Learning loop¶

The loop is the main Orion idea.

flowchart LR
  goal[Goal] --> map[Approved map]
  map --> preview[View quest details]
  preview --> teach[Begin teach-back]
  teach --> evidence[Evidence and score]
  evidence --> decision{Learner decision}
  decision -->|confirm or correct| memory[Durable memory]
  decision -->|defer| pending[Hypothesis stays pending]
  decision -->|dismiss| closed[Closed gap]
  memory --> retrieve[Restart and retrieve]
  retrieve --> changed[Changed next mission]
  changed --> teach
  pending --> practice[Optional bounded practice]
  practice --> teach

Map¶

The learner gives Orion a goal. Orion proposes nodes and relationships. The learner approves the map.

Preview¶

Each quest has its own mission ID, title, description, prompt, audience, and scaffold. A preview is not a teach-back.

Teach¶

The learner explains the selected quest by text or editable local Whisper transcript. The response is an attempt.

Evidence¶

The evaluator maps the response to required moves. The result includes a score, status, summary, and next move. Model output is typed and bounded.

Decision¶

The learner decides whether the proposed gap is useful. Confirm and correct make the gap searchable memory. Defer keeps it as a hypothesis. Dismiss closes it. This prevents a model from turning a weak inference into a durable learner label.

Memory¶

Confirmed memory is stored in CockroachDB. Its embedding is stored for filtered retrieval. S3 stores the Passport artifact and its lineage.

Return¶

On restart, Orion retrieves only memory that matches the learner, concept, status, and consent filters. The next mission can change because the learner approved the memory.

Bounded growth¶

Orion does not create an unlimited star field.

One new practice proposal per teach-back at most.
Three approved practice nodes at most in the active map.
A repeated proposal reuses an existing slot.
An accepted teach-back does not create a practice star.

If the learner remains unsure, Orion can prepare one bounded practice proposal. It stays outside the active map until the learner approves it.

The graph can preserve history without making every historical item an active star.