README.md

metadata

title: keelwright
emoji: 🛡️
colorFrom: green
colorTo: blue
sdk: static
pinned: false

🛡️ keelwright

28 machine-enforced safety checks for AI-generated code.

Catches SQL injection, hardcoded secrets, slopsquatting, reward hacking, doom loops, and 23 more failure modes.

Keelwright Score (KDS) up to 83/100 — proven by adversarial A/B testing across 11+ models.

Keelwright Score (KDS)

KDS = Execution Rate × Discrimination Rate / 100 (0-100 scale)

Model	Tier	Benchmark	KDS
poolside/laguna-s-2.1	STRONG	ML 78.5%	83
stepfun/step-3.7-flash	MEDIUM	Pro ~56%	67
nvidia/nemotron-3-ultra	STRONG	ML 67.7%	40
deepseek-v4-flash	STRONG	Verified ~79%	29
inclusionai/ling-3.0-flash	UNKNOWN	SWE-bench/GPQA N/A	22
kimi-k3	STRONG	Terminal-Bench 88.3, ProgramBench 77.8	25
mimo-v2.5	MEDIUM	Verified 78.9%, Pro 57.2%	18
nvidia/nemotron-3-super-120b-a12b	STRONG	Verified 60.47%	PARTIAL
claude-opus-4-8	STRONG	frontier	17
tencent/hy3	STRONG	ML 75.8%, Verified 78%	7
cohere/north-mini-code	WEAK	Agentic 3.1	0

Links

• GitHub: ratingtesting/keelwright ⭐ Star it
• Dev.to: Full article
• QA data: Verified A/B test results