[abrichr]

Follow-up to the 2026-11-03 external review (findings 3, 4, 8, 9 and 10). Docs and docstrings only.

• Pipelined read-back: the module docstring now describes how a linear durable run with an encrypted checkpoint store defers its last step's read-back (Replayer._may_pipeline_readback, THROUGHPUT.md).
• Production-lifecycle ratio: both result files now state the scope. The ratio comes from the cost model, with hand entry at 3 minutes per item, on a local OpenEMR with synthetic patients, and includes the record-channel duplicate check from Duplicate check asks the system of record (FHIR/REST, SQL fallback) instead of the screen list #1516.
• Model calls: README, CAPABILITIES, llms.txt and the flow and agent READMEs now say "no calls to a hosted AI model", which is what the runtime measures.
• Fleet patches: they move as files or through a local reference hub, and the docs describe only that transport.
• Induction held-out check: described as structural trace coverage. The behavioural scorer is validation.behavioral_generalization. Updated in CHANGELOG, REPAIR_LIFECYCLE.md, LEARN_FROM_MANY_RUNS.md and the learning package docstring.

llms-full.txt was rebuilt. Website copy: OpenAdaptAI/openadapt-web#598.