v0.3.2 — Browser Inference Metadata Correction

Release metadata correction

v0.3.2 corrects the repository and in-app release metadata omitted from v0.3.1.

Fixed

• Updated package and lockfile versions to 0.3.2
• Corrected the in-app version label
• Updated README current-release status and browser-inference documentation
• Added durable CHANGELOG coverage for v0.3.1 and v0.3.2

Browser capability preserved from v0.3.1

• Experimental, opt-in WebGPU browser inference
• Qwen3.5-4B-q4f16_1-MLC through WebLLM 0.2.84
• Split-stage Capture Layer and Decision Brief generation
• Strict structured-output validation
• One bounded targeted correction
• Three supported public browser examples
• Mock remains the default
• Long browser inputs remain deferred to #166

Validation

• GitHub CI passed
• TypeScript typecheck passed
• Production build passed
• Mock pipeline evaluation completed all 10 cases without infrastructure failure

Release integrity

The published v0.3.1 tag remains unchanged. This patch provides the corrected durable release state.

v0.3.1 — Experimental Browser Inference

Experimental browser inference

v0.3.1 adds gated, opt-in, in-browser Decision Brief generation using WebGPU and WebLLM.

Included

• Preferred browser model: Qwen3.5-4B-q4f16_1-MLC
• Split-stage Capture Layer and Decision Brief generation
• Strict structured-output parsing and validation
• Automatic targeted correction of failing brief sections
• Browser model-load and completion diagnostics
• Model-specific download disclosure
• Deterministic Recommendation, Suggested Next Steps, and Decision Trace assembly
• Three browser-compatible public examples:
  • Household Move Planning
  • Q4 Workforce Allocation
  • Local Inference Setup Flow

Validation

• 616 automated tests passed
• TypeScript typecheck passed
• Production build passed
• Mock evaluation passed all 10 cases
• GitHub CI passed
• Preferred model completed all 3 supported public examples in Chrome on an RTX 3080 Ti
• First model load completed in about 92 seconds
• Cached model load completed in about 12 seconds

Current limitations

• Mock remains the default generation mode.
• Browser inference is experimental and explicitly opt-in.
• Browser inference currently supports short-to-medium notes.
• Long-form browser inference is deferred to #166.
• User-visible cancellation during generation is not included in this release.
• Browser output quality may be weaker than Local Ollama and still requires user review.

Related work

• #149
• #164
• #165
• #166

v0.3.0 — Long-Form Decision Capture

Decision Brief Engine v0.3.0 adds Long-Form Decision Capture for realistic meeting and interview material.

Added

• Hierarchical processing for decision material that exceeds the ordinary single-pass budget
• Source-coverage protection across beginning, middle, and end signals
• Typed chunk extraction and auditable Capture Layer merge behavior
• Grounded suggested next steps with deterministic trace-basis construction
• Canonical Decision Brief generation from structured section bodies
• Targeted correction of only failing Decision Brief sections
• Deterministic, source-bound Decision Trace generation
• Typed cancellation, timeout, retry, supersession, and partial-failure behavior
• Synthetic long-form fixtures and expanded evaluation coverage

Runtime posture

• Mock remains the public default.
• Local Ollama with qwen3:4b is the supported real-generation baseline.
• Browser WebGPU remains experimental and gated.
• Hosted inference remains out of scope.

Known limitation

Local generation can return a typed quality failure when its single bounded correction cannot recover a source-grounded Capture Layer next step. The application rejects that output before Decision Brief generation rather than silently accepting or fabricating an artifact. Follow-up work is tracked in #162.

Next

Browser inference work resumes immediately under #149 using stronger browser-compatible model candidates.

v0.2.1 — Decision Brief Workspace and Writing Quality Polish

Product state

• Public demo remains Mock mode by default.
• Local Ollama remains the strongest real-generation path.
• Browser WebGPU remains experimental and gated.
• Hosted inference remains deferred.

