上下文比规模更重要

Context Budgeting in Practice

Every IELTS diagnostic session carries user history: prior mock bands, recurring penalty patterns, target immigration threshold. Dumping the full transcript into every request dilutes signal and slows feedback.

We rank context by recency and penalty salience — a Band 6.0 learner stuck on cohesion gets cohesion-focused examples and their recent flagged responses, not generic Band 9 essays.

Context Prioritization Framework

Think of context as stacked layers, each serving a distinct job. Public rubric language grounds the model in how IELTS actually scores. Learner history surfaces recurring penalties. Recent attempts keep feedback specific. Skill exemplars show what improvement looks like at the next band.

{
  "context_layers": [
    {"layer": "public_rubric", "purpose": "Anchor scoring to IELTS band descriptors"},
    {"layer": "learner_penalty_history", "purpose": "Surface recurring score-limiting patterns"},
    {"layer": "recent_attempt", "purpose": "Keep feedback specific to this mock"},
    {"layer": "skill_exemplars", "purpose": "Show what the next band looks like in practice"}
  ],
  "principle": "Curate before you scale — relevance beats raw model size"
}

When Scale Still Matters

Scale helps fluency and multi-step reasoning — but only after context is curated. Tight retrieval with focused context often beats a larger model fed unstructured noise.

Context engineering is the unfair advantage most teams skip because it isn't tweetable.

Designing Context Budgets

When tokens are scarce, every sentence in the prompt must earn its place. Band9AI prioritizes public band-descriptor language, the learner’s latest attempt, and the specific task constraints over generic motivational filler.

Larger models help, but they do not excuse lazy context. Domain-specific retrieval and instructions often move accuracy more than jumping a model tier — especially when latency and cost still matter.