How 10-Dimension Matching Works
How 10-Dimension Matching Works
Most philanthropy platforms that claim to "match" a donor with an organization do it by computing a single number. You see "87% match" or "strong fit" and have no way to inspect the math underneath. When the donor asks the platform why — and donors do ask, constantly — the platform cannot answer. The composite score is a black box that was convenient to build and impossible to audit.
Catylst takes the opposite approach. Every recommendation surfaces the work across ten transparent dimensions. Each dimension is defined, measured, and cited independently. There is no composite score and no proprietary algorithm dressed up as insight. If a match is strong, the donor can see why on each axis; if a match is weak, the donor can see the gap. This essay walks through each of the ten dimensions, explains the zero skip-penalty rule that protects faith-data sharing, and closes with the five-layer validation pipeline that keeps AI-generated output honest.
The ten dimensions
Every profile on the platform — both donor and organization — is structured around the same ten axes. The axes are purposefully distinct so that a strength on one does not mask a weakness on another.
1. Mission alignment
Does the organization's mission statement describe work the donor wants to support? Mission alignment is measured by matching the donor's stated priorities (in their own words) against the organization's mission statement, program descriptions, and recent annual report language. When the donor has been explicit — "I want to fund youth mentorship in rural communities" — the alignment signal is crisp. When the donor has been broad — "education" — the signal surfaces a wider set of candidates and flags the ambiguity for the donor to narrow.
Worked example. A donor profile that names "workforce development for returning citizens" matches strongly against organizations whose mission language covers reentry services, apprenticeship programs, or post-incarceration skills training. It does not match a general workforce training organization without any reentry focus — because mission alignment measures intent, not proximity.
2. Faith alignment
For donors who choose to share faith fields, the platform records denominational tradition (Protestant, Catholic, Orthodox, non-denominational, interdenominational, or prefer not to specify), stewardship posture, and specific ministries or doctrinal areas the donor cares about. For organizations, the platform records the tradition the organization operates within, whether it welcomes donors from other traditions, and any doctrinal commitments donors should know about.
Faith alignment is treated as preference, not prescription. A donor can prefer Catholic-led organizations without excluding Protestant-led ones. A Protestant donor can choose to support Catholic relief work in Latin America. The dimension is designed to surface aligned organizations, not to build walls between traditions.
3. Cause
Cause is the broad category of work: education, health, housing, relief, economic development, arts, creation care, criminal justice, and so on. Causes are multi-select for both donors and organizations. An organization that works on both housing and health is surfaced to donors who care about either — without the double-counting trick some platforms use to inflate their match counts.
4. Geography
Two geographic signals matter. The first is the geography of the donor's interest: "I want to support work in the Pacific Northwest" or "I want to support work in Sub-Saharan Africa." The second is the geography where the organization operates — which is rarely a single point and often a complicated set of program sites, headquarters, and beneficiary regions.
The match surfaces overlap honestly. When an organization operates in three countries but the donor cares about only one of them, the match flags the partial fit. When an organization's headquarters is in New York but its beneficiary communities are in the Mississippi Delta, the platform prefers the beneficiary geography over the headquarters geography — because the donor's money is going to the program, not the HQ.
5. Grant size
Every organization has a grant-size range where it operates effectively. A local community organization with a $400K annual budget cannot absorb a single $2M gift without distorting its programs. A national organization with a $40M budget cannot deploy a $5K check efficiently.
Grant-size matching respects both sides. Organizations declare the ranges they accept. Donors declare the size bands they plan to give in. The platform surfaces matches where the intended gift lands inside the organization's comfortable range — and flags mismatches clearly, so a donor who wants to make a catalytic large gift sees organizations that can actually receive it.
6. Funding type
Funding type is the structural nature of the gift: unrestricted operating support, program-restricted funding, capital (buildings, equipment, technology), capacity building (hiring, systems, training), multi-year commitments, emergency relief, or matching grants. Donors and organizations both declare their comfort zones. A donor who prefers unrestricted multi-year support does not see organizations that only accept program-restricted one-time gifts, and vice versa.
7. Outcome and outputs
Outcome alignment asks whether the organization measures and reports on the same outcomes the donor cares about. A donor interested in literacy gains, for instance, is matched to organizations that track reading-level change, not just session attendance. A donor focused on housing stability is matched to organizations that track time-to-stable-housing, not just shelter bed-nights.
When an organization reports on outputs rather than outcomes — "we served 2,400 meals" instead of "participants sustained stable nutrition for six months" — the platform discloses the distinction and offers the donor a choice. Output-only reporting is not dishonored; it is just accurately described.
8. Population served
Who the work reaches matters independently of the cause and geography. A youth-serving organization in Appalachia serves a different population than a youth-serving organization in the Bronx; both deserve support; they are not interchangeable. Population fields include age bands, demographic descriptors the organization uses in its own reporting, and any specialized populations (veterans, returning citizens, unhoused families, newly-arrived immigrants, etc.) the donor or organization declares.
9. Trust signals
Trust is measured with seven trust badges that the platform issues based on publicly verifiable criteria: timely filing of IRS Form 990, completed annual audit, a program-expense ratio above the stated threshold, documented governance practices, outcome reporting in the annual report, board diversity disclosures, and a public history of sharing both wins and failures. Each badge has an explicit definition and a link to the source evidence.
