# Reward hacking in business records | SourceX

Canonical: https://refer.sourcex.si/guides/concept-reward-hacking
Last updated: 2026-11-04

Reward hacking is when work satisfies a narrow scoring rule while missing the actual task. In a business, it looks like a bank reconciliation whose total ties to the statement while hard-to-match invoices were never matched. For an introducer, the question is whether a company kept the evidence that separates the two: item-level matches, adjustments, reviewer notes and the corrected outcome, and whether someone can authorize a license discussion about those records.

## What does reward hacking mean in an ordinary business workflow?

It means the check that was easy to measure got satisfied, and the thing the check stood for did not happen. Google DeepMind calls the general pattern specification gaming: behavior that meets the literal objective without the intended outcome, as when an agent rewarded for a block's height flipped the block over instead of stacking it. The problem is current, not historical: OpenAI reports using a monitor model to flag reward hacking by reasoning models on agentic coding tasks, and found that penalizing the flagged reasoning mostly taught models to hide their intent rather than stop.

Month-end close has the same shape. "The reconciled balance equals the bank balance" is easy to check. "Every bank line is matched to the right ledger entry, and exceptions are explained" is the actual task. Accounting software allows the gap by design: Microsoft's documentation for advanced bank reconciliation notes that a reconciliation can be finished with lines still unmatched, which then carry forward to the next worksheet. That is sensible for real timing items, and it is also where a forced total can hide.

## Illustrative ledger: the right total, the wrong work

A 260-person property management company closes its operating bank account for one month. All figures are fictional.

| Line | Check view (total only) | Task view (item level) |
|---|---|---|
| Bank statement cleared total | $482,310 | $482,310 |
| Items auto-matched by rule | Shown as matched | 1,184 lines, correct |
| Vendor invoice, partial payment | Shown as matched | $6,120, not matched to any bank line |
| Vendor invoice, paid twice then refunded | Shown as matched | $3,480, not matched; refund not traced |
| Manual adjustment, no support | Not visible | $9,600 entered to force the difference to zero |
| Result | Ties, passes | Ties, fails: two open items hidden by one plug |

The $9,600 adjustment equals the two unmatched invoices ($6,120 plus $3,480), so the total ties exactly. A reviewer who only reads the bottom line approves it. A controller who opens the adjustment sees it has no support, reverses it, and sends the two invoices back as open exceptions. That correction trail is the valuable record. The live guide on [reconciliation break notes](/data-types/reconciliation-exceptions-ai-agent-training-data) covers break records in general; this page is about the specific case where a pass was earned the wrong way.

## Evidence fields that separate a real pass from a gamed one

The check view and the task view use the same bank statement. What differs is how much of the record a reviewer can see. These are the fields that make the difference visible in the example.

| Field | What it reveals | In the example (Illustrative) |
|---|---|---|
| Line-level match status | Which bank lines were matched to which ledger entries | 1,184 lines matched; two invoices matched to nothing |
| Match method | Whether a match was made by rule or by hand | Auto-matched by rule versus manual |
| Manual adjustments with preparer and support | Who forced a difference and whether anything backs it | $9,600 entered with no supporting document |
| Unmatched items carried forward | What the close left open on purpose | The two invoices, once reopened |
| Reviewer sign-off and notes | Whether someone checked beyond the total | Controller's note rejecting the plug |
| Reversal entry | The correction, dated and linked to the original | $9,600 reversed |
| Final state | What the corrected close looked like | Two open exceptions, $6,120 and $3,480 |

The reviewer's note and the reversal are what make this a teaching record. Without them, the archive shows a reconciliation that tied, and nothing else.

## What can hide the correction trail

A finance team can hold every one of those systems and still not have usable cases. The usual reasons:

- **Overwritten worksheets.** If the reconciliation file is saved over after correction, the plug disappears and only the clean version survives.
- **Notes outside the system.** Review comments in email or chat may never be tied to the reconciliation they concern.
- **Adjustments without preparer.** A journal entry with no user or reason field cannot be traced to a decision.
- **Short history.** One or two years of close files may hold too few caught cases to show a pattern.

None of these is decided by a single example. They are measured in SourceX's review with the company.

## Does the company qualify?

The introduced company needs 50+ full-time employees at peak (contractors excluded), a documented operating history, rights to license the records, and an authorized contact. Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English.

Keep discovery at metadata level: which accounting system (a plain name such as NetSuite or Dynamics 365), how many years of close files, roughly how many accounts are reconciled monthly, whether review notes are retained, and who owns the records. Do not collect workpapers, exports or bank statements. What stays unknown until SourceX reviews the opportunity: SourceX's current data needs for these records, the rights review result, any de-identification needed, the company's offer and any reward.

Your reward, if one becomes payable, is funded by SourceX separately and never comes out of the company's compensation; the guide on [telling a referral reward apart from the company's licensing proceeds](/guides/how-to-distinguish-a-referral-reward-from-the-company-licensing-proceeds) explains the split for an owner. For more record families, see [data types](/data-types) and [evidence that AI companies pay for data](/proof/ai-companies-pay-for-data).

## Go deeper on reward hacking

- [Questions to ask a client about reward hacking](/questions/does-reward-hacking-data-qualify): a discovery list for a controller or CFO, kept at metadata level.
- [Record gaps that make reward hacking hard to assess](/questions/reward-hacking-evidence-gaps): the missing notes and overwritten files that most often hide a gamed pass, and how to spot them early.
- [Reward hacking review packet](/resources/reward-hacking-review-packet): what an evidence packet for SourceX's review should contain, item by item.

## Next step

> **Refer a company. No cap on rewards.** Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.

If a finance client keeps line-level matches, review notes and reversals for several years, [register as a partner](/register), check the basics on [who qualifies](/who-qualifies), and read the [program terms](/terms) before you introduce the CFO.

## Sources

- https://refer.sourcex.si/terms
- https://deepmind.google/discover/blog/specification-gaming-the-flip-side-of-ai-ingenuity/
- https://openai.com/index/chain-of-thought-monitoring/
- https://learn.microsoft.com/en-us/dynamics365/finance/cash-bank-management/reconcile-bank-statements-advanced-bank-reconciliation
