Project Story Bank

Pivot Story

Situation: the earlier prototype optimised a bounded answer-repair loop but did not provide the strongest deterministic coaching wedge.
Task: reset the product source of truth without carrying contradictory assumptions forward.
Action: separated the products, selected calculation-path diagnosis, froze one Kp archetype, rewrote specs/fixtures/course content, and added regression guards.
Result: [insert merged PR and validation evidence].
Learning: a pivot is an authority and evidence reset, not a naming change.

Situation: an LLM could produce plausible arithmetic or exam marking.
Action: separated responsibility across typed deterministic tools, question-specific policy, ECF rules, trace verifier, and human-review fallback.
Result: [insert eval and unsupported-claim evidence].

Situation: three learner sessions exposed [observed recurring failure].
Action: selected exactly one bounded fix, added a regression case, and reran the frozen suite.
Result: [insert before/after evidence].