Trust is not a proxy for quality — a young organization without a three-year audit history can still be doing excellent work — but it is a signal that helps donors understand the operational maturity of the organization they are considering.
10. Urgency and funding gap
The last dimension captures the current funding moment. Is the organization in the middle of an urgent campaign? Are they at risk of shrinking a program that has a waitlist? Have they just been awarded a challenge grant that requires matching dollars? Urgency is self-reported by the organization and verified against public communications (grant announcements, emergency appeals, board minutes where available). It never overrides alignment; it surfaces time-sensitivity as an additional input the donor can weigh.
The zero skip-penalty rule
Here is a design decision we have been asked about often, and we want to explain it carefully because it shapes the whole system.
A donor using Catylst can leave any field blank. Including every faith field. Including demographic fields about themselves. Including declared geographies or causes or grant sizes. The donor is always in control of what the platform knows.
When a donor leaves a field blank, that field does not contribute to the match on either side. The remaining dimensions still produce a match quality score, and the match quality is the average of the active dimensions — not a penalty against the blank ones.
This matters for faith data in particular. Faith beliefs are considered special-category personal information under GDPR Article 9 and analogous U.S. frameworks. Platforms that quietly degrade match quality when a donor withholds faith data are penalizing the donor for exercising their privacy rights. We refuse to do that. A donor who shares nothing about their faith receives matches computed from the other nine dimensions, at full quality on those nine dimensions. A donor who shares everything about their faith receives matches computed from all ten dimensions. Both paths are first-class.
The zero skip-penalty rule also extends to organizations. A young organization that has not yet completed a full audit cycle is not quietly downweighted in discovery because of a missing trust badge; the badge simply does not contribute. Donors can filter on trust badge presence if they want to, but the default view treats missing-but-honest data with the same posture as donor-side privacy preservation.
Cited reasoning per match
Every match surfaces its reasoning per dimension. If mission alignment is strong, the donor sees the specific words from the organization's mission statement that aligned with the specific words from the donor's priorities. If grant size is a partial match, the donor sees both ranges and the overlap region. If urgency is flagged, the donor sees the public communication that established it.
Nothing is editorialized into the match. The platform does not synthesize a "because this fits your values" paragraph that is not traceable back to the donor's declared values and the organization's declared fields. If the platform cannot show its work on a dimension, it discloses the gap rather than filling the gap with a plausible-sounding paragraph.
The five-layer validation pipeline
The matching math is deterministic. The AI-generated content that surrounds it — summary paragraphs, recommended-gift sizing notes, grant-package drafts — is not. We run every AI output through five validation passes before a donor sees it.
Layer 1: Advisory-phrasing removal. AI models are trained on content that contains financial, tax, and legal advice. Our models are instructed not to produce such advice, and the first validation layer scans output for advisory phrasing patterns ("you should," "we recommend you," tax-strategy assertions, legal claims) and either rewrites or flags them. Catylst prepares recommendations; it does not provide financial, tax, or legal advice. The first layer enforces that line.
Layer 2: Grounding verification. Every factual claim in the AI output is checked against the source material the model was given. If a draft paragraph asserts that an organization served "twelve thousand students last year," the layer verifies that claim against the annual report, the 990 filing, or whatever source was in the retrieval context. Claims that do not verify are removed or flagged.
Layer 3: Citation checking. Every citation emitted by the AI is checked for resolvability. If the model cites a page number in a PDF, the page must exist and contain the cited content. If the model cites a URL, the URL must resolve to a live page whose content supports the citation. Broken or hallucinated citations are removed before the output reaches the donor.
Layer 4: Sensitivity review. Faith context, politically charged issues, and potentially harmful framings are reviewed by a sensitivity classifier. The layer does not censor viewpoints; it ensures that faith commitments are represented accurately (Protestant organizations are not described with Catholic language, and vice versa), that ministries are not misclassified, and that no organization is characterized through a lens its own leadership would reject.
Layer 5: Confidence calibration. Every AI-generated paragraph carries an explicit confidence label — high, moderate, or low — based on the density and quality of citations, the age of the source material, and the model's own internal uncertainty. A donor never sees uncalibrated AI output. When the confidence is low, the paragraph says so plainly, and the donor knows to dig into the primary sources before leaning on it.
What this costs
Building a system this way is more expensive than building one that ships a single composite score. Ten transparent dimensions require ten separate data schemas, ten separate quality reviews, and ten separate UI surfaces for each match. The zero skip-penalty rule requires a matching engine that can evaluate against any subset of dimensions gracefully. The five-layer validation pipeline adds latency and compute cost to every AI response.
We believe the cost is the point. A donor making a decision with real capital deserves to see the evidence, and the platforms that have shaved that cost off their architecture have shaved the donor's agency off with it. If you want to see how this shows up in practice, walk through the matching flow or read more about the security and governance posture underneath.
Ready to see the dimensions in action for your own giving? Start with a guided match — no money moves, no card required. If you are a foundation or an advisor evaluating the platform for your clients, get in touch.
