Practice the job, on a box that answers back.

Your first morning at Halvard Freight and the nightly rate sync failed. Orient on an unfamiliar machine, find the failure in the logs, work out which config value actually ran, and prove the fix with a dry run.

Halvard's receiver has three places a timeout can come from and they disagree. Work out which value actually runs, why the one in the repo is not it, and put the right value in the layer that wins.

Halvard's nightly cut assigns the last hour of deliveries to the wrong day, but only some nights. One system stamps UTC, one Oslo local time, and the boundary moves twice a year. Find the seam, prove it with the data, and fix the cut.

Halvard Freight's carrier webhook created two invoices for one delivery. Find the duplicate in the logs, explain why the carrier retried, and make the billing write safe to repeat.

Every Monday at nine Meridian's claims assistant slows to a crawl and a third of runs fail. The provider is rate limiting, and the router's retries are making it worse. Read the storm, name the mechanism, and tame it with one config change.

Meridian's nightly claims sync reports success and the totals are wrong. The claims API returns errors inside 200 responses and a cursor the client ignores. Find both, fix the client's config, and make the dry run tell the truth.

Calder's support agent ran 47 turns on one case overnight, calling the same tool against a service in maintenance. Read the trace, bound the loop, handle the tool state the agent never handled, and replay the case to a clean handoff.

Calder's agent sometimes answers in prose where a proposal object is expected, and the pipeline counts those as successes with an empty action. Count them, make the output a schema or a hold, and stop success from meaning nothing happened.

One slow provider call became nine at Calder because three layers each retried it. Read the trace, count the multiplication, give the deadline one owner, and prove a case now fails fast instead of nine times slowly.

Calder's agent called read_support_case with a case id that does not exist, got a neighbour's record back, and proposed a replacement for it. Tighten the schema and the adapter, and replay the case to a clarification instead.

Calder's support agent proposed a policy-waived replacement because the customer's message told it to. Find the injected instruction, see where write authority actually lives, and make the evidence compiler treat customer text as untrusted.

An operator approved a replacement for Calder case C01, then the case changed and the service refused the request. Show what the approval is bound to, why the refusal is correct, and re-review properly.

Calder's release review wants evidence the support agent cannot be talked into shipping a pump. Run five recorded attacks, read how far each reached, commit to which got furthest, and report what you did not test.

A Northwind adjuster got an answer that quoted another insurer's policy. Find which layer crossed the tenant boundary, fix the key that let it, and name the second layer that must filter by tenant before it ranks anything.

Calder's action worker sent a replacement request and died before recording the receipt. Reproduce the crash, find out what the service actually committed, choose a recovery that cannot ship a second pump, and prove it.

A bad release is proposing wrong replacements at Calder and it must stop shipping now, without stopping customers from getting answers. Pick the stop that blocks effects and not lookups, apply it, and prove both halves live.

Calder's action queue has been filling for forty minutes and nothing is draining it. Contain it, find out why the consumers went away, get the backlog moving without shipping anything twice, and tell the morning shift what happened.

Calder shipped two pumps for one approved replacement after a network blip: the worker retried with a new key. Find the key shape that made the retry a new request, fix it, and show the service refusing a changed payload under the same key.

Calder wants a quality number for its support agent. Predict how many of six conversations fail, label them against the policy, mark the first turn that broke, and find the failure type the supplied automated checks miss.

Calder's new agent release scores 10 of 11 against the old one's 8. Read the paired results case by case, find the case where the better score hides a worse action, and say what should happen to this release.

Calder's model judge scores the agent's proposals and it disagrees with the human reviewer on the cases that matter. Compare the two on the labelled set, name which way the judge is wrong, fix the rubric it was given, and rerun.

An eval passes while the assistant contradicts the current coverage policy. Find stale golden labels, update their policy references, version the manifest, and rerun against unchanged model answers.

Calder's candidate passed 18 of 20 tasks in CI and the gate says ship. Rerun every task eight times, find the task that passes most runs and files an unconfirmed replacement on the rest, read why, and make the gate hold.

Calder's grounding judge agrees with the reviewer on 94% of 200 conversations. Split agreement into failures caught and passes passed, find why the judge misses half the failures, fix it, and correct the dashboard's failure rate.

Calder's candidate release scores higher than the baseline and someone wants it shipped today. Read the paired packet, find the case that should block it, turn that failure into a release-blocking test, and make the gate say hold.

Calder's live grounding judge pages so often that on-call muted it. Backtest the monitor on a recorded week, tune sample share, threshold and window to meet three constraints, then show the setting holds on weeks you did not tune on.

Northwind wants the claims model on their own two 80 GiB cards; the sheet says 70 GB of weights. Work out whether it fits with a KV cache at their concurrency, in the units the card uses, and set the admission limit that keeps it fitting.

At Monday peak Northwind's primary model sits at its rate limit while the fallback idles. Find the bottleneck, route the traffic that does not need the large model, and keep the policy reasoning where it belongs.

Northwind published a new coverage policy at nine and the assistant kept quoting the old one until lunch. Retrieval was fine. Find the layer that remembered, key it by what the answer depends on, and keep the hit rate.

Calder's agent started proposing replacements without inspections after a policy update. Nothing in the prompt changed. The compiled context did: it is over budget and the clause that governs the decision is the one being cut.

Northwind's claims assistant is timing out for a third of customers and you cannot read their code. Contain it from the console, find the phase that is slow, commit to a root cause, and write the first status update.

Meridian's claims assistant has been quietly wrong for three days. Latency is fine, errors are zero, and the evaluation on an unchanged manifest fell from 9 to 6. Find what changed, contain it, and add the alert that should have fired.

The capstone. Calder's support agent is up for release and the board has forty minutes. Gather evidence from the live system, answer the three questions a board asks, and commit to hold or advance with reasons a reviewer accepts.

Halvard's billing rejected every invoice since two in the morning and the carrier has been redelivering ever since. Contain the storm, find the rotated token, restore invoicing, and hand over evidence the next shift can act on.