Why publish a method instead of a number
Wealth-screening vendors have long marketed accuracy figures without a definition of what was counted, what the denominator was, or how the test set was chosen. Some publish no figure at all. Some prohibit customers from publishing evaluations of the service. We think the useful question is not "what percentage does the vendor claim" but "what would have to be true for this field to be wrong, and can I check it myself." This page explains how DonorGraph is built and tested so that you can.
The five commitments
| Commitment | What it means in the product | What it protects against |
|---|---|---|
| Cited evidence per field | Every asserted value links to its primary record: deed, Form 990 line, SEC filing, FEC or state campaign-finance line. | Black-box ratings that cannot be audited. |
| Identity resolution against public registries | Matches are made against IRS BMF/990, FEC, deed and SEC records using name, place, household, self-reported employer and organizational ties; the matching record is shown. | Same-name collisions, wrong-person property, mis-attributed board seats. |
| One-shot, held-out evaluation | A ground-truth set the system never trains or tunes on, scored in a single pass with no retries, field by field. | Overfitting to the test, cherry-picked reruns, headline scores that hide which fields fail. |
| Synthetic negative controls | Fabricated and deliberately mismatched identities are seeded into evaluation; a correct system must return nothing for them. | Systems that always produce a number, even for people who do not exist. |
| Refusal when evidence is thin | Fields say 'no evidence found' with the reason, rather than inferring a value from weak or ambiguous records. | Confident wrong numbers, which are more expensive than blanks. |
Where the data comes from
DonorGraph is built on primary public records rather than licensed consumer profiles. The core sources are the IRS Business Master File and Form 990 filings (organizations, officers, directors, compensation, grants), SEC EDGAR filings (insider holdings, Form D, ownership reports), federal FEC campaign-finance filings and state-level campaign-finance records, county deed and parcel records, and organizational registries used to connect people to boards and businesses. Each record type has a known structure, a known publisher and a known refresh cycle, which is what makes citation possible: a fact can be linked to a document that a researcher can open.
How identity resolution works
The hard problem in screening is not finding records; it is deciding which records belong to the person in front of you. DonorGraph resolves identity in layers. First, structured anchors: a name and location, an employer as self-reported on a campaign-finance filing, a household member listed as a co-owner on a deed, an organization the person is already known to serve. Second, corroboration across sources: a board seat on a Form 990 is more credible when the same organization appears in the person's giving or professional history; a property is more credible when the deed's co-owner matches a known household member. Third, conflict handling: when two candidates are plausible and the records cannot separate them, the profile presents the ambiguity rather than picking one. Each accepted match shows the supporting record, and a researcher can reject it in one click.
How we test
Held-out ground truth
We maintain a set of donor records with independently established facts. The system does not train on this set, and changes to matching or prompting are never tuned against it. When we evaluate, the set is run once, in one shot, with no retries or manual intervention; the output is compared field by field to the ground truth.
Negative controls
Alongside real records, the evaluation includes synthetic identities that do not exist and real names paired with the wrong location or employer. A screening system that emits capacity numbers for people who do not exist is measuring its own confidence, not the world. The negative controls tell us whether refusal is working.
Field-level review, not a headline score
A single accuracy number blends easy fields with hard ones. We review property, board service, securities, political giving and identity match separately, because they fail differently and are fixed differently. Where a field is weak, the product's default is to refuse rather than to guess until the method improves.
What we do not publish
We do not currently publish a percentage. The evaluation set is real but not yet large enough for a single figure to carry the meaning a buyer would attach to it, and we are not willing to publish a number without its denominator. When the set is large enough to support a figure with a stated method, we will publish both together. Until then, the test you can run yourself is on the trial page: three complete screenings on donors you know.
Correction is part of the method
Every field in a profile has a one-click correction. A correction is stored for your organization, applied when that donor is screened again, and fed back as a signal to identity resolution for similar cases. This matters because screening errors are rarely random: they cluster around common names, family members with shared property, and organizations with similar names. A correction loop turns each caught error into a rule.
What this means when you compare vendors
DonorGraph's answers are on this page and in the product. Pricing is published on the pricing page, and comparisons with DonorSearch, iWave and Blackbaud WealthPoint describe where each of them is stronger.
Frequently asked questions
What does 'cited evidence per field' mean?
Every value in a DonorGraph profile (a property, a board seat, a securities holding, a political contribution) links to the primary record it came from: the deed, the Form 990 line, the SEC filing, the FEC or state filing. If a field has no record behind it, it is not shown as a fact.
How does DonorGraph decide that a record belongs to my donor?
Identity resolution runs against structured public registries (IRS Business Master File and Form 990 officer lists, FEC filer records, deed and parcel records, SEC filings), using name, location, household members from deed co-owners, employer as self-reported on filings, and organizational ties. Each match carries the supporting record so a researcher can check it.
Why does a profile sometimes say 'no evidence found'?
Because the records did not support a conclusion. DonorGraph is designed to refuse rather than infer when evidence is thin: a common name in a large metro, a property that could belong to a same-named person, or a board seat that cannot be tied to the right individual. A blank with a reason is more useful than a confident wrong number.
How is the system tested?
Against a held-out set of donors with known ground truth that the model never sees during development, scored in one shot with no retries, and seeded with synthetic negative controls: fabricated or mismatched identities that a correct system must reject. Results are reviewed field by field, not as a single headline score.
Do you publish an accuracy percentage?
Not at this time. Our evaluation set is not yet large enough for a single number to be honest, and the category has a habit of publishing accuracy claims without denominators. We publish the method instead, and we let you run three real screenings before you buy.
What happens when I correct a field?
The correction is stored for your organization, applied to that donor's future screenings, and used as a signal for identity resolution on similar cases. It does not change what other organizations see.
- Primary record sources described above are public: IRS Business Master File and Form 990 (irs.gov), SEC EDGAR (sec.gov), FEC bulk data (fec.gov), state campaign-finance portals, county recorder and assessor records.
- This page describes method only. No accuracy percentage is published; see 'What we do not publish'.