censuslisted

# /census — aggregation over the corpus Answer an aggregating question by traversing the corpus, not by retrieving from it. ## Why this exists The step-1 benchmark (2026-08-13) measured the incumbent retrieval path against 15 code-computed truths. On the aggregating class the retrieval ceiling was **0.000**. All seven questions scored exactly 0.00. That is not a low score but an unreachable one. The answer sits in no single chunk, so top-K cannot return it at any depth. `/census` measured **6 of 7** on the same questions and the same corpus state. Median 0.05 s per answer, against 0.92 s for retrieval. The one miss is a known wording debt, not a capability gap. agg-09 names a reference list of countries that the traversal never receives. It therefore measures whether the list was guessed. It stays wrong, because the alternative was a program rewritten after the answer was known. Both numbers live in `outputs/operations/census-bench/` in the data overlay. ## When NOT to apply | Signal | Route to | | --- | --- | | "what do we know about X", "where did we decide Y" | `/recall` | | The corpus scope fits the context window | `/recall` — the engine REFUSES with exit `4` | | An exact string in a known file | `Grep` | | A new note to capture | `/zk` | | Advice from the Odin brain | `/odin` | | Anything about a person or company outside the workspace | `/osint` | The window refusal is not advice. SRLM (arXiv:2603.15653) reports that a traversal primitive on a corpus that alre