Skip to content

Release 0.28.0: stateful hybrid optimization and execution-funded decisions - #54

Merged
OpenUSA merged 2 commits into
mainfrom
feature/stateful-training-funded-risk-028
Nov 13, 2026
Merged

OpenUSA merged 2 commits into
mainfrom
feature/stateful-training-funded-risk-028

Conversation

@OpenUSA

@OpenUSA OpenUSA commented Nov 13, 2026

Copy link
Copy Markdown
Contributor

Native quantum layers previously required users to assemble stateless geometry proposals, empirical input batches and classical optimization separately. Version 0.28.0 adds a single hybrid optimizer and an after-cost scenario decision objective, and extends portable classical predicates with proved integer division.

  • Add Torch/JAX DistanceAwareQNM with QFIM convention F=4 Re(QGT), natural-space momentum, active polar mixing, a trajectory-dependent radius, native immutable state and checked circuit/configuration identity.
  • Add hybrid_qnm for joint classical AdamW and native QuantumLayer optimization. Backward-selected feature rows span microbatches; DDP reduces sample-weighted metric products; trainer checkpoints resume both optimizer states and validate identity before model mutation.
  • Add execution_risk_projection and ScenarioPlan.execution_decision/execution_loss. Sorted sign changes and stable quadratic roots fund fees and impact while enforcing anisotropic Wasserstein CVaR or expectile constraints. Pre-trade positions and post-cost target weights have explicit accounting units.
  • Lower bounded Uint division only when the symbolic divisor is always nonzero. Strengthen selector proofs to retain bit correlations, with work limits and explicit diagnostics.
  • Share natural solves/polar mixing and atomic checkpoint writes, remove duplicate implementations, and exclude repository-only governance files from source distributions. Preserve runtime APIs and optional framework loading.
  • Fix older Dynamo reflection on captured NumPy constants and financial dtype checks. Expand legacy, compiled, portable and distributed CI coverage.

Validation includes independent dense multi-step Fisher/momentum trajectories, exact checkpoint replay, microbatch invariance, independent risk feasibility/maximality and self-financing ledger references, first/second native derivatives, typed compiled failures, exhaustive integer predicates, and interrupted-write preservation. Physical CUDA tests require hardware; distributed Gloo validation runs in GitHub Actions because this workspace disallows its sockets. Wheel/sdist build and Twine validation pass; framework runtimes are not bundled.

Research scope and numerical/financial assumptions are documented. This release does not claim universal speedup, quantum advantage or investment performance.

@OpenUSA
OpenUSA merged commit e0b7575 into main Nov 13, 2026
17 checks passed
@OpenQML-org OpenQML-org deleted a comment from chatgpt-codex-connector Bot Nov 13, 2026
@OpenUSA
OpenUSA deleted the feature/stateful-training-funded-risk-028 branch November 13, 2026 20:08

This branch was successfully deployed

1 active deployment
pypi — b46d9275 Deployed Nov 13, 2026 by OpenUSA via publish #38
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant