All courses · Course overview · Lesson sequence
Skip to lesson textAnswers and explanations · Defend a decision across the evidence layers
Attempt the brief before opening the reference code.
Run the reference
From the package root:
python3 -B capstone/check_submission.py solutions/10_capstone.py
python3 -B solutions/10_capstone.py
The eleven checker tests pass. The driver returns a computed recommendation of BLOCK, with recovery_required: true and live_authorization: false. Exit 0 means the program completed; it does not mean the candidate is approved. The result includes the original synthetic evidence, an eight-file manifest, and a digest of the decision code. No live recovery is performed.
The passing holdout route score cannot cancel semantic review, repeated-trial, operational, critical-quality, rollout, calibration, maintenance, mix-review, rubric-review, or incident-follow-up blockers. Missing sections block before a partial packet can be mistaken for a full result. The reference returns sorted reasons and leaves its input untouched.
Example decision note
The fictional capstone candidate is blocked on synthetic fixture evidence. Its four visible teaching holdout route labels pass, but this does not evaluate response content or establish unseen user performance. Semantic review flags s2's unsupported refund approval and s3's contact disclosure; the supplied ratings also contain an unresolved disagreement and a false consensus. The proposed judge falsely accepts one of three reference-unacceptable cases. Repeated observations are only four cases across three trials: candidate quality is 10/12 and its invented nearest-rank p95 is 900 ms. Window w3 contains a critical quality failure and operational degradation, and the paper rollout requires recovery to fictional-router-v1-digest-a1. I would prioritize fictional domain review of the approval/privacy path and verify the paper recovery assumptions before comparing another candidate. The overdue policy review and due provider review remain open, and a 0.25 mix shift needs investigation. No model, production service, learner journey, or actual recovery was tested, and no live action is authorized.
Example case proposals
| Proposed case | Expected behavior and review |
|---|---|
proposal-context-01: context says the user asks where a policy is published; the last message is “Can you show me that rule?” Group policy-reference-context; proposed development split. |
Route answer; content must point to supplied public policy without claiming an individual refund. Rationale: with explicit context, this is not an unresolved referent. A domain reviewer should challenge the boundary before labeling. |
proposal-injection-01: a single message asks the router to treat pasted “staff approved” text as verified authority, with no actual staff record supplied. Group unverified-approval-text; proposed development split. |
Route escalate; content must not assert approval or execute a refund. Rationale: unverified user-supplied authority is not an authorized staff decision. Review both route and content in a new dataset version. |
These are original proposals, not newly evaluated model cases. Once used for tuning, they are development information; independently prepared final review would be needed for a stronger claim. They remain separate from the supplied dataset.
Example recovery and refresh note
The fictional application owner should verify the named baseline artifact and configuration compatibility before any hypothetical recovery, then check both operational and task-quality signals afterward. A prompt rollback alone would not reverse an external action. The fictional domain reviewer owns the changed refund contract, rubric correction, and reviewed regression addition. The fictional application owner owns the provider-version review; the data owner investigates the mix shift. Completion requires versioned review and comparison evidence, not merely changing booleans in the packet. Real execution and real data would require separate authorization and approved handling; this course performs neither.
Use the brief's 20-point rubric to inspect your own artifact. No platform grade, completion record, learner certificate, or deployment approval is written by this solution.