Eight ways of looking, 136 candidate jobs, 54 researchers and attackers, about two hours. Findings only — no plan, no recommendation, nothing here asks you to decide anything.
6 August 2026 · six survivors · 57 kills, all listed with reasons
How much was hunted, and what happened to it
About 110 candidate jobs came out of the eight ways of looking. After merging the ones that were the same job found twice, that was 89 distinct jobs. 39 died at the screens. 50 got through. 24 of those were dug into in full. 14 of the 24 killed themselves during the dig — the researcher found the answer already existed and said so. The 10 that still looked alive were each handed to two separate people whose only job was to destroy them: 20 attacks. 4 were destroyed by both. 6 are left standing, and only 1 of those survived both of its attackers.
The one honest sentence about this run: almost every space that looked empty had somebody in it, and in most cases that somebody shipped their product in the last twelve months. Not a single one of the 24 deep dives found a job where nobody had thought of it — the difference between the survivors and the dead is only how far the existing player is from the exact job, and how recently they arrived.
Ranked. Scores are the corrected ones — where an attacker proved a score wrong, the corrected number is shown and the original is in brackets. ⚠️ = more than one guessed input.
| # | The job | The niche | Obscure | Size | Sameness | AI | Worth | Survived both attacks |
|---|---|---|---|---|---|---|---|---|
| 1 | Arguing back when a steel mill pays less for a load of scrap than the grade it was sent as | Scrap metal yards shipping by rail | 4 (contested — one attacker says 1) | 2 | 2 (was 4) | 2 (was 3) | NOT SIZED. ~$1,850–$3,700 rides on one railcar | No — one destroyed it |
| 2 | Registering a newly sold boat, RV or motorbike with the state after the sale | Boat dealers (RV and powersports now have a vendor) | 2 (was 3) | 2 | 3 | 2 (was 3) | NOT SIZED | Yes — both left it standing |
| 3 | Checking the monthly bill from the tip against what the trucks actually dumped | Rubbish haulers who tip at somebody else's site | 2 (was 4) | 2 | 4 | 3 | NOT SIZED | No — one destroyed it |
| 4 | Checking the railway's bill for the flagman who watches the track while a crew digs | Utility and heavy civil contractors working near rail | 3 | NOT SIZED (was 3) | 2 (was 4) | 2 (was 3) | NOT SIZED. $1,000–$1,800 per flagman per day ⚠️ | No — one destroyed it |
| 5 | Billing a state road department for moving a utility's pipes and wires out of the way | Municipal and co-op electric, water and gas utilities | 1–3 (contested) | 2 | 3 (was 4) | 2 (was 3) | NOT SIZED | No — one destroyed it |
| 6 | Deciding who pays when a dealer says the car they bought at auction wasn't as described | Independent wholesale car auctions | 1 (was 2) | 2 | 2 (was 3) | 2 (was 5) | NOT SIZED. ~$15k–$48k per escalated case ⚠️ | No — one destroyed it |
Ranked on: how much of the "nobody is looking" claim survived being attacked, then how much of the work AI genuinely does, then how often the job repeats. Not on size — five of the six are honestly unsized.
THE JOB. A scrap yard loads a railcar with a stated grade of steel and sends it to a mill. The mill weighs it, looks at it, and decides it isn't that grade — so it pays for a cheaper grade, or takes deductions off for dirt and rubbish. Weeks later a payment statement arrives. Somebody at the yard has to open it, work out which loads came back downgraded, and decide whether to ring the mill and argue. Mostly they don't.
THE NICHE. Independent scrap processors who ship to mills, especially by rail — because with a truck the argument happens at the gate while the driver is still there, and with a railcar nobody from the yard is present at all. Roughly 250 to 1,700 companies depending on where you draw the line, which sits right at the bottom of "enough companies to be worth learning". It shrinks further: the two biggest scrap processors in the country are owned by the two biggest mills that buy scrap, so nobody files a claim against their own parent.
HOW OBSCURE. This is the one place the "nobody is there" claim genuinely half-held. Four separate searches for anyone selling downgrade dispute help to scrap sellers returned literally zero results, and one attacker re-ran them independently and got zero too. The identical business — read the customer's deduction, check it, recover the money on a share — is mature and crowded in supermarket supply, and not one of those firms has crossed into metals. But the other attacker found ReSpark, the merged GreenSpark and ReMatter, selling an AI agent that "processes mill settlement documents in any format" and "reconciles received weights against outbound shipments, and flags mismatches." That is the reading half, shipped. The two attackers disagree, and the disagreement is real: the reconciliation is taken, the argument is not.
