Company
Legal name, funding and aggregated petition fields exist. Original match provenance and fiscal-year denominators are missing.
Research-ready; not a willingness forecast.THE REALLOCATION ENGINE · RESEARCH & BUILD LAB
Sponsorship is a chain of different questions. Explore the tools already built, test assumptions, and help close the gaps between public records and useful decisions.
Educational research, not legal advice. No personal eligibility guarantees or employer promises. Check decisions with your school's international office or a qualified immigration attorney. No employer records or live listings are joined to these scenarios.
01 / ONE SHARED MAP
Read the flow in order. Each stage keeps its evidence and uncertainty attached. This is the intended architecture; the project cards below identify the pieces that actually run.
SEC funding · DOL filings · USCIS petition outcomes · linked job statements
Keep release, year, retrieval date and denominators.Legal entity → documented match → human review of ambiguity
Missing match stays unknown. Funding is context.Company history / role evidence / actual worksite
Headquarters and occupation averages cannot stand in for a role.Coverage · freshness · posting liveness · timeline assumptions
Missing evidence becomes a research task, not a zero.Historical summaries / uncalibrated rankings / explicit toy models
Show inputs and methods. No personal probability.Employer answer + qualified immigration review + your judgment
Selection, approval, status and visa issuance remain separate.These are distinct questions, not independent events to multiply. Not every pathway uses every stage. Petition approval does not guarantee visa issuance; a change of status and a visa are different processes. Department of State: temporary workers.
02 / OBSERVED FILE COVERAGE
of rows have any sponsorship field populated
The fresh recount on 2026-10-25 matches the May 2026 audit. This is an old mapped dataset: its newest funding date is September 26, 2025. The underlying government release years and original employer-match decisions are not recoverable from the file.
“Populated” means at least one of approvals, denials, approval rate, salary or sponsored titles is present. It does not mean the row is accurate, current or relevant to a specific job.
Inspect the exact source snapshot ↗Legal name, funding and aggregated petition fields exist. Original match provenance and fiscal-year denominators are missing.
Research-ready; not a willingness forecast.Some rows have sponsored job-title lists. There is no SOC field or validated link to a current vacancy in this snapshot.
No role-specific sponsorship rate.City and state describe headquarters. The separate BLS / BEA tool supplies occupation-metro wage context.
No worksite-specific selection estimate.METHODS / READ THE NUMBER BEFORE USING IT
1,557 ÷ 30,369
Describes how much of one historical file contains any sponsorship field. All rows were counted; this is not a sampled estimate requiring a sampling interval.
Unknown: accuracy, freshness and missingness bias.Inspect the data project ↓(0.35 × 1 + 0.30 × 1) × 1 × 1
The existing role scorer can return this with maximal sponsorship and fit inputs. It combines chosen weights and gates. It is not a 65% chance of sponsorship.
Unknown: whether a higher score predicts any defined outcome.Inspect the scorer project ↓250 slots ÷ 1,000 members
Correct for one equally weighted member under the stated fictional assumptions. Repeated random trials illustrate variation under that model.
Unknown: applicability to any real immigration process.Try the synthetic example ↓These are synthetic probes of the actual Engine CLI at 015843d, not employer or applicant outcomes. The source repository was left unchanged.
Actual scorer source ↗ · Tests and negative controls ↓ · Archived formula comparison ↓
Why no confidence band? Simulation noise, missing data, matching errors, model mismatch and changing rules are different uncertainties. More toy draws reduce simulation noise; they do not repair the other gaps. An honest real-world interval needs a defined target, justified sampling or statistical model, and evaluated data.
03 / CHANGE AN ASSUMPTION
These two educational tools are new to this page. They make the assumptions visible while the Reallocation Engine's full rule-aware calculators remain research tasks.
INTERACTIVE · SYNTHETIC INPUTS
Explore why the size of a pool matters. This toy model gives every member an equal chance in a single draw. It is not the current H-1B selection process and does not estimate anyone's immigration outcome.
Method: p = slots / pool. One equally weighted member; fixed slots; no duplicate beneficiaries, wage weighting, advanced-degree stage, employer decisions or adjudication. These are chosen scenario values, not observed USCIS data.
No simulation run yet. Each run repeats the same hypothetical conditions.
A verified rule version and effective date, fiscal-year beneficiary counts, wage-level / occupation / worksite distributions where applicable, advanced-degree handling, selection rounds and denominator definitions. We have not validated these inputs. Real-year forecasting is unavailable; no simulation here predicts FY 2027 or a later year.
INTERACTIVE · CALENDAR ONLY
Move a hypothetical milestone against a date window. The example dates are arbitrary. Being inside this window does not establish eligibility, lawful status or permission to work.
UTC calendar-day subtraction; boundaries included, elapsed days exclude the start day. Off-window markers stop at the edge. No OPT, STEM OPT, unemployment, registration-window or cap-gap rules are encoded. A date comparison cannot decide any of those.
