AI OPERATIONS · PROJECT

A model with an operating plan.

Collect release, monitoring and incident evidence, then explain how you would operate the system.

Start here without finishing earlier lessons. Revisit release stages, monitoring or incident response when needed. For runnable HTTP code and measured timings, use the optional P08 studio. Choose or change a learning path.

Report: draft. 0/6 saved captures. First-attempt transfer: 0/2 cases passed. Human artifact review and learner pilots: pending. Submission checks matched evidence and report presence, not reasoning or code quality.

Project evidence formatVariant

1. Run and capture your evidence

The workbenches are temporary until you click their save-to-project button. Switching workbenches starts a fresh playground; saved captures remain below. Keep at most six captures, leaving room for two different monitoring policies. No earlier lesson completion blocks these experiments.

Release: use the unchanged candidate; run a passing shadow check, advance, run staged traffic, roll back and save the record.
Monitoring: finish all eight ticks, save, change a threshold or review decision, finish again and save a distinct comparison.
Incident: observe a raw candidate failure, restore stable routing as an operator, then run healthy stable traffic and save. If you pause, obtain a fresh matching probe before resuming.

Workbench

Active experiment

Stage a release without hiding failures

This browser simulation runs real predictions from the packaged Iris classifier against 20 public synthetic regression requests. Expected outputs were recorded by the separate Python runtime. They test the original input contract and probabilities, not true species labels or generalization. No real users, network service or external deployment are involved.

Candidate to load

Starting a new practice clears only this temporary deployment trace and note. Export it first if needed. The candidate selector is a draft; the loaded candidate named below remains active until you start a new practice.

For guided practice, reproduce a timeout at 25% and restore stable routing. For implementation, specify a real local service’s request/response contract and tests. For a decision brief, justify which failure should stop rollout and which evidence is still missing. The broader P08 service studio and independent operations-project review remain separate work.

2. Inspect saved evidence

Numerical outcomes are reconstructed; imported “passed” claims are not accepted as evidence. Download the full report to inspect recomputed results. Repeated public cases are development practice, not independent tests. Removing a capture reopens submission and does not alter source lessons.

3. Write a supported operating plan

Saved report — 2,000 characters per fieldArtifact and provenance

Identify model/runtime/data versions, units and trusted sources. Distinguish byte consistency from correct behavior.

Acceptance tests and release evidence

Cite the shadow, staged traffic and rollback records. State which public checks passed and what they do not establish.

Performance and operating limits

Record measured timing files/method/sample counts where available. Label assumed costs separately. State capacity, cancellation and missing measurements.

Monitoring policy comparison

Compare your two saved policies, label selection, subgroup denominators, false alarms, missed signals and missed reviews.

Incident response and recovery

Cite the failed candidate and restored stable traffic, denied/stale actions, withheld demand and unchanged earlier consequences.

Decision, ownership and follow-up

Justify proceed/defer/reject, name an incident owner and specify a follow-up experiment. Include reproduction instructions for implementation.

Each report field needs at least twenty non-whitespace-trimmed characters before submission. This is a presence check only. Saved work stays on this browser/device; use My Work export/import to transfer a backup. No cloud synchronization is implied.

Separate objective transfer assessment

Try a different problem independently

These two fixed cases use different data and separate scoring arithmetic from the guided model/release project. Work without the walkthrough first. Opening and submissions are saved; the first submission is retained, with at most two practice retries. Reopening, resetting or importing an earlier backup does not erase known exposure. This is a local objective check, not authenticated exam supervision or human review of your implementation.

0/2 cases passed on first submission. Latest/practice passes: 0/2. Project submission and these outcomes are separate.