omega-lock 0.2.0

Post-optimizer audit gate for tuned candidates

Omega-Lock

Post-optimizer audit gate for tuned candidates. Use Omega-Lock after an optimizer or manual tuning step to check whether the chosen candidate generalizes, stays inside declared constraints, and leaves a JSON artifact reviewers can inspect.

pip install omega-lock

Omega-Lock is audit-first, not search-first. It does not try to be the universal optimizer. It sits after grid search, Optuna, Bayesian search, an internal tuner, or a candidate chosen by hand, then asks the release question:

Did the tuned candidate actually generalize, and is it still inside the constraints we declared before looking at the result?

The failure it prevents is common: the highest-scoring training candidate gets shipped even though it overfits, violates a hard constraint, or cannot be explained later from a reproducible artifact.

For normal audit and CI usage, start with:

P1Config(constraint_policy="prefer_feasible")

prefer_feasible keeps selection practical: among candidates that satisfy declared constraints, it prefers the highest-fitness one. Use hard_fail for stricter release gates. Use record only when you need backward-compatible behavior that records constraint violations without gating selection.

Quick Start

This example is copy-paste friendly and uses the public CalibrableTarget, Constraint, run_p1, and P1Config APIs.

from typing import Any

from omega_lock import EvalResult, P1Config, ParamSpec, run_p1
from omega_lock.audit import AuditingTarget, Constraint, make_report


class ToyTarget:
    def param_space(self) -> list[ParamSpec]:
        return [
            ParamSpec("x", "float", low=0.0, high=1.0, neutral=0.5),
            ParamSpec("risk", "float", low=0.0, high=1.0, neutral=0.5),
        ]

    def evaluate(self, params: dict[str, Any]) -> EvalResult:
        x = float(params["x"])
        risk = float(params["risk"])
        fitness = 1.0 - abs(x - 0.8) - 0.4 * risk
        return EvalResult(
            fitness=fitness,
            n_trials=100,
            metadata={"risk": risk},
        )


def risk_ok(params: dict[str, Any], result: EvalResult) -> bool:
    return float(result.metadata["risk"]) <= 0.6


target = AuditingTarget(
    ToyTarget(),
    constraints=[
        Constraint(
            "risk_ok",
            risk_ok,
            "Risk must stay at or below 0.6.",
        )
    ],
)

result = run_p1(
    train_target=target,
    config=P1Config(
        unlock_k=2,
        grid_points_per_axis=5,
        constraint_policy="prefer_feasible",
        stress_verbose=False,
        grid_verbose=False,
    ),
)
report = make_report(target, method="run_p1", seed=None)

print(result.status)
if result.warnings:
    print(result.warnings)
print(result.config_full["constraint_policy"])
print(result.search_settings)
print(report.best_feasible.params if report.best_feasible else None)

The P1Result is JSON-serializable. Use result.save(path) when you want to persist the run artifact.

Constraint Policy

• record: backward-compatible. Records constraint violations in the audit trail but does not gate best-candidate selection.
• prefer_feasible: recommended for normal audit and CI. Feasible candidates are preferred when selecting grid_best.
• hard_fail: strict release/CI gate. If no feasible result should pass, the run status fails instead of silently falling back.

Local Demo

There is no tracked demo video in this repository. The old video placeholder has been removed.

Run deterministic demos locally:

python examples/demo_replay.py
python examples/demo_sram.py

They require no network and no API keys. The replay shows sensitivity measurement, top-K unlock, grid search, walk-forward validation, KC gate results, and artifact output from a captured deterministic run. The SRAM demo shows a 6T bitcell audit across PVT corners with declared constraints and writes output/audit_sram.json.

Subtitle/transcript file for a possible screencast: docs/demo/omega-lock-demo.en.srt

What Omega-Lock Records

• schema_version and omega_lock_version
• config_full, kc_thresholds, and search_settings
• selected candidate, grid results, stress ranking, and KC reports
• walk-forward evidence, including pearson_status and pearson_computable
• optional holdout evidence when you pass holdout_target
• warnings when a mode records evidence but does not gate selection/status
• audit trails from AuditingTarget, including feasible vs. absolute-best candidates

These fields are intended to make run artifacts reproducible, diffable, and reviewable in CI or release review.

