01
Context Wings
Prompt material, retrieved passages, tool output, memory, and hidden assumptions waiting just offstage.
Room map
The taxonomy is arranged by what a concept changes.
The Stage Index is the front-of-house map. It keeps LLM concepts grouped by effect rather than alphabet: what enters the answer, what the model does with it, what evidence supports it, how it is judged, and what happens when the first answer fails.
02
Model Floor
Capability claims, latency pressure, refusal behavior, tone control, and family-level differences under the same prompt.
03
Grounding Gallery
Source trails, citation habits, live search, freshness signals, and the point where evidence turns into decoration.
04
Evaluation Pit
Rubrics, red-team checks, regression examples, human review, and the uneasy space between score and usefulness.
05
Repair Exit
Follow-up prompts, tool retries, fallback models, escalation rules, and the moment an answer admits it needs help.