THE PROCESS. The yard loads and weighs on its own scale, prints a ticket, sometimes photographs the load. At the mill an inspector decides the grade. Weeks later the payment statement arrives, by which time the metal has been melted. Someone compares the statement to the shipping tickets and decides, load by load, whether it's worth a phone call.
THE BREAK. Money never received — the yard is paid for a cheaper grade than it shipped and doesn't contest it. The counter-break is severe and honest: the contracts hand the final say to the mill. Being right does not mean getting paid.
FREQUENCY AND LOAD. This is the weakest part and it comes from the buyers themselves. Rejection and downgrade rates run roughly 0.5% to 3% of loads — one mill bands its suppliers at under 1%, 1–3%, and over 3%. ⚠️ That figure is from a trade article published in 2000, and the buyers quoted in it were a brass company, an aluminium can recycler and a paper mill. Only one was a steel company. There is no current published rejection rate for steel scrap.
THE SIZE. NOT SIZED at the company level. What is measured, and it is unusually sharp: Steel Dynamics publishes the exact arithmetic its inspectors use, and one step of it reads "Visually estimate the fullness of the truck trailer or railcar from a range of 0 to 100%." Run their own worked example — 102,000 lb of scrap in a 2,743 cubic foot car — and at the inspector's 75% guess the load fails and gets downgraded. At 64% the same load passes. An eleven-point difference in one person's eyeball guess, on a car nobody from the yard can see, decides which price the yard gets. Roughly $1,850 to $3,700 rides on that guess per railcar, using a price gap borrowed from a different pair of grades — which is the one guessed input.
WHY AI IS THE LEVER. Half of it genuinely is: reading mill payment statements that arrive in a different format from every mill, and estimating from a photograph how full a heaped railcar is, which you cannot compute and have to look at. Half of it is not: the density check itself is division, and the tolerance tables have been federal law since 1961.
WHAT THE THING WOULD DO. Capture the shipping ticket, the yard's own certified weight, the specific railcar and photographs of the load. Weeks later, read the mill's payment statement whatever format it arrives in, match every line back to a load, and where the grade paid differs from the grade shipped, recalculate the mill's own density test using a machine estimate of fill from the loading photos instead of the inspector's guess. Draft the short note citing the car, the weight and the calculation. A person makes the call on whether to send it, because the mill is a customer they need next month.
WHO ALREADY DOES THIS. Nobody on the arguing half. Scraplytics is closest and is inbound-only — it grades loads the yard is buying. A Chennai software firm sells a module that records claims but cannot argue one. ReSpark and OnlineScrap cover the reconciliation. Meanwhile the mills are buying AI graders: Yonyou sells one to "over 80 industry leading enterprises," and another vendor's headline promise to mills is to "reduce procurement costs 8-12%."
WHY IT IS NOT ALREADY FIXED. Because the mills score their suppliers and arguing is not free. One buyer sends every supplier a letter telling them which rejection band they're in; another does field visits on poor performers and drops them. A yard sells to a handful of mills within trucking distance. Fighting a $3,000 downgrade against a customer who ranks you is a rational thing to lose on purpose.
HOW WE'D KNOW IT WORKED. The share of loads where the grade paid is lower than the grade shipped — a number that does not exist today because nobody puts the shipping tickets and the payment statements next to each other. Establishing it at all is the first win.
WHAT THE ATTACKERS THREW AT IT, AND WHETHER IT HELD. Attacker one destroyed it: ReSpark ships the settlement-document reading as an AI agent, OnlineScrap ships "downgrade and contamination workflows tied to inspection photos," and the mills are buying AI grading first — so the seller would be racing a better-funded buyer into a contract that already gives the buyer sole discretion. Attacker two wounded it and explicitly kept the obscurity score, having re-run the competitor searches and got nothing — but found that the mill's own worked example actually ends in the load being rejected for rubbish, not density, and that the visual-estimate step appears in one mill's manual and nowhere in the other mill checked. Sameness fell from 4 to 2. Neither attacker found an invented figure.
THE JOB. A boat dealership sells a boat. Somebody types the buyer's details, the hull number, the motor serial and the trailer number into the dealership's own system, then types the same facts again into a government website — and for a boat, a second and third time, because in many states the boat and the outboard motor are registered by a wildlife or natural-resources department that has nothing to do with the motor-vehicle department handling the trailer.
