methods

three layers of processing

aysra takes filed IRS data and turns it into structured, comparable records. nothing is silently transformed. values are as filed or clearly marked as interpretive.

aysra operates in three distinct layers. each layer has a different level of certainty and is treated differently in the product.

structured

this layer reflects what can be read and organized directly from filings.

• filings are parsed into structured tables
• fields are standardized across forms and years
• grant records are normalized into comparable entries

these outputs are drawn directly from filings or computed from them in a defined way.

linked

this layer connects records that refer to the same entity.

• recipients are matched across filings
• filers and related entities are linked where identity is clear
• EIN relationships are resolved using consistent rules

not all matches are equally certain. aysra uses matching tiers and only promotes links when the evidence is sufficient. uncertain matches are not treated as resolved.

interpretive

this layer describes patterns across structured data.

• payout behavior
• grant concentration
• renewal and first-time giving
• narrative claims about funding behavior

these outputs summarize what the data shows. they do not introduce new facts.

interpretive outputs are marked with a kento symbol.

provenance

aysra makes a clear distinction between what is filed and what is interpreted.

• hanko marks values that come directly from filings or from defined calculations on those filings
• kento marks values that describe patterns or synthesize across multiple fields

if a value is not marked as interpretive, it can be traced back to a filing.

null is not zero

absence is treated as information.

• a blank field is not assumed to be zero
• a missing filing is not assumed to represent inactivity
• a behavior that does not appear in filings is not inferred

if a pattern cannot be observed, aysra will say so.

what aysra does not do

• it does not guess missing values
• it does not smooth across missing years
• it does not infer intent from mission language alone
• it does not present interpretation as fact

known limitations

aysra is built from public IRS filings. that makes the system traceable, but it also means the system inherits the limits of the filings themselves.

recipient names on grants are free text. the IRS does not require private foundations to report a recipient's EIN, so the same organization can appear under multiple name variants, such as "Harvard University," "Harvard Univ.," or "President and Fellows of Harvard College." aysra runs name normalization and uses a curated EIN override table to consolidate variants where the evidence is strong, but recipient matching is best-effort, not perfect.

tax-year lag is real. most organizations file months after their fiscal year closes, so the most recent tax year visible for any given filer usually trails the calendar year.

Schedule B donor names are redacted by the IRS for most filers. aysra does not have them either.

990-N postcard filers are not in the corpus. organizations with less than $50,000 in annual gross receipts generally file a short electronic postcard instead of a full Form 990, and the IRS does not release structured XML for those postcard filings.

boundary of inference

aysra's core principle is simple.

mission language and narrative text provide context. filings determine the underlying signal.

if a statement cannot be tied to structured evidence, it is not presented as fact.

for how aysra uses AI in its interpretive layer, see the AI use policy.