Case Study: Standard Trainer Calculation Coach

Problem

Cambridge A-Level Chemistry learners can reach a wrong Kp result through many different paths. Final-answer feedback does not reveal whether the earliest failure was chemical, mathematical, dependency-related, or presentational, and generic AI feedback may invent arithmetic or marking logic.

Product Thesis

Standard Trainer reconstructs structured working, aligns it to a canonical solution graph, identifies the first invalid step, applies a frozen marking/ECF policy, and chooses targeted practice. The first slice is one KP_FROM_EQUILIBRIUM_MOLES problem.

Decisions

Chose calculation-path diagnosis over broad tutoring.
Used curated versioned content instead of arbitrary parsing or runtime RAG.
Assigned arithmetic and marking authority to deterministic tools.
Separated mathematics, chemistry model, dependencies, policy, ECF, hints, and learner modelling.
Withheld the agent claim until runtime routing and traces exist.

Delivery

Days 4-7 stage the graph/workspace, diagnosis/ECF, traced orchestrator, and targeted follow-up loop. Day 8 consolidates. Day 9 captures three sessions and one bounded fix. Day 10 turns measured evidence into portfolio material without adding features.

Evidence To Insert

Demo PRs and commits: [pending]
Eval dataset/result: [pending]
Chemistry/curriculum review: [pending]
Three learner sessions: [pending]
Bounded change and regression: [pending]
Demo/deployment: [pending]

Claim Boundary

The current fixture and marking policy are AI_DRAFT. No CAIE marking-agreement, learning-outcome, agent, or production-readiness claim is available yet.