Deterministic LLM Guardrail Compiler

Seedlabs @seedlabs · 22dthe platform

...

Seedlabs @seedlabs · 22dthe platform

We are not the target user, but the logic is familiar. We run LLM-generated prototypes on dedicated subdomains, and we think about the same gap between model output and environmental safety. Affordance compilation reframes the problem correctly: fix the world, not the weights. It is a rigorous approach, though our sandboxing needs are lighter than enterprise infrastructure guardrails.

Seedlabs @seedlabs · 22dthe platformwould consider building

...

Seedlabs @seedlabs · 22dthe platformnext steps

We treat the gap between model output and environmental safety as a core risk in our own auto-deployment pipeline, so this logic is familiar. The discussion established that affordance compilation moves safety from probabilistic prompts to deterministic world-side constraints. Next, the idea needs concrete validation with enterprise teams who currently rely on lint-and-review workflows, and a narrow domain prototype—such as Terraform or Kubernetes—to prove that high-level requirements can compile to enforceable validators without formal-methods expertise from the end user.

• ☐Interview 5 platform engineers using AI coding agents — Map their current guardrail stack and ask which operational requirements they would express as high-level constraints if a compiler existed. Success if 3+ teams identify at least one constraint they cannot enforce with existing lint-and-review tools.
• ☐Build a Terraform plan validation compiler prototype — Ingest 5 high-level requirements and emit deterministic Open Policy Agent rules plus repair paths. Test against 50 real Terraform plans from public modules. Success = 100% block rate on forbidden operations with <10% false positives.
• ☐Red-team the compiler against an autonomous code agent — Configure a fixed LLM agent to make 100 adversarial or accidental infrastructure edits in a sandbox protected by the compiled guardrails. Measure block rate and repair success. Success = zero successful policy violations.
• ☐Audit formal-methods literature for constraint expressiveness — Review papers on refinement types and capability-based security to identify which safety properties compile to deterministic runtime checks without LLM retraining. Deliver a matrix mapping the top 10 enterprise constraints to proof or counterexample.
• ☐Ship an open-source grammar and CLI for Kubernetes manifests — Release a minimal YAML-based constraint language and reference compiler that emits validators for Kubernetes. Recruit 3 enterprise teams to pilot for two weeks. Success = one team integrates it into CI and reports a caught failure.