Provenance and rights.
The institutional buyer's second question, answered before it is asked: where the data came from, who made it, what each record carries, what you are licensed to do with it, and the short list of things this business will not do because any one of them would end it.
Public and licensed, with the licence on the record.
"Publicly available" is not "free to train on". Every source in a pack has a rights profile that says what it permits.
Labs draws on three kinds of source. Public disclosures: filings, regulatory releases, exchange notices and the documents issuers publish themselves. Licensed data: fundamentals, estimates and prices from providers whose terms permit the use the pack makes of them — recorded on each record as a machine-readable rights profile, the same one Quantify Terminal enforces at read time. Expert-created work: tasks, rubrics, gold outputs and trajectories produced by contracted practitioners under agreements that assign the rights to Labs.
Where a source permits derived use but not redistribution, the pack carries the derived record — the stamped figure, the span reference, the hash — and a pointer to the source, not the content. Where a source permits neither, it is not in the pack. A computed figure inherits the strictest rights of anything it was computed from.
Never in a pack: Quantify Terminal client data, client research or private documents, under any anonymisation; material non-public information; scraped content whose rights cannot be shown; market microstructure data.
Former and independent practitioners, screened and on a cooling-off period.
The people who author tasks are the people who have done the work. They are not the people currently doing it for a competitor of yours.
Task authors and reviewers are contracted chartered accountants, CFA charterholders and former sell-side and buy-side analysts. Each is screened for conflicts before engagement and before each assignment, and each observes a cooling-off period from any issuer they covered, held or advised. No one authors or reviews a task on an issuer they currently cover or hold.
Authors work from public and licensed material inside a controlled environment. They do not bring documents of their own, and they attest to that on every task. Every task has a second expert's review; where the two disagree, the disagreement is recorded with both positions rather than averaged away, and the task is either resolved by a third or withdrawn.
Authors are paid for their time and judgement at practitioner rates, and are credited in the pack by a stable identifier, not by name, unless they choose otherwise. How to work with us as an expert →
Nothing in a pack is unattributed.
Every record says where it came from, when it became true, and who made it.
A document carries its publication time, its effective period, its source and the SHA-256 of the stored copy. A task carries its as-of, its author and reviewer identifiers, its rubric and the engine run that produced its ground truth. A trajectory carries its author, the task it answers, the datapoint each number was taken from, and the rubric it was scored on. Verifier output carries the verifier version, so a score can be reproduced against the version that gave it.
Personal data is limited to what the authors have agreed to: an identifier and the credential that qualifies them. Screen recordings are made in a provisioned environment with no personal accounts, so nothing in a trajectory can identify anyone who did not consent to be there.
What you receive, and what you do not.
A draft is in the sample pack. The final agreement is negotiated; the shape below does not change.
- Granted: a non-exclusive licence to use the pack to train and evaluate models and agents, with the rights profile of every record stated.
- Per-record: where a third-party source permits training but not redistribution, you receive the derived record and the pointer, and the licence says so on that record.
- Held-out sets: delivered under terms that forbid training on them, forbid publication of the tasks, and set a retirement date after which the set is rotated out.
- Not granted: redistribution of the pack or any part of it; resale; use of expert-created content to train a product that competes with Labs' own data business.
- Warranties and indemnity: scoped to provenance — that each record is what its stamps and rights profile say it is. The scope and the caps are in the agreement.
Sandboxed documents, versioned verifiers, no live systems.
An environment is a sealed set of documents and a deterministic checker. It touches nothing of yours and nothing of ours.
Datasets are delivered as files with checksums. Environments expose a tool API over a sandboxed copy of the documents a task permits, and a verifier callable as a service; neither reaches into Quantify Terminal's production systems or into yours. Verifiers are versioned, and every score names the version that produced it.
Independent security attestation for Labs' own systems is on the roadmap and will be stated here when it exists, not before.
The short list.
Any one of these would end the business. None is worth a contract.
- Put a Quantify Terminal client's data, research or documents into a pack, anonymised or otherwise.
- Author tasks with people who currently cover, hold or advise on the issuers in them.
- Include material non-public information, or content whose rights cannot be shown.
- Publish a held-out test set, or deliver one without terms that forbid training on it.
- Score with a judge model, or print a verifier error rate that was not measured.
- Resell market microstructure or build labelling operations at commodity scale.