Appearance Ceiling Diagnostic Design

Purpose

The V4 evidence-residual model fits one Hypersim window well in geometry but stops at 18.19 dB RGB PSNR. Before changing the production architecture, a target-label oracle must separate three possible ceilings: optimization of the current shared color grid, spatial resolution of that grid, and view-independent appearance.

The oracle is diagnostic-only. It may optimize against hidden target RGB because its outputs are never model predictions, checkpoints, baselines, or paper results. The production encoder remains target-independent and the sealed final holdout remains untouched.

Variants

All variants reuse density and bounds produced by the accepted V4 checkpoint and optimize only color logits on the first deterministic overfit window.

• native_shared: one color volume at the current density-grid resolution, shared by all target cameras. This estimates how much of the gap is ordinary color optimization rather than missing representation.
• highres_shared: one shared color volume with every spatial axis doubled. Density and variance are fixed trilinear resamplings of the accepted state. This estimates the gain from separating appearance resolution from the 0.25 m geometry grid.
• native_per_view: one native-resolution color volume per target view. This is deliberately an inadmissible upper bound; its advantage over native_shared estimates the value available from target-direction-dependent appearance.

Every variant starts from the accepted state's color, uses the same fixed renderer and cameras, optimizes masked RGB MSE, and reports initial/best/final PSNR, SSIM, MSE, parameter count, runtime, and the checkpoint/config hashes.

Selection Rule

Choose exactly one production change after the diagnostic.

1. If highres_shared gives the largest admissible gain, add an explicit high-resolution evidence-residual appearance field while keeping density, evidence confidence, and provenance on the existing geometry grid.
2. If native_per_view has a materially larger gain than highres_shared, add a bounded low-order directional color basis evaluated analytically by the parameter-free renderer. Per-view oracle volumes themselves are never admitted into the model.
3. If neither oracle materially exceeds native_shared, do not add a representation module; investigate encoder optimization and RGB objective conditioning instead.
4. If high-resolution and per-view gains are close, prefer high resolution first because it is an explicit spatial field and introduces fewer assumptions about reflectance.

The comparison is causal diagnostics, not a publication ablation. Exact observed deltas, not a post-hoc preferred story, determine the branch.

Safety And Compatibility

• Legacy and evidence-residual checkpoints keep loading without conversion.
• Target RGB is accepted only by the standalone oracle optimizer and never by model forward.
• The renderer remains parameter-free.
• Evidence confidence and provenance remain deterministic and optimizer-invariant.
• The diagnostic uses only the existing training overfit scene; validation, diagnostic test, and final holdout data are not read.
• Outputs are written below outputs/appearance_oracle_v1 and cannot overwrite accepted runs.

Acceptance

The diagnostic is accepted when focused unit tests pass, all three variants produce finite gradients and JSON metrics on CUDA, the source checkpoint hash is recorded, and a rerun with the same seed is deterministic within floating-point tolerance. Production work begins only after the variant comparison identifies a single smallest justified change.