AMR Docking Solver v3.4

Primary Pipeline

DOCK ENGINE

AMR Standard

Single AMR search per ligand. Fast baseline.

Budget:

AMR Standard, Redock

Forces fresh AMR dock + score. Ignores cached CSV features. For receptor rescue experiments.

Budget:

AMR 4-Stack (recommended)

Conformers + Weyl rotation + Crystal seed + E8sig selection

AMR 4-Stack + Max-Pool

4-stack Weyl rotations, max-pool orientation features (aroBurial, e_disp, etc). Label-free.

AMR 4-Stack + E8 Pose Field

Physics-first 1% near-tie selector; prioritizes E8 field alignment and prefers negative aroBurial skew.

AMR 4-Stack + E8 Pose Field v2

Strict E8-good gate; otherwise searches 1% physics near-ties with e_disp hazard guard.

AMR/E8 Pose Router v1

Label-free router: use E8 Field v2 unless weak aroBurial + positive e_disp skew flags AMR fallback.

AMR 4-Stack + AroBurial Min

4-stack Weyl rotations, choose the lowest aromaticBurial pose per ligand. Label-free experiment.

AMR 4-Stack + AroBurial Max

4-stack Weyl rotations, choose the highest aromaticBurial pose per ligand. Label-free counter-test.

AMR 4-Stack Pose Study

Full pose pool export: ALL poses per ligand with features. No label selection. For PGD research.

AMR Clean Study PH-13

AMR-only dock, no label paths, signal bundle export. Zero isoScore calls.

Budget:

PHYSICS FEATURE ENGINE

V2 Legacy Physics

Current feature rows. Existing `e_coul`, `e_disp`, and `e_rep` values.

V3 Experimental Feature Engine

Same V2 feature row, but swaps only `e_coul`, `e_disp`, and `e_rep` from typed-shell V3. All scorers consume the same slots.

V4 LP+ Physics Engine LP+

V2 row shape with LP+ trust-gated `e_coul`, `e_disp`, `e_rep`; raw `hbq` suppressed while `hbq_implicit_donors` remains.

SCORING MODES (each with its own pose selection: F=Flat S=Sweep A=Archetype)

DX-08 Manual Descriptor Prototype

Frozen physics route stays primary. Descriptor overlay is bounded, optional, and post-route only. Current baseline lock: B11.

Activate overlay Type Action Band Use atlas best action if loaded Show advanced actions Load atlas JSON

Default menu follows the compact DX-07c atlas: family ladder plus `n_rotatable` and `fraction_csp3`. Turn on advanced actions for `mw` and research-only pairs.

No DX-08 run yet.

DX-09 Compact Law Search

Frozen B11 stays fixed. This sweeps a small shared descriptor-law grid and summarizes which single compact law is most useful overall, in DUD-E, and in DEKOIS.

Search space Max active families

No DX-09 run yet.

DX-10 Compact Law Search (B12 baseline)

Frozen B12 (B11 + 8 gates) stays fixed. Sweeps the same descriptor-law grid on top of B12 to find further improvement.

No DX-10 run yet.

DX-12 Physics Law Search

Frozen B11 stays fixed. Sweeps physics features (contact_density, e_rep, e_disp, insideFrac, etc.) with sign/magnitude grid. Measures real AUC delta per target.

Max active features Sign search

No DX-12 run yet.

AMR Physics Lab / MERGE07 Router

Lab mode Baseline Select the MERGE07 scorer below to run the lab lane.

Active Mode

MERGE07 canonical stack

Feature Schema

physics_v3_registry_2026_04_28

Router

voro06-js-mirror

Shader

v3_typed_shells.wgsl

Law Stack

merge07 canonical

Baseline

not run

Labels

not run

AUC / Delta

--

EF 1% / 5%

--

Branch / Regime

--

Laws Fired

--

Router audit will appear after a lab run.

MERGE07 evidence: delta +0.0063120062, DUD-E +0.0102542841, DEKOIS +0.0017991355, worst regression -0.0132972555, 0 new failure-floor targets, Gate A.

Feature signage audit will appear after a lab run. It compares VORO geometry-predicted feature signs against label-derived feature signs and, when available, IsoScreen v8.33 coefficient signs.

Batch summary will appear as lab targets finish.

Receptor fingerprint off

Run Tracker

Named benchmark runs with stage timings, dates, and ligands/sec.

Date	Run	Status	Set	Tgt	Lig	Best μ	Load s	Dock s	Score s	Total s	Lig/s	Details

Target Set

HELD-OUT:

REPAIRED (frame+H+pocket):

10 hard targets (exhaustive ceiling <0.65)

Absolute dG (PDBbind production model)

waiting for docked features

▶Docking Modes (55+ research modes)

BASELINES

▶CRYSTAL SEED VARIANTS

▶SINGLE STRATEGIES

▶COMBOS (multiconf + strategy)

▶QUINT STACKS (all 5 layers)

▶WEYL COMBOS (structured orientation sampling)

EXPERIMENTAL (E8 geometry) ▶

▶PLATONIC v2 + UNIVERSAL (chemistry-aware, fixed weights)

▶SERVER-SIDE OPTIMIZER (Wave 5 physics-only CV)

▶FFT DOCKING (WebGPU grid search)

▶TOPOLOGY FEATURES (DUD-E gap, pattern not energy)

Capture probe fingerprint XP-01 Capture receptor features

Config

AMR budget: Sweep budgets: PCA seeds: Entry rays: E8 a₀ (Å): Conformers: Seed radius (Å): Seed confs:

poseBasin Analysis

ADMET Liability Panel

Target	Lig	Dataset
MEAN	--	--

DUD-E Best Achieved ▾

Target	Lig	CSV	Crystal GPU	Best Docked	Mode	Δ vs CSV	Δ vs CG	Date
MEAN	--	--	--	--	--	--	--	--

DEKOIS2 Best Achieved ▾

Target	Lig	CSV	Crystal GPU	Best Docked	Mode	Δ vs CSV	Δ vs CG	Date
MEAN	--	--	--	--	--	--	--	--

Wave 7 Optimizer (Physics-Only)

Fast (3-fold CV) Dock a target first

AMR Docking Solver v3.4

Primary Pipeline

Target Set

▶Optimizer (GPU settings, wave modes)

Absolute dG (PDBbind production model)

▶Docking Modes (55+ research modes)

Config

ADMET Liability Panel

DUD-E Best Achieved ▾

DEKOIS2 Best Achieved ▾

Log

Wave 7 Optimizer (Physics-Only)

▶W10 Feature Hunt Harness (distribution checks, single-feature lift, greedy add)

Per-Feature Results

Batch / Target Summary

Harness Log

AMR Docking Solver v3.4

Primary Pipeline

Target Set

▶Optimizer (GPU settings, wave modes)

Absolute dG (PDBbind production model)

▶Docking Modes (55+ research modes)

Config

ADMET Liability Panel

DUD-E Best Achieved ▾

DEKOIS2 Best Achieved ▾

Log

Wave 7 Optimizer (Physics-Only)

▶W10 Feature Hunt Harness (distribution checks, single-feature lift, greedy add)

Per-Feature Results

Batch / Target Summary

Harness Log

Speed Report