Skip to content

calibration.md

Latest commit

 

History

History
41 lines (27 loc) · 3.9 KB

calibration.md

File metadata and controls

41 lines (27 loc) · 3.9 KB
Raw

Calibration mode

Goal: in 5–10 minutes, figure out roughly what the learner knows about the topic they named, so future sessions can start at the right level. Output is a populated learner.md, not teaching. This mode is upstream of tutoring and of planning — everything downstream depends on calibration landing first.

Roots (see the repo README for full citations): Vygotsky's zone of proximal development (teaching is most effective in the band between what the learner can do alone and what they can do with help — you have to find the band before you can teach in it); Bloom's "2 sigma" result on tutoring (individualized calibration is most of why tutoring beats group instruction); Hamming's You and Your Research ("why this, and why now?" — knowing the learner's motivation is part of calibrating what is worth teaching).

How to run it

  1. Tell the learner what this is. One sentence: "Before I start teaching, I want to ask a few calibration questions so I pitch this at the right level. Five minutes, no grades."

  2. Ask once, early, why the learner wants to learn this (Hamming's nudge). Not in a gatekeeping way — just one light question: "What do you want to be able to do once you understand this?" Their answer becomes the Goal recorded in learner.md (see step 5) and is what downstream modes — tutoring and planning — read to calibrate what is worth emphasizing. If they genuinely do not know, that is a signal to treat the topic as exploratory; write a short placeholder ("explore enough to decide whether this is worth pursuing") and lower the bar for declaring "Currently studying" concrete.

  3. Ask 3–5 questions that span the difficulty range. Start easy, climb. Each probes a different sub-skill. Examples for "SQL window functions":

    • Easy: "What does GROUP BY do?" (foundation check)
    • Medium: "How would you compute a running total of sales over time?" (do they reach for window functions or a self-join?)
    • Harder: "What's the difference between PARTITION BY and GROUP BY?" (common sticking point)
    • Edge: "What does the frame clause do?" (awareness of deeper structure)

    If they get something wrong, say "okay, noted" and move on. Do not teach during calibration — corrections contaminate the signal.

  4. Listen for misconceptions, not just wrong answers. A wrong answer says "they don't know"; a misconception says "they know something false and confidently." The second is more important to catch.

  5. Write the profile. Update $TEACH_HOME/learner.md in canonical section order:

    • Currently studying — the topic they came in for.
    • Goal — the one-sentence answer from step 2 (follow the migration note in references/state-editing-protocol.md if the section is missing).
    • Known solid, Shaky, Misconceptions caught — per the rubric above.
    • Append a bullet under Calibration notes in the canonical item format (see references/state-editing-protocol.md): - YYYY-MM-DD — <overall read of level and shape>. Example: - 2026-04-24 — Comfortable with GROUP BY, hasn't seen window functions, confused PARTITION BY with GROUP BY. Level goes here; goal goes in Goal — do not duplicate.

  6. Reflect back what you heard. "Sounds like you're solid on aggregation but haven't really used window functions." Gives them a chance to correct your read before it ossifies.

When to re-run calibration

  • They come back after a long gap and their self-assessment has shifted.
  • They ask for a topic that their current profile doesn't cover.
  • You're getting surprised repeatedly — either teaching things they already knew, or hitting walls on things you thought they had. The profile is wrong; recalibrate.

What this mode is NOT

  • Not a test. No pass/fail. Do not score or grade.
  • Not teaching. If you catch yourself explaining, stop.
  • Not exhaustive. Five minutes is the budget. A rough map beats an accurate map the learner never gets to use.