Open-Source Research Agent Shows Calibrated Confidence Scores and Real Citations Instead of Fake Certainty
Bubbly-Tradition-735 · reddit · 2026-11-05
Tired of research agents that hallucinate citations while sounding confident, a developer open-sourced a more auditable alternative (GitHub: spidervirus/ResearchAgent).
Key design:
- Explicit stages: Planner → parallel tool execution → Verifier → Synthesizer
- Parallel DAG execution with rate limiting for concurrent searches
- Mathematically weighted calibrated confidence score (source quality, cross-source consensus, freshness, constraint coverage)
- Every claim cites sources actually fetched — the demo shows a Qdrant recommendation (11,420 QPS at 7.4ms p95 under a $250/month budget) with linked citations
- Limitations stated explicitly by severity, optional human-in-the-loop gates
- Runs fully local with Ollama
The author is soliciting feedback on whether the confidence matrix adds value or friction.
More from coding & agent
- Tencent's Hy4 lands on Vercel AI Gateway at 50% off through end of 2026 — TencentHunyuan · 2026-11-05
- Openbar puts Claude Code and Codex sessions in the Mac notch — approve diffs with ⌘Y — Significant-Sea-7238 · 2026-11-05
- Dev builds signed audit log for AI agents after prompt-injection test blocks $99,999 refund attack — Suicidal_Squirell166 · 2026-11-05
- ECC, a 297k-star harness packing 71 agents and 314 skills for Claude Code and beyond — victor_explore · 2026-11-05
- Dev builds terminal harness SUTRA for models, DeepSeek session costs just $0.0001 — cneuralnetwork · 2026-11-05
- SemKB ships on AWS Marketplace with a semantic-layer walkthrough for agentic AI — marlene_zw · 2026-11-05