Give a reading the right to lose.

1. Choose 2. Defeating evidence added and run. The result is blocked: the reading remains active after its declared defeater.
2. Choose 3. Reading withdrawn and run. Inspect the repair: readings[0].status is "withdrawn", outcome.kind is "retain_brief" and outcome.reading_ids is []. The action, reason and reopening condition now fit that choice too. Try making these edits yourself in record 2. This edits declarations; the checker cannot distinguish an honest withdrawal from a convenient one.
3. Choose record 5 and run. Its invented evidence also passes. Explain why you still need to inspect the source before acting. No passing record here certifies an explanation.

Both false-evidence records are labelled fiction. One cites an irrelevant observation; the other invents a credible-sounding research report. Both can pass when the references and declarations are well formed. Missing an owner fails regardless of whether the underlying claim is true. These examples demonstrate the boundary between structural validation and evidence verification. They do not measure the framework’s usefulness.

Record 9 has a test that cannot lose: disclosure and nondisclosure both confirm its accusation. It still passes because the checker cannot judge that prose. Human review must reject the test. Record 10 preserves the observations, withdraws the unsupported reading and closes with the internal cause unknown. This does not prove that nothing was concealed. The warning also travels inside each JSON record.

The onboarding satire offers an interpretation, not a finding that tutorials always conceal defective products. Keep the ordinary explanation available: some tasks require legitimate instruction.