context-benchmark

Context Benchmark

Designed, integrated, independently refactored, and continuously maintained by TIKAZ.

Inputs

Accept a versioned benchmark manifest, fixed source fixtures, declared budgets, protected facts, expected anchors, optional route labels, and an optional externally scored downstream-answer rubric. Use identical inputs and settings when comparing systems.

Workflow

Run a versioned manifest of independent cases and keep raw per-case results. Report efficiency and quality separately:

• source and packed tokens;
• final-budget compliance;
• protected-fact recall;
• evidence-anchor correctness;
• deterministic repeatability;
• preparation runtime;
• optional externally supplied answer score.
• document-route correctness, informative-visual recall, decorative/duplicate skip accuracy, and complex-table fidelity warnings for multimodal fixtures.
• retrieval Recall@K, Precision@K, MRR, nDCG@K, evidence-facet coverage, and explicit insufficient-evidence fallback on independently labeled cases.
• dynamic-selection count, selected-evidence recall and precision, facet coverage, and estimated selected-token reduction; never present reduction without the corresponding selection recall.

Do not hide failures inside averages. A smaller pack with lower fidelity is a regression, not a win. Do not claim superiority until the same files, questions, model/detail settings, budgets, and blind answer rubric are used. Use the shared CLI benchmark command and read ../references/benchmark-method.md when publishing results.

Output contract

Publish summary.json, metrics.json, raw cases.json, and the generated evidence card together. Keep context efficiency, exact-repeat prompt efficiency, literal fact and anchor fidelity, multimodal routing, and pending provider, vision, or downstream evidence separate; never replace them with one composite fidelity score.

Validation and fallback

Keep failed cases visible and verify manifest version, fixture identity, budgets, settings, and denominators. Estimated tokens must be labeled estimates. If provider telemetry or blind downstream scoring is unavailable, mark it Pending; do not infer superiority from local fixtures.