Statutory deductions
PAYE is recalculated against the current SARS tax tables for the stated income and period. UIF is checked against the prescribed rate and ceiling. A fabricated payslip almost never gets both right.
PAYE R 1 120.00: off tax table
Tycho · Document verification · An ElevateIQ venture
Something is wrong with this document
Tycho is a document verification service for South African payslips and bank statements. It finds the edits, the fabrications, and the numbers that do not add up, and shows you exactly what it found.
TYCHO / TMA-1 ANALYSISDOC-0417
- DOCUMENT
- PAYSLIP / MAR 2026
- GROSS
- R 28 400.00
- NET
- R 31 200.00: EXCEEDS GROSS
- PAYE
- R 1 120.00: OFF TAX TABLE
- UIF
- R 0.00: MISSING
- GLYPHS
- TWO TYPEFACES IN AMOUNTS
- METADATA
- MODIFIED AFTER ISSUE
FRAUDULENTRISK 87/100
Documents Payslips · Bank statements
Hosting Azure South Africa North
Status In development
Detection
What it actually checks
Generic document tools were built for other countries' paperwork. Tycho is built around South African payslips and bank statements, and it tests them against rules that must hold.
Every finding is a named check you can audit: recalculated tax, reconciled balances, measured typography. Not a language model’s opinion about your document.
Arithmetic that must balance
Gross minus deductions must equal net. Statement running balances must reconcile line by line. Year-to-date figures must agree with the period. Editing one number breaks the others.
NET R 31 200.00: exceeds gross
Visual tampering
Amounts retyped in a PDF editor rarely match the surrounding text. The engine compares glyphs, baselines, spacing and alignment across the page, and detects pasted or cloned regions.
Two typefaces in amounts column
File forensics
The document's own structure is inspected: producer software, creation and modification history, incremental saves, and object-level edits left behind by common PDF editors.
Modified after issue date
Template origin
Fabricated documents usually come from a small pool of online templates and generators. Layout fingerprints are compared against known employer and bank formats, and against known fakes.
Layout does not match issuing bank
Consistency across a series
Three months submitted together must agree with each other: closing balance to opening balance, salary credits to payslip net pay, and employer references throughout. Perfect agreement is its own signal.
Closing balance ≠ next opening balance
Seven layers · 34 named checks
Structure · Typography · Pixels · Numbers and statute · Cross-document · Issuer · Anomaly
Sequence
Six stages. One verdict.
A forgery that survives one stage is caught by the next. Each layer looks for a different class of fraud, and the score at the end shows its working.
- STAGE 01
Triage
Identifies the document and routes it: payslip, bank statement, or something dressed as one.
- STAGE 02
Forensics
Reads the file itself: creation history, editing traces, and structural tampering the eye cannot see.
- STAGE 03
Extraction
Pulls every field into structured data, so the numbers can actually be tested.
- STAGE 04
Vision
Studies the page as an image: font substitutions, misalignment, pasted regions, cloned areas.
- STAGE 05
Rules
Tests what must balance: net pay, deductions, PAYE and UIF, templates, consistency across a series.
- STAGE 06
Scoring
Weighs the signals; conclusive findings stand alone. No black box.
REST API and review dashboard. Upload a batch, get a verdict with the evidence attached, typically under a minute for a digital PDF.
Honest scope
What Tycho does not do
Any vendor who says their tool catches everything is selling you a problem for later.
- It does not decide
- It returns a risk score and the findings behind it. Approving or declining stays with your team.
- It does not verify employment
- It tests the document, not the employer. Whether the company exists and employs the person is a separate check.
- It does not query your bank
- There is no open-banking connection. Findings come from the documents you submit.
- Photographs weaken it
- A phone photo of a printed payslip strips the file forensics layer. Original PDFs give the strongest result, and the report says so when they are missing.
- A clean score is not a guarantee
- A well-made forgery can pass. The score reflects what the engine found.
Data handling
What happens to the documents
Your documents are never used to train anything.
Stays in South Africa
Processing and storage run in the Azure South Africa North region. Documents do not leave the country.
Deleted on your schedule
Retention is configurable, down to deleting the source file the moment analysis completes. Only the report is kept if you choose.
Auditable by design
Every check, score and access event is recorded, so a report can be reconstructed and defended months later.
Questions
Straight answers
Is this a language model reading our documents?
No. Every finding is a named check with a code: PAYE recalculated against the SARS tables, gross minus deductions against net, running balances that must reconcile. Vision is one of six stages, and it never decides alone.
How would we integrate it?
REST API and a review dashboard. Upload a batch, get a verdict with the evidence attached, typically under a minute for a digital PDF.
What comes back?
One of four verdicts, Authentic, Suspicious, Unverifiable or Fraudulent, with a risk score and the findings behind it. Each finding cites its check by code.
Which documents does it handle?
South African payslips and bank statements. Built around this country's paperwork rather than adapted from somewhere else's, which is where the statutory checks come from.
What happens to the documents we send?
Processing and storage run in Azure South Africa North, and they do not leave the country. Retention is configurable, down to deleting the source file the moment analysis completes. They are never used to train anything.
We only have phone photographs. Does it still work?
Partly, and the report says so. A photograph of a printed payslip strips the file forensics layer entirely. Original PDFs give the strongest result.
Does it approve or decline?
No. It returns a risk score and the findings behind it, and the decision stays with your team. It does not verify employment or query anyone's bank either: it tests the document, not the employer.
Are you running with clients yet?
Not yet. Tycho is in development. The fastest way to evaluate it is to send documents and see what the report says.
Tycho
Document verification
The lab
One venture at a time
ElevateIQ is a founder-led venture lab. It builds one product properly, takes it to revenue, and only then starts the next. Every venture runs the same loop. It is deliberately unexciting, because unexciting is what survives a production incident at two in the morning.
- Analyze
- Find a problem someone already pays to solve badly.
- Learn
- Go deep enough on the domain to argue with an expert.
- Build
- Ship the smallest system that fully solves it.
- Deploy
- Put it in a real environment, with real constraints.
- Test
- Break it on purpose before a customer does.
- Observe
- Watch it in the wild and let the evidence redirect the build.
- Entity
- ElevateIQ (Pty) Ltd
- Registration
- 2024/737044/07
- Jurisdiction
- South Africa
- Model
- Founder-led venture lab
- Flagship
- Tycho (TMA-1)
- Hosting
- Azure ZA North
- Outside capital
- None
Open channel
Send us documents you believe are clean
Or ones you know are not. The fastest way to evaluate Tycho is to give the engine something real to find. You get the full report back.
Request a pilot