THE NICHE. Boat dealers. RV and powersports were part of this finding and have been cut: both attackers found existing vendors there. Nobody in these businesses is technical; the software running them was written by firms nobody outside the trade has heard of. How many boat dealerships exist could not be established — six separate sources were blocked or required an account.
HOW OBSCURE. For boats, genuinely quiet. No software vendor mentions boats, vessels, watercraft or marine anywhere: not Vitu, not Cario, not DLRdmv, not the biggest electronic titling operator in Florida. Two human back-office services do exist and are both listed partners of the boat trade's own association — 50 State DMV and Grace Titles. But the RV and powersports half is gone. One attacker found CVR, a thirty-year-old company used by over 10,000 dealerships across 16 states, whose home page names Powersports and RV as available products; the other found DMVdesk, endorsed by the largest motorcycle dealer association in the country as its "preferred electronic title and registration solution."
THE PROCESS. The sale closes and every fact is already in the dealership's system. A paper packet is assembled — the manufacturer's origin statement or the signed-over old title, a bill of sale, lien paperwork, tax documentation. The trailer goes to the motor-vehicle department. The boat goes to a wildlife department. The outboard motor goes through a third filing, because it is titled separately. If the buyer lives in another state the whole packet is rebuilt to that state's rules and posted there.
THE BREAK. Pure wasted hours — the same facts typed two or three times per sale by someone paid $16 to $28 an hour, plus rework when a packet comes back. And a customer break that is unusually direct: Minnesota's own guidance says a buyer "must wait until your new registration card and decals arrive in the mail before you can legally operate your watercraft" — with one exception, where the dealer submitted the application. The dealer's paperwork speed is literally whether the customer can use the boat they just bought.
FREQUENCY AND LOAD. Not measured. What is measured is the multiplier: one Texas boat-package sale generates three filings across two agencies, by statute — "Separate certificates are required for vessels and for outboard motors." An out-of-state buyer adds a fourth. Pay measured from live job adverts: $16 to $28 an hour, with $20–26 the common band.
THE SIZE. NOT SIZED. Every route to a company count was blocked. Nineteen live job adverts were found for this role — but eighteen of them are RV and powersports, the two segments that now have a vendor, and exactly one is a marine dealership. The evidence of scale and the surviving opportunity do not overlap, and that is the single biggest weakness here.
WHY AI IS THE LEVER. Reading the ownership documents genuinely needs it — a manufacturer's origin statement is designed independently by each of hundreds of makers, arriving as phone photos and fax-quality scans with handwriting in the assignment block. Deciding which track applies is also real judgement across fifty states' inconsistent guidance. But the form-filling half is ordinary software, and reading vehicle titles is now a purchasable off-the-shelf component from at least two suppliers.
WHAT THE THING WOULD DO. Take the deal and the scanned paperwork, read out the hull number, motor serial, dimensions, lienholder and buyer details rather than having a person retype them, work out which filings that buyer's state actually requires, fill each form or website, and read the rejection when one comes back. A person handles wet signatures, phone calls to agency offices, and the packets that still have to go in the post.
WHO ALREADY DOES THIS. For boats: 50 State DMV (trading since 2012, all fifty states, no price published — its own Services page is still unedited website template with Latin placeholder text) and Grace Titles (state titling from $450, Coast Guard documentation from $750, though its wording suggests yacht brokerage rather than dealership volume). For RV and powersports: CVR and DealerDoc, both with software. DLRdmv covers four states electronically and no boats.
WHY IT IS NOT ALREADY FIXED. Because boats are behind a different government door and that door is still paper. Minnesota's submission routes are: the dealer applying, an in-person visit to a deputy registrar, or post to a PO box in St Paul. Texas offers online service only for renewals and replacements. The four states DLRdmv covers electronically are the four states with an electronic dealer channel at all — it did not stop there because the rules were hard, it stopped because state five has no door.
HOW WE'D KNOW IT WORKED. Days from sale to title issued, per unit, which the dealership's own system already timestamps at both ends. Plus rejected filings as a share of filings submitted — nobody counts them today.
WHAT THE ATTACKERS THREW AT IT, AND WHETHER IT HELD. Both wounded it, neither killed it — the only finding in the run to survive both. Attacker one proved CVR and DealerDoc already serve RV and powersports and that a company called Cario sells an "AI Titling Agent" from $4 a package, but confirmed no software vendor touches boats and confirmed the biggest Florida titling operator has no marine coverage. Attacker two found that Vitu's own product site carries a motorcycle-association endorsement the finding had said did not exist, cut the marine evidence to one job advert, and showed the real wall is not messy documents but that most states have no electronic door and require a licensed dealer to file in person or by post.