When To Use It

Use Omega-Lock when a tuned candidate affects a real downstream decision:

• model or strategy parameter tuning
• hardware or simulation calibration with hard physical constraints
• process control or materials discovery
• optimizer governance where reviewers need an artifact, not just a score

It is probably overkill for a throwaway toy search where nobody will inspect the result.

API Surface

Core protocol:

• CalibrableTarget.param_space() -> list[ParamSpec]
• CalibrableTarget.evaluate(params: dict[str, Any]) -> EvalResult

Main runners:

• run_p1: grid/zoom-grid search with the standard audit gates
• run_p1_iterative: repeated lock-in rounds for higher effective dimension
• run_p2_tpe: optional Optuna TPE path, installed with pip install "omega-lock[p2]"

Audit helpers:

• AuditingTarget: wraps any target and records every evaluation
• Constraint: named predicate over (params, EvalResult) -> bool
• make_report / render_scorecard: produce human-readable and JSON audit summaries

Advanced Notes

Omega-Lock still exposes the deeper research machinery for users who need it:

• KC-1..4 kill criteria for time, sensitivity, action count, and walk-forward evidence
• stress measurement and top-K parameter unlock
• zooming grid and iterative coordinate lock-in
• optional random-search baseline and benchmark scorecards
• adapter patterns for bringing your own optimizer

Those concepts are implementation and review tools. The first decision remains simpler: use Omega-Lock after candidate generation to determine whether the candidate deserves trust.

Release History

0.2.0 (2026-05-22) - Public README and release-surface polish. Sharpens the GitHub/PyPI first screen, removes the misleading missing-video placeholder, makes the quickstart API-correct with Constraint, updates English and Korean docs, adds cache-conscious dynamic PyPI badges, and updates release checklist examples for 0.2.0. No runtime behavior changes beyond version metadata.

0.1.9 (2026-05-21) - README, PyPI metadata, and release hygiene correction. Sharpened the GitHub/PyPI long description around Omega-Lock's actual position: audit-first, post-optimizer, constraint-aware, and artifact-producing. Kept the PyPI badge dynamic, removed stale badge/version references, rewrote Korean documentation as valid UTF-8, and added an explicit release checklist so GitHub tags, package metadata, fresh dist/ artifacts, and PyPI uploads stay synchronized. No runtime behavior changes.

0.1.8 (2026-05-21) - Audit reliability and static hygiene release. Established a clean baseline across pytest, pyright, and ruff: 289 tests passing, pyright src tests at 0 errors, and ruff check src tests clean. Static hygiene work tightened optional optuna imports for run_p2_tpe, cleaned CalibrableTarget Protocol conformance in tests, and fixed hash-chain typing without changing JSON shape. Audit artifacts gained reproducibility metadata such as schema_version, omega_lock_version, config_full, kc_thresholds, and search_settings. Safety signals became explicit for constraint_policy="record", holdout_mode="evidence_only", iterative test reuse, and walk-forward Pearson computability.

0.1.4 (2026-04-20) - Added the omega_lock.audit module with AuditingTarget, Constraint, AuditReport, make_report, and render_scorecard, plus the SRAM audit demo.

0.1.3 (2026-04-18) - Initial public release with P1, iterative P1, optional P2 TPE, perturbation sensitivity, walk-forward validation, KC-1..4, holdout support, benchmark support, and reference keyholes.

Release Checklist

See RELEASE.md. PyPI does not support overwriting an already uploaded version: always bump the version, delete dist/, build fresh artifacts, and verify the package before upload.

Citation

If you use Omega-Lock in research or a published project, please cite:

@software{omega_lock_2026,
  author  = {hibou},
  title   = {Omega-Lock: post-optimizer audit gate for tuned candidates},
  year    = {2026},
  version = {0.2.0},
  url     = {https://github.com/hibou04-ops/omega-lock}
}

License

Apache 2.0 License. See LICENSE for details.

Copyright (c) 2026 hibou.

License history. PyPI distributions of versions 0.1.0 through 0.1.4 were shipped with an MIT LICENSE file. The repository was relicensed to Apache 2.0 on 2026-04-22; 0.1.5 and all later versions ship under Apache 2.0. Anyone who installed 0.1.0 through 0.1.4 holds an MIT license to that copy; license changes do not apply retroactively.