AI Tutor exercise coaching integrity policy (no final answer in active attempt)

DEC-0084 - AI Tutor exercise coaching integrity policy (no final answer in active attempt)

Context

A core user requirement is that AI Tutor should guide learners during practice but never do the work for them. Current discovery graph does not lock this policy for active attempt flows.

Decision

Active-attempt integrity rule:

• While learner is in active assignment/test/exercise attempt, AI Tutor must not provide:
  • final answer,
  • answer key option,
  • full ready-to-submit rewrite.

Hint ladder contract (broad -> detailed):

1. orientation hint (what to focus on),
2. concept hint (rule/pattern to apply),
3. scaffolded steps (how to solve),
4. self-check prompt (learner validates own answer).

Repeated direct-answer request behavior:

• AI Tutor restates integrity policy briefly and continues at next allowed hint level.

Post-attempt mode:

• After submission/finalization, AI Tutor may provide full worked example labeled as reference learning material.

Decision Value

• Preserves learning integrity and prevents answer-leak behavior.
• Aligns AI Tutor with teacher-support boundary in course mode.
• Keeps AI useful without replacing learner effort.

Rationale

Guidance-first with a deterministic hint ladder keeps support quality high and predictable, while enforcing a hard boundary against answer outsourcing.

Implications

• Product/UX impact:
  • chat surfaces need explicit “guided mode” behavior during active attempts.
• Data/logic impact:
  • AI runtime must detect active attempt context and switch answer policy.
• Operational impact:
  • support and QA teams get a clear policy for violation audits.

Alternatives considered

• Option A: allow full answers when user asks repeatedly.
• Option B: strict no-final-answer during active attempt with hint ladder (selected).

Open follow-ups

• Resolved by DEC-0092: sandbox exception only for non-graded/no-impact contexts with explicit sample/reference label.

Decision quality check

Decision quality check: DEC-0084

• Score: 12/12
• Weak dimensions: None
• Action: Promoted.