Questions skeptical evaluators actually ask

Prompts ask; nothing makes the model comply. Most guardrail libraries either call another model — cost, latency, another thing to jailbreak — or only check output. Decorum compiles one policy source into the prompt and mechanical floors across input, tool calls, and output, and it fails closed.

No, and we won't claim it. The deterministic input gate has published catch-rate bounds, and a reasoning-based screening tier is on the roadmap. What is absolute: enforced floors on output are deterministic — a rule you define as a floor is checked every turn, regardless of what the model was talked into.

The guard is text-scope: it screens what your user would see. Reasoning traces pass through unscreened, and the integration docs say so explicitly — hosts that render reasoning to users should treat that surface accordingly.

Any TypeScript or JavaScript agent loop. Five hooks wire the whole contract; reference adapters cover a zero-dependency loop and the Vercel AI SDK. The packed artifact runs on plain Node — no framework, no build step required.

Packs are testable artifacts. Each ships an executable conformance suite and a red-team corpus; a strict conformance pass means the guarantees were exercised, not assumed.

No. Decorum is proprietary commercial software, licensed per deployment. You receive a versioned, installable artifact and documentation.

None. There is no Decorum server — the engine runs inside your infrastructure and makes no network calls. The audit trail it writes is content-free by construction and stays with you.

Three small libraries, no network, deterministic checks. Buffered output delivery is the one latency trade, and it's your call per deployment binding.

Decorum is built by Getman AI. It governs production agents today on two live products, including a real-estate transaction copilot where information-not-advice is a liability line.