PREVIEWS
Working technologies people can use
A capability earns its claims by being usable. Before a finding is written down it runs in front of visitors, in production, where the failure modes are real ones.
Labsintelligence · lab1 of Labs
We work on intelligences that do more than describe the world — systems that perceive, reason, coordinate and act, and that can be checked while they do it.
Our current work runs along two lines: speech‑to‑action (STA), translating human intent directly into action over a live surface, and orchestration intelligence — speech‑to‑orchestration (STO), deciding which intelligences do what, when, under a real‑time constraint. We publish while the questions are still open.
Some labs search for the microscopic laws of reasoning — how intelligence emerges inside a single model. Labs works at the other scale: the macroscopic laws of coordination — how many intelligences compose into systems that are governed, verifiable and accountable while they run.
Research Notes
Generation made intelligence abundant — and abundance moved the bottleneck. Orchestration Intelligence is the discipline of the layer that is now scarce, decomposed into the five layers of coordination.
Read →Research Specs
Agentic file formats: .spec carries intent, .script a workflow, .arch capabilities and policy, .orch the append-only execution trace. Two conformance profiles, and a generated example from the shipped implementation.
Read →Research Notes
AI-native computing needs more than the API: IPI, the interface for intelligence that acts; OPI, the interface for intelligence that coordinates; PAP, the protocol for intelligences that work together. APIs do not disappear — the abstraction stack gets taller.
Read →Research Specs
The wire format, complete and versioned: pap:// and paps://, the envelopes, describe/invoke/stream, the error registry, auth and execution metadata — with the reference implementation live in LabsAI Studio.
Read →Research Notes
Maya Davis' founding interface thesis: a digital canvas where ideas are orchestrated rather than typed - layered tracks of text, voice and biometric prompts sculpting multimedia human expression - and the era it opens, from Web3 into Web4's spatial web.
Read →Research Briefs
An Actra converts speech and world state into action; an Octra converts objective and resources into orchestration policy - to action and coordination what the GPT was to language. The family presentation: capabilities, recorded demonstrations, the Actra and Octra Scores on IIB-1, and the LAIMA forms - with every unreleased rung marked as what it is.
Read →Research Notes
Every voice system can decide what to say. Almost none can decide when. These notes are the account of building the second thing — the interaction loop, its invariants, the latency campaign run against the production system, and the parts that fought back.
Read →Research Notes
A model class is defined by what it natively predicts, not by whether it contains text. Actras predict action. Octras predict orchestration policy. The interactive surface we had to build first is the specification both are drawn from.
Read →Research Benchmarks
Every public benchmark for real-time voice scores the audio channel and, at most, whether a call was emitted. IIB-1 asks whether the artifact arrived on time, whether the referent resolved, and whether the agent told the truth about what it did.
Read →| Date | Category | Title |
|---|---|---|
| August 17, 2026 | Research Notes | The Scarce LayerOrchestration Intelligence, and the five layers of coordination. |
| August 17, 2026 | Research Briefs | Introducing TARGTandem Agents, and runtime guardianship. |
| August 17, 2026 | Research Briefs | Introducing the Orchestration GraphThe record of work, and the trust layer it creates. |
| August 17, 2026 | Research Briefs | Introducing ConductorThe connective movement of Orchestration Intelligence. |
| August 17, 2026 | Research Briefs | Introducing StructsSystemic reasoning, orchestrated learning, orchestrated control. |
| August 17, 2026 | Research Specs | A2F/1.0 SpecificationAgentic file formats — portable intent, governed execution, verifiable evidence. |
| August 17, 2026 | Research Notes | Intelligence Above the APIThe three planes of AI-native computing: IPI, OPI and PAP. |
| August 17, 2026 | Research Briefs | Introducing IPIThe Interactive Programming Interface — the programming interface for intelligence that acts. |
| August 17, 2026 | Research Briefs | Introducing OPIThe Orchestration Programming Interface — the programming interface for intelligence that coordinates. |
| August 17, 2026 | Research Briefs | Introducing PAPThe Parallel Agent Protocol — the communication protocol for intelligences that work together. |
| August 17, 2026 | Research Specs | PAP/1.0 SpecificationThe Parallel Agent Protocol — the wire format for invoking agent computation. |
| August 16, 2026 | Research Notes | Interactive Perspective SpheresGateways into the Interconnected Generative Era. |
| August 16, 2026 | Research Briefs | Introducing IPSInteractive Perspective Spheres - the unified creative platforms of the Interconnected Generative Era. |
| August 15, 2026 | Research Briefs | Introducing Actras & OctrasThe Labs model families for speech that acts, and orchestration that listens. |
| August 16, 2026 | Research Briefs | Introducing STA & STOSpeech-to-Action, and Speech-to-Orchestration. |
| August 16, 2026 | Research Briefs | Introducing GAT & GOTGenerative Action Transducers, and Generative Orchestration Transducers. |
| August 16, 2026 | Research Briefs | Introducing AI Orchs & Orch ModelsOrchestration Models are intelligence. AI Orchs are execution structures. |
| August 14, 2026 | Research Notes | Knowing When to SpeakTurn-taking, presence, and epistemics in a real-time voice loop. |
| August 14, 2026 | Research Notes | Larger Than LanguageLanguage is not the primitive; the surface is the specification, and action is the prediction. |
| August 14, 2026 | Research Benchmarks | Interactive Intelligence BenchmarkScoring what a real-time voice system does beyond the audio channel. |
PREVIEWS
A capability earns its claims by being usable. Before a finding is written down it runs in front of visitors, in production, where the failure modes are real ones.
NOTES
Research Notes are working documents. Each states what it specified, what it measured, what it only bounded, and what it left unsolved — including the negative results, which are usually the expensive half.
BENCHMARKS
Where no public benchmark measures the thing we built, we write one and publish its tasks, rubric and recordings — so the number can be reproduced by someone who did not choose it.
Formal Research Papers follow the notes, and carry what notes cannot: pre‑registered protocols, repeated and blinded trials, and statistical treatment. Holding the three apart is deliberate — it is how a reader knows which claims are specified, which are measured, and which are still ahead.
Labsintelligence is lab1 of Labs. We are a small research group working on speech‑to‑action, orchestration intelligence — speech‑to‑orchestration (STO) — and the interfaces that make both legible to a person.
Explore Research Roles →