supervisor \
  --launch-agents-dir "$HOME/Library/LaunchAgents" \
  --python-executable /absolute/python3 \
  --now 2026-08-02T00:00:00Z

This proves:

• L0: independent launchd health supervisor, zero model calls, exact L1 repair;
• L1: Codex App kind=heartbeat, exact target_thread_id, ten-minute cadence;
• L2: sequenced process/session/task-bound Worker heartbeats;
• activation: immutable receipt binding identities, intervals, logs, probes, and owned resources.

L0 never reads research content or invokes a model. It can restore only the exact frozen L1 file. L1 is the model-bearing Host tick. L2 proves Worker liveness, not scientific progress.

Heartbeat operation

For every L1 run:

1. verify the exact task/run/controller binding and publish one heartbeat;
2. run inspect;
3. create at most the one closed input requested by the action;
4. run one tick;
5. run verify;
6. report the committed phase and end the Codex turn.

Typical sequence:

terminal P2 → NEEDS_P3 → NEEDS_P4 → NEEDS_P5
→ NEEDS_ENGINEERING_REVIEW → NEEDS_CHILD_P2
→ successor Worker starts → terminal P2

P6 may publish deterministic P3/P4 transitions and bounded P5 artifacts. It may automatically freeze only execution-only IR changes. A scientific delta must become WAITING_HUMAN; do not weaken the policy to keep the loop moving.

Observe and control

• inspect: read-only next controller action.
• verify: replay supervisor ticks and P1–P5 lineage.
• inspect-runtime: read-only L0/L1/L2/process/log/residual snapshot for the Dashboard projection.
• l0-health-tick: zero-model health observation and exact permitted repair.
• pause: block new work and pause the exact L1 registration; require an explicit --authority-id.
• resume: replay supervisor and full activation before reactivation.
• stop: exact-once ordered shutdown—block work, disable L0, disable L1, terminate only identity-matching Worker processes, and report residuals.

L0 must detect both a stale L1 execution heartbeat and a stale/missing active L2 Worker heartbeat. It may repair only an exactly missing frozen L1 file; drift, scheduler absence, and stale execution produce typed recovery evidence instead of an unsafe resend or an invented success state.

Codex App owns the automation updated_at value and may normalize terminal prompt whitespace when it persists a heartbeat. Treat only those representation changes as equivalent. Compare the normalized complete prompt and all stable controller, task, recurrence, name, creation, and lifecycle fields; any semantic difference is L1_DRIFT.

The Codex App Scheduled view is the operator surface for L1. L0 remains outside the App failure domain. Dashboard data is a read-only projection; it never authorizes a transition.

Completion language

Report concrete evidence separately:

• deterministic P1–P6 tests passed;
• installed resources match the repository;
• L0 is loaded and activation replays;
• L1 is visible and bound to the exact task;
• a real scheduled heartbeat returned to that task;
• a real L2 heartbeat and terminal receipt were recorded;
• the next research stage actually started.

If any item is absent, name it as an unclosed gate. Never collapse these into a claim of 24h, 7x24, production readiness, SOTA, or full paper completion.