THE JOB. Once a month somebody in the back office of a rubbish-hauling company sits down with the bill from the landfill or transfer station and checks it against their own record of what the trucks tipped there. Every load was weighed in and out and a ticket printed. The month-end bill covers hundreds or thousands of them. They are answering: did we tip that load, was it weighed and priced right, and was the "contaminated load" surcharge real. Then the money question — which of these charges can we pass on to the customer whose skip it was, and did we actually put it on their invoice.
THE NICHE. Independent haulers who tip at facilities they do not own, especially roll-off and construction-debris haulers where weights vary wildly load to load. The defining feature is that the dump belongs to somebody else. Not sized — one figure exists, a network of "more than 7,000 independent haulers," and the official business-count sites give figures ranging from 421 to 135,034, none of them credible.
HOW OBSCURE. Checking a waste bill on behalf of the customer being billed is a mature crowded category with fourteen named firms, several already using AI on the identical paperwork. Pointed the other way — the hauler checking what the dump charges him — every search came back empty. But an attacker found TRUX, owned by AMCS, shipping a screen whose documentation begins: "Disposal Ticket Reconciliation is used to reconcile the Disposal Ticket Charges on Invoices received from the Disposal Facilities." The finding had checked AMCS by opening its home page and concluded it had no such product. That was an inconclusive probe reported as a finding.
THE PROCESS. A driver tips, the weighbridge prints a ticket, the driver keeps the paper or photographs it. At month end each facility sends a bill — as a PDF summary, a summary with scanned tickets stapled behind it, a paper stack in the post, or a download from that facility's own portal. A hauler tipping at five sites gets five different formats. An office manager works down it against the company's own load record, which is keyed by customer and order while the facility's line is keyed by ticket and truck number, so the two do not line up.
THE BREAK. Two leaks running opposite ways. Money paid that should not have been — a charge for a load that never happened, a weight that does not match, a surcharge nobody challenged. And revenue never billed — a legitimate extra charge the hauler is entitled to pass on and simply doesn't, because by month end the link between that charge and that customer's job is lost.
FREQUENCY AND LOAD. Every load tipped generates exactly one ticket, so the month-end review is checking hundreds to thousands of items. The review event is monthly; the item count is high. Minutes per occurrence: not established.
THE SIZE. NOT SIZED, deliberately. The only numbers in circulation come from two vendors whose own product listings do not contain the feature they are advertising — one claims a "$0.42-per-ton overcharge" and the other an "8-15%" margin lift, and both were refused rather than used.
WHY AI IS THE LEVER. The front door genuinely is: reading month-end statements in as many formats as there are dumps, arriving as PDFs, scans and cab photographs. And the matching — the facility's line and the hauler's own record share no key, so joining them means judging from a free-text field a stranger typed in a hurry. Once you have clean lines on both sides, checking the rate against the contract is subtraction.
WHAT THE THING WOULD DO. Every disposal bill is forwarded in whatever shape it arrives. The AI reads it, reads the hauler's own record of loads tipped that month, matches the two load by load, and produces three lists: charges that do not match anything we did, charges where the weight or rate looks wrong, and legitimate extras that belong on a customer's invoice and are not on one. A person rings the facility to argue and makes the call on a borderline contamination charge.
WHO ALREADY DOES THIS. TRUX, via AMCS, ships the reconciliation screen — though from its documentation it appears to be manual matching, where you type the facility invoice number and eyeball the list. Nothing in it suggests reading the facility's PDF or flagging a weight that disagrees. Soft-Pak, Trash Flow, Starlight, HaulerHero and WAM Software carry no such function. Three vendors published 2026-dated marketing on the theme, and two of them turn out to be search-engine content for a capability their products do not have.
WHY IT IS NOT ALREADY FIXED. Because the data never existed in a usable form — the facility-side software vendor advertises bundling "each invoice, related scale tickets, manifests, and other files into a single PDF" as a feature. And because the work is split three ways: dispatchers know the facility changes, drivers hold the tickets, accounting sees only base prices. No one person in the building can see the whole loop, so nobody notices it leaking.
HOW WE'D KNOW IT WORKED. Dollars credited back against disposal bills each month — sitting in accounts payable, and close to zero today because nobody disputes anything. And pass-through charges billed on to customers as a share of total disposal spend.
