mini-omega-lock

English · 한국어 · English easy · 한국어 easy

Empirical preflight for repeated judge variability and missing measurements before interpreting evaluation results. It does not replace independent hold-out evaluation or an A/B significance test.

Repository: hibou04-ops/mini-omega-lock · PyPI: mini-omega-lock · import: mini_omega_lock · CLI: preflight · MCP: mini-omega-lock-mcp with mini-omega-lock[mcp]

0.8.5 — Verified tag-context CI audit

The sole required dependency is pydantic>=2.6.0. Necessary contracts, provider boundaries and judge code are included under internal _vendor, allowing core imports, CLI and fake-provider demos without omegaprompt, omega-lock, NumPy or provider SDKs. SDKs load lazily when a live provider is selected. omegaprompt>=1.1.0 is confined to optional extra [bridge].

Requires Python 3.11+. This checkout has source version 0.8.5. The commands below install the checkout. The previous 0.8.4 is published on PyPI. Check version-specific 0.8.5 PyPI files and Actions separately; source metadata alone does not establish publication.

On tag CI, scripts/ci_release_audit.py requires matching package/runtime versions and resolves the actual Git tag to the checked-out HEAD before setting the release-mode flag for the audit child only. Main and pull-request CI retain the pre-tag guard. Every audit step and the other release-marker checks still run.

python -m pip install .
preflight --help
python examples/demo_replay.py
python examples/cli_demo_replay.py

The fixture demos above run without API keys, live SDKs or network access. Installation may download dependencies. Live probes separately require the selected SDK, provider configuration and credentials.

compute_empirical_noise_ceiling

from mini_omega_lock import (
    compute_empirical_noise_ceiling,
    compute_minimum_detectable_effect,
)

assert compute_minimum_detectable_effect is compute_empirical_noise_ceiling
assert compute_empirical_noise_ceiling([-0.1, 0.1]) == 0.1

The canonical API is compute_empirical_noise_ceiling(paired_deltas, confidence_z=1.96), exposed from both the package root and mini_omega_lock.probes. The legacy compute_minimum_detectable_effect name is a backward-compatible alias to the same callable in both modules.

magnitudes = abs(paired_deltas)
ceiling = mean(magnitudes) + 1.96 * pstdev(magnitudes) / sqrt(n)

Default confidence_z is 1.96 and may be changed to a finite nonnegative value. pstdev is population standard deviation with denominator n. Empty samples, NaN/Inf observations and negative/nonfinite z are rejected; a singleton returns its absolute delta. This statistic is an empirical absolute-noise ceiling for the supplied repeated observations. It does not establish a sampling design, power, a distribution-free confidence interval or an A/B minimum detectable effect.

Measurement states and accounting

empirical_preflight() returns (judge_quality, endpoint, performance, warnings). Use local contracts from mini_omega_lock.contracts, providers from mini_omega_lock.providers and the judge from mini_omega_lock.judges.

Read warnings.measurement_status first and persist probe_accounting. States are measured, unmeasured, insufficient_samples, failed, estimated, not_applicable. The existing 4-tuple and warnings-list behavior remain. Human warnings are diagnostics; legacy substring fallback without structured states does not verify historical measurement coverage.

judge_noise_floor is clipped CV (1 - consistency); judge_relative_variability aliases the same value. fitness_noise_floor is the standard deviation of aggregate fitness. Relative CV cannot be compared directly with an absolute A/B score delta. Consistency and gate-flip share observations for the same (rubric,item,response), with each metric using its requested prefix. The default batch uses 5 actual observations; additional probes add calls.

Accounting includes failed-call time separately from successful-call latency projection. Unobservable provider-internal retries and failed-token costs remain unknown/null. Suggested schema fallback is advisory and does not prove runtime consumption.

CLI and optional integration

CLI exit codes: 0 means requested measurements completed; 2 indicates unmeasured, failed, insufficient or estimated states; 3 indicates a user threshold breach on a measured value; 1 indicates input/runtime errors. Exit 0 without thresholds is not judge-quality acceptance or deployment authorization.

# Live example: requires the selected provider SDK and credentials.
preflight --provider anthropic --rubric rubric.json \
  --probe-item item.json --probe-response "4" --summary

python -m pip install ".[mcp]"
mini-omega-lock-mcp

Supported MCP SDK range is >=1.0.0,<2.0.0; 2.x runtime behavior is unverified. Install [bridge] for parent integration. Local PreflightReport and adaptation work without the parent. Convert reports with mini_omega_lock.bridge.to_omegaprompt_report(report); wrap an existing parent provider with adapt_omegaprompt_provider(provider) before passing it to the local LLMJudge. Model classes are not assumed to share identity.

Local verification and limits

python -m pytest
python scripts/generate_readme_claims.py --check
python scripts/check_repo_consistency.py

Local contracts and fake-provider reproducibility do not establish external OOS utility, live-judge quality, load stability or production readiness. Paid-provider validation is not claimed. An independent OOS comparison must separately declare sampling units, splits, repeated-measurement correlation, uncertainty and decision rules.

Optional bridge contract checks

Tests compare field and validation schemas plus Enum values between vendored models and the actual installed omegaprompt, then roundtrip JSON in both directions. Invalid inputs must retain matching error types and locations. These checks make no live provider calls.

python -m pip install ".[dev,mcp,bridge]"
python -m pytest -m bridge tests/test_bridge_contracts.py

Only an absent parent skips this optional integration file. Broken imports or changed contracts in an installed parent fail. Standalone core tests still run without the parent; required dependencies are unchanged.

ProviderRequest.output_schema is a Python class, so JSON does not transmit its class identity. A concrete ProviderResponse.parsed type must be restored through an explicit schema. Tests exercise the actual provider bridge, user schemas, degradation fields and original exception causes. ProviderError and JudgeError are distinct RuntimeError subclasses, not JSON models.

Evidence and documentation

• Generated claims: English / 한국어
• Trust model · Claim ledger
• Vendored source and maintenance · Changelog
• 0.8.0 measurement-contract migration

README.ko.md is the canonical Korean document; README_KR.md is an identical compatibility copy.

Apache 2.0. LICENSE · NOTICE.