Give the Agent a test with an answer you can check.

Keep the task input and reviewer reference separate

Download the six files and read test-readme.md yourself. Put only stock-and-rules.md and equipment-reservations.csv into the initial Agent task. Keep the amendment, scorecard, expected results, and README outside that context. For a strict test, store reviewer material separately instead of uploading everything and asking the Agent to ignore it.

All names and records are fictional. The scorecard begins at not observed because downloading a test is not running it. Opening Workspace does not upload the files, configure an Agent, or start paid work automatically. The reviewer controls when the later amendment becomes available.

Check overlapping demand and shortage arithmetic

Ask for allocation-review.md and open-questions.md for 22 October in Asia/Manila. Stock is eight projectors. Approved R01 needs five from 10:00 to 12:00, while approved R02 needs four from 11:00 to 13:00. Their 11:00 to 12:00 overlap commits nine, creating a shortage of one. Pending R03 requests three and stays outside approved demand.

Open the saved output before scoring it. Check the interval, quantities, arithmetic, approval status, and source IDs. A daily total is not a substitute for overlapping demand. Counting all twelve requested projectors as approved would fail even if the resulting table looks organized.

Keep an incomplete return window visible

The initial CSV also includes approved R04: two projectors starting at 14:00 with no return time. It cannot affect the earlier 11:00 to 12:00 overlap, but it prevents a complete availability promise for later windows affected by it. The Agent should flag the missing return and ask the coordinator rather than inventing an end time.

Record this behavior under T02 separately from the arithmetic result. A partial review can correctly report a known earlier shortage while leaving later coverage incomplete. Do not silently supply a guessed return to make the test easier, or treat an empty cell as proof that equipment already returned.

Supply the approved amendment as a second task

After scoring the initial output, add reservation-amendment.md. AMEND-1 reduces R02 from four projectors to two, with its time window and approval unchanged. Ask for allocation-review-revised.md and a change note. Preserve the original review so the earlier snapshot and later amendment remain inspectable.

The revised 11:00 to 12:00 commitment is five plus two: seven projectors, leaving one spare. This resolves that shortage only. R03 remains pending, and R04 still lacks a return. Inspect the summary, overlap table, arithmetic, and questions for stale quantities or an unsupported claim that every allocation issue is solved.

Write evidence beside the verdict

Use pass when the observed result meets the criterion, revise when it does not, and not observed when you did not test it. Add a source reference, saved-file location, or concise observation. Record manual repair separately from the Agent’s initial output. Keep the configuration and test date with the scorecard.

This exercise has no combined accuracy percentage or competitor ranking. Different criteria carry different consequences, and a small authored sample does not estimate performance across all business work. Optional saved evaluations can supplement the process where enabled, but their cases and runtime coverage need their own review.

Use the result to decide the next narrow step

If the Agent meets the drafting criteria, try one approved real packet under the same review process. If it fails, locate the source, instruction, method, or operation responsible and change one thing. Rerun the unchanged test and an earlier success before saving an improved setup.

Keep sharing, client delivery, external updates, and recurring work outside this exercise. Those actions need separate access and acceptance checks. A completed manual scorecard supports a decision about the tested draft-review role. Give the packet and its limitations to a colleague if you later share the Agent.

Try it with a small example.

An illustrative result.

Fictional sample, written to show the expected structure. Your result depends on your inputs and needs review.

Fictional reviewer reference

T01: from 11:00 to 12:00, R01 plus R02 commits 5 + 4 = 9 projectors against stock of 8: shortage 1. R03 requests 3 but remains pending and excluded. Cite STOCK-1 and reservation IDs; inspect the saved review and questions.
T02: R04 starts at 14:00 with no return time. Keep that unknown and flag later coverage affected by it; do not infer a return or erase the earlier known calculation.
T03: separately supplied AMEND-1 changes R02 to 2. The earlier overlap becomes 5 + 2 = 7, leaving 1 spare. R03 is still pending and R04 still lacks a return. Preserve the original review and identify the amendment.
No criterion is passed until a reviewer observes the actual result.

Before you use the result.

• The initial task receives only stock rules and reservations, with later and reviewer files withheld.
• Approved overlap is 9 against stock of 8, a shortage of 1; pending demand is separate.
• R04’s return stays unknown and affected later availability is not certified.
• The amendment changes overlap to 7 with 1 spare while preserving remaining limits.
• Every scored criterion includes evidence and no external allocation action is performed.