references/rubric.md

# Learning rubric

Score evidence of unaided reasoning, not polish produced with assistance.

## Outcome labels

- **Independent**: correct with no conceptual hint at all.
- **Lightly assisted**: correct after hint level 1, 2, or 3.
- **Heavily assisted**: correct after hint level 4 or 5.
- **Walked through**: answer or decisive reasoning supplied at level 6.
- **Incomplete**: no correct result within the timebox.

Keep the highest hint level even if the final code is excellent.

A correct first answer cannot be recorded as heavily assisted or walked through. Conventional AI assistance cannot be recorded as an independent outcome or as an unaided transfer score.

A level-1 diagnostic question is a conceptual hint, so it costs the independent label. Only level 0 — task clarification and incidental syntax or toolchain help — preserves it. That is the same bar the transfer attempt uses in [session-protocol.md](session-protocol.md), so the two measures stay comparable.

## Explain-back score

| Score | Evidence |
|---:|---|
| 0 | Cannot explain the mechanism |
| 1 | Repeats steps without causal reasoning |
| 2 | Explains the core idea but misses an important edge or tradeoff |
| 3 | Explains invariant, correctness, and main tradeoff accurately |
| 4 | Also predicts failure modes and generalizes the pattern |

## Transfer score

| Score | Evidence |
|---:|---|
| 0 | No viable start without help |
| 1 | Recognizes pieces but applies the wrong model |
| 2 | Correct model with gaps or a significant bug |
| 3 | Correct unaided solution and explanation |
| 4 | Correct solution plus explicit comparison to the original pattern and tradeoffs |

## Session result

Use this compact template:

```text
Resultado independente: independent | lightly_assisted | heavily_assisted | walked_through | incomplete
Maior pista conceitual: 0–6
Confiança antes do feedback: 1–5
Calibração: calibrated | confident_wrong | uncertain_correct
Explicação de volta: 0–4
Transferência sem ajuda: 0–4 | não aplicada
Política da transferência: strict_unaided | standard_unaided
Tempo: N min
Capacidade: capability_id
Fase: baseline | practice | retention_7d | retention_21d | final
Política de ajuda: strict_unaided | standard_unaided | coached | conventional_ai
Avaliador: coach_scored
Pacote do desafio: package_id
Evidência observada: ...
Próximo alvo: ...
Revisão sugerida: YYYY-MM-DD
```

The transfer attempt belongs to the session that produced it and is always unaided, even when the session itself was coached. Record it as a session field with its own policy; it is not a separate phase. Leave it unrecorded when no transfer task was given rather than entering a zero.

Do not aggregate a single session into a percentage of “mastery.” For trends, compare repeated unaided transfer and delayed-retention attempts on the same skill family.

Do not label an evaluator as independent. Nothing in this package isolates scoring from the coaching context; record `coach_scored`. Deterministic tests establish behavior for covered cases; they do not independently establish reasoning quality or complete correctness.