V6a.1 Fixed Ray-Evidence Transport Design

Decision and scope

V6a step 5000 failed its frozen one-window capacity gate only on depth: RGB PSNR changed by -0.0473 dB, normal mean angle improved by 4.82%, and depth AbsRel worsened by 7.878% against the V5 step-4500 reference, beyond the admitted 5%. V6b and the 16/8 pilot remain blocked.

The context-only telemetry report at outputs/hypersim_er_v6a_dual_evidence_telemetry_step5000_v2/evidence_telemetry.json shows that the center candidate is not merely weak. With score normalization by actual valid observers and at least two other observers, top-appearance profiles retain 93.07% of positive excess on off-center candidates. The center fractional argmax rate is 7.59% versus a 21.09% chance baseline. This authorizes one new mechanism branch that transports fixed evidence to the 3D candidate that generated it. It does not validate the transported evidence in advance.

V5 evidence_residual and V6a dual_evidence remain immutable ablations. V6a.1 uses the explicit architecture name dual_evidence_transport. It preserves:

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

It adds no target encoder, target-conditioned state, learned renderer, image warp, Gaussian primitive, per-scene optimization, or target-label input.

Fixed local competition

For source voxel center x_n, reference view i, offsets r in {-2,-1,0,1,2}, and the existing minimum native voxel edge delta, candidates are:

y_inr = x_n + r * delta * d_i(x_n)

The fixed RGB-pyramid pair score is unchanged, but each candidate score is divided by its actual valid observer count. A candidate requires the reference projection and at least two other valid observers. A profile contributes only when at least three of five candidates are valid. Two-view inputs therefore produce zero transport localization support.

For valid candidates:

p_inr = softmax_valid(score_inr / 0.05)
m_inr = clamp((K * p_inr - 1) / (K - 1), 0, 1)

Flat profiles have zero transported mass. Temperature, offsets, descriptor, and observer minimum are fixed before any target-label audit.

Trilinear transport

Each candidate y_inr is converted to native voxel-center coordinates:

g(y) = (y - b_min) / ((b_max - b_min) / [W,H,D]) - 0.5

Its mass is splatted to the eight neighboring cells. Neighbors outside the AABB are discarded and the remaining weights are renormalized per candidate. Let a_n be detached V5 appearance confidence at the source voxel and w_inrq the normalized splat weight:

R(q) = sum w_inrq * valid_inr
A(q) = sum w_inrq * valid_inr * a_n
M(q) = sum w_inrq * valid_inr * a_n * m_inr

c_app_transport(q) = A(q) / R(q) when R(q) > 0 else 0
peak_transport(q)  = M(q) / A(q) when A(q) > 0 else 0
c_loc_transport(q) = M(q) / R(q) when R(q) > 0 else 0

The implementation must satisfy c_loc_transport == c_app_transport * peak_transport within FP32 tolerance and clamp all three confidence fields to [0,1]. R is exposed only as nonnegative transport coverage telemetry.

State and reconstruction contract

V6a.1 adds TransportEvidence; it does not change DualEvidence semantics. Geometry uses:

base_density_logits = 4 * c_loc_transport - 2
localization_gate = floor + (1 - floor) * (1 - c_loc_transport)
density_logits = base_density_logits + localization_gate * density_residual

The V5 appearance path remains unchanged:

appearance_confidence = native/high-resolution V5 appearance agreement
appearance_gate = floor + (1 - floor) * (1 - appearance_confidence)
color = sigmoid(logit(base_color) + appearance_gate * color_logit_residual)

The trainable parameter names and count must remain identical to V5 and V6a. A V6a.1 run starts from seed 17 rather than resuming a V5/V6a checkpoint.

Required evidence and stopping rules

Static acceptance requires mass-conserving boundary splat, off-center displacement, flat-profile zero support, reconstruction identities, detached evidence, view-permutation invariance, unchanged V5/V6a behavior, zero renderer parameters, CPU and BF16 CUDA finite forward/backward, full pytest, Ruff, and compileall.

V6a.1 then runs the same 5,000-step one-window gate against V5 step 4500:

PSNR drop no worse than 0.20 dB;
depth AbsRel relative worsening no greater than 5%;
normal mean angle relative worsening no greater than 5%.

If this capacity gate fails, the transport branch stops. If it passes, run the frozen 32-window validation surface audit before any V6b training. V6a.1 must improve scene-macro surface-vs-free and surface-vs-behind AUROC by at least 0.05 over V5, lower support in both non-surface bins, retain nontrivial support coverage, and make risk increase with held-out geometric error. Failure blocks V6b and long training. No external-baseline or CVPR-ready claim is allowed before matched multi-scene and external evaluations.