Source-led context comparison
This is an independent experimental publication. Navigation, evidence delivery and legal answerability are separate.
The report covers 40 unchanged staff question occurrences in each stage and budget, plus a separate unknown-question control.
| Stage | Budget (bytes) | Cases with evidence | Cases with requirements | Cases with relationships | Insufficient | Retrieval truncated | Assembly truncated |
|---|---|---|---|---|---|---|---|
| logical-context | 32768 | 6 | 8 | 0 | 40 | 40 | 40 |
| logical-context | 524288 | 40 | 8 | 18 | 40 | 40 | 40 |
| structured-context | 32768 | 0 | 40 | 0 | 40 | 40 | 40 |
| structured-context | 524288 | 40 | 40 | 40 | 40 | 40 | 40 |
Reading the measures
- This summarises retained assemblies; it does not execute retrieval or a model.
- Cases are known development questions. No independent relevance, claim accuracy or legal acceptance is measured.
- Path and route totals are case incidences, not distinct legal rules; shared routes can occur in several questions.
- Restored and inferred location routes remain separate from requirement support. Selected mapped units may have arrived lexically.
- Original open obligation identifiers are deduplicated, not summed across questions or budgets.
- Byte ranges describe this bounded delivery. They are not an affordability or controlled speed benchmark.
The machine-readable companion retains every case, separate requirement planes, restored and inferred routes, original open obligation identifiers, source bytes and package bytes. A larger route count is not proof of a supported answer.