Why we built this.

I’m Devon Smith. I ran a restoration shop in Vista, California — British marques were our specialty, but the work spanned the trade — until a critical motorcycle accident took me off the floor and closed the shop. Before that, I had the seat every customer in this country wishes they had — a veteran mechanic behind the counter who had no reason to upsell them. I priced jobs to keep customers for life, not to clear that month’s parts invoice. I told people to walk away from cars I would have made good money fixing. I told people to keep cars other shops had quoted to total. I lost work doing it. I would do it again.

What I saw inside the trade in those years is the thing customers never see — the gap between what the shop knows and what the customer knows. The customer walks in with one symptom and walks out with the shop’s answer. They have no way to test the answer. No way to know if the $3,800 estimate they just signed is actually the cheapest way out of the problem. No way to know if the manufacturer already published a customer-satisfaction program that covers the exact failure their car has. No way to know whether the platform they own is the kind that goes another 150k miles after this repair, or the kind that limps three months and then needs another $4,000.

That information exists. It exists in NHTSA recall data, in TSBs, in extended-warranty programs, in failure-mode patterns documented across millions of vehicles. It exists in the heads of veteran mechanics. None of it gets to the customer at the moment they need it, because nobody whose paycheck depends on a different answer is going to walk them through it.

I watched customers try to make decisions without having all the information, or pay for something that would’ve been covered for free at the dealer.

After the accident, I had time. I had a thousand conversations replaying in my head — the ones where I had been on the customer’s side of the counter and the ones where I had been on the shop’s side. I started building the tool I wished I could have handed every one of those people on their way in. More about who I am is on the about page. The rest of this is about the engine.

There is one decision that runs underneath every shop conversation a car owner ever has, and they almost never get to name it out loud: repair, or replace? Fight, or walk away? Spend $3,800 here, or take the dealer’s $2,800 trade offer? Spend $1,400 now and another $2,000 next year, or sell the car running for $6,500 and put it toward something different?

The shop cannot help you with that decision. Not because they’re dishonest — because they’re the wrong person to ask. The shop is paid to repair the car. Their job is the repair. Their expertise is the repair. The decision about whether the repair is the right move at all is a different question, and it is the one customers have historically had to make alone, at the counter, under pressure, without any of the numbers in front of them.

The math customers need to run looks like this. What is the most likely cause of the symptom at this VIN, this engine family, this mileage band? What does that repair actually cost in this ZIP code, at an independent vs. a dealer? What is the car worth running? What is it worth in trade? What is documented to fail next on this platform between now and 200k? Is there an open NHTSA recall, a CSP, a warranty extension, a goodwill program any of which would change the cost of the repair from $3,800 to zero?

No customer at a counter can run that math in their head. No shop owner can run it for them without conflict of interest. That is the gap RepairVerdict fills. The 13-stage deterministic pipeline walks every one of those numbers in the same order, on every Verdict, with the audit trail to prove it. How the keep-or-replace math actually works has a worked example.

The first version of this product I built in 2024 was a single Claude prompt. VIN goes in, paragraph comes out. Same VIN went in the next day, different paragraph came out. The model would invent TSB numbers that did not exist. It would quote labor times that were ballpark but never specific. It would hedge — “you may want to consult a qualified mechanic” — at the exact moment the customer needed a recommendation. I stopped sending that version to customers within a week.

What replaced it is the engine in production right now — v3.8.14 against prompt Engine version 41. The first eleven stages are deterministic. No language model. VIN decode against the NHTSA vPIC API. Recall + TSB + ODI pull keyed to the VIN. Failure-mode database join. Mileage-band cluster. Weibull hazard math per system. Documented mitigation crediting. Coverage triage — the mandatory five-question sweep for federal emissions warranty, extended powertrain, CSP/goodwill, lemon-law, open NHTSA investigation. Regional cost adjustment. Repair-to-value math. Score and band assignment.

The same inputs produce the same score, today, tomorrow, and three years from now. That is not a property you can get out of an LLM. That is a property you can only get out of math. The language model’s job in this product is translation, not judgment. Stage 12 — and only stage 12 — is where Claude Sonnet 4.6 takes the deterministic findings and renders them into prose a non-mechanic can read. Opus 4.7 is available via an env override for cases that need deeper reasoning. Neither model touches the score.

The model does not decide whether your car is in the green band. The math does. The model writes the explanation of what the math found.

If that sounds like a small distinction, it isn’t. It is the difference between a product you can stake a $4,000 repair decision on and a product that is fancier ChatGPT. Read the stage-by-stage architecture for the pipeline detail. /proof shows the six pillars of evidence in customer terms.

