Skip to content

Latest commit

 

History

History
27 lines (18 loc) · 1014 Bytes

File metadata and controls

27 lines (18 loc) · 1014 Bytes

Phase 5 Part 0 — Predictor Overhead

Generated: 2026-04-18T12:18:37.640Z · Node v22.18.0 · win32 x64

Batch-accumulation methodology: N = 1000 iterations per component, divided by N; 100-iteration warm-up discarded (JIT cold-start). No per-call performance.now() reads (Spectre 5 μs floor) and no performance.measure marks (Heisenberg). See METHODOLOGY.md for the full protocol.

Per-component cost

Component Total (ms) Avg (μs)
predictor.forward 3.043 3.04
trainer.push 0.102 0.10
trainer.trainStep (batch 16) 21.474 21.47

Per-frame infra cost (upper bound)

Assumes trainStep runs once per frame. Actual cadence is rIC-gated (≤1/frame).

  • Sum (forward + push + trainStep): 24.62 μs
  • Frame budget (FRAME_BUDGET_60): 16670 μs
  • Percent of budget: 0.15%

Gate decision

  • Threshold: 30% of frame budget
  • Outcome: PROCEED
  • Reason: Per-frame infra 0.15% ≤ 30% gate. Continue to scripts/measure-floor.js.