// For AI teams

How Lasker works.

How we source, what a delivery looks like, what it costs and what you agree to. Ready to ask? The request builder is on the home page.

// What you can request

Four kinds of data, one process.

Each maps to a stage of your training loop.

Evaluation

Expert evals

Agent output scored against a written rubric by credentialed reviewers.

Demonstrations

Workflow traces

Step-by-step expert work with the reasoning kept alongside each step.

Preferences

Preference and rubric data

Experts rank outputs and write the rubrics that define good work.

Post-training

Licensed data

Records from data owners, cleared by five rights checks before delivery.

Delivered as JSONL, CSV, documents or your own schema.

// How we source

From request to delivery in four steps.

Every step ends with something you can review. Nothing is billed before the pilot is accepted.

1

Scope

We confirm the capability, the schema, the volume and what your license must allow.

You

Share task and rubric

Lasker

Write the spec

Output

Signed-off spec

2

Source

Experts for evals and traces. Data owners for licensed sets, found through our data partnerships team.

You

Nothing needed

Lasker

Match and recruit

Output

Sourcing plan

3

Check and review

Rights record per dataset. Two credentialed reviewers per item, with agreement reported.

You

Review the record

Lasker

Five checks, QC

Output

Pilot sample

4

Deliver

Full set in your format, under the license you agreed, with a delivery log.

You

Accept the pilot

Lasker

Deliver in batches

Output

Dataset and license

// Sample delivery

Pipeline-ready, with its paperwork attached.

Each item carries its reviewers, their agreement and the rights record it came from. Your pipeline reads the fields. Your counsel reads the record.

evals.jsonl · 1 of 500illustrative
{
  "id": "ev_000184",
  "domain": "accounting.close",
  "task": "Reconcile bank statement to GL, March",
  "rubric": { "completeness": 4, "accuracy": 3, "controls": 4 },
  "score": 0.73,
  "reviewers": [
    { "credential": "CPA", "verified": true },
    { "credential": "CPA", "verified": true }
  ],
  "agreement": 0.86,
  "rights_record": "RR-SAMPLE-0001",
  "license": "non-exclusive / 24m / train+eval"
}
// For data owners

Five phases, each with an owner and an output.

The detail behind the six steps on the data owners page. Every phase ends with a document you keep. You can stop after any phase until the license is signed.

1

Discovery under NDA

We map what you hold and where it comes from.

You

Name one owner

Lasker

Run the inventory

Output

Data inventory

2

Rights check

Title, consent, PII, license terms and sanctions, checked and recorded.

You

Confirm title and consents

Lasker

Run five checks

Output

Draft rights record

3

Pilot release

A small de-identified sample, reviewed before anyone else sees it.

You

Approve the sample

Lasker

De-identify and review

Output

Approved sample and scan log

4

Licensing

Your counsel reviews the license and sets the buyer rules.

You

Counsel review

Lasker

Draft and negotiate

Output

Signed license

5

Delivery and oversight

Transfer, payment, buyer audits and deletion at term end.

You

Release approval

Lasker

Deliver and monitor

Output

Delivery log, deletion certificate

// Commercial terms

Priced per unit you can count.

Price is set after scoping and written into the order. The pilot comes first.

Per task

Expert evals, priced by credential, rubric depth and number of reviewers.

Evals

Per trace

Recorded workflows, priced by length and the expertise needed.

Workflow traces

Per license

Licensed datasets, priced by scope, term and exclusivity.

Licensed data

We do not publish rates until we have closed deals to base them on.

// For data owners with agents

Get value from your agents and your team.

If you already run agents or staff on professional workflows, their work is data too. Each option is separate and optional.

Expert evaluation of your agents

Credentialed reviewers score where your agent fails on real tasks, and why. You get the findings and a rights record for every item.

Ongoing

Trajectory licensing

Every agent task leaves a trail of steps. With your approval, qualifying trails can be licensed as workflow traces.

With approval

Staff demonstrations

Your accountants or analysts record how they do a task, step by step, as a project your company approves, under written terms.

Optional
// Terms

What you get, and what you agree to.

Clean rights run both ways. The seller's terms protect them; yours protect you.

What Lasker provides

Documented in the rights record and the order.

  1. Chain of title. Who created the data and who owns it.
  2. Consents on file. From contributors and owners.
  3. No personal data. Removed or transformed before delivery, with a scan log.
  4. Screened counterparties. Sellers and experts are sanctions-screened.
  5. Verified credentials. Each reviewer's license checked against its issuer.

What you agree to

Standard for every licensed dataset.

  1. Scoped use. Training and evaluation as named in the license.
  2. No re-sale. No onward transfer to third parties.
  3. Audit on request. The seller can verify use against the license.
  4. Deletion at term end. Certified in writing.
  5. Sanctions screening. Your company is screened before the pilot.
// FAQ

Questions from AI teams.

Can you source any data?

We run pilots in finance and accounting now and are sourcing legal. For anything else we reply with a sourcing plan or a clear no. We do not take work we cannot source with clean rights.

What does a pilot look like?

A small sample produced through the full process: spec, sourcing, rights check, two reviewers per item. You accept it before the full order starts.

How do you check reviewer quality?

Credentials are verified with the issuer before paid work. Each item gets two reviewers and we report their agreement. Items below the agreed threshold are re-reviewed.

Can we bring our own rubric or schema?

Yes. We deliver in your schema and score against your rubric, or help you write one.

Can we get exclusivity?

For licensed datasets, yes, for a defined term. For evals and traces written for you, the work is yours under the order.

Is Lasker delivering today?

The company is in setup and no dataset has been licensed yet. Send a request and we will tell you plainly what we can deliver now.

// Start

Where does your agent fail at real work?