Everything the brief is built on. This is a reference, not a chapter — nothing depends on reading it in order. It is here so you can check any claim against what was actually said or seen, including the 2 times I put the idea in someone’s head before they said it, and the 7 things I got wrong and had to fix.
This page and the data file behind it are built from the same source. The counts above are not typed in by hand — they are simply how many there are. Every other count in this project drifted because somebody typed it somewhere. These cannot: change the record and both the page and the file change together, or neither does.
Each observation has its own id, what kind of evidence it is, what I take from it, and which questions it helps answer. That structure is in the page itself, and in a file you can hand straight to an AI.
Download the whole record · 42KB The raw screens and documents →Harbor is not a real company. It is the imaginary insurer this brief is written for. Everything below is real — real insurance websites, real documents, real people, real accounts.
The record quotes real documents, so it has to use their words. Here is what they mean.
Forty years, insurer side and claimant side
Testimony from inside the industry, not something I watched happen — strong, and still one account.
Asked whether the incentives could be better for buyers, he proposed an insurer backed by a credit union, where members share in the dividend — “you’re participating here, you get a kickback.”
Surfaced on 17 Sep, in the session. His diagnosis: drive well and the premium still rises, so there is no upside outside a claim. His instinct is sound and the vehicle is wrong — see the note in what I got wrong.
“A lot of people stay because they fear the risk of getting dropped once they have a claim — because they don’t have any history with the company.”
His answer to what causes the gap, given without prompting.
“But you can be with an insurance company 20 years and have a claim, and they’ll still drop you.”
He demolished his own answer in the next breath. The belief is real. The protection is not.
“If a prior claim was taken care of to your satisfaction — done quickly, fast, fairly — they may say it was worth it, because I got treated right.”
Where trust actually comes from. Not the quote. The claim.
Insurers promised 24-hour adjuster contact. Does it work? “Depends how much stuff the adjuster got going on.” — “Insurance companies have fired most of their adjusters.”
He named specific insurers and an empty regional office. I could not verify any of it, so I cut it rather than print checkable claims I did not check. What stands is that he said the staffing is gone. Nothing in the finding rests on it.
Can a normal person tell whether two policies cover the same thing? “No. Sales never explains coverages — because half the time they don’t know it anyway.” And: “the insurance the sales gives you, the claims takes away.”
One of the two independent sources for the comparison claim. The other is the broker’s client, at the opposite end of the same transaction.
The tree. Wind blows your tree onto your house: they pay to lift it off and repair the house, will not haul the tree away, and you get nothing for the tree. Lightning drops the same tree: you get something for the tree. Why? Probability.
One example, and coverage is revealed as unknowable to a normal buyer.
Home, four vehicles and equipment · one insurer
Looks every six months and almost never binds. Exactly the population this brief is about.
“I look every six months, because usually on every six-month basis they take your premiums up.” Do they ever go down? “Never go down.”
He has looked every six months for years and has never once moved on price. Shops, does not bind.
The one time he switched: a denied hail claim. Every house on the block got a new roof; his was denied. He took photographs and showed them. He moved home and auto together.
Not price. A claim.
He is about to switch again, right now — a claim filed over a month ago is still unpaid. “Mine is getting ready to be broke now, and I’m getting ready to fix it.”
The decision is being made live, and again the trigger is a claim.
Any retention offer, ever? “No. They just tell me if I don’t pay it by a certain time, they’re gonna cancel.”
Second person, and again no save attempt of any kind. This is what killed the retention counter-move hypothesis.
He asked to raise the deductible on his vehicles. They raised the deductible on his house. Grandfathered at $250, changed to $2,000. “I didn’t ask them to do it.”
If he needs another roof the deductible is now 6% of what the home is worth, and wind damage is being removed. A material adverse change to one policy, triggered as a side effect of an unrelated request the customer initiated, discovered afterwards. He consented to none of it.
At his last switch he did not take the cheapest. “I went with the more reputable company… I would have a better experience if I had a problem with them.”
It denied his claim anyway.
Two quotes at that switch. The one he did not take: “I just never went back.”
Third independent confirmation that a quote is abandoned rather than declined.
“Nowadays they can put your name in, they can see what coverages you have. I get texts and emails telling me they can cover my house and my vehicles and I’d save X a year with the same coverage I got now, because they know what your coverage is.”
Coverage lining coverages up is solved and deployed — as an outbound sales weapon. A insurer can match your coverage automatically to poach you. You cannot do the same thing to shop.
