PCB debugging on a real bench

The board, the capture and the verdict in one window

Left: every net in the board graph, coloured by what YProbe knows about it. Centre: the capture. Right: each parameter judged against the expected signal, a signal-health score, and the ranked fault hypotheses. The strip under the toolbar is the assistant's next move — “Probe c:EN/sync_5V4 next (4 still open).”

The dialog is YProbe arguing with the expectation rather than the board: 2 MHz was asked of a node whose RC corner sits near 25 kHz (≈ 102 kΩ source, ~62 pF of probe and load). A clean square is not physically realizable there, so a low-amplitude capture is physics — not a defect — and you get told before it becomes a false FAIL.

And that warning is completely deterministic. Nothing was asked of a model: the source impedance comes out of the resistor network in your netlist, the capacitance is the probe and load, and the corner is arithmetic. Same board, same expectation, same answer, every single time.

Every conjecture on the record — and you can veto it

Each root-cause conjecture the session has formed, ranked by confidence, with the evidence tally that got it there and the nets and components it implicates. Nothing is a black box you have to take on faith.

Select one and it tells you how to settle it at the bench: tie EN/sync_5V4 to a known-good level through 10 kΩ, or lift the Q2 pin — each with what the outcome would mean. Record the result and the ranking moves. Refuting is a first-class action: your veto is stored as evidence and the AI stops pursuing that branch.

A bad probe touch is never a board failure

This capture never became a verdict. The classifier called the trace floating, the status bar says so in red — “No probe contact — capture not used” — and the capture was dropped: no PASS/FAIL, nothing written to the session record, nothing handed to the AI as board evidence.

That is the difference between a tool that measures and a tool you can trust with a diagnosis. A slipped tip that reads 0 V looks exactly like a dead rail, and it will cost you an afternoon.

Bus traffic decoded straight from the analog capture

No separate logic analyser, no second instrument. The Logic tab takes the captures you already have and decodes them — here two channels of UART at 19200 8N1, annotated with the bytes that came out.

And it reports what actually happened rather than what you hoped for: the lower trace decodes to seven bytes with three framing errors. Decoded traffic feeds the diagnosis, so a NACKing sensor or a garbled frame is evidence, not just pixels on a screen.

Follow the milliohms down to the shorted pad

A dead short lives on a pad, not on a net. Short Hunt lists the candidate pads on the shorted net — sixteen here — walks you through them one at a time, and ranks every reading you take: the lowest one is electrically closest to the short.

Read the blue box. It tells you plainly that a 0.1 Ω handheld meter reads ~0 Ω on every pad and cannot rank them, and that you need a 4-wire bench DMM or a low-Ω tracer. You open this tool yourself from the Analysis menu — the assistant never sends you here, because most benches don't have the meter it needs.

Build the signal you want to test against

Pick a base waveform, stack impairments on top of it, couple the channels together, and push the scenario at the simulated scope. Here CH1 is a 15 V rail carrying 10 mV of noise, CH2 is set to play back a waveform file exported from another instrument, and 20 % of CH1 crosstalks into CH2.

This is why you can evaluate YProbe with nothing on the desk. The entire workflow — capture, judge, decode, monitor, diagnose — runs against generated signals, including faults that only misbehave once every few minutes.

Exactly what your contract manufacturer sees

A signed task package, opened in the free technician client. Probe points, expectations, and a local hint — and nothing else. The nets are aliases: NET_001, NET_002. Your netlist, your BOM and the AI's reasoning are not hidden in this package; they were never written into it.

The client can hint but never decide — “Final judgement is the designer's”. Results come back signed and replay-protected, and the verdict is recomputed on your machine from the raw samples before it counts.