Web Assistant changelog

v0.1.14 — Factual extraction and source-linked drafting

Prepared September 22, 2026.

• Shorter extraction instructions focus on the source and preserve roles, dates, currencies, mentorship direction, and limits on outcomes.
• Records without a usable quotation from their source section are withheld from evidence and kept separately for inspection. Name-only introductions are read with the following section.
• Structured drafting connects each accepted sentence to supplied evidence. Invalid or unfinished responses remain incomplete and are not reused as finished drafts.
• Evidence review highlights possible role reversals and unsupported outcome claims. Source links still require claim-by-claim review.
• Browser inference has a deadline so a stalled structured-output request can be stopped and the model reloaded.

Desktop v0.1.7 supports the same structured extraction and drafting with local Ollama and llama.cpp. Updated model comparisons preserve raw outputs and remaining quality problems.

v0.1.13 — Saved work and reviewed evidence

Prepared September 22, 2026.

• Keep document history synchronized with the latest corrections and drafts; show autosave failures with retry and backup controls.
• Export and import portable workspace backups, including source passages, evidence, corrections, and all three drafts.
• Edit or remove every extracted fact, and explicitly review secondary records before they enter drafting.
• Retain incomplete extraction results without caching them as finished. Preserve distinct same-title projects and complementary evidence.
• Choose the document language independently from the interface, with English and French ranking signals.
• Review evidence, write drafts, and export without WebGPU or a model download using “Continue without AI.”

Desktop v0.1.6 also validates local-only model configurations, retains incomplete drafts with continuation, and includes refreshed fallback installers verified against the release source.

v0.1.12 — Extraction recovery and current evidence review

Prepared September 8, 2026.

• Resume interrupted extraction using completed sections while preserving manual corrections and removed contributions. Optional browser saving also restores checkpoints.
• Review every draft sentence, including missing citations, changed evidence, and sources outside the current selection. Readiness checks use the selected evidence.
• Keep incomplete drafts visible and exportable, with explicit continuation after a response limit or interruption. Only completed generations are reused as finished drafts.
• Preserve PDF lines, headings, page boundaries, and common column layouts in the browser and desktop source readers.
• Batch streamed screen updates and report local time to first text and runtime-reported generation speed.

v0.1.11 — Model setup and download recovery

Prepared September 7, 2026.

• Fixed model selection so start and retry use the chosen model ID.
• Allowed source and agency preparation while local AI loads, with a persistent status and guarded AI actions.
• Separated download, memory loading, and GPU preparation progress; completed downloads are reused on retry.
• Made cancellation immediate, ignored late results from cancelled attempts, and reused a model already in memory.
• Kept automatic fallback limited to memory and compilation failures, and added cache and offline guidance.

v0.1.10 — Launch safeguards and complete workflow checks

Released July 27, 2026.

• Added complete save, reload, restore, export, and deletion coverage, plus checks for all three evidence gates, unsupported details, language page estimates, and combined exports.
• Added release checks that require the configured analytics tag on every public page.
• Replaced instructions to bypass operating-system security warnings with safer browser, IT-review, and signed-installer guidance.
• Documented why WebLLM remains the browser engine while the native desktop route provides a fallback for unsupported devices.
• Labelled French as a partial preview, explained which text remains in English, and verified that key interface text and the page language switch correctly.

v0.1.9 — Academic recognition and accessibility

Released July 27, 2026.

• Expanded publication, funding, supervision, SSHRC, NSERC, and CIHR terminology used during evidence discovery.
• Fixed section detection so short academic evidence lines are not mistaken for headings and discarded.
• Added selected-state announcements, deletion confirmations, progress labels, skip links, reduced-motion support, and a searchable plain-language glossary.
• Made saved CV documents strictly opt-in and made “Remove all saved CV work” clear both the workspace and document history.

v0.1.8 — Public beta launch readiness

Released July 26, 2026.

• Added visible public-beta compatibility and work-saving guidance.
• Made the lower-memory 1B model the safer first download while retaining automatic recommendations for capable devices.
• Clarified privacy wording, completed the French previous-document pathway copy, and declared the Node.js test requirement.

v0.1.7 — Guided drafting workflow

Released July 26, 2026.

• Divided drafting into four steps: choose target, review evidence, rank elements, then draft and review.
• Added a visible workflow stepper and focused navigation between drafting decisions.

v0.1.6 — Faster model setup

Released July 26, 2026.

• Added persistent IndexedDB model caching with visible cached/not-cached status.
• Added background model downloads, cancellation, retry, and smaller-model guidance.
• Added optional CV context compression with a user-ranked priority review before agency selection.

v0.1.5 — Forced runtime refresh

Released July 26, 2026.

Added a new runtime and service-worker release identifier so browsers cannot reuse the earlier incompatible application shell or model worker.

v0.1.4 — WebLLM runtime and cache repair

Released July 26, 2026.

• Updated the official WebLLM runtime from 0.2.79 to 0.2.84.
• Added release-specific URLs for the application, model worker, and AI runtime.
• Changed JavaScript and worker updates to network-first loading with offline fallback.
• Added a runtime-specific recovery action that removes an outdated application shell before reloading.

v0.1.3 — Source-first agency positioning

Released July 26, 2026.

• Reordered the workflow to load a model, add a CV or source, and then choose the funding agency.
• Added an “Other funding agency or program” option with opportunity-specific prompts.
• Added local-model suggestions that prefill only source-supported agency evidence for user verification.
• Expanded end-to-end coverage for the new workflow, all agency modes, and guided evidence review.

v0.1.2 — Local model runtime repair

Released July 26, 2026.

• Replaced the fragile esm.sh WebLLM rebundle with the official pinned npm browser distribution.
• Cache-busted both the page and worker runtime to prevent stale incompatible modules.
• Added a real bundle-import and model-worker startup test in Chromium.
• Classified missing runtime exports separately from memory, WebGPU, download, compilation, and extension failures.

v0.1.1 — Workflow reliability and evidence review

Released July 26, 2026.

• Added compatibility checks, automatic model fallback, retry/cancel controls, cache recovery, and copyable diagnostics.
• Separated compatibility, model, agency, and source selection into one-decision screens.
• Added live structured SSHRC, NSERC, and CIHR fit reports and follow-up questions.
• Added previous-document workspace restore, rename/delete controls, storage reporting, and a privacy dashboard.
• Added responsive extraction and drafting progress with cancellation and recovery guidance.
• Added mocked full-journey Chrome tests plus keyboard and dynamic-status coverage.

v0.1.0 — Initial web release

Introduced private in-browser model execution, CV extraction and verification, evidence ranking, narrative drafting, review checks, and text/Markdown export.