Show HN: Forge takes an 8B model from 53% to 99% on agentic tasks

Show HN: Forge – Guardrails take an 8B model from 53% to 99% on agentic tasks

Forge adds five guardrail layers to self-hosted LLM tool calling, raising Ministral 8B to 99.3% across 18 multi-step agentic scenarios, with the accepted ACM CAIS ’26 paper covering 97 model/backend configurations and 50 runs per scenario.

Why it matters: HKR-H/K/R all pass: the 53%→99.3% jump is clickable, the test setup has concrete numbers, and self-hosted agent reliability is a live practitioner pain. Single-source Show HN/GitHub evidence keeps it in the 78–84 open-source-tool band, not P1.