Methodology
A guided loop, and the machine behind it.
Debrief is not an evaluation tool with a score at the end. It is a learning loop grounded in retrieval practice and generative learning: explain, find the break, teach to it, apply the fix.
The loop
Explain
You write the idea from memory, in your own words. Retrieval practice: no notes, no multiple choice. The act of producing it is where the learning starts.
Probe
Debrief maps the claims your explanation must carry, judges each, and asks one curious question aimed at the single weakest point.
Reveal
It shows the exact sentence where your reasoning stops holding, with a short, just-in-time correction. The fracture, made visible.
Repair
You use the corrected idea in a new situation. Application, not repetition, so the fix becomes something you can transfer.
The understanding map
Understanding as a beam through the strata of an explanation.
Every concept decomposes into the claims a complete explanation must carry. Each claim is a stratum; a single beam threads them top to bottom. Its state at each crossing is what your explanation is actually holding: clean, porous, fractured, or not yet reached.
The map is driven entirely by the session state. It is read-only: it never asks for input, it just shows, at a glance, where the understanding is load-bearing and where it breaks. To the right is a real closures debrief, mid-session, with one claim fractured.
Understanding map
Closures
What a strong explanation must carry, and how yours holds.
- Retains surrounding scopesolid
- Captures the variable, not a copyneeds attention
- Each keeps its own environmentunclear
- Powers callbacks & handlersnot reached yet
- solid
- unclear
- needs attention
- not reached yet
What a claim can be
Clearly and correctly explained. The stratum is dense; the beam passes clean.
Mentioned, but vague or partial. The band is porous, and the light diffuses.
The explanation contradicts the claim. The beam refracts: the amber fracture.
Never addressed. A faint, dotted trace the beam hasn't arrived at yet.
The machine
The model proposes. The reducer decides.
Each turn runs a narrow model call, validates the response against a strict schema, then hands the result to a pure state machine that owns every transition. The model never drives the session directly. It only supplies judgments the reducer chooses to act on.
Teaching is anchored on curated notes, so the model adapts wording but cannot invent a wrong correction. When a call fails validation, the turn falls back to authored content rather than faking a judgment.
- decomposeAndEvaluateclaims + states
- selectFocusClaimthe weakest point
- generateProbeone curious question
- evaluateFocusAnswerre-judged claim
- generateTeachingthe distinction
- generateRepairapply it anew
> every response is schema-validated; the pure reducer is the single source of truth for stage, focus, and verdict.
Grounded in
Retrieval practice
Recalling from memory, then checking against the source, beats rereading. Debrief begins with explanation from memory and ends the repair round with application, not a restated definition.
Generative learning
Explaining, visualizing, and applying are complementary sense-making modes. Debrief sequences all three: the explanation, the understanding map, and the applied repair.