Profila Sign up

Licensing

The shapes we think fit the data.

42-True Nothing here is an offer

Read this first

Proposed, not offered

No pricing is stated on this page and nothing here is a commitment. These are the shapes we think fit the data. We would rather agree the model with early licensees than impose one — and one of them, the training licence, depends on a permission that does not exist on any record today.

42-True

Model 1 of 4

Evaluation licence

For: teams wanting to measure an existing intent model against human-verified ground truth.

A fixed release snapshot, delivered as JSONL with its datasheet and signing key. Internal evaluation and benchmarking only; no redistribution; no derived-model distribution.

Why this exists: it is the fastest way for a buyer to establish whether the data is worth more to them than we claim, using their own systems and their own metrics. Often the fastest path to value is to use the corpus to measure before using it to train.

42-True

Model 2 of 4

Training licence

For: training or fine-tuning models on declared intent with verified relevance.

Rolling access to successive releases. Likely tiered by category and market coverage rather than by raw record count, because record count is the least informative dimension of this dataset.

This one is gated on something that does not exist yet. It requires the third-party training consent surface to be live and the relevant permission actually granted. We will not license training rights against records that do not carry them, and that is enforced in code rather than merely in contract.

42-True

Model 3 of 4

Demand-intelligence subscription

For: brands and analysts who want the unmet-demand and correction-chain signal rather than training data.

Periodic aggregate reporting: unfilled declarations by category and market, correction-chain resolution rates, category-level demand movement. Aggregates only, subject to the same k-anonymity floor.

Why it is separate: the buyer is different, the sensitivity is lower, and it needs no permission beyond what already exists. It also has the shortest readiness path of anything here — see For Brands.

42-True

Model 4 of 4

Benchmark and referee licence

For: supply-side participants who want to know how their inventory performs against declarant-verified relevance rather than against clicks.

Depends on supplier diversity, which does not exist yet — there is one search provider today. The value here is Profila as a neutral referee: the scoreboard is meaningful only because we do not own the inventory being scored.

42-True

Terms

Terms we would expect in any agreement

These follow from the architecture rather than from negotiating posture.

  • No attempted re-identification, and no joining of records across releases. The per-release pseudonym makes the second unavailable rather than merely prohibited.
  • Honour the erasure model. Because releases are rebuilds, a person who deletes their account is absent from the next release. A licensee holding an older copy is expected to move to current releases on a defined cadence.
  • Scope adherence. Records may only be used within the consent scope they carry.
  • Attribution boundaries. Named-supplier attribution in derived benchmark products is a specific right, not implied by a training licence.

Candidly, what we would want from an early licensee is feedback on which slices actually carry value. The datasheet tells us what we can measure; it does not tell us what is useful. Early agreements would be priced — when they are priced — with that exchange in mind.

42-True

Diligence

Diligence: what we will show you

We would rather be checked than believed. Available for inspection now:

  • The full release datasheet for any release, including the readiness verdict and every unavailable metric with its reason.
  • The signing public key, so records can be verified independently and offline.
  • The validation rule set, and the dead-letter count for any release.
  • The most recent scheduled re-validation report: what was checked, when, and every failure by rule.
  • A sample release under NDA, so the schema can be evaluated against real records rather than the illustrative one.
42-True

Diligence

Questions we expect, with the honest answers

How many records?

We are not publishing release counts until the corpus passes 100 contributors — and record count is the least useful number here anyway. Ask about contributor count and the readiness verdict (not licensable), which are the ones that bind. We will give you both directly.

Is it representative?

No. Not yet. A handful of contributors across a handful of markets — which is exactly why the figures are held back until there are enough of them to describe a population.

Can you prove it is anonymous?

We can show you the mechanism, the enforcement and an independent re-check. We would encourage you to have the claim tested rather than take it from us.

What happens when a user deletes their account?

They are absent from the next release. Not suppressed, not tombstoned — absent, because every release is rebuilt from current state.

What is the biggest weakness?

Contributor count. Everything else on our list is either engineering with a known path or a deliberate choice. Scale is the one thing that only time and product growth fix.

42-True

Start a conversation

We would rather agree the model with early licensees than impose one

What we would want in return is candid feedback on which slices actually carry value. The datasheet tells us what we can measure; it does not tell us what is useful.

Before any of that, read Readiness — it states exactly what is blocking a licence today, and why the answer differs depending on which product you want.

Email [email protected]