NIXFRED WORKBENCH

Operate Released v0.1.0

Failure Investigator

Work backward from a bad output to the mechanism that produced it, then write the fix down.

Runs in your browser. Nothing you paste here is transmitted, logged, or stored unless you explicitly save it to this device. Treat the page like a local scratchpad.

Describe what happened

This orients the walkthrough. It does not filter which mechanisms can rank.

Nothing you paste below leaves this browser tab. It is never transmitted, logged, or stored anywhere unless you explicitly export it to this device with the export bar further down.

Answer what you know

Answer as many as you can, in any order. A skipped question simply carries no weight. Answering more narrows the ranked list below, question by question, rather than all at once.

Try it two or three times with identical input before answering.

This is the single fastest way to rule sampling randomness in or out.

A budget or positional limit only shows up once there is enough length to hit it.

Pick the shape that matches best, even if more than one seems close.

Compare an early turn against a late one in the same session if you can.

Pick the closest match. Pick nothing changed if you genuinely do not know of one.

A web page, an uploaded document, or a tool result all count.

Includes retrieval systems, since a retriever is a tool call from the model point of view.

This is the question that tells a retrieval miss apart from an invented detail.

Ranked hypotheses

Scored against the answers above. Every entry shown carries at least one matched piece of evidence, for or against it. This tool proposes containment steps below; it never performs them. Select a hypothesis to draft the postmortem around it.

Idle

Stale Inputs changed after this result was produced. Run it again to see numbers that match what is on screen.

Postmortem draft

Symptom, mechanism, evidence, fix, and the regression test that would catch it next time.

Idle

Stale Inputs changed after this result was produced. Run it again to see numbers that match what is on screen.
Note

This tool proposes. It never performs.

Every containment step and every fix below is a proposal for a person to weigh and carry out, never an instruction to run without thinking. Nothing on this page executes a remediation, calls a tool, or changes any system on its own. A step marked destructive says so plainly, in words, and states what it costs to undo it, not only that it can be undone.

Prefilled with a suggested test for the selected mechanism. Edit it to match your own test suite.

Your own running notes. Carried into the export as written.

Export

Failure mechanism catalog

Every mechanism this tool knows about, browsable on its own, independent of the investigation above.

Idle

Stale Inputs changed after this result was produced. Run it again to see numbers that match what is on screen.
Note Why will this not just tell me the cause?

A questionnaire is evidence, not a debugger attached to your system. Two mechanisms often produce the same visible symptom, a stale index and a stale cache both read as a confident wrong answer, and the only way to fully tell them apart is the check named in the next diagnostic field, not another multiple choice question.

So this tool ranks instead of concludes. A hypothesis with strong evidence and no contradiction still carries a next diagnostic, because a ranking built from a handful of answers is exactly that, a ranking, and treating it as certainty is how a team fixes the wrong thing and ships the same incident again.