ADR-001: Pivot Standard Trainer To Calculation-Path Diagnosis

• Status: Accepted
• Date: 2026-07-14
• Owners: Product and engineering

Context

The original Dynamic Equilibrium short-answer prototype centred on missing-word detection, dangerous-phrase repair, Standard Nodes, and a forced rewrite. The exploratory implementation is preserved in standard-trainer-demo PR #1 and commit 83b5dff. It established useful workflow and evaluation lessons, but it is not the active product direction.

Separately, the Memorization Bank is an active-recall and spaced-repetition product in student-site. Combining it with Standard Trainer would blur product purpose, data ownership, and evidence.

Decision

1. Standard Trainer and Memorization Bank are separate products.
2. Standard Trainer is the flagship product.
3. Standard Trainer focuses on CAIE 9701 Chemistry calculation-path diagnosis and targeted practice.
4. Equilibrium Calculations is the first curriculum module.
5. Kc/Kp is the first engineering epic.
6. KP_FROM_EQUILIBRIUM_MOLES is the first runnable calculation archetype.
7. Memorization Bank remains independent in student-site.
8. Standard Trainer will not replace, migrate, embed, or inherit the Memorization Bank MVP, its routes, IDs, or learning state.
9. The former short-answer prototype is a superseded exploratory spike.
10. Curated, versioned problems, solution graphs, and marking policies are the initial source of truth; RAG is not the core MVP engine.

Consequences

• Active specifications, fixtures, course manifests, portfolio examples, and interview stories use the calculation-coach definition.
• The first slice accepts a curated structured problem and structured student steps; it does not parse arbitrary questions.
• Deterministic tools own arithmetic and marking operations. LLM assistance is optional and bounded.
• The term “agent” is withheld until traced runtime routing exists.
• Git history plus this ADR is the archive for superseded product assumptions.

Retained Infrastructure

The repository boundary, sibling demo repository, source-authority discipline, local learning console, canonical Day Markdown, answer-first reveals, response storage, progress, reflection, export, and auditable review-status boundary remain valid.

Risks And Revisit Conditions

The key risk is encoding an incorrect solution graph or marking policy with false confidence. All initial content remains AI_DRAFT until qualified review. Expansion beyond the first archetype requires passing eval gates, three learner sessions, and an accepted plan grounded in evidence.