Private by design: these inputs stay in component memory. They are not saved, uploaded or read from a fellow roster. Reloading resets them.
04 / DIFFERENT PATHS, DIFFERENT EVIDENCE
This is a research map, not an eligibility screen or an exhaustive list. A pathway's existence does not make it available to an individual. We do not rank routes by invented approval odds.
First establish which process applies. A selection model needs the applicable fiscal-year rules and population data. Cap exemption requires its own evidence.
Model state: no current-year forecast. Official indexed USCIS notices reference FY 2027 process changes; the full rule text could not be fetched in this audit, so no weighting or rates are encoded.
USCIS registration process ↗Research work authorization, program requirements, extension conditions, unemployment rules and any cap-gap provisions separately. The date tool above only subtracts calendar days.
Model state: no automatic extension or lawful-status conclusion. Detailed current rules remain unverified here.
USCIS OPT reference ↗O categories concern extraordinary ability or achievement. L categories concern qualifying intracompany transfers. Evidence about accomplishments or corporate relationships differs from employer sponsorship history.
Model state: no eligibility classifier or probability model found. Detailed criteria need case-specific review.
State Department category overview ↗TN concerns Canadian and Mexican citizens in qualifying USMCA professions; E-3 concerns Australian nationals in specialty occupations; H-1B1 concerns Chilean and Singaporean applicants in specialty occupations. Nationality alone is insufficient.
Model state: comparison research only. No pathway calculator found in the audited Engine tree.
Official overview pages reviewed 2026-10-25. USCIS and DOL full-page fetches returned access errors during this audit; references remain available for human review. No fees, deadlines, unemployment limits or current lottery rates are asserted here.
05 / BUILD ON WHAT EXISTS
There are 10 workstreams below, not 10 finished calculators. Status distinguishes code, tests, historical data and proposals. Every card includes a useful next contribution and versioned source.
Choose a small component, use fictional fixtures, include a deliberate failure test and record what you did not test. The repository defines contributor namespaces and review requirements. No individual maintainer or invitation link is assumed.
The original SEC funding → DOL / USCIS employer-history project. The surviving mapped CSV has company, headquarters, funding and aggregated sponsorship fields. It is a foundation for research, not a live employer directory.
Inputs → method → output
Presence audit: populated rows ÷ all rows. Approval_Rate is an inherited aggregate with no recoverable fiscal-year denominator in this file.
Checked: Recounted all 30,369 rows. 1,557 have any of five H-1B fields; 28,812 have none. SHA-256 matches the May audit. Latest funding date: 2025-09-26.
Cannot conclude: No raw LCA / USCIS inputs, fiscal years, SOC, worksite, match method or match score in this snapshot. Historical copy paths are not extra datasets. Missing fields do not mean no sponsorship.
Recover source releases and join lineage; separate initial and continuing petitions, filing years and worksites before showing employer comparisons.
Chapter 7 proposes combining filing activity, petition approvals, funding recency and company size. Its illustrations explain an intended ranking method; no maintained end-to-end implementation of this formula was found.
Inputs → method → output
0.40 × filing signal + 0.30 × approval signal + 0.20 × funding recency + 0.10 × size signal.
Checked: Chapter, design document and illustrations inspected. The design draft proposes Proven ≥ 0.65 and Likely ≥ 0.35; feature normalization and outcome calibration are unspecified. No predictive validation found.
Cannot conclude: The manuscript calls the result P(sponsorship), but a weighted sum is not an estimated probability. Funding and filing history cannot establish current willingness for a particular role.
Define the target outcome and denominators, preserve unknowns, reconcile the chapter’s tier warning with the four-tier design draft, and design held-out calibration before predictive claims.
Matches SEC entities to supplied LCA employer records, then records the method used. A separate script prepares a deterministic review sample from populated sponsorship rows.
Inputs → method → output
FEIN exact → normalized name exact → fuzzy match at threshold 0.88 → unknown. Sample defaults: 200 rows, seed 80, first 30 for manual review.
Checked: Synthetic FEIN, normalized-name and unmatched cases passed with the difflib fallback. The review script prepares a packet; it does not validate identity against government records.
Cannot conclude: Raw source tables are absent. False-match rate is unknown. rapidfuzz and difflib use different matching methods. Exact-name collisions still require review.
Add ambiguous-parent and subsidiary fixtures, preserve candidate matches, measure precision/recall on independently labelled entities and repair obsolete default input paths.
Consumes supplied role-evidence JSON and optional profile data, then writes Apply / Consider / Skip and a per-term audit trace. It does not collect the components or fit a Bayesian model.
Inputs → method → output
(0.35 × sponsorship input + 0.30 × fit input + 0 × role quality) × liveness × timeline. Apply threshold 0.30; Consider floor 0.20, with additional demotions and overrides.
Checked: Five-row chapter fixture rerun: 2 Apply, 1 Consider, 2 Skip. Current and Summer archive scorer blobs are identical. No hold-out outcome validation found.
