Introducing IPI

Interfaces, not features. A model class is defined by what it predicts; an interface class is defined by what it exposes. APIs expose functionality. IPIs expose interaction — and once interaction is an interface, acting on an environment becomes something intelligence can be programmed against, measured on, and held to.

01The closed loop

Issuing an action is easy; knowing it landed is the discipline. An IPI is therefore not an input channel but a closed action loop — and every stage of it is part of the interface:

Observe→Reference→Act→Verify→Update state

Spoken, the work an IPI carries arrives as Speech‑to‑Action Voice Commands — the unit of spoken work an STA Model interprets and executes against the surface (their coordination twin, the Speech‑to‑Orchestration Voice Command, is the OPI’s concern). A voice command like “pause the second one” exercises the whole loop: observe what is on stage, resolve the referent, act on it, verify the player actually paused, and carry the new state forward. The Interactive Intelligence Benchmark makes verification the heaviest action-side track for exactly this reason — the loop is only closed when the report about the world is true.

02What a surface exposes

Under an IPI, an interactive surface is not pixels but a contract:

surface
  ├── objects — what exists, each with a stable identity
  ├── capabilities — what can be done to each, and whether it verifies
  ├── state — what the surface holds right now, queryable
  ├── references — how spoken and written names resolve to objects
  ├── actions — the vocabulary of change
  ├── events — what just changed, including changes the human made
  ├── permissions — what requires holding, confirmation or refusal
  └── receipts — the record of what was attempted and what landed

The last line is the one that separates an IPI from tool calling. A receipt is a data structure, not a hope: every action returns verified, attempted, blocked, failed or held, and a claim of success is only ever made from a receipt that says so.

IPI is the programming interface for intelligence that acts.

03The first generation

The first-generation IPI runs in LabsAI Studio. Its surface publishes a live capability catalog scoped to object lifetime — an action exists exactly while the thing it acts on is mounted; spoken referents resolve through a token-overlap scorer with an explicit ambiguity band, and a genuinely ambiguous reference is asked about rather than guessed; every action is verified against the real state of the surface and returns its outcome as a receipt; destructive actions are held for confirmation before anything happens; and the surface’s state — what is staged, what is playing, what is selected — is queryable rather than assumed. It is one surface, honestly described; the interface is written so the next surfaces arrive under the same contract.

The full argument — three planes, one stack — is in the note.

Read Intelligence Above the API →

Contributors

Labsintelligence

Contributing authors: Maya E. Davis · Duránd F. Davis Jr.

Tell us what you think, join us

This research is published while the questions are still open, and the systems it describes are live. We would love for you to join us — and please share your thoughts at research@labsintelligence.ai.

Citation

Please cite this work as:

Davis, Maya E., and Davis, Duránd F., Jr., “Introducing IPI.” LabsAI Research Briefs, Labsintelligence — lab1 of Labs Companies, Inc., August 2026.

Or use the BibTeX citation:

@article{labsintelligence2026introducingipi,
  author  = {Davis, Maya E. and Davis, Duránd F., Jr.},
  title   = {Introducing IPI},
  journal = {LabsAI Research Briefs},
  publisher = {Labsintelligence, lab1 of Labs Companies, Inc.},
  year    = {2026},
  month   = {august},
  url     = {https://labsintelligence.ai/research/labsai/introducing-ipi/},
}