Wave-2 freshness measurement: pre-registration
This document was written and frozen before the measurement ran. It fixes the cohort, the method, the metrics and the success thresholds in advance, so the result cannot be re-shaped after the fact.
What a pre-registration is
Before running a measurement, we state in public what we will measure, on which records, with which method, and what result would count as success. Because the cohort, metrics and thresholds are frozen in advance, the outcome cannot be quietly re-framed after the numbers arrive.
This wave is the out-of-sample test. The instrument was tuned on earlier data. Wave 2 tests it on changes that had not yet happened when it was built.
Questions under test
Precision, out of sample
When the detector says a contact changed company, how often is that true?
False negatives and recall
How many real changes does it wrongly call unchanged?
Decay rate
Confirmed change per month at a ninety-day horizon.
The cohort, frozen
The cohort is 1,361 records covering 1,346 unique profiles. It is our own data only, with no customer data. It was closed on 2026-07-17 and no record may be added or removed after that date.
5bb65af0f02897092d9e82bcf7efd1790037ec05f63ede4d8998cf45458124ad 9e5e06e201eb8d1d251acca7a98e81d7716b25c2574a3fcaacbb2937d9dbddf4 When records age past ninety days
| Date | Records aged past 90 days |
|---|---|
| 2026-08-18 | 485 |
| 2026-08-31 | 852 |
| 2026-09-08 | 1,014 |
| 2026-09-13 | all 1,361 |
Frozen instruments
The detector decides whether a contact changed company in a fixed order. None of this logic may change between now and the run.
- 1 Company-page link match first. If the old and new records point at the same company page, no change is claimed.
- 2 Name-matching cascade. Normalized company names, with acronym and brand-alias handling for entities that trade under more than one name.
- 3 Rename-relabel guard. Same normalized title and same start month and year is treated as a relabel of the same role, not a move.
One logged amendment (A1, 2026-07-18). A utility change to the tooling that does not touch how anything is classified. Frozen data was run through the tooling before and after the change and all outputs were byte-identical. The comparator itself is byte-untouched.
Run parameters and budget
We fetch recent-cached copies from a leading contact-data enrichment provider, identical to production semantics. The provider copy can be up to twenty-nine days old, so any change we measure is a floor on the true rate.
Fetching is paced at one third of our account ceiling, so live customer work keeps full headroom throughout.
Run window
Exclusions, stated in advance
These classes are removed mechanically before analysis. Every excluded count will be published.
- •Records we could not fetch after retries
- •Profiles that are gone, reported as their own class
- •Provider-lag payloads with no usable history
- •Records with no comparable company on either side
- •Duplicates, collapsed to one row per unique profile
- •Known data artifacts in company-name fields
Metrics and success thresholds
These decision rules are fixed now, before any result exists.
| Outcome | Rule |
|---|---|
| “Verified” wording unlocks | Precision point at least 90%, and the 95% lower bound at least 80%, and at least 30 flagged claims. |
| “High confidence” soft label only | Precision point 85 to 89.9%. |
| No accuracy wording | Precision point below 85%. We lead with the mechanical layer instead. |
| Claim count too thin | Fewer than 30 flagged claims. We report the numbers and defer any wording change. |
| Recall gate | Separate gate at 85% estimated recall. |
| Data too thin | More than 15% of reviewed rows unverifiable. We declare the data too thin and ship no external claims. |
How we will count
- •Wilson 95% intervals for all proportions.
- •Exact Poisson intervals for rates.
- •Rule-of-three upper bounds when zero events are seen.
- •No significance testing.
- •Every quoted number carries its interval.
Known limits, stated up front
Source freshness floor. Provider copies can be up to twenty-nine days old, so measured change understates true change.
Invisible multi-role moves. A person who adds a role while keeping the old one listed can look unchanged.
Cohort bias. A founder network with a UK and sales skew, not a neutral customer sample.
Short observation window. Annualized figures carry wide error bars and are stated as ranges.
Prior context
An earlier in-sample pass estimated about 96.8% precision. That figure is exactly why we are now testing the instrument out of sample, on changes it has never seen.
Commitment
We will publish the result whether it is strong, weak or null, alongside both precision and recall.