# Goddard leveling after Scooter dismount 2026-11-04. User reports that Goddard retains the dismount angle. Reproduced in managed integration with actual VR05 steering/lean and normal selection cleanup. The cleanup copy itself is recovered behavior; the missing later AI phase caused pitch and roll to remain unchanged indefinitely. ## Original contract and implementation - 4281B3..4284FB (file offsets281B3..284FB) copies vehicle root and full Euler to the restored companion. Preserve that handoff; do not flatten at unequip. - 4096E0..40988D (86E0..888D) is the AI game-interface facing update. It checks gameplay gate475CA0, interface+5 enabled and+A dirty, reads current Euler224/228/ 22C and destination274/278/27C, wraps each delta strictly>180 or<−180, multiplies axis factors810/814/818 by dt, and calls334→458F90→472480. The wrapper repeats the gameplay gate. All three axes participate. - 41C8DB..41C90C (1C8DB..1C90C) overrides Goddard's complete8D0/8D4/8D8 factors to (.3,3.2,.3). These are interface810/814/818; the complete/game displacement isC0. - Ordinary supported moving AI4093BD..4093D0 (93BD..93D0) calls36C with pitch0, target yaw and roll0. Airborne and other AI branches differ; they are not added by this fix. The prior base/facing phase consumes the preceding destination. - 472480..472570 (72480..72570) adds Euler with the source asymmetric negative truncation and strict>360 wrap. The existing RocketOrientation.Add reproduces that shared leaf, including canonical360 and unusual negative-input rounding. ActorPatrol previously changed Y alone. Its opt-in Goddard follow bridge (VR05/ LEV1) now applies the full original axis step toward its ordinary upright X/Z follow target. Heading selection/order and non-follow lab/Carl patrol stay in their existing bounded adapters. Disabled companions remain frozen. Cleanup still copies the actual lean; enabled subsequent updates gradually level it. At60Hz the.3 factor gives about22% of the original tilt after five seconds. ## Evidence and validation Original EXE SHA256 f5281076e1a2492f2d61f87aae3f4544b6d0f535ebe33f9fbb6f78710fe1aea8. `tools/binary/goddard_leveling.py` guards the complete identity and records four bounded spans/hashes/PE offsets. Committed fixture `tests/GamePort.ScooterTests/original-leveling-reference.json` records source, probe generator/shared harness/executable/log hashes and11 copied-original observations. Nine active cases compare against the managed facing step: tilt, opposite signs, seams, negative angles, exact/over180, one60Hz tick, five seconds, and upright. Two additional native enabled/dirty gate observations remain separate from the pure step. The synthetic probe executes copied4096E0 and472480 only, with source Goddard factors; no original body/AI target selection/game entrypoint. The initialized-prefix run exits0 and prints GODDARD_LEVELING_COMPLETE. An earlier fresh-prefix15-second run did not finish startup and establishes no observations. REA/Ghidra requests4096E0/409140 returned explicit procedure lookup failures. Complete replies, exported bundle and closed snapshot are retained under artifacts/scooter-dismount-followup/rea. No successful pseudocode or Evidence ID is attributed to these failed selectors and they were not repeated. Supporting instruction extracts and successful native numeric traces are retained alongside that directory. This does not claim whole-original gameplay comparison. Scooter integration adds actual tilted dismount, exact cleanup copy, gradual next-phase decay, five-second recovery and disabled freeze, plus nine native numeric comparisons (13 groups total). Existing10 Patrol groups still pass. Native rendering/build/package follow-up is recorded in scooter-live.md. Native Windows/user-visible confirmation of this correction remains separate.