Holding is your personal immigration strategist for O-1A and EB-1A. The night before you pay a lawyer, it reads your case against USCIS's own instrument and hands you every criterion's cited verdict, the one link your whole case turns on, which "credentials" do no work on your record, and the questions to bring to counsel. Every line opens to the government's own page, and it never sells you a petition. It is bilingual, Chinese-first for its first market.
We are building the independent decision layer for high-skilled immigration: the standard that decides your future should not live only in the heads of lawyers, agencies, and adjudicators. Immigration is only chapter one.
Holding does not predict petition outcomes, determine eligibility, score a case, recommend what to file or argue, calculate deadlines, or draft legal submissions. It is not a law firm, creates no attorney-client relationship, and is not a substitute for the advice of an attorney.
The read leads with the criterion-by-criterion cited verdict, names the one necessary link the case turns on, and closes on the applicant's own ordered next move; the deliberate no-score stance is a below-the-fold boundary, not the headline. Implementation scope and safety contracts are unchanged.
Evaluating for Build Week? No API key, and no data download, needed. From a clean clone, run the three commands below, open the app, pick the EB-1A engineer/researcher fixture, and click Generate my strategist read. The two deidentified demo profiles ship with their reads precomputed (the read is deterministic), so the full cited-verdict experience renders from a clean clone with no secrets and no large corpus to fetch.
pnpm install
pnpm build
pnpm startOpen http://localhost:3000. (pnpm dev also works for local iteration.) The root journey contains two explicitly deidentified synthetic profiles. The EB-1A engineer/researcher profile is the demo-critical path: it separates originality from “major significance in the field,” reproduces the exact published USCIS deficiency taxonomy, cross-cites a page-pinned AAO passage, and quarantines planted paid-award and unverified-publication items.
Anonymous structured evidence stays in the browser and is processed only in-process by the app's own bounded endpoint. Nothing is sent to an external service, and there is no raw-document upload.
The demo runs from a clean clone with no setup. The verifiers below and the rebuild scripts operate on the immutable official source snapshot under data/ (a large build artifact, not committed); rebuild it with the pnpm *:sync scripts to run the full pipeline live. The shipped demo reads do not need it.
pnpm judgment:verify
pnpm verify
pnpm corpus:verify
pnpm authority:verify
pnpm structure:verify
pnpm dossier:verify
pnpm evaluation:verifyThe judgment verifier runs both deidentified profiles through the intake, proof chart, Step 1, Step 2, challenge-exposure, integrity, consultation, bilingual, exact-citation, and outcome-permutation contracts. Failed citation verification suppresses a claim; it is never replaced by an uncited summary.
The production-browser protocol lives at qa-reports/playwright_slice11.py. Against a production build it exercises the real EB-1A/O-1A submissions, focus and skip behavior, accessible names/tree, citation expansion, contrast, 375/768/1280/1920 reflow, a 200%-zoom proxy, print, loading, error, and double-submit states.
Next on the roadmap: applicant-value testing with real users, two-attorney calibration, and a production UPL review with counsel. These are the human checks that take a verified engine to a trusted product.
lib/corpus: immutable, official USCIS decision corpus and matter linkagelib/retrievalandlib/dossier: outcome-blind candidate ranking and post-lock researchlib/authority: exact regulation, Policy Manual, precedent, and federal-opinion sourceslib/structure: neutral criterion/evidence extraction with exact spanslib/judgment: controlled evidence inventory, proof charts, Step 1/Step 2, challenge exposure, integrity quarantine, sequencing, and consultation compilationapp/strategist-app.tsx: Chinese-first officer's-ruler read experienceapp/api/judgment: bounded, no-store transient compilation endpoint
The ranker can receive benefit, criterion, posture, regime/time, controlled evidence types, and controlled quality facts. It cannot receive disposition, criterion findings, source identity, or outcome-bearing prose. AAO source text joins only after the candidate set is immutable.
The founder supplied the product thesis, lived user context, legal-methodology constraints, and increasingly sharp falsification tests: prove the corpus first; distinguish adjudication research from outcome prediction; keep every claim on an official page; make abstention useful; and finally turn the research desk into the officer's ruler a real F-1/OPT user can carry into counsel.
Holding was built with Codex, running on GPT-5.6, throughout.
GPT-5.6 did the load-bearing reasoning: decomposing each USCIS criterion into the necessary links that decide it, mapping vague legal standards onto exact evidence structure, generating the published-deficiency taxonomy and the outcome-blind ranking, and writing the code. Codex ran the loop: reading the immutable corpus, planning each vertical slice, and pushing it through deterministic source, leakage, citation, bilingual, build, and security gates. Whenever the available AAO text supported less than the ideal claim, the product claim was narrowed to exactly what the sources prove, and the outcome firewall and citation gate were kept intact. The judgment layer was built on the verified corpus, authority, structure, dossier, and consultation foundations, never around them.
This collaboration preserves the frozen intent at hash 96ce7fdad027. Codex session (/feedback): 019f6945-6aae-7060-be97-83342cb0b690.