Sample · fabricated data

A finished audit, start to finish.

This is what an audit returns. The company, the customers and every figure below are invented — no NemulAI customer's data appears on this page. The arithmetic is real: every total is derived from the rows, so you can check it.

Margin by customer — July 2026

Worst first, because that is the order the decision gets made in.

Customer margin for July 2026, worst first. Fabricated sample data.
CustomerRequestsRevenueInference costMargin%Attributed by
Northwind Healthbelow cost412,000$4,200.00$7,180.00−$2,980.00-70.95%Tagged ±2%
Lattice Legalbelow cost88,500$1,900.00$2,410.00−$510.00-26.84%Tagged ±2%
Vantage Retail1,240,000$12,000.00$10,940.00$1,060.008.83%API tag ±5%
Cobalt Logistics305,000$3,600.00$2,760.00$840.0023.33%Scheduler ±10%
Harbor Point Media96,400$2,400.00$1,180.00$1,220.0050.83%Tagged ±2%
Ardent Studio51,200$1,500.00$610.00$890.0059.33%API tag ±5%
Fernwood Labs22,800$900.00$290.00$610.0067.78%Heuristic ±35%
Attributed total2,215,900$26,500.00$25,370.00$1,130.00
Cost with no customer attached42,000$3,120.00not attributed
Provider statement$26,500.00$28,490.00−$1,990.00

Attributed margin says $1,130.00. The month was −$1,990.00.

Add up the seven accounts and the period looks profitable. The provider billed $28,490.00 against $26,500.00 of revenue, and $3,120.00 of that carried no customer identifier at all. Counted, the month is −$1,990.00. A report that quietly dropped the unattributed cost would have shown the first number and stopped.

What this audit will not do

It will not spread the $3,120.00 across the seven accounts. Splitting it by revenue, or by request count, would produce a per-customer number for every row and a different answer for at least one of them — and there is no evidence in the data saying which account incurred it. A figure arrived at that way looks identical to a measured one on the page and is not the same thing.

So the gap stays a line of its own, and the report says which decisions it is not strong enough to support.

Where the gap comes from
Shared batch endpoints and warm-pool time, invoiced per GPU-hour rather than per request.
What would close it
Emit the tenant id on batch submissions, and record which tenant reserved each warm pool.

What the audit could and could not establish

Both figures are checked against fixed thresholds. This dataset clears one and misses the other, which is stated rather than smoothed over.

98.1%
Requests carrying a customer id

Clears the 95% threshold. Per-customer figures describe substantially all of the period.

89.0%
Cost attributable by value

Short of the 90% threshold by 0.95 points. Enough to rank the accounts; not enough to state a company-wide margin as fact.

Getting this for your own numbers

Four commands run entirely on your machine — inspect the SQL before it connects, aggregate inside your own warehouse, then verify the result offline with no network and no credential. One file crosses, and you drag it in yourself.

No raw rows, no credentials, no content, and nothing you have not seen listed and approved. NemulAI never holds a warehouse key, never syncs in the background, and sits nowhere near your request path — it reads inference it does not run.

← Back to home