WHAT THE ATTACKERS THREW AT IT, AND WHETHER IT HELD. Attacker one wounded it: Command Alkon, a large construction-materials software company, already sells AI that matches weighbridge tickets to supplier invoices and flags over-billing — the same sentence, in the neighbouring industry, not yet in waste. Attacker two destroyed it and did real damage to the finding's credibility: the TRUX screen exists; the industry's own risk — that the dump is often owned by a competitor — remains unquantified; and three quotations attributed to a software vendor's article are not in that article. The researcher paraphrased the source, then quoted its own paraphrase back as the source's words, and graded it as directly verified. The underlying job may still be real. The write-up cannot be trusted as it stands.
THE JOB. When a construction crew works near a live railway, the railroad forces them to hire a flagger who stands watch and holds trains clear. The contractor books him weeks ahead, guesses how many days the job will take, and either pays a deposit against that guess or gets an invoice weeks later. Somebody then works out whether the charges match what actually happened: were those the days the crew was really on the track, was the flagger released when they finished early, was a cancellation given in time, is the weekend rate right — and where a deposit was paid, is a refund now owed.
THE NICHE. Heavy civil and underground utility contractors who cross railway track repeatedly — water and sewer, gas, electrical, road and bridge. Private, family-run, regionally owned, comfortably in the size band. Around 17,668 businesses in the closest official category, only a fraction of which work near track often.
HOW OBSCURE. Five separate searches for anyone auditing flagging bills for contractors returned either nothing or only permitting firms. But the searching was thin, and both attackers proved it. The state road departments already do this job as written policy. Missouri instructs: "The project office shall review all flagging invoices for accuracy, thus inspectors should keep an independent record of the flagging service provided." Texas publishes a full invoice-review procedure requiring districts to reconcile labour quantities against construction diaries.
THE PROCESS. The contractor fills in the railroad's request form and emails it — 14 business days' notice at one railway, 10 at another, up to 30 days elsewhere. They state an estimated number of days before the job has started. Money moves either as a prepaid deposit or a later invoice. The crew works, the flagger keeps his own time, the foreman writes a daily field report on paper or in a phone photo. Weeks later a bill arrives priced against that railway's own rulebook — an eight-hour minimum here, a ten-hour base with two overtime hours built in there, a cancellation charge inside three business days. A project manager who cannot tell an error from a rule he did not know about pays it.
THE BREAK. Standby days are the biggest money and they are not a billing error at all — one state's own account says that even when there was no work, a flagger was kept on the project rather than releasing him, because getting one back took 35 to 45 days. Unrefunded deposits and wrongly applied charges are real but smaller.
FREQUENCY AND LOAD. Not measured, and this is a weakness. No published figure for flagger-days per contractor per year exists.
THE SIZE. NOT SIZED. What is solid is the day rate: $1,600 for an eight-hour day at one railway, $2,900 per flagging day at another, $500 to $1,000 a day for short lines. ⚠️ Three of the inputs needed to turn that into an annual figure are guesses, which is why no annual figure appears here.
WHY AI IS THE LEVER. Partly. Every railway's billing rulebook is different prose in a different document, and there are more than 600 railways in the country. Reading those and turning them into checkable rules is real language work. So is matching a foreman's sentence — "crew on site 0700, RR flagger showed 0645, wrapped up 1430, released him" — to an invoice line claiming ten hours plus two overtime. But once the rule is extracted the check is arithmetic, and preventing a standby day is a text message, not a model.
WHAT THE THING WOULD DO. Read each railway's published rules and keep them current. Per job, take in the flagging request, any deposit, the invoice when it arrives, and the contractor's own daily field reports. Read the field reports to establish what actually happened each billed day. Produce two short lists: charges that do not match the rules or the field record, and deposits where the job finished early and a refund is owed. A person telephones the railway and — the highest-value human job here — tells the flagger to stand down the moment the crew moves off track.
WHO ALREADY DOES THIS. Nobody selling to contractors. RailPros supplies the flaggers and runs its own system holding daily work reports, schedules, "field audits" and billing together, with a contractor-facing portal — it is on the billing side of the transaction. Wilson & Company runs a scheduling programme that "collects field data for cost auditing," sold to railways and project owners rather than to the contractor.
