[domijin]

Summary

Adds the first reusable learning-checkpoint capture protocol for Open Learning Lab, revised after the first dry run exposed an orchestration-integrity failure.

Workflow

Primary learning thread -> branch conversation -> run checkpoint capture -> verify package/evidence/PR -> review -> return to primary learning thread.

Adds

• prompts/checkpoint-capture.md — reusable, agent-neutral capture prompt;
• prompts/README.md — sidecar checkpoint workflow;
• Marin study checkpoint instructions;
• Marin study publication policy: study-relevant learning progress is public by default, while unrelated/private context remains excluded;
• top-level README discoverability.

v0.2 completion invariant

The first dry run (capture/marin-cp001) created only a manifest, referenced missing evidence V003, omitted required checkpoint files, and opened no PR. That failure is preserved in #8 rather than retroactively cleaned up.

The protocol now MUST verify before claiming completion:

• all five required checkpoint package files exist;
• every referenced ledger ID resolves;
• goal/hypothesis/experiment references resolve;
• integrity checks pass;
• a PR exists when GitHub write access is available.

Otherwise it must report CHECKPOINT INCOMPLETE.

Methodological guardrails

• capture mode does not continue tutoring;
• verbatim learner trace is separated from normalization/interpretation;
• observations are not automatically promoted to findings;
• active/time-sensitive domain claims are re-verified from current sources;
• project-development/orchestration failures are not learner evidence;
• the prompt itself is a versioned experimental intervention (checkpoint-capture/v0.2).

Related: #8