note-lm v0.1.2 - Three languages, archive-index design, localized errors

Third packaged release - the app speaks three languages and reads like a research archive.

New since v0.1.1

• Deutsch / English / Español: the whole UI switches language live (settings or onboarding), persists across restarts, and formats dates/sizes per locale. German remains the source of truth; adding a fourth language is one dictionary.
• Engine errors localize: failures surface in your language with the raw engine detail always visible as a muted line; stable error codes (sub-codes for package import, sha mismatches, notebook-exists) power the mapping - nothing renamed, unknown codes fall back honestly.
• Archive-index design: numbered sections (01 QUELLEN...), sources and notes as index rows, claims as filed cards, the evidence matrix as a printed table, a heavy notebook title, ruled chat entries - neo-brutalist inside the notebooks, red reserved for actions and citations.
• Dark mode audited: every text/border pair measured (table in tokens.css); placeholder 3.30->4.92, structural hairlines 1.14->3.04, all text now meets WCAG.

Verification

19-gate pipeline green (exit 0): 344 unit tests, engine e2e, installed smoke/recovery/pause gates, and the full walkthroughs (research+revision, CSV, audio, voice) from the installed package. UI verified by the 17-step automated drive plus dedicated dark-mode screenshots audited for legibility.

Download

note-lm_0.1.0_x64-setup.exe (~99 MB), Windows 10/11 x64, verify with the .sha256 file. No runtimes needed; model downloads on first use. Unsigned (SmartScreen may warn) - signing on the roadmap.

note-lm v0.1.1 - Onboarding, real architecture docs, package fixes

Second packaged release. Same install path as v0.1.0 - upgrade by installing over it.

New since v0.1.0

• First-run onboarding: welcome wizard with live resource recognition (chat/embeddings/whisper runtime/models on disk), AI setup in three paths - local models from the catalog with one-click download, API provider with connection test and OS-keyring key, or honest text-only - and the sample notebook offer. Reopenable from settings, fully keyboard-driven.
• Packaging fix (critical): installed windows previously failed to spawn the engine ("Motor nicht erreichbar"); resource resolution now covers every layout. Proven by the automated real-window harness.
• Honest failure states: material generation without an AI provider lands in an error instead of spinning; engine e2e hardened.
• Architecture page now documents the real Tauri/local system.
• Convex remnants fully removed.

Verification

19-gate pipeline green end to end (exit 0): typecheck, 343 unit tests, engine e2e, ABI-matched bundle, clean-PATH installed smoke, kill-and-recover, pause-survives-restart, and full walkthroughs (research + revision, CSV calculation, audio, voice) from the installed package. The version-recording race in the walkthrough gate was root-caused (three reproduction chains) and fixed in the test, not papered over.

Download

note-lm_0.1.0_x64-setup.exe (~99 MB), Windows 10/11 x64. Verify with the .sha256 file (c8bbe73f...3119de). No Node.js/Python needed. SmartScreen may warn (unsigned); signing is on the roadmap.

Known limits: unchanged from v0.1.0 - model downloads need internet on first use, E3 pilot pending.

note-lm v0.1.0 - Research notebook with verifiable evidence

First packaged release of note-lm: a local-first research and study notebook for Windows.

Download

• note-lm_0.1.0_x64-setup.exe (~99 MB) - NSIS installer, Windows 10/11 x64. No Node.js, Python or any runtime needed: everything the app uses is packaged or app-managed. Verify with the .sha256 file.

What you can do

• Import PDFs, web URLs, CSV tables, audio and video - everything becomes an immutable, versioned source.
• Ask questions with local AI (llama.cpp - chat and embedding models download from Settings; Whisper speech-to-text installs itself on first use) or bring your own API provider (OpenAI, OpenRouter, or any OpenAI-compatible endpoint; keys live in the OS keyring, never in files).
• Save claims anchored to the exact source version, page or time range - open any citation and see the original.
• Re-import a document as a new version: the app deterministically flags which claims need review (moved or vanished quotes), you accept or reject, history is kept.
• Run reproducible calculations over CSV versions - strict validation, never LLM arithmetic.
• Evidence matrix: claims x sources grid showing linked evidence, pending review, and honest "not found in the recorded search".
• Export research as portable packages (hash-verified, with or without originals) and restore them anywhere.
• Everything runs inside the one installed program: workers persist with pause/resume/cancel and recover from crashes (proven by automated tests that kill the engine mid-work).

Verification

This release was built by a 19-gate automated pipeline: unit suite (343), engine e2e, clean-PATH installed smoke, kill-and-recover gates, and full walkthroughs (research, revision, calculation, voice) driven from the installed package. Quality measurements (retrieval benchmarks, speech WER, change-review precision) live in the repo under eval/.

Known limits: no code signing yet (SmartScreen may warn), in-app PDF page rendering shows the stored original, and the first-run model downloads need internet.