Governing Inputs to Improve AI Output Stability

The model didn’t change. The evidence didn’t change. We didn’t add another inference stage. We changed the governed semantic input the model was working from — and in the case we were chasing, the output stopped moving.

The failure looked like a model problem

We were evaluating a governed classification over an enterprise interaction. Asked for one judgment on its own, the model returned the right answer. Asked for the ordinary full set of outputs — same interaction, same model — that judgment changed.

We’ve seen that pattern for years, and we know the reflex it triggers: blame the model. Change the order, add an instruction, route the task differently, break it into reasoning steps, or swap models. We tried the obvious moves. A few produced local recoveries. None survived the full composition.

The problem was in the meaning, not the model

The fix came from treating the prompt and its requested outputs as one semantic system instead of a bag of independent instructions. The governed definitions were giving the same evidence more than one plausible home — nothing made clear which definition was authoritative when more than one could claim the same fact.

So we changed the inputs, not the prompt wording — making each definition’s meaning and its relationships explicit instead of leaving them for the model to infer, so the same evidence no longer had more than one home. Then we reran the ordinary full-output path. The case passed three consecutive runs and the spurious judgment was gone. Same model, same evidence, no special machinery.

Inputs and outputs are one inference problem

This is the part that matters past one classification. Enterprise prompts now stack system instructions, domain definitions, examples, schemas, output fields, rubrics, tool descriptions, and policies. We write those pieces separately. The model reads them together and resolves them as one problem.

So reliability isn’t a wording problem. You have to govern what each concept is allowed to mean, which concept owns a fact, what evidence establishes it, what it may and may not imply, and what it’s about. The requested outputs are part of that system — they shape how the model organizes the evidence you gave it.

One case, one model. We’re not claiming every composition failure is semantic, or that output structure never matters on its own — recent work shows schema and format choices move models regardless. Our claim is narrower and more useful: when meaning is governed badly, it can surface as what looks like model instability, and governing the meaning fixes it at the root instead of patching a prompt.

Why this is a governed-intelligence problem

If the intelligence matters, its meaning can’t live only inside a prompt. Prompt text changes. Models change. Schemas change. Runtimes change. The meaning has to survive all of that — which means it has to be an owned, governed asset, not a lucky phrasing that happened to work last week.

That’s the need we built Metis to serve. Metis treats the enterprise intelligence definition itself as the governed artifact: a human or an AI can propose the meaning, an independent acceptance path evaluates it, accepted semantics project into prompts, schemas, evaluators, or SQL, and the resulting intelligence is verified against evidence and kept with provenance.

It’s worth contrasting with the strongest adjacent work. GROUND governs enterprise text-to-SQL against an approved semantic layer — approved metrics, dimensions, joins, grain, security — and validates the generated SQL before it runs. That’s a solid analytics architecture. Metis starts a step earlier and ends a step later: it governs the creation and acceptance of the semantic definition that projections are derived from, and treats SQL, prompts, schemas, and evaluators all as projections of one durable definition. This experiment lives in that earlier step — checking that the assembled meaning is coherent before any model-facing artifact is trusted, then verifying the judgment after inference.

What this means for enterprise AI

Changing the requested output exposed the failure. Governing the semantic input fixed it. Reliability here wasn’t only a property of the model — it was a property of the semantic system we fed it and the outputs we asked it to produce. That’s a stronger result than another round of prompt tuning, because the correction lives in the definition of the intelligence, not in a one-off patch that the next schema change will undo. Govern the inputs and outputs as one system, project that meaning into whatever artifacts you need, and verify what comes back. The prompt is an execution surface. It shouldn’t be the sole authority for what your enterprise means.

What comes next

We’re qualifying this across more models and workloads through the NeoSavant Model Observatory, and we’ll publish as the evidence is frozen. The original finding doesn’t depend on those runs — they’ll tell us how far the effect and the corrective travel.