RecognitionFold Benchmarks

Honest comparison of RecognitionFold against native PDB structures. RecognitionFold is a physics engine, not a statistical predictor. The value is the glass-box mechanism, not the RMSD number.

What RecognitionFold is: A first-principles protein folder where every geometric parameter (Cα-Cα = 3.85 Å, helix i→i+4 = 6.23 Å, contact budget = N/φ²) is derived from the golden ratio. Zero training data. Zero free parameters. Machine-verified in Lean 4.

What RecognitionFold is not: A replacement for AlphaFold in production structure prediction. AF achieves ~1 Å RMSD using deep learning on evolutionary data. RecognitionFold achieves ~8-16 Å RMSD from pure physics. The point is not the number — it’s that the physics works at all with zero parameters.

Benchmark Results (2000 iterations)

Protein	PDB	Residues	RMSD (Å)	Rg Predicted	Rg Native	Rg Error	Time
Trp-cage	1L2Y	20	7.83	7.11	7.00	1.6%	0.8s	Try it →
Villin HP36	1VII	36	8.79	8.77	8.82	0.5%	2.2s	Try it →
Engrailed homeodomain	1ENH	54	16.58	12.08	10.09	19.7%	4.8s
Protein G B1	1PGB	56	15.40	11.49	10.27	11.9%	5.7s
BBA5	1T8J	22	7.77	7.66	8.54	10.3%	1.1s	Try it →

Helix Design Validation

10 helical sequences designed from RS φ-geometry alone. Cross-validated with ESMFold and AlphaFold (no circular dependency — RS predicts the helix, AF confirms it independently).

Test	Result
Designed helices formed	10 / 10
Negative controls disrupted	10 / 10
PDB geometry match	< 2% error
Contact quantization (E38)	Signal detected at φⁿ intervals

φ-Lattice Geometry

Every distance is derived from the golden ratio φ = (1+√5)/2 and measured bond lengths. No fitting. Lean proofs: IndisputableMonolith/ProteinFolding/Derivations/D2_PhiGeometry.lean

Cα–Cα backbone

3.85 Å

φ² × 1.47 Å

Helix i→i+4

6.23 Å

φ × backbone

β-sheet interstrand

4.90 Å

√φ × backbone

Helix bundle

10.08 Å

φ² × backbone

Contact budget

N / φ²

≈ 0.38N contacts

R_g target

(N/φ)^1/3 × 3.85

Compact globule scaling

Falsified Claims

Intellectual honesty requires disclosing what did NOT work. These dead ends guided the theory toward the current 8-tick dynamic clock model.

Claim	Experiment	Result
W-tokens predict native contacts	E1–E7 ablations	Falsified
Q6 trajectory strain predicts quality	E27, E37	Falsified
Static encoding determines fold	Multiple	Falsified — must be dynamic

Install & Run

pip install recognitionfold

recognitionfold fold --sequence NLYIQWLKDGGPSSGRPPPS --output result.pdb
recognitionfold benchmark --output results.json