StandardsForge v0.1.0a7

StandardsForge a7 makes the prepared library easier to install and connects models to the current eleven read-only tools.

• Windows x64 Python 3.12: extract the archive and double-click setup.cmd. Library and MCP install offline.
• Linux/macOS Python 3.11+: run sh setup.sh --mcp-online. Only MCP dependencies use the network.
• Setup prints generated Claude Desktop/Cursor JSON and Codex TOML with your actual paths, after a real stdio query.
• Includes structural browsing, evidence selection, reviewed-reference navigation and verified original PDF delivery.
• Query launchers no longer trigger setup; failed revalidation invalidates old readiness.

The 441-package content baseline is unchanged: 438 source page packs, MIL-STD-810H outline-v3 and two bounded MIL-STD-1661 recovery packs. Setup replays 52 real-document cases. This remains an alpha release; corpus-wide semantic review, applicability and compliance are not claimed.

Code-only installation: python -m pip install "standardsforge==0.1.0a7". Standards content comes from the prepared ZIP.

Prepared archive: 1,080,468,608 bytes; SHA-256 8555fcafa683c60186575193c7cc504331ca6d49c5d64b76bda431b83b1bda9b. First and repeated setup passed on Windows, Ubuntu, Intel macOS and Apple silicon macOS, including generated MCP configuration and real stdio queries. All 52 runtime-bound real-document cases passed. Two prepared builds produced identical bytes.

Native installer acceptance · Source CI and attestation · PyPI publication · Install guide

StandardsForge v0.1.0a6

StandardsForge a6 brings the current source qualification and reviewed MIL-STD-1661 recovery into one prepared installation.

• 441 packages: 438 acquisition-pinned page packs, the 8,319-record automated MIL-STD-810H outline-v3, and two separate 1661 recovery packs.
• All 18 scanned 1661 pages retained with reviewed transcription and raster evidence; 41 reviewed semantic records include five separate 4.2.4 directives.
• First and repeated setup replay 52 source-bound real-document cases. Semantic checks bind full statements and typed qualifiers with exact evidence spans.
• Query startup is read-only; edition comparison verifies unchanged evidence; reviewed blank pages retain evidence without invented text.
• The offline Windows MCP dependency bundle includes patched PyJWT 2.15.0. The PyPI MCP extra requires at least that version.

Install the code CLI with python -m pip install "standardsforge==0.1.0a6". The PyPI package contains code only. For the precompiled library, download and extract standardsforge-ready-0.1.0a6.zip, then run python setup.py (Windows) or sh setup.sh (Linux/macOS). On Windows CPython 3.12, setup.ps1 also installs the bundled offline MCP runtime.

The acquisition baseline remains 912 PDFs and 35,235 physical pages. Recovery packs are additional representations, not additional source documents. The 810H outline remains automated and unreviewed; 1661 assertions are bounded agent review, not independent corpus-wide qualification, project applicability, compliance, or human approval. This release makes no claim of globally optimal compression.

Validation completed before publication:

• CI and release attestation passed the release commit on Windows/Linux Python 3.11/3.12 and Intel/Apple-silicon macOS Python 3.11/3.12.
• Native prepared-candidate acceptance passed first/repeat core installation and real evidence retrieval on Windows, Ubuntu, Intel macOS, and Apple-silicon macOS. Windows also passed first/repeat offline MCP setup.
• The 197-test local suite passed from an empty working directory (two platform skips). All 52 real-document cases passed; the full archive and wheel each built twice identically.

Prepared ZIP: 1,080,443,923 bytes. SHA-256: 06bccbb0c3839507b1c7b6147887cd394e08ab17695ddec554b7e82c27c69983.
Core wheel SHA-256: 72b442852da8eb651154e91aa0a7b915bf711e8c290cb18e9154724cb90f085a.
Source commit: 7974fe8b9b91a048ec0c2b0d001529283e633f4f.

PyPI trusted publication passed eight native wheel installations across Windows, Linux, Intel macOS, and Apple-silicon macOS on Python 3.11/3.12. The prepared archive embeds the exact published PyPI wheel.

Public acceptance completed after publication: the ZIP and checksum were anonymously downloaded and matched the recorded identity. The PyPI wheel matched the embedded wheel, installed into a fresh environment, and exposed a working standardsforge console command. All 41 recovery cases passed from that installation with socket creation denied. See the durable release record.

StandardsForge v0.1.0a5

StandardsForge prepared MIL-STD library with 438 acquisition-pinned page-text packs (35,218 text records from 912 verified PDFs), a separately labeled automated and unreviewed MIL-STD-810H outline, and offline core setup for Windows, Linux, and macOS. The PyPI 0.1.0a5 package carries code only, not standards content. This release fixes repeat Linux setup by normalizing only the standard owned-venv lib64 alias while retaining the no-links integrity check.

