V6a.1 Transport Failure Seal

Decision

V6a.1 is sealed as FAIL_CAPACITY_DEPTH_NORMAL. Stop the complete V6 localization-gating branch: no longer training, no 32-window surface audit, no V6b, no 16/8 pilot, and no external-baseline or CVPR-ready claim.

This is a mechanism failure, not a runtime or implementation failure. The implementation remains as a versioned negative ablation.

Frozen capacity gate

The candidate and reference use the same seed, Hypersim scene/window, 4 context views, 2 target views, 128x160 images, 48x32x48 metric state, and 64 fixed-renderer samples.

V6a.1 also worsens over the already failed V6a by 2.9599% AbsRel and 12.4236% normal angle.

Authoritative artifacts:

• outputs/hypersim_er_v6a1_transport_capacity_gate_step5000/capacity_gate_report.json
• outputs/hypersim_er_v6a1_transport_overfit_eval_step5000/evaluation_report.json
• outputs/hypersim_er_v5_appearance2x_overfit_eval_step4500_reference/evaluation_report.json
• V6a.1 checkpoint SHA-256: ff9f796cdc87252205dcedd29a30744031d996db0f33ba6b85357cbb014e16fb

Mechanism diagnosis

Transport fixes where off-center evidence is written, but it does not turn the fixed local RGB competition into a dense or reliable surface likelihood.

• Only 6.636% of source profiles contribute after requiring two other observers and at least three valid candidates.
• Transport coverage is nonzero in only 12.515% of native voxels; its median is zero.
• Surface localization support has mean 0.001950, median zero, and p90 0.002506.
• The localization completion gate has mean 0.998245, so the learned residual remains almost fully open throughout the grid.
• The final density residual has mean absolute magnitude 2.25, while localization-weighted rewrite energy is only 0.004991.
• Candidate competition has median normalized entropy 0.9979. Center fractional argmax is 7.49% versus a 21.09% chance baseline.
• 78.0% of unique winners lie on the -2 or +2 search boundary. Transport therefore moves a signal dominated by boundary, visibility, occlusion, or appearance-gradient effects rather than a centered local surface likelihood.

The resulting density base is almost uniformly empty (4 * c_loc - 2 has mean approximately -1.9922), and the predictor must reconstruct geometry through a large residual. This can retain RGB while shifting integrated surface depth and destabilizing normals.

What is ruled out

The following are not justified follow-ups:

• more V6a.1 steps;
• changing temperature, offset range, or observer minimum after seeing target metrics;
• another mapping from the same local RGB competition into the density base;
• weakening the frozen 5% thresholds;
• selecting an intermediate checkpoint;
• claiming the V6 evidence path improves the V5 typed state.

Two opposite placement interventions have now failed: V6a leaves the same evidence at the source center, while V6a.1 transports it to the winning candidate. The failure is no longer explained by placement alone.

Remaining project boundary

V5 remains the current engineering baseline for the original project claim:

context RGB + calibrated cameras
-> target-independent typed metric 3D state
-> fixed zero-parameter renderer
-> RGB / depth / normal / point / visibility / evidence risk

A future sparse mutual-triangulation operator would be a separate research branch, not V6a.2. It would need context-only mutual epipolar correspondences, multi-view reprojection consistency, and a label-blind surface audit before any training. That direction is not authorized by this failure seal and was not implemented here.