Stage 13 of the pipeline is where the engine reads its own output back, six different ways, before any customer sees it. Each lens is a separate Claude call with a tightly scoped system prompt. The lenses don’t see each other’s findings. Their pooled findings drive a revisor pass. The cycle runs up to three times. Every Verdict carries the audit log.

The six lenses each catch a different category of mistake. factual_grounding catches fabricated identifiers — any TSB, recall, CSP, or NHTSA number in the prose that does not resolve to the audit trail gets flagged as critical and the Verdict never ships. anti_punt catches the 19 banned phrases — “a second opinion,” “consult a specialist,” the hedges that LLMs hide behind when they don’t want to be wrong. If the prose recommends a second opinion without first naming a specific hypothesis and a specific discriminator test, the audit pass rejects it.

mechanical_reasoning checks the chain — would a veteran mechanic actually order the steps the prose recommends? customer_experience checks the read — is this legible to a non-mechanic at the kitchen table? output_discipline holds the structure — every required section (§11.7 Bottom line, §14.5 What to do next, §15.7 Coverage opportunity, §15.11.1 What we don’t know, §16.5 Cost summary, §17.3 Predictive maintenance roadmap, §25.5 Diagnostic tells, §29.5 Mechanic walkthrough) has to be present. credibility_of_claims is the newest lens — it checks whether the prose reads like an expert wrote it or like an LLM hedged.

Any one of those lenses run in isolation would still produce a better Verdict than the industry default. Run together, with revisor cycles, they produce something I can stake my name on. The 6-lens audit section of /architecture has each lens written out in full.

The engine doesn’t hypothesize failures out of thin air. It joins against a database of 730+ documented failure-mode patterns, each one platform-specific, each one carrying make / model / year-range, mileage band, symptom tells, repair cost range, and the diagnostic test that confirms or rules out the pattern.

These aren’t generic categories. They are specific named patterns — Subaru CVT torque-converter shudder, Ford 3.5L EcoBoost intake-valve carbon, Honda V6 timing belt tensioner failure, GM 5.3L AFM lifter collapse, BMW N63 oil-burn, Toyota oil-gel sludge on the 2AZ-FE. The full encyclopedia is browsable — every pattern, every source citation, every mileage band, every confirmation test. The same data the engine reads. Read-only, on the public site, because hiding it would be the same kind of asymmetry the product exists to close.

Where do the patterns come from? Real NHTSA recall data. Real TSBs. Real CSPs and warranty extensions. Real ODI investigations. The sources pagelists the named data sources in full. Anything in a Verdict that points at a specific failure mode also points at the specific pattern entry — so the customer, or the customer’s mechanic, can read the source.

If a Verdict can’t name the pattern, name the source, and name the discriminator test, it’s not a Verdict. It’s a guess. The engine is built to refuse to ship guesses.

The 730+ catalog isn’t a marketing target. It’s the count today. It grows when a real-world Verdict surfaces a pattern that isn’t in the file yet. Every addition has to carry the same evidence load as the originals.

Every Verdict carries an immutable audit_trail_id and an inputs_hash. Same inputs against the same engine version produce the same Verdict, byte for byte. That is not a marketing promise — it is a property of the deterministic pipeline. Stages 1–11 are math. Math is reproducible.

What that buys the customer is credibility you can show to a shop. When a Verdict says “the most likely cause of your symptom is a Honda VCM piston-ring failure, here is the TSB number, here is the diagnostic test, here is the repair cost range, here is the warranty-extension program that may cover it,” you can walk into the shop with that printout and the shop can look up the audit trail. Every claim resolves to a source. Every source is named. Nothing in the prose is unverifiable.

The audit trail is the answer to the engineer or coder friend who asks “how do I know the AI isn’t just hallucinating this?” The answer is: the AI didn’t generate it. The math generated it. The AI wrote the explanation of what the math found. Every claim in the explanation is back-referenced into the audit trail. Anything that doesn’t back-reference into the audit trail is flagged by the audit pass and never ships. More on security and data handling.

This is the part of the business model I get the most pushback on, mostly from people who have run businesses. “You could double revenue overnight if you took referral fees from shops,” the argument goes. “Quote a fair repair, send the customer to a partner shop, take 15%. Everyone wins.”

Everyone except the customer. Because the moment RepairVerdict gets paid more if you say yes to a repair, it stops being the third-party voice and starts being another voice in the same broken room. The reason the product works is not the AI. It is the structural independence. We make the same amount whether you fix the car, sell the car, walk away from the car, or keep driving it past the warning light. That single property is what makes the Verdict trustworthy to a skeptical car owner.

The moment anyone other than the customer is paying us, the product is structurally compromised. So nobody else does.

