# Askable Labs — Capabilities (AI-readable)

> Askable Labs captures four data types that frontier models can't generate for themselves, from real people doing real work, produced to a bespoke brief and not picked off a shelf. The four are listed in the order the site presents them. Human page: [/data/](https://labs.askable.com/data/). Machine-readable: [/agents/catalogue.json](https://labs.askable.com/agents/catalogue.json).

Generated: 2026-07-21

## How to use this (for agents)

Match the data type to the signal the model needs to train:

- agents / SFT / tool-use / multi-human coordination / long multi-step work → **Long Horizon Captures**
- voice / ASR / TTS → **Conversational Audio**
- reward models / evals / preference / ranking → **Evals and Preference**
- embodied / video / world models → **Physical Skill**

Each type below lists what it trains and a few example captures. The example captures are public sample pages; the full pattern library is internal, and every engagement is scoped to the lab's brief, not assembled from a published list. To scope one, see [/agents/engage.md](https://labs.askable.com/agents/engage.md). Delivery schema, capture stack, and provenance: [/agents/catalogue/keys.md](https://labs.askable.com/agents/catalogue/keys.md).

## Long Horizon Captures

Complex multi-step work over hours or days.

- **Trains:** Agents, SFT, Multi-agent, Tool-use
- **What it is:** The most capable humans navigating the most complex real world multi-step tasks over hours or days. Solo or multi-human with handoffs, reasoning, generated assets and every action captured.
- **Detail:** Solo or multi-human, screen + reasoning, timestamped.

Example captures:

- [Expert solo task with think-aloud](https://labs.askable.com/sample/?id=s-038)
- [Complex booking and coordination across systems](https://labs.askable.com/sample/?id=s-041)
- [Real human use of AI tools embedded in workflow](https://labs.askable.com/sample/?id=s-044)
- [Expert demonstrates, user watches](https://labs.askable.com/sample/?id=s-001)
- [Expert teaches by deliberate failure](https://labs.askable.com/sample/?id=s-003)
- [Expert troubleshoots a user who is already stuck](https://labs.askable.com/sample/?id=s-005)
- [Multi-expert debate on a contested question](https://labs.askable.com/sample/?id=s-023)
- [Adversarial expert pairing (red team / blue team)](https://labs.askable.com/sample/?id=s-028)
- [Brief writer to executor](https://labs.askable.com/sample/?id=s-031)

## Conversational Audio

Multi or single human conversations.

- **Trains:** ASR, STS, TTS, Voice
- **What it is:** Full-duplex audio: overlap, backchannel, emotion, dual-channel isolation, paired with on-screen context.
- **Detail:** Multi-language, dual-channel, 48 kHz.

Example captures:

- [Solo spontaneous non-native speech](https://labs.askable.com/sample/?id=ca-solo-nonnative)
- [Code-switched bilingual speech](https://labs.askable.com/sample/?id=ca-codeswitch)
- [Incentivized dual-human negotiation](https://labs.askable.com/sample/?id=ca-negotiation)
- [Contested-topic multi-human debate](https://labs.askable.com/sample/?id=ca-debate)
- [Human-AI interface conversation](https://labs.askable.com/sample/?id=ca-human-ai)
- [Expert-novice knowledge transfer](https://labs.askable.com/sample/?id=ca-knowledge-transfer)

## Evals and Preference

Human preference and decision data.

- **Trains:** Reward models, Evals, Preference, Ranking
- **What it is:** Human judgment at scale: preference rankings, side-by-side comparisons, rubric-based grading and red-team annotation, from verified, demographically precise participants.
- **Detail:** Structured labels, quantifiable, named experts.

Example captures:

- [Expert grades and ranks user attempts](https://labs.askable.com/sample/?id=s-016)
- [Expert rewrites the user's approach](https://labs.askable.com/sample/?id=s-017)
- [Multiple experts review the same user session](https://labs.askable.com/sample/?id=s-019)
- [Expert reviews and predicts what happens next](https://labs.askable.com/sample/?id=s-020)
- [Without-AI versus with-AI comparison](https://labs.askable.com/sample/?id=s-021)

## Physical Skill

Real-world tasks performed on camera.

- **Trains:** Video models, World models, Embodied
- **What it is:** Real-world tasks on camera: technique and outcome captured. The embodied signal you can't scrape or generate.
- **Detail:** Multimodal, on-camera, technique + outcome.

Example captures:

- [Expert-novice walkthrough](https://labs.askable.com/sample/?id=s-063)
- [Skill demonstration](https://labs.askable.com/sample/?id=s-064)
- [Remote coaching session](https://labs.askable.com/sample/?id=s-065)
- [Sign language exchange](https://labs.askable.com/sample/?id=s-066)
- [Guided physical task](https://labs.askable.com/sample/?id=s-067)

## Briefs are bespoke

The captures above are examples, not a menu. Every Labs engagement is scoped to the lab's brief: domain, population, tool stack, language, participant credential floor, session length, modalities, volume, and the consent scope the lab needs. Scope one via [/agents/engage.md](https://labs.askable.com/agents/engage.md) or [/contact/](https://labs.askable.com/contact/).