What would move him: “a few hundred dollars cheaper” with the same coverage — and without the 6% roof deductible. “I just don’t think that’s fair.”
What he wants out of a claim: “I just want to be made whole. I paid into this. I have a claim, I want to be made whole. I don’t want any more than what is required. I’m not trying to run a game.”
Surfaced on 17 Sep, in the session rather than in the write-up. It corrected an assumption I had not noticed I was holding — that a claimant wants to extract more than the loss. He wants restoration, and he volunteers the distinction from people who do not. The adjuster models this relationship as adversarial. The claimant does not.
The wind-versus-lightning tree example also came up in this conversation. ⚠ I introduced it.
I had heard it from the adjuster hours earlier and carried it in. It is therefore not a second independent source and is not counted as corroboration. Recorded rather than dropped, because a single source quietly becoming “two people said it” is exactly how a four-conversation study produces a false finding.
Logged into an active auto account, read end to end, with the customer’s permission
This corrected what I had written from the public buying surfaces. The correction runs in that direction, and it cost me a claim.
The Coverages screen is genuinely good. Every coverage with its limit and its individual price: liability $100k/$300k — $405 · comprehensive $1,000 deductible — $949 · collision $1,000 deductible — $733 · rental $40/day — $71 · six coverages at No Coverage, $0. Total $2,190.
Better disclosure than most industries manage — and it corrects what I had written from the marketing pages. Inside the account, nothing is vague. This is the disconfirmation.
Comprehensive and collision are $1,682 of a $2,190 premium — 77%. The liability that covers catastrophic loss is $405.
Visible, itemised, and never surfaced as a decision.
The landing page is titled “Policy and Coverages” and contains no coverages — premium, vehicle, people, contact details, and two upsells. Coverage is a link.
Policy Activity logs everything. Two mid-term premium changes in six weeks, four figures apart, both labeled only “Policy change” — plus three more reading “Policy change $0.00.”
Amounts and the full sequence under S4.
In three months of logged email: 8 billing · 5 marketing · 3 driving tracker · 3 paperless · 2 confirmations. Zero about coverage.
“Scheduled Activity: Nothing to report here! You don’t have any activities scheduled for a future date.” The policy expires in 85 days.
The renewal is not an event in the customer’s view of their own policy. Third confirmation, now from the insurer’s own forward calendar.
A returned bank payment in July produced a $20 fee and a cancellation notice with an effective date twelve days out. Reinstated when payment cleared.
It sits in the log with exactly the same weight as a paperless-settings email. A near-loss of coverage, published and unflagged.
The policy summary — the one document that states what you bought — is filed under “ID Cards and Other Proof of Insurance.”
Three of the four items in that section are proof you produce for someone else: your car, your lender, the state. Categorised as evidence for third parties, not as your own record.
In the app it renders as a flat document image — not searchable, not selectable, not readable by the page around it. The downloaded PDF does have a text layer. The actions offered are Save, Print, Email and Fax.
Corrected 16 Sep. I first recorded the document itself as having no text layer. It has one — the preview is the image, not the file. The legible version exists and you have to leave the interface to get it.
The policy summary states: “This coverage summary replaces your prior one.”
Supersession announced, with no way to see what changed.
And: “Your insurance policy and any policy endorsements contain a full explanation of your coverage. The policy contract is form 9611D AR (12/15). The contract is modified by forms 4884 (10/08), Z357 (01/07), A239 AR (07/17), A264 (02/22) and A331 (11/21).”
Six form numbers. The contract itself is retrievable — under Documents, behind a filter, on a different page from the summary that names it. The five endorsements that modify it are not: listed as plain text, with no date, no link, and not bundled in the 40-page contract. This is where the answer to “is the fallen tree covered” actually lives.
Two policies on a prior vehicle show “We rescinded this policy. This means you’re not covered by it.” — and the interface still offers “Report a claim” and “Get your ID card” on both.
The most ordinary event in insurance: swapping one car for another
I have the customer’s account of it and the insurer’s timestamped log of the same weeks. One subject, two kinds of evidence — not two sources.
What the insurer recorded. 30 June: policy change +$1,115.26, the new vehicle added. 16 July: policy change −$696.15, the old vehicle removed. Net +$419 on a $2,190 policy.
Two halves of a single transaction, sixteen days apart, each logged as “Policy change” with no explanation and no link between them. Identical labelling to a paperless-settings update.
