Skip to content

Track 22

Theory Checkpoint Sheets

This track is a review layer, not a first theory course. Use it to rehearse core model families cold, connect each family to a visible pattern, and route yourself back to the weak topic when the explanation breaks.

Examples Solved Questions Topics Study Plan

Primary Goal

Explain The Family Cold

The point is not to recite definitions. The point is to connect a model family to a toy pattern, a failure mode, and the simplest baseline that still matters.

Best For

After The Core Workflow Exists

Use this when the workflow is familiar but the explanation still collapses under timing pressure or oral questioning.

Exit Rule

One Sharp Explanation Per Family

You are done when you can say the assumption, the revealing pattern, and the failure mode in under 30 seconds without hiding behind a score.

Use This Track When

  • you already know the core workflow and need review, not first exposure
  • your solved-question answers are still vague or too long
  • you want a compact bridge from theory recall into timed drills or IOAI-style revision
  • you can run the workflow but still struggle to explain why one family fits better than another

What This Track Is Training

This track trains one practical rule:

  • explain the family, the revealing pattern, and the failure mode before you mention the score

That means every good checkpoint answer should include:

  • the family assumption
  • the toy pattern or artifact that reveals it
  • the failure mode that breaks the story
  • the simplest baseline that should still be named

First Session

Use this order:

  1. run academy/.venv/bin/python academy/examples/theory-checkpoint-recipes/family_checkpoint_demo.py
  2. run academy/.venv/bin/python academy/examples/theory-checkpoint-recipes/calibration_checkpoint_demo.py
  3. inspect the summary table and checkpoint cards
  4. answer three prompts cold before opening notes
  5. write one line on which family explanation broke first

Full Track Loop

For the complete workflow:

  1. run the two checkpoint demos from repo root
  2. run academy/.venv/bin/python -m pip install -r academy/labs/theory-checkpoint-sheets/requirements.txt
  3. run academy/.venv/bin/python academy/labs/theory-checkpoint-sheets/src/theory_checkpoint_workflow.py
  4. inspect the comparison table and prompt cards before rereading topics
  5. finish the matching exercises in academy/exercises/theory-checkpoint-sheets/
  6. route back to the weakest topic only after the cold answer fails

What To Inspect

By the end of the track, the learner should have inspected:

  • checkpoint_summary.csv for which family wins where
  • family_comparison.csv for side-by-side assumptions and pattern fit
  • checkpoint_cards.md and concept_prompts.txt for short cold prompts
  • phase_extension_prompts.txt for later-family review without pretending everything needs a full toy benchmark
  • checkpoint_comparison.png only after the written summaries are already clear

Common Failure Modes

  • giving a definition without naming the revealing pattern
  • quoting a score without saying which setup produced it
  • calling a family "better" without a failure mode
  • overreading an embedding or geometry view
  • treating calibration or threshold logic as optional when it controls action
  • reaching for advanced family names when a simpler baseline already fits the evidence

Exit Standard

Before leaving this track, the learner should be able to:

  • explain the main model families cold
  • connect each family to a visible pattern or toy setup
  • name one failure mode for each
  • distinguish a score from the story behind the score
  • route themselves back to the weakest topic without guessing