Concept
Large language models are probabilistic by nature - ask twice, get two answers. That's a feature for writing and a hazard for software. Deterministic AI generation is the architecture that resolves the tension: let AI interpret, never let it improvise the thing that runs.
Software has one non-negotiable property: it must do the same thing every time. LLMs have the opposite property. When an AI writes application code directly, that mismatch lands in production - regeneration changes behavior, two identical requests yield different bugs, and security controls exist only if the model included them on that particular run. The 2025 incidents around AI-generated apps - missing access controls at scale, agents deleting production data - are this mismatch playing out, not a temporary model weakness.
Deterministic AI generation splits the pipeline at the point where consistency starts to matter:
One sentence: the AI decides what, the engine decides how. The full implementation is on the architecture page.
Determinism applies to generation, not to the product's intelligence. Inside a running Chromoly system, AI steps score leads, triage inboxes, extract invoice fields, and draft replies - each inside a frame of cost caps, retries, fallback, and human approval gates. The same principle at a smaller scale: probabilistic judgment, deterministic guardrails.
An architecture where AI output is constrained to a structured, verifiable artifact - like Chromoly's AppSpec - and a deterministic compiler turns that artifact into the running software. The AI supplies interpretation and judgment; the machinery that actually builds and runs the system behaves identically every time.
Because LLMs are probabilistic: the same request produces different code on different runs. For a prototype that's harmless. For a business system it means rebuilds can change behavior, security depends on what the model generated that day, and no one can promise the system tomorrow matches the system today. Determinism restores that promise.
It trades open-ended flexibility for reliability - deliberately. A code-writing AI can build anything, including things that quietly don't work. A spec-constrained AI builds within a domain (data models, workflows, screens, AI steps) where every output is valid by construction. For business software, the domain covers what's actually needed.
Yes - as bounded workflow steps: classify, extract, summarise, score, draft. Each runs with cost caps, retries, provider fallback, and optional human approval gates, with low-confidence results routed to a review queue. Judgment at the edges, determinism in the middle.