Documentation

Docs / Reproducibility

Reproducibility

LatticeZero is committed to reproducible computational results. This page documents our determinism guarantees, known sources of variance, and versioning practices.

Determinism Guarantees

IsoScore (Rescoring)

IsoScore rescoring is deterministic within a single GPU and browser session:

Same input SDF + same scoring grid + same profile weights = identical scores and rankings
No random number generation is involved in rescoring
Results are computed in a single GPU dispatch - no race conditions

Cross-session guarantee: Given the same input files, profile version, and engine version, IsoScore will produce the same ranked order on the same hardware.

IsoPose (Docking)

IsoPose uses a genetic algorithm for pose search, which introduces controlled randomness:

The GA uses a seeded pseudo-random number generator - same seed = same poses
Default behavior: a fixed seed is used, producing deterministic results
When conformer enumeration is enabled, the set of starting conformers is also deterministic for a given input

Practical impact: Re-running the same IsoPose job on the same machine will produce the same poses and scores.

Sources of Variance

Cross-GPU Variance

WebGPU scoring involves floating-point arithmetic on the GPU. Different GPU architectures may produce slightly different results due to:

Floating-point rounding: IEEE 754 allows different rounding modes. GPU vendors (NVIDIA, AMD, Intel, Apple) may implement fused multiply-add operations differently.
Reduction order: Parallel sum reductions may accumulate in different orders on different GPU architectures, affecting the least significant bits.

Measured impact: In testing across NVIDIA (RTX 3090), Intel (Iris Xe), and SwiftShader (software), absolute score differences are typically < 0.001 for individual terms. Relative rankings are preserved - the same ligands appear at the top regardless of GPU.

Our guarantee: Same input + same profile + same engine version = same ranked order across GPUs, even if absolute scores differ at the epsilon level.

Browser Differences

All modern WebGPU-capable browsers (Chrome, Edge, Firefox) use the same underlying GPU driver and should produce identical results on the same hardware. Safari's WebGPU implementation (WebKit) may show minor floating-point differences.

Versioning and Provenance

Engine Versioning

LatticeZero follows semantic versioning. The current engine version is displayed on the platform's health endpoint (/health) and in exported data.

Major version changes: Scoring function modifications that may change absolute scores
Minor version changes: New features, profile additions, UI improvements
Patch version changes: Bug fixes, performance improvements

Profile Versioning

Each scoring profile includes version metadata:

Profile ID: Unique identifier (e.g., hmgr_score_v1)
Validation method: e2e_validated (full docking pipeline) or score_validated (rescoring only)
Validation date: When the profile was last benchmarked
Engine version: Which engine version was used for validation

Profile Pack Provenance

Exported Profile Packs (v2.0 format) include full provenance:

{
  "profile_id": "hmgr_score_v1",
  "version": "2.0",
  "tier": "platinum",
  "auc": 0.980,
  "validation": {
    "method": "5-fold CV, tie-corrected AUC",
    "dataset": "DEKOIS2 HMGR",
    "date": "2026-02-13",
    "engine_version": "0.9.7"
  },
  "weights": { ... }
}

This allows any exported profile to be traced back to its validation conditions.

Best Practices for Reproducibility

Record your profile version when reporting results. Different profile versions may weight terms differently.
Use the same browser and GPU for comparable runs. While ranked order is preserved across GPUs, absolute scores may differ slightly.
Export results in SDF or CSV format - these include all per-term scores for independent verification.
Note the engine version displayed at /health. Major version updates may change scoring behavior.