Cannot conclude: Missing gates default to 1. Out-of-range synthetic inputs yielded 5.2 and Apply. Role quality has zero weight; profile authorization uses broad text matching. These are engineering defects / assumptions, not measured odds.
Validate input ranges, make unknown gates explicit, replace authorization substring matching, and agree on weights and outcomes before advising use.
The system design document and plain-language summary predate the implemented chapter scorer. They are earlier attempts at the same system, not separately deployed calculators.
Inputs → method → output
The design document writes a product of four inputs and weights 0.35, 0.30, 0.20 and 0.15, with Apply ≥ 0.45 and Consider ≥ 0.25. Its overview uses a simpler product; the later CLI instead sums weighted votes and multiplies two gates.
Checked: Read the archived design formulas. With inputs bounded by 1, the literal weighted product cannot exceed 0.00315 and can never reach either threshold. No implementation of that literal product was found.
Cannot conclude: The proposals conflict internally and with the shipped scorer. Neither multiplication nor a Bayesian label establishes independence, calibration or a personal success probability.
Document a single target quantity and retire contradictory formulas through review; do not port the archived expression into a public probability calculator.
One harness invokes the actual scorer CLI and mutates a temporary copy as a negative control. The other implements its own scoring functions as a teaching demonstration.
Inputs → method → output
Strong votes must not override closed liveness or timeline gates. Test normal arithmetic, threshold edges, source labels and deliberate additive-gate failures.
Checked: CLI harness: 10 baseline cases passed and all 5 gate-sensitive mutant cases failed as expected. Copied-arithmetic harness: 22 checks passed across 5 examples.
Cannot conclude: The copied harness does not prove the shipped scorer behaves correctly. Both use synthetic cases and establish no predictive accuracy.
Extend the CLI tests for absent, malformed and out-of-range inputs; consolidate duplicate teaching logic without losing the negative control.
Chapter 10 sketches authorization, hiring and registration windows. Four timeline figures and the Chapter 7 / 11 visuals are explanatory artifacts, not individual eligibility engines.
Inputs → method → output
Compare supplied dates and route a timeline constraint into the role scorer. The current scorer accepts a factor; it does not calculate it from visa dates.
Checked: Chapter and four timeline HTML files inspected. No executable rule-aware date calculator or current cap-selection simulator found in the audited tree.
Cannot conclude: Some chapter language is overly categorical. Extensions, cap-gap, status and permission to work require separate, current review. The manuscript cannot serve as a legal rules database.
Build versioned rule fixtures with expert review, including boundary dates and exceptions. Keep dates on-device and show unknown rules explicitly.
The Summer case set covers backend, data/ML, full-stack, NLP, post-offer sponsorship audit, OPT runway, STEM OPT, targeting and career pivots. These are workflow specifications sharing a template, not 14 working calculators.
Inputs → method → output
Source → scope → data shape → script readiness → approval → report gates, with named input/output contracts still containing TODOs.
Checked: All 14 relevant Summer case files are DRAFT, each with 13 or 14 declared open TODOs. Referenced component scripts and anonymized sources are not supplied by those recipes.
Cannot conclude: No end-to-end verified run for these case variants was established. A recipe title does not prove that its implementation exists.
Choose one narrow recipe, define its schema, implement one component in the contribution namespace, and add a fictional worked run plus deliberate failure case.
Proposes comparing LLM-inferred occupation and wage level from job text with filed LCA values, then studying employer variation and petition-denial associations.
Inputs → method → output
Proposed mismatch statistics, employer-level random effects and regression. The draft leaves results, model details and validation values as TO DO.
Checked: Draft inspected; empirical results and a runnable pipeline were not found. Its results-shaped prose is not evidence of completed research.
Cannot conclude: A model disagreement is not proof of legal misclassification. No measured error rates or causal interpretation are supported.
Define a lawful text corpus, build blinded expert labels, test classifier error and confounders, and replace placeholder claims with measured findings.
A separate BLS / BEA tool compares occupation-metro wages with local price levels. This supplies defensible location context, not a worksite sponsorship rate or a DOL wage-level assignment.
Inputs → method → output
Adjusted wage = nominal wage ÷ (regional price parity / 100); exact metro-code join. Missing rows remain missing.
Checked: Source inspected; DOMAIN.md records a historical sample with 100 of 112 occupation-metro pairs available and 12 missing. The cited generated report is absent in this checkout; that sample was not independently rerun.
Cannot conclude: Occupation-metro survey estimates do not describe a company offer. The role scorer currently gives role quality zero weight.
Keep survey vintage and suppression visible; do not use cost-of-living adjustment to infer H-1B wage levels or selection weights.
06 / THE AUDIT TRAIL
Download the ten project records, fourteen case links, source commit, coverage fingerprint and reproducible toy examples as JSON. This fixed public file contains no visitor inputs or employer-level records.
eccdee2addf472b1639269f42eec693b083b7ce251347d5fd0b2856cfdae6270 · SHA-256 of the surviving mapped CSV. Only aggregate coverage is shown here.