RBR-style scoring for Gauntlet Inspector findings #6

New issue

Closed

opened 2026-05-23 20:54:22 +00:00 by coilysiren · 1 comment

coilysiren commented

2026-05-23 20:54:22 +00:00

Owner

Originally filed by @coilysiren on 2026-05-03T19:53:38Z - https://github.com/coilysiren/gauntlet/issues/7

🤖 Filed by Claude Code on Kai's behalf.

Inspiration: openai/safety-rbr-code-and-data - rule-based rewards: a small-data way to fit weights for policy-shaped LLM outputs.

Joins against: gauntlet's Inspector role (currently emits Findings as free-form text), the AI/ML literacy gap.

Gauntlet's Inspector role is the obvious place for RBR. Findings today are unstructured text. Reframe as: each Finding scored against a small rule set (severity, blocker class, replayability) using the RBR weight-fitting code. The dataset ends up small enough to hand-label, the math ends up legible enough to write up, and the result is a real ML artifact tied to a real adversarial-testing system. AI/ML literacy gap closer that lives inside an existing repo, no new domain to learn.

🤖 Filed by Claude Code on Kai's behalf.

Moved from coilysiren/coilyco-ai#40.

_Originally filed by @coilysiren on 2026-05-03T19:53:38Z - [https://github.com/coilysiren/gauntlet/issues/7](https://github.com/coilysiren/gauntlet/issues/7)_ > 🤖 Filed by Claude Code on Kai's behalf. **Inspiration:** [`openai/safety-rbr-code-and-data`](https://github.com/openai/safety-rbr-code-and-data) - rule-based rewards: a small-data way to fit weights for policy-shaped LLM outputs. **Joins against:** `gauntlet`'s Inspector role (currently emits Findings as free-form text), the AI/ML literacy gap. Gauntlet's Inspector role is the obvious place for RBR. Findings today are unstructured text. Reframe as: each Finding scored against a small rule set (severity, blocker class, replayability) using the RBR weight-fitting code. The dataset ends up small enough to hand-label, the math ends up legible enough to write up, and the result is a real ML artifact tied to a real adversarial-testing system. AI/ML literacy gap closer that lives inside an existing repo, no new domain to learn. > 🤖 Filed by Claude Code on Kai's behalf. --- *Moved from coilysiren/coilyco-ai#40.*

coilysiren commented

2026-05-30 05:44:11 +00:00

Author

Owner

Iceboxed in the 2026-05-29 backlog burn-down: Explore-openai-RBR inspiration, speculative future feature. Reopen anytime if it becomes real.

coilysiren

2026-05-30 05:44:11 +00:00