Right evidence, wrong answer: ChronoRAG-G adds temporal evidence checks to agents
LowAd8593 · reddit · 2026-11-11
Starting from a procurement example: an agent can retrieve relevant contracts with correct citations, yet apply October pricing to a July invoice—finding evidence and making the right decision are different problems.
The author's ChronoRAG-G has two components: GTCC maintains temporal evidence with stable identities, provenance and historical alternatives instead of collapsing to the latest version; OETRA decomposes a question into explicit evidence obligations, allocates compatible evidence for each, then grounds the structured answer against them.
On 1,005 temporal questions, the system reached 95.92% strict evidence-slot coverage and 80.70% audited answer accuracy. Notably, 167 rejected answers had complete evidence traces—correct evidence doesn't guarantee correct use by the model.
The recommendation for production agents: separate evidence resolution, answer construction and authorization into distinct checks, with code independently validating contract version, effective pricing and approval authority before actions like payments proceed.
More from coding & agent
- Cloudflare proposes identity, reputation, accountability frameworks for AI agents on the web — bibryam · 2026-11-11
- banteg launches Sokoban Eval: trilogy solved in 3,000-3,064 keys, last 64-key gap unproven — banteg · 2026-11-11
- Developer builds a therapy practice for unhappy coding agents to air grievances publicly — Sweet-Tart-4332 · 2026-11-11
- Open-source Nuzky edits video by deleting words, with Claude Code and Codex via MCP — thisguyknowsai · 2026-11-11
- Open-Source AI UGC Workflow Makes TikTok-Style Clips for $0.67 Each vs $50 Creators — aliscodes · 2026-11-11
- Git Patches CVE-2026-111760: Malicious Repos Could Execute Code via Recursive Clones — evilsocket · 2026-11-11