OpenRater 0.1.3 — the trust release

The trust release. An independent 35-problem audit of 0.1.0 reached one verdict worth keeping — the rating engine's arithmetic is honest and exact — and thirty-five findings about everything around it. This release fixes all of them.

Highlights

Wrong numbers, presented as right — fixed.

• Rating a spreadsheet now applies the minimum-premium floor exactly as quoting does; one composition path serves both.
• Yes/no eligibility rules fire (a stored-answer vs compared-literal seam, now pinned by a conformance vector).
• Declared input defaults are honored on every path, and omitting a required eligibility-deciding input refuses by name instead of rating.
• Plans with failed verification checks quote with their own health caveats attached.

The desktop extension actually runs. 0.1.0's bundled Node runtime was signed without JIT entitlements, so the scoring engine died on launch on every macOS install — while node --version still passed, which is how it shipped. The runtime now signs with committed entitlements, a gate executes real JavaScript through the exact shipped bytes, and the release smoke fails loudly if the engine boots degraded. This release's pipeline also refuses to go green unless Apple's notarization verdict is Accepted — the proof run for these artifacts shows Accepted on both macOS architectures.

New since 0.1.0

• Two-run compare — "same book, what changed": per-row deltas joined on your own identifier column, totals, biggest movers, newly refused rows — in the app's run drawer and as the compare_runs tool.
• compare_plans — committee-shaped plan-to-plan differences with canonical territory-membership counts and a shareable Exhibits deep link.
• Check-time detectors for silent workbook holes: a declared level with no factor row (R-174) and tolerance-less multi-coverage expectations (R-175).
• A documented schedule-rating door, a current-state workbook export that carries in-app repairs, book-import transforms (@times: column scaling, bare-number schedule judgments), and a transcription spec readable in sections from chat.

Reading experience — traces read like the manual (no float dust, no NaN arithmetic, no zero-seed sums, plain words instead of reserving jargon), numbers format like money, the muted text meets WCAG AA in both themes (enforced by a CI contrast gate), and the recurring 8px horizontal scroll is gone.

Full detail: CHANGELOG.md · supported boundary: LIMITATIONS.md

0.1.3 is 0.1.2 with one metadata fix: the bundle manifest now declares all 18 tools (the two compare tools were missing from the listing metadata; runtime behavior is identical).

Artifacts

Platform	File	Signature
macOS Apple Silicon	openrater-0.1.3-darwin-arm64.mcpb	Developer ID, notarized (Accepted)
macOS Intel	openrater-0.1.3-darwin-x64.mcpb	Developer ID, notarized (Accepted)
Windows x64	openrater-0.1.3-win32-x64.mcpb	Azure Trusted Signing

Install: Claude Desktop → Settings → Extensions → drag the .mcpb in. Everything runs locally; the engine binds to 127.0.0.1 and no telemetry leaves your machine (privacy policy).

A note on versions: v0.1.1 was tagged but never released — its build pipeline reported success while Apple rejected both macOS notarization submissions, which is itself one of the defect classes this release closes. Do not use artifacts from that tag.

Disclaimer. OpenRater outputs are reconstructions for analysis. The carrier's filed and approved rates govern; verify results against the source document. This repository ships only synthetic rating content. OpenRater is an independent project with no affiliation to NAIC, SERFF, ISO/Verisk, AAIS, NCCI, or any insurer.

OpenRater 0.1.2 — the trust release

The trust release. An independent 35-problem audit of 0.1.0 reached one verdict worth keeping — the rating engine's arithmetic is honest and exact — and thirty-five findings about everything around it. This release fixes all of them.

Highlights

Wrong numbers, presented as right — fixed.

• Rating a spreadsheet now applies the minimum-premium floor exactly as quoting does; one composition path serves both.
• Yes/no eligibility rules fire (a stored-answer vs compared-literal seam, now pinned by a conformance vector).
• Declared input defaults are honored on every path, and omitting a required eligibility-deciding input refuses by name instead of rating.
• Plans with failed verification checks quote with their own health caveats attached.

The desktop extension actually runs. 0.1.0's bundled Node runtime was signed without JIT entitlements, so the scoring engine died on launch on every macOS install — while node --version still passed, which is how it shipped. The runtime now signs with committed entitlements, a gate executes real JavaScript through the exact shipped bytes, and the release smoke fails loudly if the engine boots degraded. This release's pipeline also refuses to go green unless Apple's notarization verdict is Accepted — the proof run for these artifacts shows Accepted on both macOS architectures.

New since 0.1.0

• Two-run compare — "same book, what changed": per-row deltas joined on your own identifier column, totals, biggest movers, newly refused rows — in the app's run drawer and as the compare_runs tool.
• compare_plans — committee-shaped plan-to-plan differences with canonical territory-membership counts and a shareable Exhibits deep link.
• Check-time detectors for silent workbook holes: a declared level with no factor row (R-174) and tolerance-less multi-coverage expectations (R-175).
• A documented schedule-rating door, a current-state workbook export that carries in-app repairs, book-import transforms (@times: column scaling, bare-number schedule judgments), and a transcription spec readable in sections from chat.

Reading experience — traces read like the manual (no float dust, no NaN arithmetic, no zero-seed sums, plain words instead of reserving jargon), numbers format like money, the muted text meets WCAG AA in both themes (enforced by a CI contrast gate), and the recurring 8px horizontal scroll is gone.

Full detail: CHANGELOG.md · supported boundary: LIMITATIONS.md

Artifacts

Platform	File	Signature
macOS Apple Silicon	openrater-0.1.2-darwin-arm64.mcpb	Developer ID, notarized (Accepted)
macOS Intel	openrater-0.1.2-darwin-x64.mcpb	Developer ID, notarized (Accepted)
Windows x64	openrater-0.1.2-win32-x64.mcpb	Azure Trusted Signing

Install: Claude Desktop → Settings → Extensions → drag the .mcpb in. Everything runs locally; the engine binds to 127.0.0.1 and no telemetry leaves your machine (privacy policy).

A note on versions: v0.1.1 was tagged but never released — its build pipeline reported success while Apple rejected both macOS notarization submissions, which is itself one of the defect classes this release closes. Do not use artifacts from that tag.

Disclaimer. OpenRater outputs are reconstructions for analysis. The carrier's filed and approved rates govern; verify results against the source document. This repository ships only synthetic rating content. OpenRater is an independent project with no affiliation to NAIC, SERFF, ISO/Verisk, AAIS, NCCI, or any insurer.