Image: Untitled, licensed under CC0 1.0.
A carrier's billing rep calls about a short pay. Your team moved $1,840 off a detention invoice three weeks ago and he wants to know on what basis. Your analyst opens the audit tool and reads back what it shows: flagged for review, confidence 0.91.
That is not a dispute position. It describes how a piece of software behaved, and the rep is entitled to ignore it. A finding is only worth the money attached to it if someone can rebuild the reasoning later, in front of a person whose job is to make you pay.
Three things every finding has to carry
A defensible finding is not a verdict with a probability stapled on. It is a short argument with its parts visible, and the parts are always the same three.
- The rule invoked. The contract clause, tariff item or regulation the charge is being tested against, cited to the document and the version in force on the shipment date.
- The facts it was applied to. Arrival timestamp, departure timestamp, free time granted, appointment window. Values, not adjectives.
- Where each fact came from. System of record, field, retrieval time. "The TMS" does not qualify. "Trip stop record, geofence exit, pulled 14:02 on 3 March" does.
Drop any one of the three and the finding stops travelling. A rule with no facts is a policy statement. Facts with no rule are trivia. Facts with no provenance are your word against the carrier's, and the carrier is the one holding an invoice.
Signals rank the queue, evidence settles the argument
Agent systems produce both, and treating them as interchangeable is how a team ends up quoting a confidence score to a billing desk.
A signal is anything that makes a finding more or less worth attention: variance against the lane's rolling average, a carrier's history on this charge code, a model's own estimate of how sure it is. Signals earn their keep by deciding what gets worked first and what gets escalated to a human. None of them belongs in a dispute letter, because none of them is a reason the charge is wrong.
Evidence is the opposite kind of object. It is checkable by someone who has no reason to trust you: a signed delivery receipt, a geofence trace with timestamps, the rate table clause, the published index value a surcharge formula consumes. The test is practical. If the carrier's analyst can pull the same artefact and land on the same conclusion, you have evidence. If reproducing the result means running your model, you have a signal.
That distinction is also an architectural one, which is why the arithmetic belongs in deterministic code rather than inside the model. A demurrage day count recomputed from gate records can be repeated by anyone holding those records. A day count a model produced cannot, and no amount of after-the-fact explanation converts one into the other.
What belongs in the decision record
Write the record once, at the moment the finding is made, and write it for two readers who will never be in the same room.
| Element | Example | Who asks for it |
|---|---|---|
| Rule invoked, with version | Master contract 4.2(c), detention, effective 1 Jan 2026 | Carrier |
| Facts, as values | Arrival 08:14, departure 13:46, free time 2h, appointment 08:00 to 09:00 | Carrier |
| Source and retrieval time per fact | Geofence entry, visibility provider, read 3 Mar 14:02 | Auditor |
| Computation performed | 5h32m dwell, less 2h free, 3h32m billable against 4.5h billed | Both |
| Conclusion and amount | Overbilled 0.97h, $1,840 disputed | Both |
| Signals considered, labelled as signals | Variance 310 percent above lane average for the code | Internal only |
| Who approved, and at what autonomy setting | Auto-cleared under L4 ceiling, or named reviewer | Auditor |
The last two rows are the ones teams skip, and they are the ones that decide whether the file survives scrutiny. Keeping signals in the record but marked as signals is what stops them leaking into the dispute. Recording the approval path is what lets you answer the question an auditor reaches eventually, which is who was accountable for this particular dollar.
The clock you are defending against
Reconstruction is not an open-ended exercise, because the right to argue expires.
Under 49 U.S.C. 13710(a)(3)(B), a shipper must contest the original bill or a subsequent bill within 180 days of receipt in order to have the right to contest those charges. Subparagraph (A) imposes the same 180 days on a carrier issuing a bill for charges beyond the ones originally billed. Both sides of the conversation run on a timer.
Ocean runs tighter. Under the Federal Maritime Commission's billing rules at 46 CFR 541.8, the billed party gets at least 30 calendar days from the invoice issuance date to request fee mitigation, refund or waiver, and the billing party must attempt to resolve that request within 30 calendar days of receiving it, absent agreement on a later date. Section 541.7(a) separately requires the billing party to issue its invoice within 30 calendar days of the date on which the charge was last incurred.
Those windows are the real specification for your record keeping. A finding whose basis an engineer has to excavate from application logs is a finding you will not file at volume inside 30 days. It will quietly become a charge you pay.
Your auditor asks a different question
The carrier wants to know whether the charge was wrong. Your external auditor wants to know whether the number in your freight accrual can be relied on, and that question now has standards text sitting behind it.
PCAOB AS 1105 gained paragraph .10A, effective for audits of financial statements for fiscal years beginning on or after December 15, 2025. Where a company provides external information in electronic form, the auditor has to obtain an understanding of the source and of the company's process for receiving, maintaining and processing that information, and then test either the information itself or the relevant controls. The Board issued a policy statement on 18 September 2025 and staff guidance with illustrative examples on 1 October 2025.
Carrier EDI, visibility telemetry and published index values are precisely that category of input. If your accrual rests on charges an agent validated against a geofence feed, someone will ask where the feed originates and what controls sit on it. A model score answers the first question as poorly as the second.
The AI frameworks arrive at the same destination from another direction. NIST's AI Risk Management Framework pairs accountability with transparency, ties traceability to both, and observes that transparency is often necessary for actionable redress related to AI system outputs. Article 12 of the EU AI Act requires high-risk systems to technically allow the automatic recording of events over the lifetime of the system. Whether or not either instrument binds your deployment, what they describe is what a dispute file needs anyway.
How Penny keeps the file defensible
Penny audits every invoice before payment, which means the record gets written while the evidence is still warm rather than assembled from memory a quarter later. Each finding carries the governing clause, the facts as values, the source and retrieval time behind every fact, the computation, and a drafted dispute that cites all of it.
Signals live alongside that record and stay labelled. They route the work and set the priority order. They never appear in what goes to the carrier. The detention example above is worked end to end in what a detention charge has to prove before you pay it, where the causality test is the part a rate check never reaches.
What to do next
Take five disputes your team filed last quarter that the carrier rejected or ignored. For each one, write out the rule, the facts and the source of every fact, using only what your systems hold today. Note how long each reconstruction takes.
Most teams find two things. A few disputes were correct but unprovable, which is the same as wrong once the 180 days run out. And the reconstruction cost per item is high enough that nobody will ever do it for the small charges, which is where the volume lives.
Book a walkthrough to see the decision record Penny builds behind a finding, on your own invoices.