# NTF-v2 six-way loss and Gaussian-mechanism experiment

The clean-control guardrail and candidate selection rule are fixed before the
queue starts.  All v2 objectives are exactly the standard linear 3DGS loss when
the supplied shot/read noise coefficients are zero.  The automatic strength is
the 10 dB SNR signal threshold divided by 0.05 and clipped to one, giving
approximately 0/0.27/0.53/1.0 on x0/x4/x16/x64.

1. `adaptive_blend`: globally interpolate between the linear and NTF-v1 losses
   using a strength derived from the calibrated SNR threshold.
2. `additive_anchor`: retain the complete linear loss and add NTF only as a
   regularizer; this is deliberately conservative about fine detail.
3. `frequency_split`: apply NTF to a 5x5 low-pass image, while retaining
   noise-significant target edges and a linear highlight anchor.
4. `confidence_linear`: remove the asinh transform and test SNR confidence
   weighting alone in the original linear domain.
5. `curriculum`: use robust NTF during geometry growth, then decay toward the
   linear objective after densification stops at 15k.
6. `variance_deadzone`: remove residuals that fall within one predicted sensor
   noise standard deviation and optimize a Charbonnier excess residual.

Stage A runs all six methods at x0 and x64.  A method passes the clean guardrail
when its x0 PSNR is no more than 0.2 dB below the existing Vanilla result.  Among
passing methods, x64 PSNR gain selects the best two, with SSIM and LPIPS as tie
breakers.  Only those two run x4 and x16 in Stage B.

The mechanism report includes Gaussian count, geometric scale quantiles,
aspect-ratio tails, low-opacity fraction, opacity-weighted count, and a combined
tiny-and-needle proxy.  "Tiny" uses the first scale percentile of the clean x0
Vanilla reconstruction as a frozen threshold.  These are mechanism proxies,
not ground-truth labels of erroneous Gaussians.  Held-out edge/flat-region PSNR
is reported separately to test whether suppressing shape outliers costs genuine
high-frequency detail.