Context Adherence
Chapter: Context adherence
Context engineering gives you control. This chapter is about what happens when the model doesn't use it.
Instruction following
Groundedness
Completeness vs. correctness
Diagnosis → control stack
Final version
Merged chapter
One chapter, nine sections. Practitioner pacing with full science depth.
Final
Context Adherence
The merged chapter. Opens with Air Canada and the Premium-plan scope violation, builds the two-contract mental model, walks the diagnosis loop with worked examples, defines minimum viable evals (including atomic claim decomposition), presents the control stack, and goes deep on grounded data, SFT, and RL with a fully computed reward example. Closes with operational checklists.
- 9 sections: Opening, Mental model, Diagnosis, Evals, Control stack, Grounded data & SFT, RL, Practical moves, Summary.
- 4 figures: Two-contracts diagram, diagnosis loop, data pipeline, cost tiers.
- 6 worked examples: Requirement-level diagnosis, claim-level diagnosis, atomic claims decomposition, four training artifacts, reward computation, escalation decision.
Earlier drafts
Source material
The earlier drafts that fed the merged version. Kept for reference.