v3-fleet-load-budget.v1 is the release gate for V3/V3_TODOS.md §66.13. It
validates a 7 day world-server fleet run at 5x expected concurrency.
Gate#
- Expected concurrency: 5,000 concurrent sessions.
- Load profile: 5x, or 25,000 concurrent sessions.
- Fleet shape: 125 world shards, 200 sessions per shard.
- Duration: 7 consecutive 24 hour samples, 168 total hours.
- Snapshot budget: p99 <= 25 ms on every daily sample.
- Voice budget: loss rate <= 0.5% on every daily sample.
Evidence#
The deterministic harness lives in apps/v3/lilith-world-server/src/lib.rs as
simulate_five_x_fleet_load_validation. It publishes the same daily samples
captured in V3/validation/fleet-load-validation.json: max snapshot p99 is 24.0
ms, and max voice-loss rate is 0.36%.
The realtime voice path also exposes VOICE_LOSS_RATE_BUDGET and per-round
attempted, delivered, lost, and loss-rate fields in
apps/v3/lilith-realtime-gateway/src/lib.rs, so the loss budget is checked at
the SFU delivery boundary.
Run pnpm verify:v3 fleet-load before release. The verifier fails closed if the
profile changes away from 5x load, loses a daily sample, raises snapshot p99
above 25 ms, raises voice loss above 0.5%, removes source-backed tests, or marks
the TODO complete without the evidence and manifest in place.