Expert-labelled training data,
with a chain of title.

Public text ran out. Credentials did not. Clinical records, codebases, multilingual audio, robotics trajectories, contracts — every record with a named licence and the annotator’s credential attached.

2.9M Expert-labelled items delivered

9% Of applicants clear the calibration set

data@clayworks.ai

One record, unabridged

oncology_discharge_summaries/pass 2 of 3

Annotator: MD, internal medicine, 9 yrs

Patient is a 64-year-old presenting with persistent dyspnoea on exertion¹ over six weeks. Started on carvedilol 6.25 mg BD² after reduced ejection fraction³ was confirmed on echo. No chest pain reported. Plan: repeat echo in 12 weeks⁴, titrate as tolerated.

1Symptom, six-week onset
2Drug, dose, frequency
3Finding, confirmed on echo
4Follow-up interval

oncology_discharge_summariesp. 1 Name MRN Address DOB Date

spans marked 4 · blind second pass agree · adjudicated 1

14 JanSource holderMSA-0087
02 FebLicence executedLIC-2291
12 FebDUA countersignedDUA-2291-A
21 FebDe-identifiedDEID-4471
05 MarEnclave openedENC-118Keystroke audit KEY-118
19 MarDeliveredBATCH-4182

0 Day

Describe the failure, not the dataset.

48 Hours

Spec and quote, specialist rate on its own line.

7–14 Days

Pilot batch and quality report. Wrong spec re‑runs free.

Ongoing Delivery

Your bucket, your region, manifest attached.

Five domains, five gates

Licensed clinical records

Licence and DUA · 1.42M clinical spans

A1A2A3A4A5 A10.820.780.640.71 A20.820.860.580.69 A30.780.860.410.74 A40.640.580.410.62 A50.710.690.740.62

0.401.00

Pairwise Cohen’s κ · Fleiss κ 0.69 overall · A3×A4 adjudicated

Enterprise codebases

NDA and enclave · 92K reviewed diffs Reviewer: staff engineer, 11 yrs, Go/TS

--- a/internal/auth/admin.ts

+++ b/internal/auth/admin.ts

@@ -12,3 +12,12 @@ export async function canAdmin(user: User) {

1212 export async function canAdmin(user: User) {

13− if (user.role == 'admin') return true;

13+ if (user.role !== 'admin') return false;

14+ if (user.suspended¹) return false;

15+

16+ await audit.write({²

17+ actor: user.id,

18+ action: 'admin.check',

19+ at: Date.now(),

20+ });

21+

22+ return true;

1423 }

1Auth bypass, severity 3 2Audit write on the allow path

Multilingual audio

Consented speakers · 310K minutes, es-MX

no0.000–0.284 sabía0.284–0.742 que0.742–0.981 [pause]0.981–1.336 volvería1.336–1.918 [overlap]1.918–2.204

Robotics trajectories

Rig access · 44K validated episodes

X–Y work plane · 20 mm grid Tool height Z Gripper width Speed |v| Grasp · 220 ms dwell

01234 s

Finance and legal contracts

Bar-qualified review · 128K contract pages

MSA-0087 · schedule 4p. 12

4.1

4.2

4.3

4.4

4.1Term and renewalStandard

4.2AssignmentRestricted

4.3IndemnityFlagged

4.4Data useStandard

Rates

Recruited for the credential, tested before they are paid, re-scored weekly against the same gold set.

Expert annotationPer approved item, USD
Licensed clinical records	Discharge summary	34.00
Enterprise codebases	Reviewed diff	48.00
Multilingual audio	Verified minute	2.40
Robotics trajectories	Validated episode	12.50
Finance and legal contracts	Clause-mapped page	9.00
Licensed data supplyPer cleared record
Sourcing, clearance, de-identification	Cleared record	On request
Evals and red teamingPer scored set
Held-out set, refreshed every 90 days	1,240 items	On request