Eli
What we solveHow it worksWhy EliSecurityTrust center
Log in
Open the demo

Where the boundary is, before you share data.

Every aid file holds student PII, and verification pulls in federal tax information. This page separates what the synthetic demo implements from what production still requires. The demo uses 100% synthetic data. Raw FTI is outside Eli's boundary. The full packet (FTI boundaries, AI governance, FERPA, GLBA, RBAC, HECVAT draft) is in the Trust center.

Data handling
In transitTLS 1.2+ (Vercel edge + Supabase)
At restProvider-managed storage encryption. Application-layer field encryption is not yet implemented.
Tenant isolationRow-Level Security in the repository schema, plus server-side institution scoping.
Access modelLeast privilege. Server-only service role. No client database access.
RolesAdmin / Reviewer / read-only Auditor
Case historyWorkflow and staff actions are recorded and exportable. Later changes do not erase earlier events.
Demo data100% synthetic. No real student, IRS, or FAFSA records.
Model card
Modelsclaude-sonnet-4-6 (via Anthropic / Vercel AI Gateway)
AI doesReads documents, narrates reasoning, drafts outreach, answers questions.
AI does NOT decideConfigured rules handle dollar calculations, the $25 trigger, all-or-nothing corrections, Per Computer precedence, and SAI.
Human in loopIdentity, conflicts, non-filers, PJ, and low confidence always go to a person.
Measured accuracy100% field accuracy on a synthetic corpus (n=52), 0 hallucinated fields (last eval 2026-06-09)
GuardrailA self-check runs before any autonomous correction. Flags go to a person.
Known limitsThe accuracy corpus is synthetic documents, not scanned IRS originals. SAI and Pell figures are directional until reconciled against the federal tables.

Accuracy is measured against a labeled synthetic corpus, summarized below. A scanned-document benchmark is in progress.

Regulatory posture
  • Intended FERPA role. A school official under the institution's direct control, with access limited to that purpose. The institution makes the designation and writes it into a counsel-reviewed agreement before any student record is processed. [34 CFR 99.31]
  • Raw FTI excluded. The demo does not ingest raw FA-DDX or transcript FTI. Any future tax-data scope needs a fresh control, contract, and legal review. MFA, field-level encryption, and independent testing are not represented as complete. [NIST SP 800-171 r3]
  • TPS status is scope-dependent. Human review and decision-support design do not settle third-party-servicer status. The institution and counsel classify the statement of work before an operational Title IV pilot. [34 CFR 668.2]
  • Records retention. Verification records are retained for at least 3 years after the end of the award year, then eligible for secure destruction. [34 CFR 668.24(e)]
Fraud & identity controls
  • Explainable demo scoring. Sample applications get a rules-based 0–100 score, and the page names the signals behind it. Production validation and monitoring remain open.
  • Identity-review workflow. Each demo review records a synthetic method, reviewer, date, and evidence state. No identity provider is connected, so nothing here establishes production IAL2 assurance. [34 CFR 668.57]
  • OIG referral packet. The demo prepares a synthetic evidence packet and a hash-chained register entry. It does not transmit a referral to OIG or claim an immutable external record. [34 CFR 668.16(g)]
Auditability

The demo keeps a replayable, hash-chained record of workflow actions, extracted values, reconciliations, reviewer decisions, and cited rules. The trace and a sample audit packet are exportable. A pilot still has to confirm the record is complete in production, that retention runs, and that an auditor accepts the packet.

Audit trail

A replayable record on every case.

What Eli read, the rule behind each result, and who signed. A conflict shows as a line routing to a person. Redacted and exportable.

case traceredacted / synthetic
✓identity verified · method=live-video · IAL2 · 34 CFR 668.57
✓extracted AGI $52,000 · source=IRS transcript (Per Computer) · conf 0.98
→conflict: amended return (1040-X) · routed to a human
✓SAI recalculated · configured rules · change logged
✓decision of record · approve & correct · by aid administrator
·retained per policy · 34 CFR 668.24(e)
Sample proof artifacts
AUDIT.PDFAudit package sample›TRACE.JSONReasoning trace JSON›EDE.TXTEDE correction sample›IDV.CSVV4/V5 identity CSV›
Measured accuracy
100%
field accuracy (n=52)
$0
dollar mean-absolute error
0
hallucinated fields
21
labeled fixtures in the corpus

Last run 2026-06-09, against a labeled synthetic corpus (clean and OCR-degraded renders, not scanned IRS originals). The signed-in demo shows the full eval detail and the live override rate.

Open the demoHow Eli works

Synthetic demonstration environment. No real student, IRS, or FAFSA records.

Eli

The Financial Operations Desk for financial aid offices. Software, or our team working it with you. Staff keep every sign-off.

The front door

  • Forms and guided intake
  • The hotline
  • Chat
  • Front desk assistant

Casework & judgment

  • Professional judgment
  • SAP appeals
  • Verification of income
  • FAFSA fraud & V4/V5 identity

Money movement

  • R2T4 returns
  • Schedule of Reductions
  • Fiscal close

Reporting & assurance

  • Annual reporting
  • Assurance
  • Aid dashboard
What we solveHow it worksWhy EliSecurityTrust centerThe 5-minute proofThe 90-day pilotLog in to the demo
Email pat@tryeli.comSynthetic-data demo. No real student, IRS, or FAFSA records. © 2026 Eli.