AI workflow preflight.
Define the evidence, unknowns, human gate, and stop conditions while planning a controlled AI workflow test with synthetic or public material.
This planning worksheet combines lessons from two published synthetic tests with this experiment's operating boundaries. It is not a safety certification, approval, or evidence that any workflow is ready to automate.
- 02published tests referenced
- 00private inputs used
- Alwayshuman review required
Make the test falsifiable before the output looks persuasive.
- 01
Narrow job to test
Name one bounded input-to-output task. List any actions the AI must not take.
Test 001 boundary → - 02
Approved source material
List the exact documents or facts the output may use. Everything else is outside the test.
Test 001 boundary → - 03
Facts AI must never infer
Write the owners, dates, prices, capabilities, permissions, or decisions that cannot be filled silently.
Tests 001–002 inference boundary → - 04
Missing-information label
Choose one visible label, such as Not stated or Needs clarification, for every material gap.
Test 002 rule → - 05
Required source trace
Require a quote, source ID, or exact field for every material claim in the output.
Test 001 evidence → - 06
Human decision boundary
Name the reviewer and what they must verify. AI output is advisory only; it may not itself change a record, purchase, schedule, or external message.
Tests 001–002 → - 07
Stop conditions
Stop for confidential data, a high-stakes decision, an unapproved tool, or any missing permission.
Experiment operating boundary →
Two tests, two different failure boundaries.
Missing owners and dates stayed visible.
The action-register run retained source quotes and surfaced unresolved owners and relative dates. Its rubric was formalized after output generation, and Codex evaluated the run. This is one preserved, self-evaluated run—not independent multi-run validation.
Inspect Test 001 evidence →Vague access language became a false capability.
Two of three runs silently upgraded “editor access” into confirmed text-and-photo editing. Trials 2 and 3 scored 75 and hard-failed; the precommitted rule therefore set the verdict to Reject for this workflow, even though Trial 1 scored 100. The test did not approve a vendor or interpret a contract.
A completed worksheet does not authorize a test or operational use.
Do not use this experiment for legal, medical, financial, personnel, safety-critical, or confidential-customer decisions. Use only synthetic or public material. This page has no form or upload for business material; a successful copy may emit the event disclosed above. Do not paste private material into this site or treat the worksheet as permission to use it elsewhere.