Olea component workbench

Olea component workbench dev tooling

The product's own screens, running outside Obsidian against an invented practice vault.

What this shows

Everything below is the real product code — the same screens a student would see and answer, driven end to end. Two things about it are not real:

The material is invented. Coined words and made-up courses. Never a real student's notes.
The answers are pre-recorded. Real model calls, made once and saved. Every render since replays them, so clicking around costs nothing.

One consequence worth holding on to: because the material is nonsense by design, none of the numbers here — rankings, match scores, generated questions — say anything about how well Olea works. They show that it works at all.

Not built yet

The list below is everything Olea is planned to do that does not exist yet — not here, not anywhere in the product. There is nothing to click into for any of it. It is listed so it is obvious these are missing on purpose, not broken.

Explain-back — the student explains a concept in her own words and the AI grades the answer. Phase 4
Misconceptions — the AI notices patterns in what she keeps getting wrong. Phase 4
Mastery overview — a Today-panel summary view of what she has and has not mastered (F6.2). The mastery rollup itself now runs on the Oracle surface below; this is specifically the Today-panel summary card. Phase 4

Review states

Today panel states

Due summary, per-course counts and streak (F6.1) only. The mastery overview on this same pane (F6.2) is Phase 4 and does not render here — see "Not built yet" above.

Oracle states

The ranking, study-plan and gap/coverage chain (F4.2, F4.3, F4.5, F4.9, F4.10), driven by a synthetic curriculum and corpus over the same coined vocabulary a persona's stream already uses — never the fixture vault. Dev tooling only (D-021); zero model spend.

Retrieval states

Query -> hybrid retrieve -> grounded context or refusal (C4.7, INV-5), replaying a pre-computed embedding cassette over a synthetic corpus — zero model spend at render time. No product view exists for raw retrieval results, so this surface reports through the inspector only; see its own note there.

Generation states

Retrieve -> generate -> accept (C4.7, INV-5, INV-6), replaying a pre-computed generation cassette over a synthetic corpus — zero model spend at render time. The "Accept" control is a REAL gate: nothing becomes vault-ready `McqFields` until it is clicked. No product view exists yet for reviewing a generated card (P3-T07b's accept step is still open in the product), so this surface reports through the inspector plus one live control; see its own note there.

Timeline states

The day loop (`ol-opmb.5` [TB-4]): the SAME mastery -> rank -> plan -> gap -> queue chain as the oracle states above, replayed across a simulated semester rather than at one fixed instant. Time is an axis, not another state — scrub the day with the controls at the top of the inspector. Zero model spend; dev tooling only (D-021).