Stop 10 of 10 · Weekly
Nour's capstone
The scheduled runs and the routines that read them, then one system built end to end with tests, guardrails, an evaluation score and a cost line, graded out of a hundred.
The quarter that runs without you
Three schedules carry Nour's agency, and none of them changes an account.
| When | What runs | What it produces |
|---|---|---|
| Nightly | pull.py across twenty accounts | Fresh metrics.json and a manifest of every failure |
| Weekly | The audit assistant over every account | Proposals, unsigned, with evidence |
| First working day | write.py, check.py, publish.py | Twenty drafts in the review queue and refreshed portal pages |
On top of them sits one routine in Claude that reads the outputs and writes the Monday brief. It holds the reader credential and nothing else.
Every Monday at seven, read last week's run manifests, the open proposals, the guardrail refusals and the current ledger rows. Write one brief: accounts where a pull failed more than once and the platform involved, drafts still unapproved and how long they have waited, proposals expiring this week, every guardrail refusal with the rule id and the reason, and any ledger row that is negative or unknown. Sort by dollars where a dollar figure exists. Change nothing. Sign nothing. Contact no client. Where you are unsure whether something needs a person, list it rather than deciding.
The capstone
Take one system and build it properly, end to end, on real accounts. Not all six. One, complete.
You are submitting seven artefacts: the code with its tests passing, the guardrail module and its paired test suite, an evaluation run with its score and the gate result, one signed proposal with its signature, its dry run and its change log line, the rollback written before the apply, one client-facing output with its leak test, and the ledger row with the baseline behind it.
[upload the seven capstone artefacts: the system code and test output, guardrails.py with its test run, the evaluation results file, the signed proposal with dry run and change log line, the rollback proposal, the client-facing output with its leak test, and the ledger row with its baseline] I run a paid media agency. Grade this against the rubric below out of 100, scoring each criterion separately and quoting the evidence you used. Check the gates first and say plainly if either fails. For every point not awarded, name what is missing and what would earn it. Be strict about any figure with no source, any test that cannot fail, and any path to a platform write that does not pass through a matching signature.
The rubric
| Criterion | Points | Full marks means |
|---|---|---|
| Credentials and client data | Gate | No token, spend export, margin or client customer data in the repository, in any transcript or in any client-facing output. A breach fails whatever else scored |
| Signature integrity | Gate | Every applied change has a signature whose hash matches the proposal bytes, and the change log records from, to, approver and platform response |
| Guardrails as code | 20 | One imported module, every rule with a blocking and an allowing test, fail closed on unknown accounts and missing thresholds, and one red-team input found and fixed |
| Measured, not assumed | 20 | An evaluation set with expected answers written first, three axes scored, a gate that has actually stopped a version, and the score stamped with the prompt version |
| Numbers traceable | 15 | Every figure in every output exists in the run data, proved by a checker that fails on a seeded bad draft |
| Failure handling | 15 | One account or one platform failing leaves the other runs intact, writes a visible missing marker and records the failure in the manifest |
| Client surface hygiene | 15 | Output built from an allowlist, with a leak test over a poisoned payload covering margin, cost and a second account |
| Ledger honesty | 15 | Cost, machine minutes, review minutes, errors caught and errors shipped, a baseline behind the saving, and unknown written where an input is missing |
Eighty passes. Under eighty it is nearly always the same two: a test that could never have failed, and a saving with no baseline behind it.