Tickmark

Finance controls that show their work. A portfolio case study and working prototype for month-end close, payroll review and financial-model evaluation. It grew out of the earlier "Run Replay" prototype.

Live: https://tickmark-martinzhang.vercel.app · Notes: https://tickmark-martinzhang.vercel.app/notes

A static site: open index.html, or run node scripts/serve.mjs and visit http://localhost:5173. It deploys to Vercel as-is, with no build step; every push to main redeploys.

Use it on your own data

Open the workbench and choose Your data (or press Ctrl K / ⌘K and type "your data"). It runs the same rules as the sample, on your own CSV files, in a separate tab:

• Payroll check: an HRIS file and a payroll register (required), plus an equity vesting file (optional), for a pay period you choose.
• Bank reconciliation: a bank statement and the ledger cash account, matched on amount within a date window.

Files are read in your browser and saved only there (localStorage). Nothing is sent to a server. Decisions, the audit log and your files survive a reload; Reset session and Remove my data clear them. Every table exports to CSV.

File	Columns (header row, any order; bold = required)
HRIS	employee_id, department, status, hire_date, termination_date, currency
Payroll register	employee_id, gross, prior_gross, currency, note, equity_value
Equity vesting	employee_id, units, fmv
Bank statement	date, description, amount
Ledger cash account	date, reference, amount

Dates are YYYY-MM-DD. Amounts may use thousands separators, a currency sign, or accounting brackets for negatives ((1,000.00)). Each upload box has a Download a template link with the sample rows in the right shape. Errors name the row and column, as a spreadsheet would. Files are limited to 2 MB and 5,000 rows.

Rules

All rules live in engine.js, which has no page code: the site and the tests both load it.

Rule	Fires when	Severity
MISSING_EMPLOYEE	A payroll row's employee is not in the HRIS file	High, blocks release
DUPLICATE_PAYMENT	An employee appears twice in the register for the period	High, blocks release
INACTIVE_PAID	Status is terminated / inactive / left / separated and the leaving date is before the period	High, blocks release
OUTLIER	Gross differs from prior gross by more than 40%	Medium
CURRENCY	Paid currency differs from the HRIS currency	Medium
PRORATION	Hired inside the period; expected pay = gross × days worked ÷ days in period	Medium
EQUITY_MISSING	A vest in the equity file has no equity value on the payslip	Medium
INJECTION	A note contains instruction-like text (English or Chinese, after undoing lookalike characters)	Low; the note is withheld from any model

Employee IDs are matched without regard to case or surrounding spaces. Bank matching: each statement line takes the first unused ledger line with the same amount inside the date window; an unmatched line shows the nearest candidate within 10%.

Tests

node --test "tests/*.test.js"

25 tests cover every payroll rule and its boundaries, bank matching, the SEC controls, the FinQA grader, journal checks, the FX revaluation, CSV parsing and validation, and a round trip of the sample through the CSV templates (same results either way). GitHub Actions runs them, plus a syntax check of the page scripts, on every push.

What is real and what is synthetic

Part	Data	Source
Close, payroll, bank lines	Synthetic	data.js. The rules in engine.js run on it in the browser
Payroll engine evaluation (74 cases, recall 98.5%)	From the original Run Replay harness	carried over, not re-run here
Controls on public filings	Real	SEC EDGAR XBRL companyfacts, 8 companies' latest 10-K
FX revaluation (JE-1037)	Real rates, synthetic balance	Bank of Canada Valet API, EUR/CAD and USD/CAD
FinQA grader	Real benchmark	FinQA test set, 1,147 questions (MIT licence, Chen et al., 2021)

Findings the real data produced:

• All 8 filings satisfy Assets = Liabilities + Equity and the cash roll-forward, except that Walmart does not tag total Liabilities, so the identity can't be tested from tags alone.
• Revenue grew more than 15% year over year at Paychex (+16.9%) and Microsoft (+17.8%). The page deliberately writes no explanation: a reviewer would take one from the filing itself.
• The deterministic FinQA grader reproduces 1,108 of 1,108 executable gold programs. The other 39 need table operations, which aren't supported yet.
• One gold label (RSG/2018) calls an 82/18 asset allocation a "debt to equity ratio", which is worth questioning.
• 30 Sep 2026 had no Bank of Canada rate, so the September close rate is 29 Sep.

Refresh the public data

node scripts/fetch-real.mjs
curl -L https://raw.githubusercontent.com/czyssrs/FinQA/main/dataset/test.json -o finqa-test.json
node scripts/finqa-grader.mjs finqa-test.json
node scripts/build-real.mjs

The SEC asks for a contact address in the User-Agent; fetch-real.mjs sends Martin's.

Files

• index.html: the Work page (intro, hero, results, workbench, live control, contact).
• notes.html: every explanation (principles, one exception end to end, scores, what broke, close calendar, data sources, history, letter, credits).
• engine.js: every rule, plus CSV parsing, validation and export. No page code.
• app.js: renders the workbench, records decisions, saves the session, handles uploads and exports.
• motion.js: intro, page wipe, reveals, progress bar, counters, magnetic buttons, jump palette, toasts, segmented indicators, spotlight and tilt.
• blender/mark3d.py: the 3D mark in the hero (Blender 5.4). blender -b -P blender/mark3d.py -- still <out> and -- anim <frames dir>/f_. The frames are transparent; composite them onto #0a0a0a with ffmpeg before encoding.

Credits

Pictures: "Pittsboro Merchant Business Ledger", public domain, via Wikimedia Commons · "HollerithMachine.CHM" by Adam Schuster, CC BY 2.0. Full records are in assets/credits.json.