This page describes how a make-ready invoice gets checked, what the check deliberately does not cover, and how we handle being wrong. It does not disclose the rule logic itself, which is the part that took the work.
Where something is not yet built, this page says so. We are early, and a methodology page that described a system which does not exist yet would defeat its own purpose.
Four independent checks. Each catches a different category of error, and they compound, so a charge can pass one and still fail another.
Line items against the rates, cost-share terms and scope in your signed attachment agreement.
Whether a charge is allocated to the correct party under FCC Part 1 Subpart J, your agreement, or the applicable state standard. This is a question about who owes a cost, not about whether the work was necessary.
Where a charge sits against what comparable work actually bills at, by pole owner and region. No benchmark figure is ever shown to you unless it draws on at least three distinct pole owners across at least three distinct clients. Below that floor we omit the comparison rather than show you a thin one.
The same pole and the same charge appearing twice across documents, and lines that reappear on a true-up having already been billed once.
It will not tell you a pole did not need replacing. We take the pole owner's engineering determination as given, in full, and audit the cost consequences of it. We do not evaluate whether a loading analysis was correct, whether a clearance determination was sound, or whether the work was over-scoped. Where an answer would depend on that judgment, we say so and recommend review by a licensed professional engineer.
That boundary is fixed. It is not a matter of current capability and it will not move as the engine improves.
It will not give you a legal conclusion. PoleProof is not a law firm and its findings are not legal advice. You are the party of record in any dispute, and you send it. Where a matter needs counsel, we would rather hand you a name than pretend the service reaches further than it does.
Only a deterministic check may create a finding. That means a lookup, a comparison, or arithmetic against something documented: your agreement, a tariff, a published formula, or a quantity stated on the invoice.
The language model layer extracts, classifies, explains and writes plain English. It is never the thing that decides an overcharge exists. That is the design rule everything else rests on, and it is the answer to the obvious question about an automated auditor.
The pass that tries to prove us wrong. Every candidate finding then goes through a second, separate pass whose only job is to defeat it. Is there a contract provision authorising this charge. Is there a reading of the formula under which it is correct. Is the extracted value ambiguous on the source document. Does a state rule differ from the federal one. Anything it defeats is downgraded or dropped. It runs as a distinct pass with its own logs rather than as extra instructions inside the first pass, because the point is to be able to show that it ran.
Deterministic, fully sourced, and adoptable by you. The only tier that enters a dispute packet, and the only tier we are ever paid on.
Shown to you and labelled as not findings. Never billable and never in a packet.
Logged internally for tuning and never shown.
The threshold starts high and comes down as the evidence base grows. An engine that under-reports builds a reputation for being right. An engine that over-reports hands the other side an argument on the first bad packet, and you do not get that back.
Every finding shows its work. A finding cannot be emitted unless it can populate all of: the source document, the page or line reference, the value extracted, the comparison value, where that comparison value came from, the arithmetic, and the identifier of the rule that fired. An item that cannot fill every field is not a finding. Every deliverable also carries the engine version, the rule library version and the benchmark snapshot date, so a finding can be reproduced later against the logic that actually produced it.
The federal rules and your own agreement drive findings from day one.
State law is different. Pole attachment rules vary enormously by state and by pole owner class, and getting one wrong is worse than not having it. So a state rule is researched, then reviewed, and only enters the engine as a source of findings once a licensed communications attorney has cleared it and that clearance is recorded and dated. A rule with no current clearance cannot originate a finding, and the lookup defaults to refusing rather than allowing.
To be clear about what this is and is not. It is a check on the rules going in, performed offline, once. It is not a review of your findings. No person reviews an individual finding before you see it, and that does not change.
Today, that clearance process is not yet running, so the engine will originate findings on the federal baseline and your own agreement only. State rules can inform an Observation, never a Finding. We would rather tell you that than let you assume otherwise.
A finding withdrawn after a pole owner rebuts it is not billable, and it is logged.
In a service with no human reviewer, that is the most credible accountability mechanism available, and it is the only one that is fully automatable. It also costs us very little, because a finding that does not survive a rebuttal was never going to be recovered. We would rather be the auditor that does not get paid for being wrong.
More broadly, we are paid only on savings the pole owner actually makes, evidenced by a document the pole owner itself issued. Never on a charge that merely looks wrong.
We track three things from the first finding onward: findings you adopt, findings sustained after a pole owner responds, and findings withdrawn.
The withdrawal rate is the one that matters, because it is the direct answer to anyone arguing we are incentivised to over-claim.
We are not publishing those numbers yet. A rate computed on a dozen findings would be worse than publishing nothing, and taking a number down when it dips would be worse than both. The floor we have set for ourselves before anything goes on this page is roughly 250 findings across at least five clients and three distinct pole owners. Once it is up, it stays up.
It does not disclose the rule logic, the scoring, or the contents of the benchmark dataset. Those are the work.
What it does tell you is enough to check us. Every finding you receive names the clause, the rule or the benchmark it rests on, and you can go read that source yourself.
One free check. No account, no commitment. Send one invoice or estimate and the attachment agreement it runs under.
Run a free check →