StarPilot/roadscore/SHOWCASE.md

Replay showcase operator flow

Scope as of September 20: one showcase route, current ACE/Prism, replay only. Freeze live-car work and unrelated feature development. Preserve the gold excerpt and accepted soundtrack archives. The operator is back at the event. Hardware inspection is allowed for the designated owner; current access must be verified before claiming readiness. No device behavior should be inferred from the local rehearsal recording.

On the event comma, while offroad:

./onroad --roadscore route1

The private local favorites file resolves route1 to showcase E and route2 to backup A. No private route identities belong in this document. Normal launches use a fresh logged session seed. For this staged showcase, the user permits choosing and fixing a seed with --roadscore-seed <seed>; the same override supports reproduction and diagnostics. Current native operation uses prepared generic role conditioning and fresh ACE sampling, not a fresh semantic-planner pass per launch.

Mac seed auditions use roadscore/experiments/ace_chestnut_20260916/audition_cached_initial_mps.py with the matching native conditioning bank, exported weights and noise policy. They shortlist initial passages only. The verified baseline closely matches native PCM but is not bit-identical across backends. The selected seed still needs a complete native listening rehearsal; an accepted opening does not establish continuation or full-song quality. Audition audio and provenance stay in the Desktop rehearsal package.

In Galaxy's RoadScore section, Simulate disengage gives the replay UI its native blue AOL appearance and contains the music; Simulate engage restores enabled UI/music. Recorded engagement resets that selection. Left signal, Right signal, Signals off, and Use recorded signals separately drive native arrows and the existing rhythm-aware shaker. The controls change only the isolated replay UI view and music: no recorded messages, vehicle controls or real Params are rewritten. They follow acknowledged audible-time state. Restore both selections independently. Device code checks passed; an audible/visual operator rehearsal remains required before judging. Proposed engagement sequence: contain at0:10, open at0:15, recorded at0:19.

The next presentation revision uses native Steer Left/Right prompts labeled Replay simulation, with existing alerts taking priority. Conservative-v4 strengthens the engagement contrast and mixes the shaker after containment so its high frequencies remain audible. Manual direction changes can restart the motif after a one-bar cooldown. Stale inputs and uncertain beat timing still suppress percussion; a button press is not proof that a shaker pulse rendered.

For this explicitly staged showcase, an optional private assets/showcase_curve_plan.json identifies measured route events and the exact replay start. Only a matching native fresh v4 replay may apply it; live input and official judging ignore it. It shapes the existing song's bass during the approach and restores full bandwidth at the next suitable eighth-note boundary around the measured apex. The original composer state is unchanged. The plan and its hash remain in the captured archive. Ordinary routes without a matching plan retain their causal conductor timing.

Use the protected recorded fallback if preparation or hardware is unreliable. The local rehearsal package is Desktop/RoadScore/demos/showcase-rehearsal/; SHOW DEMO.command opens the route E presentation cut over matching camera footage: 5:00.35, unchanged soundtrack until a two-second closing fade. The unedited exact soundtrack remains protected. The full judge briefing, evidence inventory, timestamp script, and cheat sheet are in that package. This is recorded playback, not fresh generation. The original 315.4-second recording includes an approximately 15-second historical EOF tail; the source app ended with an input-stall error despite uninterrupted audio. It is not the latest device run and has neither native UI nor the newest presentation refinements.

Do not infer audible accents from source alerts alone. In the locked soundtrack, the first strong signal motif begins about 0:24.6, curve payoffs occur around 0:52.7 and 2:01.3, and recorded engagement switches at 2:10.8 and 2:15.9. Fresh generations may have different beat phase. Let the music establish and leave silence around those demonstrations.

If anything fails, switch to the labeled recording or the approved 30-second stopped-motion comparison; do not change vehicle controls, model settings or power limits during judging. Current live-car readiness is not established. Physical Bluetooth synchronization and new replay-button operation still need an operator check. The inherited-session-lock fix is installed; the recovered worker must cold-start once, and subsequent canceled sessions should release their receiver lease without losing the resident worker.

Presentation

Lead with: “RoadScore makes the drive part of the music.” Explain the model/runtime split before playback, then leave silence for the melody, signal motif and curve payoff. The core line is: “The model composes ahead; the presentation layer reacts now.” Use the local SPEAKING_CARD.md and TODAY_RUN_SHEET.md for the timestamped presenter flow.

Advance knowledge is allowed for this staged showcase. Keep the recorded source events real and disclose manual music simulation. Do not change ordinary fresh-seed behavior or official judging policy to make this presentation repeatable. A route-specific demonstration does not prove arbitrary-route or live-car readiness.