Interleaved with those two entries: a returned bank payment, a $20 fee, a cancellation notice with a twelve-day effective date, three further policy changes at $0.00, and a reinstatement. Eleven logged events across three weeks, in one flat list.
What the customer remembers: “We switched because they were cheaper… about $200 cheaper… how long did it take? My day.”
One decision, one day, clean.
Both accounts are accurate. She is not misremembering — nobody ever narrated the ten weeks to her, so there was nothing to remember. She experienced a one-day switch; the system recorded three policies, two cancellations and five policy changes.
The gap between those two is the finding, stated as a lived case: everything published, nothing interpreted — to the point where a customer’s own account of her own decision does not match the record of it.
Whether the two May cancellations on the prior vehicle were ever registered by the customer.
“We rescinded this policy. This means you’re not covered by it” is a serious sentence. Not established. If it was never noticed, it is the sharpest data point here — and I would rather leave it open than assume it.
Whether the cancellation notice influenced the decision to stay — it landed three days before the vehicle came off — and what the three “Policy change $0.00” entries were.
Sequence observed, causation not established; the rest are insurer-side events the customer may never have seen.
The test that mattered most: does any of this generalise, or is it one insurer’s interface?
Amounts and dates are withheld; what follows is structure and counts.
After the policy ended, a residual balance went unpaid. The insurer sent, in sequence: a first collection email · a first collection letter · a second collection email · a second collection letter · a third collection email · a final collection letter · a final collection email. Seven escalating, timed, multi-channel contacts, then referral to a third-party collections vendor.
Every step logged, every step automated, every step on schedule.
The same log shows the payment machinery in full: a reversed electronic payment, an insufficient-funds fee, a payment restriction placed on the bank, three declined card attempts, pending-cancellation notices automatically rescinded the moment money arrived, and cancellation when it did not.
The system reacts to payment state within minutes and records every reaction.
In the same account, across the same period: no communication explaining what the policy covered. None explaining a premium change. None stating that a renewal decision was due.
A cancelled policy at this insurer offers a one-click “Restart My Policy” button and links to seven other products it would like to sell. It offers no way to see what the policy had covered.
At the first insurer, cancelled policies still offered “Report a claim” and “Get your ID card.” Both insurers keep dead policies alive and actionable — in one direction.
This removes the only sympathetic explanation. “They lack the tooling to explain things” is not available: there is a seven-touch, multi-channel escalation ladder for a few hundred dollars owed, and no touch at all for a premium that moved by four figures.
The capability is not missing. It is allocated.
And it is the same at both insurers — the same fee, the same automated cancellation notice, the same instant cancellation on payment, the same silence everywhere else.
Said precisely: both insurers this household used. Two is better than one and it is not an industry. Wherever this brief says “both insurers,” it means these two — hold it at that size. The second-insurer test was the one that could have killed the finding. It did the opposite.
A real auto renewal invoice · September 2026 · paper
What is on it: an amount, a due date, four vehicles, a named local agent with photo and direct line, a detachable payment stub, and a promotion for a deductible-reduction program.
What is not on it: any coverage limit. Any deductible. Anything about what is covered. Any comparison to the prior term’s premium. Any statement that a decision is due.
He told me premiums rise every six months. The document that arrives does not tell him his premium rose. To know, he would have to find the last statement and compare it himself.
The renewal moment is not presented as a decision. It is presented as a bill — and a bill invites payment, not evaluation.
This is also the high-touch end of the market: a insurer with local offices and adjusters you can meet. Even here the routine communication carries no coverage information, and the one piece of free space is used to sell.
Three real buying surfaces — two insurers and a comparison site — driven end to end by an agent
No personal information entered, no quote submitted. Every screen captured. This is the repeatable half of the record.
The precision is inverted. Pet injury $1,000. Trip interruption $500. Custom parts $5,000. Liability, comprehensive and collision: “up to your policy’s limits.”
The coverages that cannot hurt you carry exact numbers. The ones that decide whether you are ruined carry a placeholder. Not one exclusion appears on the page.
The exclusions a insurer does publish: routine maintenance, wear and tear, personal items, ride-sharing, mechanical breakdown, pet injuries — prefaced with “varies by policy and insurer.”
Every item is something a reasonable person already assumes. The exclusions that decide claims are published nowhere.
A insurer’s own advice: “To best compare quotes from different providers, ensure you select the same coverage, policy limit, and deductible amount.”
Correct — and it describes a task their surface does not support. No export, no import, no shared vocabulary.
