qig-warp

Physics-based navigation for expensive computation — screening, cost prediction, convergence stopping.

Any expensive computation has three questions: What can I skip? What will it cost? When should I stop? qig-warp answers all three from a small pilot, before the main computation runs.

The problem

You have an expensive function to evaluate across many parameters. Running everything takes hours. But most of the computation doesn't contribute to your answer — perturbations decay exponentially, cost scales predictably, and estimates converge long before you finish.

The solution

from qig_warp import WarpBubble

# Self-calibrating: discovers structure from 5 pilot probes
bubble = WarpBubble.auto()
result = bubble.navigate(fn=my_expensive_function, params=param_list, budget_s=3600)

# Result: computed 12/20 params in 30 min instead of 60 min
# Skipped the expensive ones that wouldn't change the answer
# Predicted values for skipped sites from decay profile

Three operations

Four modes

bubble = WarpBubble.auto()                      # discovers constants from pilot probes
bubble = WarpBubble.qig_regime(h=3.0, J=1.0)    # regime-aware (physics calibrated)
bubble = WarpBubble.qig_frozen()                 # single calibration (v0.3 compatible)
bubble = WarpBubble.general(screening_length=0.5, bridge_exponent=0.8)  # user-specified

Use cases

Molecular simulation: Interatomic potentials decay with distance (screening). System-size cost scales as N² or N·log(N) (bridge). Energy minimization converges (convergence).

Drug discovery: Binding sites are local — only nearby residues matter (screening). Conformational search cost scales with flexibility (bridge). Docking scores stabilize (convergence).

ML hyperparameter search: Learning rate perturbations have limited range (screening). Training cost scales with model/data size (bridge). Loss curves flatten (convergence).

Climate ensemble forecasting: Weather patterns have finite spatial correlation (screening). Resolution scaling is predictable (bridge). Ensemble convergence tells you when to stop adding members (convergence).

Materials science: Grain boundary physics concentrates at the interface (screening). Simulation cost scales with supercell size (bridge). Elastic constants converge (convergence).

Performance

Validated on quantum physics lattice experiments (L=3 through L=6):

• Screening: 36% site reduction with <2.1% error
• Cost prediction: matched actual runtime to R²=0.999
• Bridge: predicted J-sweep cost within 5% across 7 coupling values

On a molecular dynamics benchmark:

• Auto-discovery found cost exponent within 0.5% of truth from 5 probes
• Budget-constrained: 12/15 evaluations, 29% time savings

Install

pip install qig-warp

Contact

Built by Braden Lang. For partnerships and research collaboration: braden.com.au

Security gate

security-scan runs on pull requests and development/main pushes. It uses Bandit 1.9.4 for common Python security hazards in all shipped src/ code, and pip-audit 2.10.1 for known vulnerabilities in the complete installed dependency set, including runtime, development, optional QIG and scanner dependencies. CI records and upgrades the runner packaging tools before scanning (setuptools 84.0.0, wheel 0.48.0); those tools remain in the audited dependency inventory. Failure summaries name the package, version, advisory IDs and fixed versions without dumping advisory descriptions. Only the local unpublished qig-warp root is omitted from the advisory lookup; its source is scanned. No Bandit finding severities are excluded, and nosec annotations cannot bypass the gate.

Both tools must first detect deliberately unsafe fixtures: dynamic evaluation and an old vulnerable Requests version. These fixtures live only in a temporary directory, are never executed or installed, and are removed automatically. Missing reports, scanner failures, unauditable dependencies and unavailable advisory services fail the check. Returned file metrics must cover every enumerated source Python file; returned dependency names and versions must exactly match the installed dependency inventory. Bandit failure output excludes source snippets and issue text that could contain credentials. Run locally after installing the same extras and pinned scanners: python scripts/security_gate.py.

This replaces unavailable hosted CodeQL scanning for this private repository with an actual check named security-scan. It does not claim CodeQL equivalence: Bandit is a Python AST pattern scanner, without CodeQL's interprocedural data-flow analysis. Dependency checks cover known advisories, not unknown defects, malicious-package detection or every possible dependency resolution. Tests, code review and required branch checks remain necessary; this gate provides no blanket claim that code is safe.

Scanner references: Bandit and pip-audit. Failed black-box evaluations now log their exception class without revealing user parameters or exception text. Existing numerical fallbacks and results are preserved.