Tell us the input your model gets wrong.

Recorded calls between a caller and an agent, both sides on separate channels.

• Scripted scenarios or free-form, your choice of sector and outcome mix.
• Both channels recorded separately so a model can be trained on one side.
• Consent captured on the recording and stored with the file.

Unscripted speech from real speakers — the hesitations, restarts and overlaps scripts never produce.

• Prompted topics, no read text. Fillers and false starts are transcribed, not cleaned.
• Speaker metadata: age band, region, first language, device.

Targeted read and prompted speech to fill the accents your model drops.

• You name the locales; we recruit to a quota and report against it weekly.
• Code-switching sets available where the sector actually code-switches.

The same script recorded in a car, a warehouse, a call centre and a kitchen.

• Matched pairs, so you can measure degradation instead of guessing it.
• Far-field, hands-free, headset and handset variants of the same utterance.

Your audio, transcribed and speaker-attributed to a written standard.

• Two independent passes with a senior adjudicator on disagreement.
• Verbatim or clean-read conventions, versioned in the guide we agree on.

The tokens that carry the transaction — account numbers, amounts, drug names, SKUs, intents.

• Schema authored with your team, not pulled off a shelf.
• Gold tasks injected at 3–5% throughout the batch.

Side-by-side judgments on model replies: which one a customer would accept.

• Raters are sector operators, not general crowd workers.
• MOS and MUSHRA runs follow ITU-R BS.1534 procedure.

Callers who lie, push, spoof and hand the phone to someone else mid-call.

• Severity tiers agreed before the run; nothing is published without review.
• Includes replay, impersonation and authority-pressure attempts.

First-person video of people doing everyday tasks in real rooms, from a chest or head mount.

• GoPro-class or phone rig with a timestamped IMU stream.
• Task families and environments recruited to a quota, protocol agreed first.

Photographs to a spec — shelves, interiors, objects — on the devices and in the light your users actually have.

• Same subject across devices and lighting when you need matched pairs.
• Locales recruited the way we recruit accents.

Receipts, forms and IDs photographed the way people actually photograph them, redacted to spec.

• Synthetic or consented documents only — no scraped material.
• Field schema authored with your team.

Multi-turn conversations and instructions authored by people who work in the sector, in its register.

• Scenario briefs agreed first; authors pass the same calibration gate as annotators.
• Available in every live locale.

Transcripts paired with the audio they came from, or written to a scenario where no audio exists.

• Verbatim or clean-read, versioned in the guide.
• Entity and intent spans on request.