Both insurers have built a coverage calculator that asks nine questions — assets over $50,000, whether you are underwater on the loan, whether you could fund a replacement car, access to a substitute vehicle. Neither puts it in the quote flow.
The purchase path asks for a ZIP code and returns a price. And the tool says it recommends a “minimum set of coverages.” This is the observation ALT–2 attacks.
The comparison layer is paid “when you purchase a policy or get a quote through us” — and on whether the compared price is the real one: “It depends!”
The only comparable number on the surface is provisional, and the channel earns on the quote either way.
Three insurers in a few years
Someone who actually moves.
“Whoever’s the cheapest, to be honest.” Any loyalty to a provider? “No.” — “My thought process on insurance is that it’s a scam.”
Said calmly, as a settled fact about the category.
$50 a month cheaper: “doesn’t seem like a sufficient enough reason to go through all the paperwork.” Told that is $600 a year: “maybe if you say it like that.”
Same money. Different answer. Her stated threshold is $100/month. The industry quotes monthly.
She has never declined a quote. “They send follow-up emails — your quote is waiting — and then they stopped sending them.” Unsure a decline option even exists.
She bought elsewhere. That insurer never learned it. The quote is still, formally, undecided.
The only thing she had to leave the flow to find was the VIN. Everything else she answered from memory, because the insurer verifies against the motor vehicle report on the back end and corrects it.
The friction story is smaller and more specific than expected.
Switched insurers for ~$200/month. Took one day. Does she understand how her price is set? “It was just like a slot machine.”
The incumbent quote required confirming stored details; the challenger quote required entering everything from scratch.
A structural re-quote advantage belonging to whoever already has you — observed once, not yet general.
Auto, home, business, rentals, family
The one person here who does not buy insurance the way this brief measures it. I went in expecting an intermediary to dissolve the problem. It relocates it, and it produced the sharpest objection in the study.
They switched auto insurers this year. Two things happened at once: twin teen drivers pushed the premium past what they would pay, and their agent left the brokerage. “I followed the agent, actually.”
A third trigger class, and one I did not register. Not a claim. Not a rate. Not a life event on its own. The distribution channel moved and the customer moved with it — insurer unnamed in the reasoning, and irrelevant to it.
Can an ordinary person read a policy and tell what they are covered for? “You know what to look for, like deductibles… but there’s so much insurance language now to where there are gaps in that knowledge, and that’s why I think you need an agent.”
The adjuster said this from inside the industry. This is a buyer saying it from outside, unprompted, in a different conversation — and naming the remedy. Two sources at opposite ends of the same transaction. This is the one thing in the study that is independently corroborated.
How did they know the new policy covered what the old one did? “It’s down to my agent.” The agent shops a panel of insurers and returns the two best.
The comparison problem is real for this person, and they have already bought a solution to it. The solution is a human being.
The quote they did not take: “I don’t personally tell the other company that I’m not going with them. You just go with the quote you move forward with.”
Fourth independent confirmation — and the strongest, because this buyer did not even generate the quotes and still never closed the loop. If a broker will not decline on your behalf, nobody will.
They filed a storm claim. It was paid. “Took a long time… some of those claim processes should be more streamlined… but it ultimately happened.” They did not switch.
Set against the customer whose hail claim was denied and who moved home and auto together in response: slow-and-paid does not move people. Denied does. Two subjects, opposite outcomes, opposite behavior.
They have never switched home insurance since buying the house. It was bundled; they unbundled it to move the auto. They are now considering re-consolidating for a discount — “we haven’t done it yet.”
The only home data point in this brief, and it is inertia. It does not test the finding on home. Home stays open.
What would move them: price first — “insurance is a price-sensitive market” — then an agent’s recommendation, then coverage. They pay roughly $700–800 a month across multiple vehicles and teen drivers.
Inside that one answer they said it would take “considerably cheaper” and then, seconds later, “if it was a little cheaper, I would do that.”
I cannot resolve the threshold from this and I am not going to pick the half that fits. The switcher’s $100/month remains one stated number from one person.
On shopping direct: “you fill out a form, and then you have a million companies that reach out to you… I hate that process.”
Consumer-side corroboration of what the comparison layer’s own terms say — it is paid “when you purchase a policy or get a quote through us.” The lead is the product.
Their explanation for why the category does not change: insurance is legally compulsory and the industry spends to keep it that way. “When you can hide behind the government policy, you make it required — and then they don’t have any incentive to innovate.”