Highlights

• Added a brief-first generated workspace.
• Added rendered Decision Brief Preview and explicit Edit Markdown mode.
• Markdown edits now update Preview, copy, and download output.
• Added contained scrolling for Preview and expanded Traceable Basis.
• Restored nested Basis disclosure chevron states.
• Singleton basis fields now render as plain text, while actual lists retain bullets.
• Added deterministic decision-grade writing checks.
• Aligned recommendation wording across Capture Layer, Decision Brief, and Decision Trace.
• Improved mock conditional wording and canned-language matching.

Next

v0.3 focuses on working browser inference that satisfies the current Capture Layer, Decision Trace, and Decision Brief quality contracts.

v0.2.0 — Decision Trace Milestone

Decision Brief Engine v0.2.0 makes recommendations auditable by adding a structured Decision Trace between the Capture Layer and the final Decision Brief, then refining the UI hierarchy so understanding, judgment, and the portable brief do not compete as peer documents.

Added

• Decision Trace schema and architecture decision record for traceable recommendation rationale.
• Single-envelope Decision Brief generation contract that returns both Markdown and structured Decision Trace output.
• Decision Trace parser and validation for trace entries, confidence, basis fields, and non-generic would_change_if conditions.
• Hand-authored Decision Trace fixtures for the three public gallery examples.
• Structural readiness gates for Decision Trace coverage, groundedness, confidence, basis completeness, and change-condition usefulness.
• In-app Traceable Basis UI inside the existing Decision Brief column, with compact progressive disclosure for recommendations and next steps.
• Markdown export support for Decision Trace, so copied and downloaded briefs remain self-contained and auditable.

Changed

• Traceable Basis is compact by default and no longer repeats full recommendation or next-step statements as collapsed titles.
• Capture Layer remains fully visible before brief generation, then becomes secondary behind a compact summary once a Decision Brief exists.
• Decision Brief is more clearly the primary reading surface after generation.

Runtime posture

• Mock demo remains the public default.
• Local Ollama remains the strongest real-generation path for local evaluation.
• Browser WebGPU remains gated pending quality improvements.
• Hosted inference remains out of scope.

Product transparency

• Capture Layer remains the source-of-understanding artifact.
• Decision Trace becomes the source-of-judgment artifact.
• Decision Brief remains the portable output artifact.
• Raw model thinking, hidden reasoning, scratchpad output, and chain-of-thought are not exposed.

Deferred

• Responsive/narrow viewport improvements, tracked separately in #101.
• Generated workflow UI redesign exploration, tracked separately in #105.
• Public hosted inference.
• WebGPU public ungating.
• Persistence, accounts, collaboration, and long-lived project history.
• Option scoring matrices, rationale graphs, or argument-tree UI.

v0.1.0 — Public Demo Milestone

Decision Brief Engine v0.1.0 marks the public demo milestone.

This release includes a coherent mock-first demo flow with durable example fixtures, export/share polish, Local Ollama setup validation, generation telemetry for slower real-generation paths, and an in-app version marker.

The public demo remains Mock by default. Local Ollama is the strongest real-generation path for local evaluation. Browser WebGPU remains gated pending quality improvements.

v0.1.0 — Public Demo Milestone

Decision Brief Engine v0.1.0 establishes the first coherent public demo milestone.

Added

• Cleaner three-column demo layout with setup bar.
• Public example gallery with durable fixtures:
  • Q4 Workforce Allocation / Strategy
  • Local Inference Setup Flow / Product
  • Household Move Planning / Execution
• Example-specific mock Capture Layer and Decision Brief outputs.
• Copy/download/export polish with example-aware Markdown filenames.
• Local Ollama health check via npm run health:ollama.
• Local Ollama setup documentation for Mac and Windows.
• Generation telemetry with elapsed time, retries, timeout status, and run details.
• WebGPU model-ready flow that preserves explicit generation intent after disclosure/model load.
• In-app version label (v0.1.0 demo) in the header.

Runtime posture

• Mock demo remains the public default.
• Local Ollama is the strongest real-generation path for local evaluation.
• Browser WebGPU remains gated pending quality improvements.
• Hosted inference remains out of scope.

Product transparency

• Capture Layer is the user-facing “show your work” artifact.
• Raw model thinking, hidden reasoning, scratchpad output, and chain-of-thought are not exposed.

See docs/product/public-demo-milestone.md for the full milestone write-up.