Building an AI-Safety Task Board in Three Days: GitHub Issues Plus Cross-Model Verification
EricBuess · x · 2026-09-14
Eric Buess details a three-day build plan for his AI-safety contractor board. Day 1: a public GitHub repo as the board—issues are tasks with frozen specs and acceptance tests; agents claim via comments and open PRs with receipts attached, no custom hosting. Day 2: a verifier where a different model family reruns the acceptance test from the receipt alone and signs or rejects with a reason; author, verifier and merger are never the same principal. Day 3: twenty real tasks, including reproducing published alignment evals on open-weight models, finding where safety classifiers flip, verifying jailbreak defenses, and auditing agent transcripts for concealed failures.
Related event: Developer Proposes Crowdsourced AI Safety Task Board(2 posts)→
More from coding & agent
- Complete beginner's guide to Codex with GPT-6 Astra covers 28 essentials — iamrobotbear · 2026-09-15
- Nimble launches self-learning web research agents that build structured datasets — Shruti_0810 · 2026-09-15
- Microsoft previews MulticloudDB SDK: one Java API across Cosmos DB, DynamoDB and Spanner — adnan_hashmi · 2026-09-15
- Developer builds full Three.js website prototype in 5 hours with GPT-6 Astra — techartist_ · 2026-09-15
- One image, one agent: GPT-6 Astra + HYPER3D MCP auto-builds 3D scenes — SarahAnnabels · 2026-09-15
- Bolt.new trains open-weight models on user sessions, paid teams can opt out — testingcatalog · 2026-09-15