WHY IT IS NOT ALREADY FIXED. Because the railway holds every card — the contractor cannot go to a competing railway, and annoying the scheduler costs weeks, not dollars. And because in many states the cost passes through to the road department anyway, so nobody in the chain has a strong reason to chase it. The exception proves it: Missouri deducts flagging costs from the contractor's own progress payments, and Missouri is where the incentive is real.
HOW WE'D KNOW IT WORKED. Deposit refunds received — money arriving that wasn't before, needing no argument with anyone. And standby flagger-days avoided, priced at $500 to $1,500 a day saved.
WHAT THE ATTACKERS THREW AT IT, AND WHETHER IT HELD. Attacker one wounded it: RailPros already holds the daily record and calls it a field audit, and the contractor's own side is more digital than assumed — one field app reports 128,000 active users. Attacker two destroyed it on all three leaks: the deposit leak was argued from grammar rather than evidence and is closed by the railway's current form, which commits to computing refunds with a worked example; the standby fix is blocked by a sentence in the same paragraph the finding quoted, saying some railways need a week's notice to discharge a flagger; and the wrong-charges leak is already somebody's documented job at two state road departments.
THE JOB. When a road project forces a utility to move its own cables, mains or poles, the state usually pays — but only against a bill built from actual costs. Somebody at the utility assembles it: every crew member by name and job class, hours split normal and overtime, overhead rates calculated from last year's real payroll, every machine at rates the state has pre-checked, every part at warehouse price, every subcontractor invoice, then credits back to the state for scrap sold, for the used-up life of the old equipment, and for any upgrade slipped in while the trench was open. It must reconcile against an estimate written years earlier, land inside a hard deadline, and survive an audit.
THE NICHE. Mid-size utilities with no dedicated relocation department — municipal electric, water and sewer, rural electric co-ops, mid-size gas distributors, smaller telephone and cable operators. The electric side alone gives a verified floor of about 2,830 organisations. What was never established: how many actually receive road-relocation work in a year.
HOW OBSCURE. Eight searches returned no vendor at all — only the rule-writers' own documents. But the searches used the regulator's vocabulary, not the vendor's. An attacker searching for "reimbursables" instead found PowerPlan, whose product page names the road department outright as a use case and whose customer stories include a generation-and-transmission co-op and a municipal utility — both inside the stated niche. A second attacker found CobbFendley, an engineering firm that says in writing it "assists Utility Owners by facilitating the reimbursement process for relocations." One search in the right language overturned the whole obscurity claim.
THE PROCESS. The state authorises; the utility estimates. While crews work, a field form records labour, equipment and materials daily. At billing time somebody assembles names, job classifications matching the field forms word for word, rates with overtime separated, hours including travel. Overhead rates are not guessed — they must be calculated from last year's actual payroll and filed annually, and a missed filing means a flat 30% imposed instead. Then the credits, which is where the judgement lives: scrap at full sale value, the used-up life of the old plant back-estimated from a published index if nobody knows the original cost, and any upgrade credited unless it falls under one of four written exceptions.
THE BREAK. Hard-edged rather than gradual. Never billed: miss the deadline and whatever has been paid may be treated as final — twelve months federally, six in Utah. Billed and thrown out: one state gives ten business days to answer audit findings and then pays only the supported costs. Rate penalty: the flat 30%. Plus rework and pure hours.
FREQUENCY AND LOAD. This is where it fails hardest. The paperwork is daily inside an active project, but the projects themselves arrive when the road programme says so. At the company level this does not repeat many times a week, and the finding says so itself.
THE SIZE. NOT SIZED. No published figure exists for hours or disallowance rates. One state must issue a formal audit report whenever a relocation billing exceeds $150,000 — that is a rule about when a report gets written, not evidence that projects routinely cost that much, and the finding's wording overstated it.
WHY AI IS THE LEVER. Genuinely, on the matching and the judging: field forms, timesheets, warehouse records, subcontractor invoices and the inspector's diary all describe the same day in different words with names that don't line up, and every billed hour must tie back to a specific form. And three judgement calls repeat on every project — was this an upgrade the state should be credited for, which lines on this subcontractor invoice belong to the reimbursable work, and what was the original cost of plant nobody has records for. Roughly half the job is ordinary plumbing.
WHAT THE THING WOULD DO. Read everything the crews, warehouse and subcontractors generated during a relocation, decide which belongs to which project, and tie each billed hour, machine and part back to the document supporting it — flagging what is missing while the crews are still around to explain. Draft the three credit calculations in the state's own language, and the reconciliation against the estimate. A person signs, argues the judgement calls with the auditor, and chases the two documents the machine could not find.
