ArgBridge v0.1.0a1 — Experimental alpha

Experimental alpha

ArgBridge preserves omitted, explicit-null and ordinary JSON values through a strict-tool schema representation, including typed maps within its supported subset. This MIT-licensed research release does not establish production readiness, superior LLM accuracy or commercial demand.

Release commit: 10556c99833149059ea9b86621a6b41acd7c0684. Package version: 0.1.0a1.

The alpha adds installable packaging, CLI entry points, resource-bound checks and regressions, developer documentation, and an offline-first OpenAI proposal example. The final commit changes only README and changelog wording. See the changelog.

Installation

Download argbridge-0.1.0a1-py3-none-any.whl and SHA256SUMS.txt below. Compare its SHA-256 with the manifest (Get-FileHash <file> -Algorithm SHA256 in PowerShell, or sha256sum <file> on Linux).

With Python 3.11 or later, create and activate a separate virtual environment, then install the downloaded wheel:

python -m venv .venv
# Activate: .venv/Scripts/Activate.ps1 (PowerShell)
#       or: source .venv/bin/activate (Linux/macOS)
python -m pip install ./argbridge-0.1.0a1-py3-none-any.whl
python -m argbridge --version

Expected version: argbridge 0.1.0a1. Pip installs the declared runtime dependencies; no model or API key is required. This release is not published to PyPI. See the quickstart for API and CLI examples.

Verification

Main CI run 40861516715 passed 13/13 jobs across Windows/Ubuntu CPython 3.11–3.15, minimum dependencies, offline evaluation and packaging: 84 core plus 29 evaluation tests, 113 distinct tests. CI made zero live requests.

The three attached archives were rebuilt twice from this exact commit in clean directories with identical SHA-256 hashes. Reproducibility was verified on the same Windows/CPython 3.12.14/zlib 1.3.2 toolchain with pinned build tools; archive metadata was normalized without changing file payloads. This is not a cross-platform or independent-party reproducibility claim. The source ZIP matches all 107 tracked files; the sdist passed 84 core tests; strict package checks and a fresh wheel installation with isolated smoke tests passed. All 35 historical evidence files retain their hashes.

Unfavorable model results and limitations

The preserved 995-request real-world evaluation used 35 schema probes and 960 intent trials on gpt-5.4-mini-2026-03-17. Original ArgBridge achieved 153/168 exact correct intents (91.1%), versus 162/168 (96.4%) for non-strict tools with local validation, and cost 46.1% more per correct intent. Original schema-byte expansion was 5.16×. Structurally valid responses still contained wrong intentions.

• These are historical real-model results, not fresh live tests of this release's example or revised manual comparator. The latter has deterministic results only; the original manual baseline had a documented nested-mapping limitation.
• One model snapshot, four synthetic workflows, 56 English instructions and three correlated repetitions; no blinded customer holdout or measured customer demand. Provider acceptance applies only to the submitted schemas/configuration.
• Only a bounded JSON Schema subset is supported. The external deterministic suite exercised 87 assertions and excluded 256 unsupported assertions.
• Validation does not establish authorization, correct intent or safe mutation. Decimal multipleOf inherits Python/jsonschema behavior. Resource bounds are not a complete denial-of-service proof.
• API stability, macOS behavior, alternate interpreters and every dependency combination remain unverified. Passing CI does not establish production suitability.

Assets: exact-commit source ZIP, Python source distribution, universal Python wheel and SHA256SUMS.txt. GitHub's automatically generated tag archives are separate from the explicitly checksummed source ZIP.