The published ZIP is 1,058,683,424 bytes, SHA-256 dd1def3e48c83c2e0ea54d27171177f8135065260591f8b60aac78bd089dc109. A full anonymous download and separate checksum download matched that identity. Native published-asset acceptance passed first and repeat setup, full-integrity doctor, exact resolve, natural-language search, and replayed source evidence on Windows, Ubuntu, Apple-silicon macOS, and Intel macOS: https://github.com/DDNA-Engineering/standards-memory/actions/runs/35781741371 . The exact tagged source passed main CI: https://github.com/DDNA-Engineering/standards-memory/actions/runs/35780609865 . Trusted PyPI publishing and eight native wheel-install checks passed: https://github.com/DDNA-Engineering/standards-memory/actions/runs/35781313767 .

The prepared corpus has no document-wide reviewed obligation classification. Source evidence and rights boundaries remain explicit; neither channel decides applicability or compliance.

StandardsForge v0.1.0a4

Superseded by v0.1.0a5: https://github.com/DDNA-Engineering/standards-memory/releases/tag/v0.1.0a5 . This prepared archive includes the 438 acquisition-pinned page-text packs and the separate unreviewed MIL-STD-810H outline, while its PyPI package contains code only.

Known limitation confirmed by native published-release acceptance: first Linux setup succeeded, but repeat setup rejected CPython virtualenvs standard lib64 -> lib alias. The a4 acceptance workflow also failed on Windows before setup because it wrote a doubled carriage return to GITHUB_ENV. Do not rely on a4 for repeatable Linux setup. The corrected a5 release passed native published-archive acceptance on Windows, Ubuntu, Apple-silicon macOS, and Intel macOS: https://github.com/DDNA-Engineering/standards-memory/actions/runs/35781741371 . Neither release classifies corpus obligations or decides applicability or compliance.

StandardsForge v0.1.0a3

Maintenance release adding bounded natural-language standards-evidence discovery. Search retains explicit exact_phrase, all_terms, and any_terms modes; natural_language uses a separate Porter index, fixed engineering-aware stopwords, bounded inputs, deterministic row-local ranking, and one disclosed strict-to-relaxed fallback. MCP tools now expose titles and descriptions for every input. The prepared archive retains the accepted 438-pack public DLA corpus and derived MIL-STD-810H outline; search remains offline candidate discovery and does not decide applicability or compliance.

StandardsForge v0.1.0a2

StandardsForge v0.1.0a2 adds authorized installed-document discovery to the ready-to-use Windows distribution.

Highlights:

• list_documents is now available through the CLI and MCP, with authorization filtering, exact metadata, bounded prefix filtering, stable signed pagination, and explicit corpus-coverage fields.
• resolve_document keeps exact resolution semantics and now returns bounded, authorization-filtered identifier hints when an exact match is absent.
• The MCP surface contains seven read-only tools and remains offline/local-policy bound.

Prepared corpus:

• 438 DLA MIL-STD page-text packs plus one derived MIL-STD-810H outline pack
• 35,218 page-text records representing 35,235 physical pages from 912 source PDFs
• 17 physical pages have no extracted text record; this maintenance release does not change corpus fidelity
• unchanged corpus compiler 0.3 package set from v0.1.0a1

Windows quick start:

1. Download standardsforge-ready-0.1.0a2.zip and its .sha256 sidecar.
2. Verify SHA-256, extract the archive, and run setup.ps1.
3. Run standardsforge.ps1 list-documents --identifier-prefix MIL-STD-810 --principal local-user.

Archive SHA-256: 39d8b8260de74a1d0f95e3aba329b54e7646cdf2a9d53556689d4fb66258150d
Bundled wheel SHA-256: 170b5c076e3b7e128e5563ecff1c0608add2c2ffa8f20ae5fcde8d76c7aaacbe

The included standards documents remain subject to their source publishers' terms and notices. StandardsForge does not decide applicability or certify compliance.

StandardsForge v0.1.0a1 — Prepared MIL-STD Corpus

First prepared StandardsForge release for immediate local standards exploration and model-host access.

Includes 438 precompiled public DLA MIL-STD packs sourced from 912 verified PDFs, 35,218 source-linked page records, the qualified MIL-STD-810H derived outline, the exact offline MCP dependency closure for 64-bit CPython 3.12 on Windows, and a distribution-rooted MCP launcher. Extract the bundle and run .\setup.ps1; no standards acquisition, PDF compilation, or package-index access is required. Setup validates the closed inventory, installs all runtime wheels offline, checks the exact environment, installs the corpus, and proves a six-tool stdio MCP round trip before marking the release ready.

Claude Desktop and Cursor can invoke standardsforge-mcp.ps1 through powershell.exe; Claude Code can register the same launcher with claude mcp add. Copy-paste configuration is included in the archive and repository documentation.

The included evidence is limited to components recorded as Distribution Statement A. Search results are evidence candidates, not applicability, compliance, test-plan approval, or certification decisions. See CONTENT-NOTICE.md inside the bundle.

ZIP SHA-256: df09a42b1a9d7f574666937f44048cb599dd0ccab53b182fdc59bd3c7ca5ac27