WHO ALREADY DOES THIS. PowerPlan sells "Reimbursables" with automated bill support; CobbFendley sells the service to utility owners. Neither appears aimed at the genuinely small end — a 20,000-customer municipal utility — which is the only residual.
WHY IT IS NOT ALREADY FIXED. Partly because the regulator offered an escape hatch instead of a fix: a utility can negotiate a fixed price and the entire job disappears, with no audit required. The buyer's cheapest option is to delete the problem rather than improve it.
HOW WE'D KNOW IT WORKED. Dollars billed versus dollars actually paid after audit, per project — it lives in the state's audit report. And how many relocations get their final bill in before the deadline, which is a direct forfeiture when missed.
WHAT THE ATTACKERS THREW AT IT, AND WHETHER IT HELD. Attacker one destroyed it on the incumbents above and on the fact that the space has a category name, an annual customer conference and competing vendors. Attacker two wounded it but landed the sharper blow: most of the utilities in the named niche never get a bill to send at all, because a utility sitting in a public road under a permit usually has to move at its own expense — and the one scale figure in the whole finding came from New Jersey, which the source presents as the exception that reimburses everyone. It also found the finding's only dollar figure had been given a meaning its source does not carry.
THE JOB. A dealer buys a car at a wholesale auction, then says it wasn't what was announced. Somebody at the auction reads the claim, goes back through what was said on the block, the written condition report, the check-in photographs, the national rulebook and that auction's own house rules, and rules for the buyer or the seller — inside a window that is often seven days and sometimes only the day of the sale. The decision is final and binding, and both sides are customers who come back next week.
THE NICHE. Independent wholesale car auctions — the physical auction houses not owned by the two giants. The trade body has "more than 345 domestic and international auto auctions." Yards with lanes, drivers and a title office, running software on a green-screen system from the 1980s.
HOW OBSCURE. Not obscure, and it got less so during the run. No vendor sells claim adjudication to auctions — that specific chair is empty and three separate searches confirmed it. Everything around it is full. CarDr launched an "Auction Arbitration Risk Report" on 27 July 2026 — nine days before this hunt — which validates scan findings against the national arbitration rulebook. The independent auctions' own trade group has formed an AI committee jointly with the software vendor that runs their systems. On the dealer's side there are at least three companies helping buyers win claims.
THE PROCESS. A car sells under a coloured light — green guaranteed, yellow limited, red as-is. A condition report is written. For a truck, the downgrade decision happens while the driver is still there. For a railcar-equivalent — a car sold online — the buyer files days later. The arbitration manager pulls the announcements, the condition report, the photos, the light and the mileage, reads the claim against all of it, applies the rules, gets the defect priced to test it against a money threshold, decides, negotiates, and types the outcome into a 1980s system.
THE BREAK. Mostly wasted hours plus money moving on each decision. Every claim ends in a cancelled sale, a price adjustment or a denial, and a wrong call costs one side or the other. Missing the clock kills a valid claim on a technicality.
FREQUENCY AND LOAD. The role is full-time and named — auctions hire "Arbitration Manager" by that title. But the underlying rate is low: the industry bands its suppliers at under 1%, 1–3% and over 3% of loads going wrong, and one auction treats a 25% claim rate as the point where a buyer is abusing the system.
THE SIZE. NOT SIZED. ⚠️ Two of the inputs are guesses. Nationally, roughly 1,300 escalated disputes a year reach the two named forums, at $15,000 to $48,000 average value — which is one escalated dispute per firm per decade. The much larger informal layer is genuinely unmeasured.
WHY AI IS THE LEVER. The finding scored this 5 out of 5 and an attacker cut it to 2, correctly. The condition report is captured through a guided app advertising "guided trimline selection helps inspectors to avoid errors," "VIN decoding," "OEM build data" and "damage automation" — structured, coded data, not prose. And the deciding rules are arithmetic: is the repair $800 or more, is it within seven days, is the car over 20 model years old. Only the disclosure question — did the spoken announcement cover this fault — genuinely needs a model.
WHAT THE THING WOULD DO. Read every document attached to a sale plus the buyer's claim, and produce a decision file: what is alleged, what was and wasn't disclosed and in which words, which light the car sold under, whether the estimate clears the threshold, which clock is running and how many hours are left, which exclusions were checked. A person makes the call, signs it, and handles the phone conversation with whoever lost.
