Profila
Sign up
What this dataset cannot do yet, in detail.
The verdict
Every release carries a machine-generated verdict that defaults to not-licensable and names each unmet criterion. Two separate questions, because they fail for different reasons.
| Verdict | Question | Current |
|---|---|---|
safe_to_publish | Is it safe to hand to anyone at all? | ✅ true |
licensable | Is it worth someone paying for? | ❌ false |
"NOT licensable: distinct_contributors, deep_funnel_records.
Privacy criteria pass; the gaps are scale and depth."
The privacy criteria all pass, on every record: timestamps coarsened, no free-text notes exported, every record signed and fully scrubbed, consent scope within grant. The gaps are scale and depth, not safety.
The flag is computed for the training buyer, and it is stricter than a brand product needs. Demand intelligence is blocked on one of the four criteria rather than all four, and that one is growth. See For Brands rather than reading one red flag as three different answers.
Limitation 1 — scale
The corpus is below its 100-contributor floor, and that is the single criterion everything else waits on.
Below roughly a hundred people, a corpus encodes individual habits rather than a population. Region and category distributions are dominated by a handful of contributors, and anything trained on it overfits them. This is a growth problem, not an engineering one — no amount of code moves it.
It is also why a large share of records currently have their region suppressed: with few contributors, most regions sit below the k-anonymity floor of 10. That reverses on its own as the corpus grows.
This is a demonstrably real but small dataset, concentrated in a few categories. Nobody should model population behaviour from it today — which is also why we are not publishing record counts, category entropy or per-category shares yet. Those figures are published once the corpus passes 100 contributors, and we will give you the current ones on request in the meantime.
Limitation 2 — depth
Rungs with no data: dwell, lead, conversion, verified_resolution. Five of nine rungs have fired. The top three are the differentiator, and they are empty.
| <code>conversion_source</code> | What it is |
|---|---|
admin_attested | Someone here recorded what a brand told them |
brand_confirmed | The brand said so, in their own verified account |
postback | Server-to-server confirmation — the only one that is measurement |
The mechanism now exists — a lead lifecycle with forward-only transitions, and a rule that a conversion must state how it is known.
Today only the first is reachable. The other two require brand-side account verification, which is real work not yet done. We report this split in the datasheet as measured versus attested rather than letting a buyer assume a conversion was observed.
Until these rungs populate, this corpus should be valued as declared intent with verified relevance — not as verified commercial outcomes.
Limitation 3 — measurement
Reported in every datasheet as a reason rather than a zero.
| Metric | Why not |
|---|---|
| Cohen’s κ (inter-rater agreement) | No second-rater track. Every verdict today comes from the declarant, so there is no disagreement to measure |
| Expected calibration error | A calibration map needs P(yes | confidence bucket) estimated per bucket, which needs far more verified records than exist. Publishing one now would be noise wearing a curve’s clothes |
| Supplier scoreboard | No supply integrations. One search provider today |
| Auction join rate | No auction — programmatic demand is disabled |
| Declaration uniqueness | Requires a corpus-wide pass; a one-of-a-kind declaration is more identifying than a common one |
| Compensation pool | No payout job yet; the ledger exists, the disbursement does not |
This table is the point, not an apology. "unique_suppliers": 0 and “we have no supply integrations” look identical as a number and mean opposite things, so we never write the number.
Coverage
not_evaluated rather than as rejections, because no judgement was made. Scope
Growth
Each verdict a user gives produces one record. There is no annotation step, no labelling vendor, no crowd-worker queue.
| Contributors | What becomes possible |
|---|---|
| ~100 | The scale floor for licensing. Region suppression relaxes as cohorts clear k=10 |
| Hundreds | Enough verdicts per confidence bucket to fit a real calibration curve. Cohen’s κ becomes measurable once a second-rater track exists |
| Thousands | Per-category and per-market slices become individually useful; the unmet-demand analysis becomes a genuine market signal rather than an anecdote |
The labels are produced by the people whose intent is being recorded, as a normal part of using the product — the only labelling process that gets cheaper per record as it grows rather than more expensive. A conventional labelled dataset costs more the larger it gets; this one costs the same per record at any size, and the marginal cost is approximately one database row.
Honest counterpoint: three things need work rather than time. Deep funnel rungs need brand-side verification. Supplier diversity needs more than one integration. Second-rater agreement needs an auditor sample track that does not exist.
Figures policy
Every release is measured — counts, rates, coverage and the readiness verdict are computed on every build. We are simply not putting the arithmetic on a public page yet, and it is worth saying why rather than leaving a gap.
| Status | |
|---|---|
| Record, declaration and contributor counts | Published at 100+ contributors |
| Unmet-demand rate, by category and market | Published at 100+ contributors |
| Correction-chain counts and resolution rate | Published at 100+ contributors |
| Counterfactual coverage and candidate volumes | Published at 100+ contributors |
| Category entropy and per-category share | Published at 100+ contributors |
| Every record declarant-verified | True today — not a sample |
| Every record signed, scrubbed and hour-coarsened | True today, enforced at export |
| Free-text notes exported | None, ever — by design, not by policy |
| Regions surviving below the k=10 floor | None |
| Last scheduled re-validation | ok: true — no rule reporting an error |
| safe_to_publish | true |
| licensable | false — blocking on contributor count and funnel depth |
Below roughly a hundred contributors, a corpus describes a handful of individuals rather than a population. A rate drawn from it would be precise and misleading at the same time — technically correct, and a poor guide to anything you would decide with it. Publishing one on an open page invites exactly that misreading.
Ask us and you get the numbers. Every figure above exists today and ships in each release datasheet; withholding them from a marketing page is not the same as withholding them from you. Under NDA you get the datasheet, a sample release, the signing key and the validation rule set.
Where a number cannot be computed at all, the reason is given rather than a zero. And if any claim on these pages cannot be reproduced from a release we hand you, we would consider that a defect in the corpus rather than in the document.
Check it
Every figure described on these pages is measured on each release, whether or not it is printed here. The full datasheet — including the counts we are holding back until 100 contributors — plus the signing key, the validation rule set and the latest re-validation report are all available for inspection. See Licensing, or just ask.
If you are a brand rather than an ML team, the blocking picture is different and shorter: For Brands.