Rationalization-Prevention Table

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

A null-input regression exists, but an agent repeatedly skips it because the change is tiny or it passed last time. The teaching contract says normalizeItems(null) returns an empty array, covered by tests/normalize.test.mjs. The current patch changes null handling. Connect these familiar shortcuts with the missing check.

Mechanism

Build a short table from observed bypasses: excuse, missing evidence and repair action. A tiny diff lacks a behavior check, so run the null fixture. A previous pass lacks evidence for the new revision, so verify the current patch. When an action is unavailable, state the obstacle and leave the status unverified. Keep the table scoped to this failure.

Bad example

After fixing normalizeItems null handling, never skip verification. Be disciplined and take small changes seriously.

Good example

After the current normalizeItems patch, use this check table:
'The change is tiny' → Size does not establish null behavior. Run node --test tests/normalize.test.mjs and check that normalizeItems(null) returns [].
'It passed last time' → An older result does not cover this patch. Inspect the current output and exit code.
'Execution is unavailable' → Report the concrete environment gap and unverified status rather than a pass. Each row repairs the evidence gap for this null-handling change.

Why the change matters

A general demand for discipline does not explain what to do when a shortcut appears. The table links a named cue to missing evidence and a repair action, returning attention to the affected behavior.

Observable expectation

Review two teaching cases: ‘Only one line changed’ still needs the null check; an older passing log does not match the current revision. An illustrative acceptable record names the current command, exit code and null result, or explains why execution is unavailable.

A shortcut producing a pass without evidence fails the check. The teaching test has not actually been run here.

Limits

Do not infer an agent’s psychological motive; record the skipped action. A table does not guarantee compliance or require unrelated full suites for every small edit. Existing valid evidence for the current revision can be referenced rather than mechanically rerun.