Methodology

Where the data comes from

Every decision in this archive is a real, published USCIS Administrative Appeals Office (AAO) non-precedent decision, downloaded directly from uscis.gov. We link the original PDF on every case page — nothing here is paraphrased without a citation back to the source document and page.

Coverage

Which cases get auto-cited as examples

Separate from the 2,369 cases you can browse and filter in the Explorer, a few places on this site pick a case automatically without you asking for a specific one: the AI research assistant’s “Related published decisions,” the signed-in dashboard’s default update feed, and one example card on the NIW page. Those draw from a narrower pool, gated on whether a case has enough real content to cite responsibly — not on the same 2,369 the Explorer uses.

The 543 excluded aren’t missing from the Explorer — you can still find and read every one of them there. They’re excluded from automatic citation because at least one of: the case’s subject-matter field is a single, possibly-ambiguous word rather than a clear phrase (380 cases — e.g. a real but terse value like “law”), the Prong 1 finding’s quoted reasoning is 60 characters or shorter, i.e. it states the conclusion without the reasoning behind it (109 cases), the Prong 1 quote has no recorded page number (85 cases), or the case has no extractable topic field at all (22 cases — some cases hit more than one of these). We check Prong 1 specifically because 91.9% of Prong-1-FAIL cases never reach Prong 2 or 3 at all — USCIS explicitly declines to address them — so requiring content from all three prongs would exclude the majority case, not a minority.

This pool used to be 16 cases (0.7%), gated on a per-document data-quality flag that turned out not to measure what its name implied. That flag has been retired from display decisions; it’s kept on record but no longer determines what gets shown.

Official USCIS-wide statistics, for context

These are not derived from this archive — they describe USCIS’s entire national filing population for these categories. Neither this data nor the archive above can estimate approval odds for an individual case.

EB-2 NIW approval rate

71% → 55.2%

FY2024 to FY2025, full-year USCIS data.

EB-2 NIW rate within FY2025

62.7% → 35.7%

Q1 to Q3, showing the decline wasn’t gradual across the year

EB-1A approval rate

66.9% → 41.7%

FY2025 full year to FY2026 Q2.

Cost to file

$1,015 gov’t + attorney

I-140 filing fee is fixed; attorneys commonly quote $6,000–$8,000 for a self-petition, more at established firms

Processing time

~24 months

Approximate, from USCIS processing-time data — varies by service center, check your own case type

A denial follows you

Must be disclosed

Refiling after a denial requires disclosing the prior petition on the new I-140

Sources: USCIS Immigration and Citizenship Data (approval rates) · USCIS Fee Schedule (G-1055) (filing fee) · USCIS Processing Times (varies by service center — check your own) · Form I-140 (prior-petition disclosure question). Attorney fee range is commonly quoted across immigration law firm sites, not a single official figure.

How it was processed

1. PDFs are downloaded once and archived locally.
2. An LLM extracts structured fields (dates, outcomes, legal findings, evidence types) — every field is required to come with a verbatim quote from the source document.
3. Each quote is checked against the source text with fuzzy string matching (to tolerate OCR noise), and scored HIGH / MEDIUM / LOW. Fields that don’t match are set to NULL, never guessed.
4. Every quote is separately matched against each page of the source PDF to find its page number, using the same fuzzy-matching approach. This succeeds for 98.2% of case-field citations and 95.6% of evidence-graph citations; the rest show “page not extracted” rather than a guessed page number.
5. Every decision is classified into a case type (NIW merits, NIW threshold-only, EB-1A, misrepresentation/fraud, misfiled) using statute citations and case-law references — because the raw USCIS archive folders mix these together (see below).

Known limitations

• Selection bias. This dataset only contains petitions that were denied and then appealed. It cannot be used to estimate an overall approval rate for NIW or EB-1A petitions, and it cannot predict the outcome of any individual case.
• The USCIS archive folders mix case types. The NIW archive directory also contains EB-2 cases that never reached the national-interest question, and a small number of misrepresentation/fraud disputes. We classify every case before it enters any statistic — see the case-type breakdown above.
• Publication history. We measured monthly decision volume directly rather than relying on secondhand reporting about a publication gap: volume was normal through April, then declined gradually from May onward — not an abrupt stop. A federal court order requires USCIS to resume full publication by 2026-11-30, so this archive’s coverage will keep changing.
• What’s not here yet. EB-1A decisions are archived but not yet searchable — we validate field-extraction accuracy for a classification framework before exposing it, and haven’t completed that for EB-1A’s ten-criteria framework. Topic pages by evidence type or legal element (e.g. “recommendation letters”, “national importance”) are also not live yet: we measured our category-matching accuracy at full scale and found it isn’t reliable enough to publish statistics from — roughly half of denial-reason labels didn’t match our category list on exact comparison. We’d rather ship these late than publish a wrong count of how many decisions discuss something.

Field accuracy

On a manually-audited sample, source quotes scored HIGH match confidence 96.6% of the time; every low-scoring and a random sample of high-scoring quotes were individually checked against the source PDF by hand, and none were found to be fabricated. Case-type classification was checked against source text on an 81-case audited sample with no errors found after one rule was corrected mid-review. Full detail on both audits is available in this project’s public methodology reports.

Correction (2026-08-21): the 96.6% figure above describes the audited sample used to design this scoring method. The per-record “confidence” value actually stored for each field turned out not to reflect that method when applied to the full archive -- checked across five unrelated fields on all 2,369 records, the value was identical within each record regardless of field, which means it was recorded once per document rather than assessed per field. We independently re-checked 3 quotes from records carrying that flag against their source PDFs byte-for-byte; all 3 were accurate. That field has been retired from every place it affected what gets shown (a “match confidence” label that used to appear on case pages has been removed, not just relabeled) and no longer determines what’s displayed. It doesn’t change the audit conclusion above -- the sampled quotes really were checked and found accurate -- it means that conclusion was never being correctly surfaced per record.