# g50t Level 4 Frontier — g50t-L4-a3

Identity: subagent `g50t-L4-a3`, level 4.

Play prefix (from `harness/`): `LIVE_DIR="../games/codex-gpt5.6sol/g50t/live" TRACE_FILE="../games/codex-gpt5.6sol/g50t/traces/g50t-L4-a3.jsonl" SUBAGENT=g50t-L4-a3 LEVEL=4 python3 play.py send`

Last NEW finding: turn 4195 (N575 terminal repeated-renewal replay composition; timing unresolved). Turn4286 reproduces C158 only.

Compactions this baton: 0.

research-manager is ON — every closure moment gets a Skill invocation (rule 3); last invocation: explicit stop retirement at turn4286 (N576-N577); pending: none.

retire ONLY on clear or true death; stuck and context pressure stay in-baton (rules 2c/4); a stuck handoff will be rejected and bounced back

Working state: BANKED by explicit user stop at turn4286/input90, `state=NOT_FINISHED`, `levels_completed=3`. WM106 inputs1-90 reproduced C158; the predicted suffix91-117 was not executed. The play agent was interrupted at the authorized stop boundary and no successor may be dispatched.

Resumability note:
1. A future explicit resume must start a fresh harness/recovery according to protocol and re-read this frontier plus the complete ARA/GAME.md.
2. Reconstruct C158, then continue N576/WM106 with inputs91-117: C167 clear, UP/UP/LEFT delayed replay, submit visible composition, exact route, audit lower mapping.
3. Preserve a live LEFTx3 goal suffix and submit only after confirmed lower live.

Ruled out / active constraints: never RESET unless `state=GAME_OVER`; never restart the live harness; absorbed replay that visibly rematerializes before remote entry blocks lower transfer; all knowledge writes use research-manager; every WM consult is prediction-ledgered before action and receives a later verdict.