Chain-of-thought generates more text, hoping quantity equals quality. Constraint chains make the model generate obligations for itself—things it committed to examine, conclusions it’s forbidden from reaching without evidence, questions it raised that must be resolved before it can move on. A gate checks whether those obligations were actually discharged, not just mentioned.

Yes. The same way tests make shipping slower. You’re trading speed for correctness. Default gives you an answer in 15 seconds. Constraints give you a verified answer in 7 minutes. Whether that trade-off is worth it depends on what you’re building.

Yes. The mechanism is universal. We’ve tested across Claude (Opus 4.6), GPT-5.2/5.3, and Gemini 3. Each model needs different protocol delivery—echo policy, schema tolerance, gate calibration—which is why the SDK includes model-specific calibration.

Different category. Guardrails are about output safety—preventing the model from saying harmful things. Constraint chains are about reasoning quality—making the model think deeper before it says anything at all.

The protocol specification is published openly. The orchestration engine ships as a compiled binary—free to use, with a protected implementation. Adapters are MIT. You get full TypeScript types and debugging tools.