AI diagnostics

Open a complaint issue and start a diagnosis. CarSignal works the problem the way a good diagnostic tech does: gather evidence, rank hypotheses, and confirm on the car before committing money to a repair.

What goes in

The vehicle's saved spec, the customer's reported symptoms, and any DTC codes. CarSignal reloads those facts from the job on every turn, so a stale browser cannot silently change the vehicle. Codes are grounded against canonical SAE definitions, and hybrid or EV code ranges force the correct powertrain assumptions.

Deep diagnosis searches manufacturer bulletins in the NHTSA archive, platform-specific repair reports, forums, and useful video walkthroughs. Two model providers research independently. Video discovery also uses YouTube's official search API and the same platform-relevance filter used by DIY; transcripts come through the configured transcript provider. Search results only nominate possible sources, and their snippets never influence the diagnosis by themselves. Only government, licensed, or direct automaker material can enter through the technical-reference path; generic blogs and parts-sales pages remain discovery leads. References must be fetched in full, and their vehicle, complaint or DTC, claim, and deciding test must match the document text. A real bulletin for the same model but a different problem is discarded. Before a forum or video can influence the diagnosis, CarSignal reads the complete thread or a complete supported transcript and verifies that the same reporter documented the matching complaint, deciding test and result, confirmed fault, repair, and post-repair outcome. Each material claim must match text in the fetched source. Incomplete threads, partial transcripts, mixed-commenter stories, search snippets, missing transcripts, and “this fixed mine” replies are discarded. Repeated versions of the same finding are collapsed instead of counted as extra evidence.

What comes back

• Ranked hypotheses, each with a working probability
• A recommended repair only after the evidence supports quoting
• The next test to run, with what a pass or fail would prove
• An evidence chain: the "because" behind every claim
• An expandable Evidence panel with the retained bulletins, articles, forum threads, and videos
• Suggested parts and labor only after the diagnosis becomes quote-ready, ready to become an estimate

The session stays conversation-first: log test results and observations, answer its follow-up questions, and watch confidence move. Alternatives and sources stay collapsed until you open them. The same evidence remains attached as the session progresses. Repair Buddy contributes licensed DTC/manual data when available; it is not presented as an OEM diagnostic tree.

The trust gate

A diagnosis is not quote-ready until a confirming test has been run on this car, and quotes will not send below 70% confidence without an explicit override. If the recommended repair contradicts the top-ranked cause, CarSignal fails closed and keeps the diagnosis in testing. Web and community sources guide test order; they never confirm a fault on the vehicle by themselves.