v1.5.0 — Trust & Quality
v1.5.0 — Trust & Quality
Three new opt-in features that make generated ideas sharper and more trustworthy.
Grounding verification for Deep Dive
Each Deep Dive section is now compared against the source paper using semantic similarity. Sections get a badge:
- Grounded (green) — aligns with source paper (≥65% similarity)
- Partial (yellow) — some claims may be inferred (40-65%)
- Caution (red) — may not be directly supported (<40%)
Enable in Settings → "Verify Deep Dive grounding". Adds ~5s per Deep Dive.
Self-distillation refinement
After generating ideas, the LLM re-reads its own output and refines: removes redundancy, sharpens titles, improves feasibility. Same ideas, better quality.
Enable in Settings → "Self-distillation refinement". Adds 1 extra LLM call per batch.
Prompt sensitivity check
Trend analysis now optionally runs with 2 different prompt framings (consensus vs contradiction). Compares keyword agreement and adjusts confidence score accordingly.
Enable in Settings → "Prompt sensitivity check". Adds 1 extra LLM call per category.
Other changes
- Idea card footer fixed — "Check Novelty" and "Show more" always visible (never clipped by overflow)
- Version auto-injected from VERSION file (no more hardcoded version in dashboard)
- All three features off by default — zero behavior change for existing users
From community PR (thanks @Bortlesboat)
- Fixed setuptools build backend for editable installs
- Integration test mocks retargeted to blueprint modules
- Testing docs updated for
pip install -e ".[dev]" - Hypothesis cache added to .gitignore
Upgrade: git pull && pip install -r requirements.txt
Full changelog: CHANGELOG.md
