Skip to content

Releases

OriginSpan 0.5.2rc2 — Offline Discovery Windows Pilot

OriginSpan 0.5.2rc2 — Offline Discovery Windows Pilot Pre-release
Pre-release

Choose a tag to compare

View all tags
@kunaldrafts25 kunaldrafts25 released this 04 Nov 11:47
v0.5.2rc2
51c89c2

Published package downloads

python -m pip install originspan-mcp==0.5.2rc2
docker pull ghcr.io/kunaldrafts25/originspan-mcp:0.5.2rc2

Container validation passed as a nonroot user with real MCP stdio and networking disabled. The image is public; anonymous OCI manifest access was verified. Container digest: sha256:bbe71a9122e9cdba9677d4775e186124d69befe345e1f6f68bc3273d02ca51d7. No local model is included.

OriginSpan 0.5.2rc2 pilot release

OriginSpan exposes persistent source evidence through MCP and adds optional local-model discovery over an already completed analysis. The Windows pilot bundle installs the application and pinned dependencies without a package index. It includes public-sample evidence and installation instructions; model assets are provisioned separately.

Public distribution repository

This release is maintained in kunaldrafts25/OriginSpan-MCP, separate from the original deployment repository. It adds package metadata and a checksum-verified PyPI publishing workflow. PyPI publication requires the owner to configure a pending trusted publisher; no registry publication is claimed before that connection.

Included

  • Standalone MCP, approved source/schema/metadata discovery and readable evidence-backed reports.
  • Optional managed loopback llama.cpp reasoning with no online provider fallback.
  • Typed unapproved field interpretations, bounded context, source-reference validation, explicit uncertainty, checkpoint/resume and recorded usage.
  • Windows x64/Python 3.13 installer, hash-pinned dependency wheelhouse, file manifest and installation receipt.
  • Approved public Bank of Anthos sample and original license; pinned model/runtime reference; technical reproduction and prior validation records.
  • Clean-install verification script exercising isolated installation, real sample indexing and actual MCP stdio. Its executed result is attached separately when available.

Evidence and scope

The prior local feasibility run covered 321 files and 4,076 evidence items, proposed 12 unapproved field meanings, and completed six sections with gaps. Its controller took 120.035 seconds after model startup inside a 3 GiB container with external networking blocked. Neural embeddings were disabled. Model weights were unchanged.

The earlier-to-final local pipeline comparison reported 70.9% fewer model input tokens and 63.74% fewer output tokens. This is not a measured Copilot comparison, a hidden-reasoning-token saving, an independent business accuracy result or a production SLA. The bundle's evidence entry point is docs/evidence/README.md; the repository record is docs/validation-results/offline-feasibility/README.md.

Release validation

The prerelease regression suite passed 149 tests, with one external PostgreSQL integration test skipped. The installer integrity tests reject tampered bytes and paths outside the package. The separately attached fresh-install report records the executed packaging check; the historical model feasibility report is not a substitute for fresh installation validation.

Known limitations

  • Pilot prerelease; no universal market-win, terabyte qualification or office-model certification.
  • Windows dependency bundle is for x64/Python 3.13. macOS and Linux require separately qualified runtime/dependency assets.
  • Python, model weights and llama.cpp binaries are not included. No automatic model download or fine-tuning.
  • Current controller owns a single-slot loopback runtime; shared remote model hosting needs an authenticated adapter and additional qualification.
  • Business meanings and target mappings remain unapproved; live customer-record profiling and target-system writes are outside scope.
  • Summary context is bounded: ORM and documentation continuations remain visible. Model suggestions are not parser facts.
  • External PostgreSQL integration and actual office identity/deployment behavior need their own tests.

Installation and feedback

Follow PILOT-QUICKSTART.md in the extracted release. Start with the public sample and current small-model baseline. Use an approved source slice for a domain-owner-reviewed pilot. Retain the installation receipt and provide sanitized reproduction details for issues.

Attached release verification

  • Source commit: 51c89c2932d59bdce5351d0dde6c02ac43df77ed.
  • Fresh Windows x64 / Python 3.13 isolated installation passed using bundled wheels, --no-index and required hashes.
  • Public banking sample indexed: 321 files, 4076 evidence items.
  • Actual MCP stdio initialized, 25 tools listed, and authorized context call passed.
  • Regression: 149 passed, one external PostgreSQL test skipped.
  • This packaging check did not run new local-model inference or assert OS-level network isolation; earlier model/network proof is included separately.

Download the Windows ZIP and checksum file, extract, and follow PILOT-QUICKSTART.md. Python 3.13 x64 and approved model/runtime assets are provisioned separately. The standalone wheel is for managed Python deployments.

Full Changelog: v0.5.2rc2...v0.5.2rc2

Loading