How VehicleMD Generates Reports

Overview

A VehicleMD report combines verified government data, public regulatory datasets, and an AI-generated analysis constrained to a strict schema. The intent is a fast, honest pre-purchase snapshot: what is known to have gone wrong on this generation, what the market expects it to be worth, and whether it warrants a second look by a mechanic. Every report is generated fresh or served from a 30-day anonymous cache keyed on vehicle specifications, not on the person asking.

How a Report Is Built

Verified government data (NHTSA)

Recalls and owner complaint records come directly from the U.S. National Highway Traffic Safety Administration via its public API. Recall entries show the campaign number, affected component, summary, consequence, and remedy. Complaints are aggregated by component and failure description. This data is authoritative for what it covers, though NHTSA completeness and timeliness depend on reporting from manufacturers and consumers.

EPA fuel economy and ownership data

Combined, city, and highway fuel economy figures are sourced from the U.S. Environmental Protection Agency. Annual fuel cost and 5-year ownership cost ranges are derived from EPA figures combined with national average fuel prices and typical maintenance schedules for the vehicle's class and age.

AI-generated analysis

The narrative portions of the report, including known issues, owner complaint summaries, the fair market value range, the inspection checklist, and the final verdict, are AI-generated. The vehicle specs, the NHTSA and EPA data above, and a strict JSON schema are passed to the model on every request. Output is validated before it is shown. If the response is malformed, the request fails rather than degrades silently.

How the Score Is Calculated

The 0 to 100 health score is a weighted composite. It is produced by the AI with guidance from the underlying data, not a hand-tuned formula, but the intended weighting is:

• Recall severity and count (approximately 30%): unresolved safety campaigns, especially powertrain or fire-risk, weigh heaviest.
• Complaint frequency and severity (approximately 25%): volume of complaints relative to production, and whether complaints cluster around expensive failures.
• Known model-year reliability (approximately 20%): chronic issues documented for the specific generation and trim.
• Age and mileage context (approximately 15%): how the mileage compares to typical lifespan for the powertrain.
• Projected ownership cost (approximately 10%): expected maintenance and fuel burden relative to the class.

The Buy, Buy with Caution, and Avoid verdicts correspond to high-confidence, mixed-signal, and high-risk score bands. The boundaries can shift based on severity weighting: a single open powertrain recall can push a verdict from Buy to Buy with Caution even at an otherwise strong score.

What a VehicleMD Report Is Not

A report describes the vehicle class and generation, not the specific unit in front of a buyer. It has no way to see rust, crash repair quality, maintenance history, or how previous owners treated the car. It does not pull VIN history, title records, accident reports, or odometer verification.

It is not a substitute for a pre-purchase inspection by a qualified mechanic. The inspection checklist in each report exists to help the buyer and mechanic focus on the components most likely to fail for that generation, not to replace the inspection itself.

How We Handle Uncertainty

The system prompt instructs the AI to avoid fabricating recall numbers, bulletins, or statistics it cannot ground in the supplied data. When NHTSA returns no complaints or recalls for a given year and model, the report is marked as such rather than filled with invented details. Fair market value ranges are presented as ranges, not point estimates, and are flagged as estimates.

AI output can still be wrong. Verdicts should be treated as one signal among several, and any meaningful purchase decision should be validated against a mechanic's inspection, a title check, and the buyer's own inspection of the actual vehicle.

Continuous Improvement

The prompt, schema, scoring weights, and surfaced data change as we learn where the tool under- or over-weights certain signals. Meaningful methodology changes are logged in the project's decision history. If you spot a report that seems off, please send it to hello@vehiclemd.com or through our contact form. Concrete counterexamples are the single best way to improve the model.