# RemakeBench Tech Review 001 — Public RTX evidence audit summary

Audit completed: 2026-07-14T15:04:59Z

Evidence version: `posthoc-20260714-v3`

## Verdict

All four model-labelled shader runs passed independent source-byte, compatibility-wrapper, compile, hardware, actual-renderer, GPU-completed performance, sustained-60, and deterministic-video checks. Each run is classified as `qualifying_posthoc_reproduction`.

This verifies post-hoc reproduction evidence only. It does not turn these records into original-run measurements, complete records, publish-ready records, or blind-verified records.

The prepared anonymous review package was rejected because public package metadata could disclose an assignment. It is not published here. No voter, vote, preference, winner, or blind-evaluation outcome was created or inferred.

## What was independently verified

- The four archived source files matched their recorded SHA-256 values byte for byte.
- Desktop GLSL wrappers were compatibility-only; no shader logic, quality, precision, or workload was simplified.
- Compilation succeeded for all four wrapped shaders.
- NVIDIA RTX PRO 6000 hardware and the actual NVIDIA EGL/OpenGL renderer used by the benchmark were attested.
- Performance used a fixed 1920×1080 framebuffer and GPU-completed timing around render plus `glFinish`.
- Each artifact received three nominal 30-second uncapped trials after a 10-second warm-up.
- Each artifact passed a separate 3,600-frame, approximately 60-second headless deadline test with zero late frames and zero deadline misses.
- Each public MP4 is a deterministic offline 1920×1080, H.264, yuv420p, 60 fps render with exactly 900 frames and 15 seconds duration.
- Independent frame decoding found no duplicate or frozen frames, corruption, blank output, compile-error display, or embedded overlay.

## Evidence timing and sample size

The hardware, renderer, performance, and video evidence was reproduced after the original model-generation session. Three post-hoc performance trials measure one generated artifact; the model-generation sample remains n=1 per model and task.

Across the four runs, trial-sample coefficient of variation ranged from 0.037% to 0.207%. Displayed FPS values use conventional two-decimal rounding. Independent CSV-only cross-checks differed from recorded aggregates by less than 0.012% because terminal timing bookkeeping is represented slightly differently.

The original ledgers' supplied cost, token, and end-to-end workflow measurements are unchanged. Repeated performance measurements are not repeated model generations.

## Still unavailable

- Original initial local FPS
- Original optimized local FPS
- Completed blind evaluation

The public snapshot intentionally excludes source code, exact protected prompts, private packages, private machine identifiers, private storage identifiers, and blind assignments.