README.md

Entropy-Gated Bank Ensemble Solver

Purpose

This experiment resolves the focused solver bottleneck in the tested regime by combining the two bank-adaptive candidate generators that remained complementary under fresh-bank evaluation:

• baseline bank-adaptive candidates
• density_ablation bank-adaptive candidates

The solved slice is:

• sparse_full_noisy
• sparse_partial_high_noise
• moderate anisotropy
• low_skew, mid_skew, high_skew

The solver keeps the forward model and latent control object fixed. It changes only the inverse policy:

1. fit a frozen four-way logistic chooser on calibration blocks only
2. default to the stable dense_support candidate
3. open an observability gate only when d_joint_entropy is high enough
4. when the gate opens, trust the frozen four-way chooser

Why This Solver

The single-variant bank-adaptive chooser still split:

• the density branch was stronger on one evaluation block
• the baseline branch was stronger in several fresh-bank confirmation cells

The merged candidate set exposed the real solver opportunity:

• the four-candidate oracle was much lower than any single path
• the missing piece was a calibration-only rule for when to trust the richer chooser instead of the stable default

The selected gate is:

• feature: d_joint_entropy
• default candidate: dense_support
• gate condition: d_joint_entropy >= 0.3655148794

That threshold is selected from calibration only. It is not tuned on holdout or confirmation.

Main Result

From entropy_gated_bank_ensemble_solver_summary.json:

• holdout solver mean alpha error: 0.1050
• holdout best single cached candidate: 0.1091
• confirmation solver mean alpha error: 0.1064
• confirmation best single cached candidate: 0.1104

So this solver clears both evaluation blocks:

• holdout improvement: about 0.0040
• confirmation improvement: about 0.0040

The experiments show a solver-policy result on the tested slice.

The gate rule is fixed from calibration only:

• open the chooser when d_joint_entropy >= 0.3655148794
• otherwise return dense_support

The solver selects one cached candidate when the gate opens. It does not average bank outputs.

Under the matched frozen shadow protocol, the ambiguity-gated alternative is worse on fresh combined data:

• entropy gate fresh combined mean alpha error: 0.105721
• ambiguity gate fresh combined mean alpha error: 0.116399

Plain-Language Read

The dense-support candidate is the safest default. It keeps the low-observability cells from blowing up.

The extra gain comes from trusting the richer four-way chooser only when d_joint_entropy is high enough to justify leaving that default.

Plainly:

• low observability: stay with dense_support
• higher observability: let the chooser pick among all four cached candidates

Per-cell results remain mixed, but the aggregate gain holds on both independent fresh evaluation blocks.

BGP Impact

This strengthens BGP.

The result shows the remaining failure in the focused slice was in solver policy, not in the latent control object. This is a practical observability gate in the tested regime, not a new control law:

• the control backbone stayed usable
• the solver bottleneck yielded to a frozen observability-gated bank policy
• no larger latent object or theory rewrite was needed

Remaining Limits

This is a focused-slice result, not a blanket claim that the entire anisotropic solver stack is finished.

The remaining open work is:

• broader regime coverage outside the solved slice
• broader fresh-bank validation outside the solved slice
• unknown anisotropy-axis orientation
• richer media
• outward extension to harder families

Artifacts

Data:

• entropy_gated_bank_ensemble_solver_summary.json
• entropy_gated_bank_ensemble_solver_trials.csv
• entropy_gated_bank_ensemble_solver_model.json

Code:

• run.py