A capstone is ready for handover when another engineer can understand its boundaries, reproduce its evidence checks and tell when it failed. A screenshot of an attractive answer is insufficient. Deliver the frozen inputs, request contract, original responses, validation records, score sheet and effort ledger together. Mark fabricated responses as synthetic and actual observations as actual. If no model or simulator was used, say so directly in the handover rather than leaving readers to infer what happened.
Package: evidence-triage-course-v1
owner: learner / receiving reviewer (roles only)
scope: read-only notes for short synthetic RTL/DV/STA/PD packets
inputs: k1-v1, k2-v1, k3-v1, h1-v1, h2-v1, h3-v1
workflow: packet validation -> draft -> contract validation -> human check
contract: teaching-v1; optional validator source included if used
evidence: original notes plus correction history and score sheet
cost: actual learner effort; no model usage measured on core route
failure policy: reject wrong identity/evidence; bounded repair; incomplete
deployment status: classroom exercise; production readiness not assessedPerform a cold handover: give the receiver the package without a verbal explanation of the answers. Ask them to select one practice and one held-out packet, identify the workflow version, trace a claim to its source and explain the remaining uncertainty. They should also demonstrate that a missing response produces incomplete rather than a clean pass. Record the receiver's corrections, because a result that cannot be reproduced from the package is a handover defect even when the author understands it.
The safety review checks concrete behavior. Did any role try to write a source, upload a packet, obey an embedded instruction or claim a tool result without evidence? The evidence review checks whether each reference exists and actually supports the claim. The independent verification review asks a human to reproduce the relevant reset trace, checker comparison or report-scope interpretation. Re-reading an assistant's explanation and agreeing with its tone is not a substitute for doing those checks.
Prepare a final note with the accepted scope, unresolved failures, per-case rubric outcomes, effort comparison and receiver's reproduction record. Suppose all field checks pass but H2 is described as signoff-clean and the receiver cannot find a final-route report. Decide whether the package passes the course gate. Then describe how a future approved MCP reader or multi-agent review could be added without making either necessary for the existing offline workflow.
The package fails acceptance for H2: valid fields do not support a signoff claim, and P2 explicitly lacks the necessary reports. Record the failed initial attempt and correction, change the workflow only under a new version, and use fresh held-out cases to evaluate that version. An optional MCP reader should expose only approved read operations and preserve packet provenance. An optional spec/RTL/coordinator branch must keep budgets, references and human checks. Both branches must fall back to the original supplied-packet route when unavailable.
The final recommendation may be to keep the manual route. That is a successful learning outcome when supported by the evidence. A different recommendation may permit a narrow assistant for drafting notes while keeping all acceptance decisions with an engineer. State which claim your evaluation actually supports and what would need a new trial: longer inputs, another tool, production data, new report types or changed requirements. A new adapter is a new boundary to verify, not just a configuration label.
This capstone does not establish signoff coverage, production access approval or model reliability. It leaves a small, inspectable workflow and a record of its limits. For risk-management background, revisit NIST AI RMF 1.0, January 2023. To deepen the exercise, have a colleague prepare new synthetic packets and repeat the frozen evaluation; preserve the original attempt so improvement remains distinguishable from answer memorization.