Their analysis, not an observation of the market, and they do not work in insurance. Recorded because compulsory demand removes the pressure to change and a legible renewal removes revenue. Both point the same way, and neither binds a challenger.
On building trust as a software product: “you’re missing the interpersonal. I’m not saying it can’t be done, because it can — but where do you incorporate that?” Their working model of trust is Chick-fil-A: “you trust that they’re gonna be consistent with what they provide every time.”
The strongest live objection to the intervention. Trust here is not built by a document. It is built by a person, and by repetition. An intervention that is only a better disclosure has to answer this, and saying it is a better disclosure is not an answer.
They have been with the same agent since buying the house — home, business, rental houses, sister’s house, mother’s house. “Because we just trust them.”
Six policies and three households, held by one relationship. That is what Harbor is actually competing with in this channel, and it is not a price.
Late in this call I described my own findings to the subject — the denied-claim story from the third interview, and the account-log observation that everything is published and nothing is explained. ⚠ Second contamination event in the same study.
Everything quoted above was said before that point; everything about transparency afterwards came back in my own shape and is not used, including the line I would most have liked to quote. And one remark I nearly attributed to them — that the industry is built on risk, not trust — is mine. The transcription tool labels every speaker “Me:”, which is how a researcher ends up quoting himself as a source.
A record that only shows what survived is a sales pitch. These are the 7 places the evidence went against me — including the one that killed how I had framed the whole thing.
| Kind | What happened |
|---|---|
| Withdrawn | A driving tracker comparison stood in the finding and was removed because it traced to no observation. It is not in this record and it is not in the brief. |
| Killed by fact-check | H4 — loyalty pays and history does not port. Built on a case I had wrong: I believed a household had run driving-tracker-priced insurance and left anyway. They never had it. They were quoted, it came in higher, they stayed put. The driver did not know the vehicle measured her driving at all. |
| Reversed by evidence | I had written that pre-purchase surfaces invert precision and that the industry is opaque. The authenticated account disconfirmed the second half: inside a policy, disclosure is close to exemplary. The finding changed from a claim about quality to a claim about time. |
| Cut, not printed | The adjuster’s claims about insurer staffing named specific companies and an office. I could not verify them, so they are cut rather than printed as checkable claims I did not check. |
| Checked after the session, 17 Sep | The adjuster proposed a credit-union-backed insurer sharing dividends with members. The vehicle is wrong and the mechanism is real: mutual insurers already do this. State Farm Mutual paid a $5bn dividend to auto customers in 2026 — about $100 a vehicle, 4–10% of premium by state, issuing from August. TruStage is the credit-union insurance program, but on auto it is an agency arrangement underwritten by Liberty Mutual and I found no member dividend. And the check cuts against the idea rather than for it: State Farm is a mutual, its policyholders are its owners, and it still shows you no comparison before you buy. Aligned ownership did not produce legibility. The dividend is also roughly $100 a year, against the $100 a month this record found as a switching threshold. |
| Corrected by reading the file, not the picture | Three errors, all mine, all from reading the summary’s on-screen image instead of the downloaded PDF. The contract is form 9611D AR, not 9611AD AR. The endorsements are Z357 and A331, not Z857 and A681. And I recorded the document as having no text layer — it has one, and reading it is what would have given me the right numbers. I also wrote that the contract is one “you do not have.” It is in the account, behind a filter. The five endorsements are the part you cannot get. |
| Contaminated ×2 | Two interviews where I supplied the finding. Both tagged, counted as single sources, downstream answers discarded. |
The field work came first and this came second. It did not produce the finding. Its job was to try to break it.
Once the brief was written and live, I ran the whole thing through the review method I use on client work — eight stages, each with checks that have to pass with evidence, not with an opinion. I did it because a conclusion you reached by research and a conclusion that survives being attacked are not the same object.
It found six errors in a brief I had already published — a count that disagreed with itself in three places, a claim traced to a source that did not say it, a figure restated wrongly, and a page that advertised a fault that had already been fixed. All six are corrected. That is the argument for doing it.
And it still fails three of its own checks, which I have not fixed. The largest is this: four tests decided the central idea, and I was supposed to score them without knowing which one was mine. I never did. The decision was still mine to make and I made it — but the one check designed to keep me honest about it did not run, and no amount of scoring afterwards fixes that.
It is worth being exact about what a score of 74 means here. It means the argument is consistent and every part of it has a named owner. It does not mean anything is built. Nothing is.