# Method, evidence and limitations

## Current population and eligibility

The October 18, 2026 update accounts for 107 canonical records: 73 qualified workflow/practice cases (61 retained baseline plus 12 additions), 7 social process/mechanism references, 4 research leads and 23 excluded records. The `svc-21` alias resolves to `comm-19`. Eight formal/reference sources remain separate methodological prior art.

The [eligibility policy](POLICY.md) separates vendor/customer promotional stories from ordinary operating documentation and firsthand practice. Targeted corroboration searches are not exhaustive. Selected boundary cases were reread; this was not a fresh full-page audit of all 107 sources. Eligibility is not measured usefulness. Six expansion records lacked sufficient documented mechanics: four remain eligible research leads and two are excluded sources.

## Original study: preserved, not retroactively expanded

The first desk study began with 72 workflow records across 12 function labels and 12 social records. Merging a duplicate DACI process left 71 canonical workflows. The original 54 training and 18 held-out coding units are unchanged. Design/research, people/HR and procurement/operations were set aside before the codebook was frozen. This was exploratory analyst coding, not a human study or statistical validation.

The initial 84 records were screened for evidence labels and obvious duplication. An independent check selected 14 cases; 13 primary-source checks succeeded and one was inaccessible. Corrections distinguished recommended from reported checks, predeployment from postdeployment testing and demonstrated steps from suggested human controls. That audit did not verify every claim against every source.

[Original snapshots](audit/README.md), the unchanged [codebook](pattern-codebook.json), [training](training-coding.json) and [held-out coding](heldout-coding.json), and [frozen-study links](frozen-study-links.json) retain historical membership. Some historical sources are now excluded. Do not use those files as the active reference population. Expansion assignments in the current [pattern map](pattern-map.json) are explicitly post-codebook exploration, not additional held-out validation.

## Evidence depth

The expansion broadens non-software operations, creative collaboration and relationship work. Some social mechanics were inspected in public static artifacts, creator transcripts or screenshots. No workflow was executed, no video playback observed, no human transfer study performed and no outcome independently measured. No full upstream templates/code, transcripts or screenshots are redistributed.

SOC01's recovered template requests invented metrics; it is cautionary only. SOC02 separates a reported consistency comparison from recommended held-out evaluation. SOC03 separates proposed ingest, query and lint jobs. SOC06 distinguishes complete-page retrieval from freshness and discloses sponsorship. SOC07 makes notification/calendar side effects and missing approval/duplicate controls visible. SOC11 enrichment remains excluded and audit-only.

## Reading dates and interpretations

Source mechanics and evidence labels must be read with their provenance notes, unknowns and limitations. Researcher classifications, codebook challenges and hypothetical transfers are not observed outcomes. Recommended checks must not be mistaken for reported behavior.

`source_dates` preserves publication, update and explicitly typed other dates, evidence URL/note, check time and status where supplied. JSON null means unknown. Partial dates stay partial; event, policy, proceedings, repository-commit or filename dates do not become publication dates. A populated value can remain indexed-only or disputed. Original and companion sources retain separate date entries. Access dates and metadata-check dates are neither publication dates nor evidence of current adoption. [SOURCES.md](SOURCES.md) exposes these distinctions; it does not claim a new source audit by the repository preparation step.

## Remaining limits

The sample is purposive, not representative of all knowledge work. The original sample favored transparent software organizations, formal procedures and vendor-selected stories; the expansion and eligibility filter do not establish representative coverage. Original concentration figures (13 of 72 workflow records from GitLab; 33 from the top five source-host families) describe the initial sample only.

Shared structure does not transfer decision authority, privacy permissions, domain expertise or acceptable risk. Generated action items are not agreed commitments; reviewer agreement is not factual truth. Test labels and cross-function adaptation with people doing real work before claiming transfer or navigation success. No schedule, automatic promotion or reviewed-core designation is enabled here.