Workshop: epistack-competition

Summary of the sibling epistack-casebooks repo and how the two repos exchange ideas. The deliberation that produced this (framing, tooling build list, case strategy) now lives in that repo's AGENTS.md and kb/work/casebook-build/README.md; this workshop is just the framework-side pointer.

What the sibling repo is for

../../../../epistack-casebooks/ (sibling of this repo under /home/zby/llm/) is a consuming project of Commonplace — a candidate entry to the FLF Epistemic Case Study Competition. It builds agent-operated epistemic casebooks: turning messy sources into source-grounded, typed, reviewable notes that expose what is known, what is contested, what depends on what, and where the gaps are — without adjudicating truth.

It exists separately, not as work inside this repo, so the entry demonstrates Commonplace's method transferring to a real consuming project rather than only documenting itself. It also keeps this repo's scope clean: application-specific casework belongs in consuming projects, not here.

Structure there: three isolated top-level case namespaces (kb/lhc/, kb/eggs/, kb/covid/), each with a sources/ collection (captured material) and a notes/ collection (the casebook, plain note type), plus a global kb/notes/ for transferable theory. The sibling editable-installs this repo's llm-commonplace package, so framework changes here are live there without reinstall.

How we move ideas between the repos

The single channel is epistack-casebooks/backlog-to-commonplace.md. The rule is one-directional with a clean boundary:

• Casework → framework. The sibling repo never edits its kb/commonplace/ mirror (read-only installed framework). When casework reveals a framework need — a missing type, a validator/gate gap, a methodology insight — it builds a local version first (collection-local types), proves it on a worked case, and logs the upstream need to the backlog, classified: bug / type gap / instruction gap / validator gap / review-gate gap / methodology note.
• Framework side (this repo). We watch that backlog and promote upstream only the primitives that survived contact with a worked case. Build-local-first, upstream-what-survives keeps this repo from accumulating speculative machinery.

So this repo stays the framework; the sibling stays the casework; the backlog is the interface. The competition entry's value to Commonplace is whatever lands back here through that backlog.

Status

Sibling repo scaffolded with a filled-in control plane and a kickoff briefing for its agent. First experiment queued there: provenance with plain notes on the black-hole case (does file-level linking suffice, or do claims need a source-span locator type?). Watch the backlog for the result.