LAB PROTOCOL · CONTEXT-BUDGET BENCHMARK
Agent Instruction-Budget Allocation: Context Fidelity Under Pressure
A 20-case protocol for testing whether an agent preserves task-critical instructions when system rules, user requests, retrieved material, tool output, and examples compete for a finite context budget.
Why this benchmark matters
A long prompt can fail without a single instruction being disobeyed: critical constraints may simply be displaced, truncated, or overshadowed before the agent begins the actual work.
Scope note: This protocol defines a test design, not a performance claim. Results depend on implementation, model configuration, permissions, tool behavior, and the test fixtures used.
Measurement matrix
| Test surface | What to observe | Evidence-led assessment |
|---|---|---|
| Budget accounting | System, user, retrieved, tool, and output tokens | Record allocation and identify which source displaced a task-critical constraint. |
| Truncation order | Early versus late constraints under pressure | Score whether requirements survive according to a declared priority policy. |
| Oversized examples | Few-shot material that crowds out task context | Measure task fidelity and omission rate with and without the examples. |
| Partial delivery | Tasks that cannot fit safely in the remaining budget | Score whether the agent states the limit and preserves the highest-priority deliverable. |
| Late constraints | Requirements introduced near the context boundary | Check whether the agent acknowledges, retains, or safely escalates the constraint. |
Protocol steps
- Construct 20 fixtures with known task constraints distributed across system, user, retrieved, and tool-result context.
- Vary total token pressure, insertion order, example size, and output reservation while holding the requested task constant.
- Score task fidelity, constraint omission, unsupported completion claims, and the quality of any partial-delivery explanation.
- Inspect summaries and compaction artifacts to distinguish a missing instruction from an instruction that was retained but misapplied.
- Publish only aggregate findings with fixture version, model configuration, and clear limitations.
Lab disclosure
This page was developed from a July 22, 2026 lab-bench brief supplied by the Content Site Consultant. It is a proposed evaluation protocol, not an independent benchmark result or product-security certification.