VEE and Quest 2 qualification

Execution plan accepted on 2026-10-15. The WebXR ADR and body-wiring ADR remain the decisions. This page records operational criteria; it does not promote a synthetic experiment to an animal result.

The subsequent accepted rendering ADR and implementation status address the failed Quest workload. The collector now requires an explicit representation for diagnostic comparisons; its default requests full surfaces and never silently substitutes skeletons.

Run the headless engineering fixture

From clean committed source, with Python 3.13 and the locked dependencies:

uv sync --locked
uv run flymsg closedloop examples/closedloop-nmf.yaml --check
uv run flymsg closedloop examples/closedloop-nmf.yaml --out runs/closedloop

The versioned config selects nmf-optics-chain/v1: two invented neurons, four test optical axes, a synthetic identity eye registration, the coverage-to-rate and torque functions of the existing NeuroMechFly fixture. Their numerical values are fixed in the fixture source; they are not biological calibrations. The config explicitly names the clock, duration, seed, checkpoint period and all five arms. A biological config refuses before constructing the body. --check loads neither the connectome nor a physical body.

The command reuses CpuRuntime, Clock, FlyGymBody, closedloop/v1 and runstore.Run. Unlike dataset CLI writers wrapped by managed, it has no connectome input; the same runstore source, environment, locking, finalization and corruption policies apply directly. Identity includes the config bytes and every FlyGym asset. Each arm records exact input event counts, spikes, commands, contacts, sampled and resulting poses.

Repeat the identical command to resume checkpoints or verify completed output. A changed source, config, asset or environment selects another execution. A corrupt completed result refuses. Checkpoints live outside output staging, one directory per arm. The replay must reproduce the intact spikes exactly. Silencing is applied to the runtime, not just labeled.

summary.json reports the observed pose/optics difference, input and spike event counts, per-arm wall time, newly computed simulated seconds per wall second, tick-time percentiles and process peak RSS. Resumed/cached ticks do not count as new work. Tick timing includes checkpoint work on control boundaries, but excludes setup and final publication; per-arm wall time includes runtime construction and body settling. These are tiny-fixture figures, not full-connectome throughput. Timing never enters a scientific record or clock.

Quest 2 development and measurement

Use Quest Browser standalone. Enable Developer Mode for this headset in Meta Horizon, connect USB data and authorize this computer. An enabled developer account alone does not enable the device's ADB interface. On this host systemd's existing Android uaccess rule grants the active seat access when the headset exposes ff4201; no custom /etc rule is needed. See Meta device setup and remote browser debugging.

adb devices -l
uv run python scripts/quest_xr.py --serial QUEST_SERIAL --scene looming --seconds 1800

For a long trial launched from an agent, run the collector as an owned transient user service so it survives the launching command's lifetime (use the checkout's absolute paths):

systemd-run --user --collect --unit=flymsg-quest-trial --slice=background.slice \
  --property=WorkingDirectory=/absolute/flymsg \
  --property=KillSignal=SIGINT --property=TimeoutStopSec=90s \
  /absolute/flymsg/.venv/bin/python /absolute/flymsg/scripts/quest_xr.py \
  --serial QUEST_SERIAL --scene looming --port 8890 --seconds 1800
journalctl --user -u flymsg-quest-trial.service -f

Stop only that owned service with systemctl --user stop flymsg-quest-trial.service. No persistent unit or system configuration is installed.

The script refuses another model and dirty source. It builds the current site in an owned temporary directory, serves only on loopback, adds a device-specific reverse mapping and opens Quest Browser. localhost is the trusted context; security bypass flags are not used. It never restarts the shared ADB server. An occupied reverse port refuses; use another --port. Normal exit removes only the mapping the script created. A forcibly killed collector may leave that mapping; inspect adb -s QUEST_SERIAL reverse --list before cleanup.

