Skip to content

mj-agentlab/learning-evidence-kit

v0.1.0MIT

Explicit learning diagnostics, recoverable local evidence and manual task record/application checks.

GitHub

Learning Evidence Kit

Two explicit skills connect task-related learning observations with recoverable records and manual checks. Current candidate version: 0.1.0; this candidate replaces Understanding Kit for the planned marketplace 8.2.0 release.

  • $learning-evidence-kit:pop-quiz diagnoses necessary knowledge and produces a complete transferable record. Diagnosis stays read-only; an explicit handoff ends the quiz and saves through common capabilities.
  • $learning-evidence-kit:learning-record-check checks a specified record's validity, then its task application using existing verification materials and your explanation. Each conclusion is reported separately. Save, check and publication require their respective explicit scope.

Native question entries are deferred in this candidate. Codex and ChatGPT desktop routes have not passed continuous pre-submission window acceptance. Pop Quiz will report that gap and preserve supplied observations, without starting new interactive knowledge questions on those routes. The two skills remain registered; metadata discovery is not interactive acceptance. See the packaged entry ledger. This candidate has not completed the original two-client plan or Owner final acceptance.

The independent local library retains complete JSON, original intention/requirement content, identity confirmation, original and supported observations, sources, versions and flow receipts. Markdown is generated for reading. Worktree manifests store locally Git-excluded references only. Cross-PR reuse retains learning ID; corrections/retests/checks append history. Full export/import and retained archive support recovery.

Sharing is manual through complete ordinary PR comments, including segmented recovery and read-back of uncertain results. Manual historical collection preserves source snapshots and coverage gaps; it does not make management decisions. There is no automatic learning, publication, periodic collection, personnel change or cloud synchronization.

Common capabilities require an actual Python 3.12 interpreter; GitHub flow reuses authenticated gh. Resolve packaged scripts/lek.py from the actual skill locator and send JSON stdin. The record/common contract specifies fields, operations, scope and receipts. Installed execution is independent of a marketplace repository or worktree.

Local originals, shared copies, record validity, knowledge application, engineering outcomes and long-term learning evidence are separate. Synthetic engineering checks do not establish human learning or promotion readiness.