That means: no referral fees from shops. No affiliate commissions on parts. No data sales to insurers, advertisers, or used-car dealers. No sponsored failure modes. No “recommended partner” networks. One transaction with you is the entire revenue model. If that ever changes, the product I built stops existing and a different product takes its name — and I’m not interested in building that one. The about page says this the same way; this is the same commitment, written once and held across every surface.

A shop diagnostic costs $150–$300 in most U.S. markets, depending on the platform and how aggressive the labor rate is. A dealer diagnostic is $200–$400. The Verdict costs less than one shop labor hour. That is the anchor I built the pricing around.

$79 Vehicle Verdict covers reliability, symptom analysis, ranked failure-mode hypotheses, coverage hunt, and the discriminator tests to confirm the hypothesis. It is built for the customer who has a symptom or a pre-purchase question and needs the math before they walk into the shop. $129 Full Verdict adds shop-quote parsing (every line scored against regional pricing baselines), the 12-month maintenance forecast, and repair-to-value math against private-party + trade-in value. It is built for the customer who already has a shop estimate in hand and is trying to decide whether it’s fair. $199 Full + Personal Video Review adds a personal video walkthrough recorded by me. It is built for the customer who wants to hear me explain the call on their specific car.

Why not $19? Because at $19 the engine costs more to run than I make on the transaction, and any product that loses money per unit either gets cheap enough to ship at scale (which means cutting the audit pass and the personal review) or it disappears. Why not $399? Because the anchor that matters is the shop’s diagnostic fee. If the Verdict costs more than that, it stops being a no-brainer for the customer who needs it most. The full pricing page lays out what each tier covers.

Every Verdict ships with a straight 30-day refund. If something is wrong with the analysis — a missed coverage program, a misread shop quote, a hypothesis the customer can document is impossible — you email me within 30 days and I review the refund request alongside the audit trail. That review is done by me, personally, against the same data the engine used. If approved, the full purchase price posts back to your original payment method within 3–5 business days. No partial refunds, no apology credits.

The refund is real skin in the game. Most customers don’t request one. The ones who have, I’ve honored. The handful of disputed cases got my actual attention — I read the Verdict, the audit trail, and the customer’s reasoning, and either issued the refund or sent an explanation of why the analysis stood. The refund policy spells out exactly how requests are handled.

Every email to smithperformanceproductions@gmail.com is read by me. Not a support team. Not a ticketing queue. Not a triage AI. Me, in Vista, California, on the same laptop that ships engine versions. If something is broken, write us. If a Verdict played out and you want to tell us how, write us — those are the emails that make the engine smarter, because the regression suite gets a new anchor entry every time a customer reports back on whether the prediction held.

That single channel is also the responsible-disclosure channel for any security finding, the bug-report channel for site issues, the source for new failure-mode patterns in the file, and the way I find out when a part of the prose isn’t reading right. There is no second inbox. If I miss something, it’s because I miss it — not because it got triaged into a queue I never see. The /contact page has the same address.

The Verdict is informational analysis. It is not a hands-on diagnostic, not a legal opinion, and not a guarantee of any future mechanical outcome. A qualified mechanic with the car in front of them sees things we cannot — physical wear, fluid quality under a flashlight, odd noises during a road test, the smell of a slowly burning clutch. Cars are mechanical systems with real variance.

Specific places the engine is honest about its limits: vehicles built before NHTSA’s structured recall era have thinner platform-level data than newer cars and the Verdict carries a confidence floor that reflects that. Brands with limited U.S. data — niche European, grey-market, low-volume EV programs — get the same treatment. Symptoms that require live OBD-II data to confirm get a discriminator test in the prose, not a confirmed cause. Anything that requires a tear-down to see — a head gasket leak, a piston-ring confirmation, a CVT internal failure — is reported as a high-probability pattern, not a confirmed finding.

Saying that publicly is part of why the audit pass exists. An LLM left to its own devices will hedge everywhere, because hedging is free for the model. The audit pass forces the engine to say where the limits are, not just that there are limits. The methodology page has the explicit scope-limits section. We say it publicly because hiding it is the same kind of asymmetry the product exists to close.

The consumer Verdict is the foundation. The next surface is a Shop License tier — same engine, shop-side UI, customer-shareable shop Verdict for the cases where the shop wants to hand their customer the same kind of third-party math they don’t want to be the one to deliver. That work is in beta with Cal West Auto. No specific dates — I don’t ship roadmap promises. Sometime this year.

After that, the failure-mode encyclopedia keeps growing and the engine keeps hardening. The product won’t expand into things that compromise the structural independence — no partner network, no insurance integrations, no advertising. The pricing won’t move out of the sub-labor-hour band. The audit trail is permanent. Everything else can change. Those three things won’t.