Skip to content

Curvature Full-Coverage Mini Experiment

Hermetic closure validates transport completion within a known scene contract. It does not establish physical correctness.

Runs Compared

run output directory preset resolution frames warmup full_frame_requested render mode full-frame coverage
Baseline output/curvature_fps_benchmark/20260607T185034Z mini 160x112 10 2 false partial-frame diagnostic snapshot partial
Experiment A output/curvature_fps_benchmark/20260607T191143Z mini 160x112 50 5 false full-frame render complete
Experiment B output/curvature_fps_benchmark/20260607T191820Z mini 160x112 50 5 true full-frame render complete

Summary Table

All percentages below are minimums across the five curvature cells unless otherwise noted.

run traced % beauty written % hit % over traced pixels miss count rows_completed FPS label mean FPS avg p95 frame ms avg pass2_phys_ms avg frame_coverage_map curvature_signature_ladder
Baseline 37.7 24.6 100.000 0 42-43 mean diagnostic FPS 0.599 1759.08 n/a partial, 37.7-38.2% filled present
Experiment A 100.0 100.0 100.000 0 112 mean FPS 0.712 1622.71 33018.66 full, 100% filled present
Experiment B 100.0 100.0 100.000 0 112 mean FPS 0.728 1557.77 31346.32 full, 100% filled present

Per-Curvature Metrics

Baseline: 10 Frames, 2 Warmup, No Full-Frame Request

curvature % traced % beauty written % hit % miss count rows_completed mean diagnostic FPS p95 frame ms pass2_phys_ms
0 37.7 24.6 100.000 0 42 0.558 1827.23 n/a
25 38.0 24.6 100.000 0 43 0.605 1777.29 n/a
50 38.2 24.6 100.000 0 43 0.647 1681.06 n/a
75 38.0 24.6 100.000 0 43 0.607 1708.42 n/a
100 38.2 24.6 100.000 0 43 0.580 1801.37 n/a

Experiment A: 50 Frames, 5 Warmup, No Full-Frame Request

curvature % traced % beauty written % hit % miss count rows_completed mean FPS p95 frame ms pass2_phys_ms
0 100.0 100.0 100.000 0 112 0.647 1779.27 36226.54
25 100.0 100.0 100.000 0 112 0.744 1538.43 31318.17
50 100.0 100.0 100.000 0 112 0.710 1681.69 33707.66
75 100.0 100.0 100.000 0 112 0.741 1519.99 31523.71
100 100.0 100.0 100.000 0 112 0.720 1594.17 32317.24

Experiment B: 50 Frames, 5 Warmup, Full-Frame Request

curvature % traced % beauty written % hit % miss count rows_completed mean FPS p95 frame ms pass2_phys_ms
0 100.0 100.0 100.000 0 112 0.680 1630.57 34913.39
25 100.0 100.0 100.000 0 112 0.746 1631.00 30894.18
50 100.0 100.0 100.000 0 112 0.754 1453.34 30151.87
75 100.0 100.0 100.000 0 112 0.751 1499.31 29989.90
100 100.0 100.0 100.000 0 112 0.707 1574.63 30782.27

Key Findings

More frames alone did produce full mini coverage. Experiment A moved from the baseline's 37.7-38.2% traced coverage and 24.6% beauty-written coverage to 100.0% traced and 100.0% beauty-written coverage at every curvature level.

The full-frame request did not materially change final coverage behavior relative to Experiment A. Both 50-frame runs completed 112 rows, reached 100.0% traced coverage, reached 100.0% beauty-written coverage, kept 100.000% hit rate over traced pixels, and reported zero misses. The full-frame request is still useful because it records the stricter intent and keeps the full-frame visual gate explicit.

The frame coverage maps visually fill in with 50 frames. Baseline maps are partially filled at 37.7-38.2%; Experiment A and Experiment B maps are 100.0% non-background for every curvature level.

Sealed closure remained intact. All three runs report zero misses across all five curvature cells. In the 50-frame runs, traversal budget stress also dropped from about 30.9% in the baseline to 0.0%, and max traversal steps dropped from 701 to 299-304.

FPS did not drop in these measurements. Average FPS rose from 0.599 in the baseline to 0.712 in Experiment A and 0.728 in Experiment B. That is about +18.9% for A and +21.4% for B versus baseline. This likely reflects the benchmark reaching a different completed-frame regime, so it should not be interpreted as an optimization result.

pass2_phys_ms was unavailable in the partial baseline because the latest perf frame report was not known. It was available in the full-coverage runs: Experiment A averaged 33018.66 ms and Experiment B averaged 31346.32 ms across curvature cells.

Comparisons

Baseline vs Experiment A

Experiment A confirms Grok's row-cursor/band-stepping hypothesis. Raising the measurement window from 10/2 to 50/5, without requesting full-frame mode, was enough to complete the 160x112 film: traced coverage rose from 37.7% minimum to 100.0%, beauty-written coverage rose from 24.6% to 100.0%, and rows_completed rose from 42-43 to 112.

Baseline vs Experiment B

Experiment B also reaches complete mini coverage, with full_frame_requested=true recorded in metadata and summary. It preserves sealed closure, generates complete frame coverage maps, and allows the run to be called full-frame at mini resolution because full-frame coverage actually passed.

Experiment A vs Experiment B

A and B are functionally equivalent on coverage and closure: both reach 100.0% traced, 100.0% beauty written, 100.000% hit rate, 0 misses, and 112 rows completed. The full-frame flag did not visibly change row completion, budget stress, coverage maps, or curvature-signature generation in these artifacts. Its main value is semantic: it makes the run declare and enforce full-frame intent.

Use Experiment B as the official mini full-coverage validation command:

CURVATURE_FPS_FULL_FRAME=1 \
CURVATURE_FPS_PRESET=mini \
CURVATURE_FPS_FRAMES=50 \
CURVATURE_FPS_WARMUP=5 \
bash scripts/run_curvature_fps_benchmark.sh

Experiment A proves full coverage is achievable by configuration alone. Experiment B is the better official validation command because it records full-frame intent and preserves the stricter visual gate semantics.

Grok Handoff

Grok can begin performance analysis for the mini full-coverage configuration. The recommended target is Experiment B, especially pass2_phys_ms, p95 frame time, traversal-step distribution, and why full coverage eliminates the baseline budget-stress signature.

Remaining blockers before calling anything full-frame FPS:

  • The run may be called mini full-frame FPS because traced, beauty-written, and classified coverage are all 100.0%.
  • Do not generalize this to larger presets yet; mini 160x112 is the only validated full-frame scale in this experiment.
  • Treat the FPS increase versus baseline as a regime change, not an optimization result.
  • pass2_phys_ms appears only once full coverage/perf reporting becomes known; Grok should verify whether it is cumulative per run or normalized per reported frame before comparing it as a per-frame stage cost.