A small logbook of sketches, notes, and unfinished ideas — the rawer layer under the writing.
| FILE | KIND | DATE | SIZE |
|---|---|---|---|
| agent-routing-sketch.md Agent routing sketch | sketch | 2026-09-05 | 1.2 KB |
| design-trace-v2.md Visual system v2 notes | note | 2026-09-05 | 0.7 KB |
| eval-rubric-notes.md Rubric vs score notes | note | 2026-09-04 | 0.9 KB |
| context-window-idea.txt Idea: show what fell out of context | idea | 2026-09-03 | 0.4 KB |
Idea: classify each user question into "needs a tool" vs "answerable from context" first.
The best route is rarely the cleverest route — it is the one you can evaluate.
(Demo fixture for the prototype — not a real note.)
Living AI Trace: USER → RETRIEVE → AGENT → TOOL → EVALUATE → OUTPUT
Rule: nothing important may depend on the visual — the SVG is enhancement only.
See DESIGN.md and docs/redesign/10-living-ai-trace.md
(Demo fixture for the prototype — not a real note.)
The same test set scores differently under a different rubric.
Lesson: never report just the pass count — report the rubric too.
See EXP 003 in the LAB for the simulation.
(Demo fixture for the prototype — not a real note.)
If the UI said "this document just fell out of the context window", users would guess less.
Built as EXP 004.
(Demo fixture for the prototype — not a real note.)