# Collection research and selection

This expansion surveyed 20 distinct pinned public repositories and two cached published npm artifacts. The [source register](collections.json) records revisions, inspection depth, license findings, runtime boundaries and rejected procedures. Earlier source work and installed-skill ancestry remain in [SOURCES.md](../SOURCES.md). The count describes this expansion, not every repository ever inspected.

Four research agents worked separately on engineering, design/content, data/operations/security, and adversarial evaluation. They wrote original drafts in isolated scratch directories. They did not install collections, execute upstream helpers, change live configurations or run their own behavioral evaluations. Cross-reviews inspected the drafts and challenged both missing safeguards and unnecessary workflow ceremony.

## What the inventory means

Engineering inventoried seven pinned repositories, with 260 physical skill entrypoints, plus the cached `@1aboveio/skills@1.0.21` and `@1aboveio/engineering-skills@0.1.21` artifacts. Design/content inventoried five main collections with 285 physical entrypoints plus 19 paths in a separately checked marketing ancestor. Data/operations inventoried nine pinned repositories with 744 physical entrypoints and deeply read 26 selected entrypoints and supporting material.

These lane counts overlap. They include host variants, fixtures and unselected specialties. They must not be summed into a unique-skill or deep-read claim. Inventorying a complete tree is not reading every procedure or verifying its behavior. Cached artifact comparisons did not independently verify the original archive integrity or recover unavailable source history.

## Selection rationale

| Sources compared | Useful contribution | What was not adopted |
| --- | --- | --- |
| Matt Pocock, Bolens, Superpowers, 1AboveIO | Contract-aware design, independent test expectations, useful test boundaries, scoped delivery, candidate and attempt identities | Mandatory agent launchers, competing controllers, compliance as correctness, review passes carried to unreviewed candidates |
| Vercel agent skills, Sentry, Trail of Bits, OpenAI skills/plugins | Framework-specific scope boundaries, source-to-sink security reasoning, exact reviewed candidates, explicit host capabilities | Production-pushing recipes, current-checkout base switching, scanner silence as security, unlicensed evaluator/runtime imports |
| Impeccable | Task/evidence distinctions and bounded interface review | Engine downloads, hooks, native assets or a claim that prose equals the full engine |
| Marketing, UX and scientific collections | Writing modes, real participant evidence, source lineage and report reconciliation | Disguised recruitment, biased sampling, universal thresholds, simulated customers as observed evidence |
| dbt, Dagster, Supabase, HashiCorp, Ansible | Grain/cardinality, incremental state, partition/replay identity, native-tool capability boundaries | Automatic connector installation, warehouse mutations disguised as previews, infrastructure discovery that upgrades dependencies |
| Public Pandoc, openpyxl and W3C contracts | Operation-specific preservation limits and accessibility evidence | A bundled converter, formula engine, standards certification or tested native-document integration |

The result is [21 canonical owners](../CATALOG.md): five existing skills plus 16 new original procedures. Consolidation combines related stages and modes rather than importing every upstream entrypoint. All required instructions stay inside the selected folder. Optional specialists are contributions, not permission to start another delivery workflow.

## Rights and provenance

The existing design/debugging/verification companions remain attributed forks with their complete ancestor notices. The discovery pair carries its Matt Pocock notice locally as well as the project license. The new 16 procedures record original expression and conceptual comparisons, not copied upstream scripts/assets. A zero exact-text-overlap screen is not legal clearance. Neither a source URL nor a blanket MIT label repairs a derivative's missing terms.

| Boundary | Disposition |
| --- | --- |
| Anthropic DOCX/PDF/PPTX/XLSX | All implementation, helper, prompt and asset material excluded. Only identity/license exclusions and original public-engine contracts retained |
| Unlicensed UX material and unresolved doc-coauthoring grant | Observation recorded, expression import held |
| Impeccable | Apache-2.0, not MIT. Native ancestry and asset notices would need separate treatment before copying |
| Trail of Bits | CC-BY-SA-4.0 expression not copied into the MIT procedures. Any later adaptation needs a compatible attributed boundary |
| HashiCorp | MPL-covered files not transplanted or relabeled MIT |
| OpenAI plugins | No blanket root grant. Proprietary earnings material and unresolved evaluator/finance grants excluded |
| 1AboveIO | Cached packages used for comparison. Incomplete per-skill notice text is not treated as a complete redistribution grant |

Research also found that Bolens' TDD `tests.md` and `mocking.md` match Matt Pocock despite a local-original declaration. They were not imported into `testing-work`. The earlier three companions retain their separately established ancestry and grants.

## Adversarial corrections applied

Independent static review identified integration defects rather than treating every draft as ready. Integration preserved existing-import provenance, replaced external delivery/TDD handoffs with optional owned roles, separated test-only work from repair, and limited delivery blockers to in-scope defects. It also assigned editorial ownership and recorded each deferred OS specialist instead of calling generic Linux guidance an equivalent replacement.

Cross-review added explicit literal-versus-formula handling for spreadsheet/text export, a candidate-code execution boundary for dependency updates, and the distinction between scanned PDFs with a usable text layer and image-only pages needing OCR. Operations now requires sensitive filtering before output reaches a hosted agent, not only before the final report. These are source corrections, not observed model failures.

## Behavioral evidence and next checks

The [entrypoint pilot](../evaluations/entrypoint-pilot/RESULTS.md) compares eight frozen tasks with no skill versus a forced entrypoint for two registered model IDs. It is a development screen, not a retention trial or proof of 21-skill effectiveness. Labels are withheld during grading. Task outcomes are not points for repeating skill wording or emitting a preferred ledger. Generic-instruction controls, repeated samples, untouched tasks, real tools, routing and cross-host behavior remain separate work.

The source collection is deliberately broader than automatic activation. Ordinary clear tasks can finish directly. A specialist should earn attention by solving a recurring failure or supplying concrete domain knowledge, not by restating generic caution. A tie on this pilot supports neither global activation nor permanent rejection.

For updates, inspect the pinned-source delta, per-file rights, side effects and support closure before adapting it. Keep unchanged source identities and historical uncertainty. Give every revised candidate a new digest. Run offline checks and inspect the actual tarball. Use fresh tasks for confirmatory comparisons after development exposure. Publishing, installing or transferring live ownership remains separately governed by the user's authority.