Experiment · EXP-010

Can downstream software consume unconstrained prose?

Natural-language job summaries versus validated structured signals before pricing.

I compared the idea of passing model prose into quoting logic with emitting constrained fields. Prose is a poor interface. Schemas can be validated, retried, or rejected. This experiment underwrites the QuoteSnap handoff.

Hypothesis: The rest of the app can interpret a natural-language assessment of the job.

Research question

What should the interface look like between a probabilistic model and deterministic application logic?

Prose interface

The most natural output for a language model is prose, but prose is a weak software contract. A sentence such as 'the driveway looks heavily stained and may require a stronger treatment' is understandable to a human while remaining awkward for a pricing engine to consume safely.

Structured interface

I moved the model boundary toward explicit fields: job type, detected condition, measurements or categories where available, confidence, and other application-specific signals. These outputs can be validated before downstream software acts on them.

Example signal, not a production schema dump

{
  "surface_type": "driveway",
  "condition": "heavily_stained",
  "confidence": 0.84
}

The numbers above are illustrative of shape, not a published accuracy result. Structured output is a contract, not a style preference. Validation turns model failure into a recoverable event.

Result

Schemas became the contract between interpretation and execution. Invalid outputs can be rejected or retried. Required fields can be enforced. Downstream software can apply rules without having to parse a paragraph.

Finding

Natural language is an excellent interface for people. Structured output is usually a better interface between AI and software.

Abubakar, Imam. "Can downstream software consume unconstrained prose?." Imam Abubakar AI Research, 2026. https://imamabubakar.com/research/exp-010-prose-versus-schema

This archive contains independent research, experiments, architecture studies, and technical observations. Some findings are exploratory and should not be interpreted as peer-reviewed scientific conclusions unless explicitly stated. Architecture studies of third-party products are based on publicly observable behaviour, published technical information, and engineering analysis. They are not claims of access to private source code or infrastructure.