Error policy

01The four kinds of error

They are handled differently, so it is worth separating them before saying what happens.

We will tell you which kind we believe an error is, and our reasoning, rather than defaulting to whichever category is most convenient for us.

02How errors are detected

The point of the controls below is that an error should be found by the system rather than by you at close.

• Continuous reconciliation. Subledgers tie to control accounts and connected systems tie to the graph daily. A divergence raises an exception with the underlying transactions attached rather than being absorbed.
• Independent checks. Reconciliation logic is written separately from the code it checks. A check that shares assumptions with the process it verifies will confirm those assumptions rather than test them.
• Property-based testing. Beyond example-based tests, we assert invariants — that journals balance, that a period’s movement reconciles opening to closing, that no posting mutates a closed period — and generate inputs attempting to violate them.
• Evaluation gates. Any release that regresses accuracy or calibration on held-out evaluation sets does not ship. Not a warning, not a review item.
• Production monitoring. Reversal rate, override rate, and escalation rate are tracked per workflow per customer. A rising reversal rate triggers investigation and an automatic reduction in agent authority.

03How errors are corrected

Nothing is corrected by editing history. The ledger is append-only. A posted transaction is never modified or deleted, including by us.

Corrections are made as reversals and re-postings, which means:

• the original entry remains visible, with its date and actor;
• the reversal is a distinct entry with its own date, actor, and stated reason;
• the corrected entry is a third record; and
• anyone reviewing the account later can see that something was wrong, what it was, and when it was fixed.

This produces a messier-looking account than silently amending an entry. That is the intent. An audit trail that conceals corrections is worse than one that shows them.

Where an error affected a closed period, we will not reopen it unilaterally. We will report it, quantify it, and let you decide with your accountants whether to reopen, adjust in the current period, or restate.

04When and how we tell you

We disclose errors we find, including ones you would probably never have noticed. That is a deliberate commitment and it occasionally costs us.

Every critical and high notification includes:

• what went wrong, in plain language;
• the period and records affected, listed rather than characterised;
• the financial impact, quantified;
• what we have done and what we recommend you do; and
• what we are changing so it does not recur.

For any critical error we publish a post-mortem describing cause and remedy. We do not publish customer names or figures.

05Agent errors specifically

An agent operating within a granted authority will sometimes be wrong. Published straight-through rates range from 94% down to 66% by workflow, which is another way of saying that between 6% and 34% of cases are escalated — and that a residual share of completed actions will be incorrect.

What we commit to

• Reversibility. Every agent action is reversible and logged, including the reasoning behind it.
• Measurement. Reversal rate is reported to you monthly per workflow. Reversal rates across our customer base run between 0.1% and 0.6%.
• Automatic de-escalation. Where reversal rate exceeds an agreed threshold, authority is reduced automatically rather than after a discussion about whether it should be.
• Root cause. A pattern of agent errors is treated as a defect in configuration, data, or the agent — not as an inherent cost of automation to be tolerated.

Where responsibility sits

You are responsible for reviewing and accepting output before relying on it, and for the authority levels you grant. We are responsible for the agent performing within its stated accuracy, for detecting when it does not, for reversibility, and for telling you.

We think that division is fair, and we would rather state it plainly here than have it emerge during a dispute.

06What we owe you

Where an error is our fault, we will, at no charge:

• correct it, including reprocessing affected periods;
• produce the analysis your accountants or auditors need to assess the impact;
• provide a written explanation suitable for sharing with an auditor, lender, or board; and
• fix the underlying cause and tell you what we changed.

Where an error caused a material service failure, we will apply service credits under any applicable service level agreement in your Order Form.

Beyond that, our liability is governed by our Terms of Service, which cap aggregate liability at the fees paid in the preceding twelve months, with stated exclusions. We would rather point you at that clause here than have you discover it later.

07Reporting an error to us

Write to [email protected]with “Error report” in the subject line. Include the records or period affected and what you expected instead. You do not need to diagnose it.

We acknowledge error reports within one business day and give an initial assessment within three, including which of the four categories in Section 1 we believe it falls into and why.

If you disagree with our assessment, say so. We will re-examine it and, where we still disagree, set out our reasoning in writing so you have something to take to your own advisers.

Nead, LLC (d/b/a DEV.co), 1425 Broadway 22689, Seattle, WA 98112, United States.