[domijin]

Checkpoint

Captures Study 001 Marin checkpoint CP001 under checkpoint-capture/v0.2.

What the learner demonstrated

• L1 immediate retrieval of the three synthesized open-development mechanisms: external knowledge, internal commitment, sustainable economy (V003).
• In-session causal/system reasoning about open development, governance, resource sustainability, cheap PoCs, scaling, and experiments as learning.
• In-session far transfer to clinical MRI remains represented by existing V002; this is not upgraded to delayed/unassisted evidence.

What remains interpretation

• Claim-centric research-memory reasoning is promising but scaffolded.
• Causal-boundary discipline, exploration-vs-confirmation, scaling uncertainty, and proper use of reputation remain untested after correction.
• Initial Marin overview contaminated later 'rediscovery'; the trace records this explicitly.

New evidence

• V003 — immediate retrieval (L1). No finding promoted.

Hypotheses affected

• H001 remains active and unresolved. CP001 does not establish comparative advantage over explanation-first learning.

Contradictions / open questions

• Scaling extrapolation was underqualified.
• The MRI transfer compressed several downstream causal links.
• Highest-value next tests are delayed unassisted reconstruction and a fresh exploration-vs-confirmation case.

New external information

• Current Marin primary sources show the 2025 Speedrun framing is historical: issue #4918 states the old Speedrun system is no longer on main and is retained as a frozen optimizer benchmark; the current README emphasizes Delphi scaling and frontier MoE work.

Privacy / publication check

• Uses only study-relevant visible learner material permitted by studies/marin/publication-policy.md.
• Excludes unrelated personal context, hidden memory/system context, credentials, and private third-party material.
• Third-party Marin material is linked/paraphrased rather than copied.

Proposed next learning test

1. L2 delayed unassisted reconstruction of the three mechanisms and counterforces.
2. Fresh test distinguishing exploration from confirmation after multiple tries / post-hoc selection.

[domijin]

CP001 is now operationally complete and scientifically conservative enough to preserve.

What passed

• all required checkpoint package files exist;
• V003 exists and the repository validator passes;
• the PR exists and references resolve;
• prior-exposure contamination is explicitly recorded rather than hidden;
• no finding is promoted;
• current Marin correction is source-backed: the present README emphasizes Delphi/frontier MoE, and issue #4918 explicitly says the old Speedrun product surface is not the current official workflow;
• learner-state cleanly separates demonstrated / emerging / misconception-risk / unknown.

Non-blocking methodological issues exposed by CP001

1. Evidence strength is multi-dimensional (#12). "L4-type in-session far transfer" shows our L0-L5 ladder conflates transfer distance, delay, assistance, novelty, and authenticity.
2. Evidence target != evidential relation (#14). V003 targets H001 but is diagnostic capability evidence, not evidence that judgment-first beats explanation-first.
3. Checkpoint provenance needs stronger pinning (#13). Record the exact repo base commit/context used by capture, distinguish prompt blob SHA from repo commit, and use stable source-interval refs when available.

Recommendation

Keep CP001 as the historical first complete capture. Do not rewrite its observations to fit future schema changes. Migrate/annotate through later schema-compatible records if needed.

From the orchestration perspective, this is a successful v0.2 capture with three useful protocol/schema learnings—not a validation of H001.