Device/browser versions, source commit, power/wakefulness, battery, thermal status and browser memory are captured before, every 30 seconds and after the trial. Browser reports are appended and fsynced in runs/quest-xr/<timestamp>/sessions.jsonl. The opt-in xrcollect=1 flag posts measurements to this local collector; ordinary viewer pages make no such request. CPU/GPU overrides, resolution overrides, disabled thermals and disabled headset safety systems are not part of the profile. Wireless ADB is unnecessary. Record charging state and keep it the same across comparisons. USB debugging authorization is persistent; reverse mappings are trial resources.

Wireless ADB is also usable through an explicitly selected transport; see the issue-resolution procedure. A transport change starts a new trial. Record temporary user-requested keep-awake settings and restore their original values; such diagnostic runs must not silently inherit the original qualification profile. Page visibility alone does not prove that the headset is awake.

Wait for scene loading before entering VR. The user activates the session and performs head motion/menu actions. Capture three two-minute active sessions per scene, looming and anatomy, then twenty minutes on the heavier measured scene. Retain cold-load observations separately. Compare the desktop baseline at the same source and scene configuration.

The one existing animation loop collects complete active-session frame intervals, including stalls longer than 100 ms and samples older than 600 frames. Visibility interruptions are events and split active timing segments. CPU time covers the frame callback; GPU time is available only with valid timer queries. Resource samples include geometries, textures, programs, draws and triangles. LOD uses the XR eye viewport and projection, not desktop CSS.

For observed refresh Hz, the predeclared budget is B = 1000 / Hz ms:

• interval p95 ≤ 1.1 B, p99 ≤ 2 B;
• estimated missed slots ≤ 1% of observed plus estimated missed slots;
• unavailable refresh or mixed refresh produces indeterminate, never a pass;
• missing-slot estimates from intervals are not compositor drop measurements;
• no crash, incorrect stereo/clipping, failed state restoration or monotonic resource growth.

The report's timing verdict covers timing only. Full hardware qualification also requires the operator checklist: scale/orientation and stereo, LOD under head motion, ten entry/exit cycles, headset-menu exit, reload, suspension/resumption, and WebGL context loss/recovery. Record the action, observed result and session identity for each. Do not qualify another device or an untested scene from this result. Page unload or collector disconnection can prevent delivery of the final report: partial periodic reports are retained, but are not a completed trial. No report means no hardware verdict.

Gate ledger

G2's geometric implementation is not its complete qualification. This distinction also applies to G5's tested synthetic integration versus biological readiness.

The cold-clone record verifies that all five scientific traces are byte-identical across the original run and a fresh clone at source 1223551; timing differs as expected. The first Quest 2 diagnostic records a failure of the default looming view: median active stereo interval 118 ms, p95 239 ms, p99 297 ms, with approximately 13 million triangles per stereo frame. It is an early failure diagnostic, not the full two-minute qualification matrix. The collector process was terminated; the final record was recovered directly from the device. A transient collector service is used for subsequent trials. The initial-refresh event preceded the first frame; independent scoring uses the runtime's recorded 90 Hz and distinguishes estimated missed slots from unavailable compositor drop statistics.

Evidence audit and scientific no-go

The checked model remains fail-closed. NeuroMechFly's primary project describes the body as based on a female micro-CT; it does not establish registration to the MaleCNS eye frame. Ache et al. 2019 models giant-fiber responses from size/velocity terms; those terms do not themselves calibrate a measured retinal-column → LC4/LPLC2 spike generator. Card & Dickinson 2008 is the sourced takeoff-speed comparison already recorded in leads, not an actuator assignment. The existing eye, motor and GF results remain unchanged. These sources and the repository audit supply no newly qualified mapping.

Do not rerun or retune already scored protocols to make these gates green. A new transfer, actuation model or scientific hypothesis needs cited parameters, an explicit protocol version and a prospective lock before confirmatory execution. Missing evidence is recorded as blocked, and existing negative or uninformative results are preserved.