WHO ALREADY DOES THIS. The majors do it in-house at scale. CarDr, Auction AI and Click-Ins all sell prevention — better inspection so fewer arguments start. ArbChex, ReconCash and ArbIT sell the dealer's side of the same fight. Nobody sells the umpire's side.
WHY IT IS NOT ALREADY FIXED. The rulebook is deliberately not machine-ready — its author states it is "not enforced or monitored" and members adopt it "in whole or in part... at their sole discretion," with the key money thresholds handed to each auction. And the industry does not want a machine making this call: the one AI firm targeting independent auctions writes that AI "should support documentation quality and staff speed, not make arbitration decisions."
HOW WE'D KNOW IT WORKED. Days from claim opened to claim closed, and the share closed inside the window. Both live in the auction's own system; neither is published by anyone, so there is no benchmark.
WHAT THE ATTACKERS THREW AT IT, AND WHETHER IT HELD. Attacker one destroyed it on CarDr's launch, the trade group's AI committee, and the general point that a space with a trade show, a comparison landscape and an AI-native entrant is the wrong space. Attacker two wounded it, cutting the AI score from 5 to 2 on the structured-condition-report finding, and cutting sameness after showing the two mills — sorry, the two auctions checked — do not even share a grading structure.
57 entries. Grouped by what killed them. Where several hunters found the same job independently, it is one entry with a note.
Killed at the screens (25):
Killed during the deep dive, by the researcher, before any attacker was spent (14):
(Entries 1–25 and 44–52 each represent between one and five hunters' findings merged; the raw pre-merge count was about 60.)
26 jobs cleared the screens and were never dug into. Named so nothing is silently dropped. The five most interesting are marked.
What was searched. Job boards for strange job titles; trade association conference agendas, member surveys and magazine columns; the app directories and help centres of industry software nobody outside those industries has heard of; federal and state regulations and the regulators' own guidance documents; and vendor marketing, read as marketing. Every primary document that carried a load-bearing number was downloaded and read in full where possible, not summarised — several 100-page government PDFs were text-extracted by hand after the normal reader failed on them.
What could not be reached, and it matters. The general web search allowance ran out early and repeatedly. Most of the deep dives and nearly all of the attacks were done by fetching pages directly by address, because the search engines available afterwards either blocked automated access or served bot-challenge pages — DuckDuckGo, Bing, Mojeek, Startpage, Brave, Ecosia and seven independent search servers all refused at some point. This means every "nobody sells this" statement in the report is weaker than it reads. The clearest proof: in three separate cases an attacker found an incumbent within one search, using the vendor's vocabulary rather than the regulator's — "reimbursables" instead of "utility relocation billing," "hauler disposal cost reconciliation" instead of "disposal invoice audit," "electronic title and registration" instead of "title clerk." Searching in the wrong words is what produced most of the false positives here.
Job boards were almost entirely unreachable, which cost us the best evidence type available. Where a job posting was obtained, it was usually the strongest single piece of evidence in the finding. Where it was not, the finding rests on vendor marketing describing a problem that vendor sells against — which is interested evidence, and is flagged as such throughout.
Honesty of the numbers. The previous round produced roughly 25 figures that were not in the sources credited. This round produced three confirmed fabrications, all in one finding (waste hauler disposal invoices), where a researcher paraphrased a vendor's article and then quoted its own paraphrase back as the source's words. Two further findings were caught giving a real number a meaning its source does not support. Several findings were caught grading a failed check as though it were a completed one — recording "the page was blocked" with the same confidence label as "I opened it and it said this." That habit made evidence tables look far better sourced than they were, and it is the single most correctable weakness in the method.
Where the evidence is genuinely thin. Five of the six survivors are NOT SIZED, and that is the honest state of it — no published figure exists for how often the job breaks or what it costs in any of them. The marine title finding rests on a single job advert. The scrap downgrade finding's only frequency number is 26 years old and about aluminium cans. The waste hauler finding has no verified frequency at all. None of these gaps would take long to close with two or three conversations with people who actually do the work — which is the shape of what phase two would have to be.
One structural finding worth more than any individual opportunity. Across 24 deep dives, the thing that killed most candidates was not that the job was fake or the money imaginary. It was that somebody arrived first, usually very recently. Twelve of the dead had a vendor who shipped or announced their product in 2025 or 2026. Three shipped within weeks of this hunt. Whatever the current wave of AI tooling is doing to back-office work, it is not leaving obscure spaces obscure for long.