Not a blame chart. Each axis is a lever on the future: how much would a change at this level keep this failure from recurring? Drag a lever out = more leverage there. The shaded shape is the failure's leverage fingerprint.
The lab's humans — who chose the corpus, the RLHF, the architecture, and shipped the weights.
Lever: a differently-built model.
Made of humans making choices — and the axis no industry taxonomy will name. Adverse-to-deflation by design.
The dice at runtime — the same prompt resampled.
Lever: re-run, self-consistency, an output gate.
A thing, not a person — no persona. High only for genuinely run-unstable failures.
The team that built the deployment around the model.
Lever: the RAG corpus, system prompt, tool wiring, guardrails, model-version choice.
A team with a persona. Distinct from B. (The "K4" finding that earned this axis.)
The person who prompts and operates it — the doctor typing the query.
Lever: a better-formed request, added context, a user-side check.
A person with a persona. Distinct from D.