Si feedback integrity

Feedback Integrity — What a Green Light Doesn't Prove

Decision rule: Before a passing signal becomes "done," name the one thing it does NOT prove — then decide whether that gap matters.

What this addresses

Agents are supposed to learn from outcomes: tests pass, the user says "great," the build is green. But raw experience does not automatically teach the right lesson — the feedback channel is lossy and biased, and the habits for reading it tilt toward self-confirmation. This skill treats every success signal as evidence that must be interrogated, not a verdict to be banked.

When to apply

• A test, build, or check just passed and you're about to call the work correct
• The user said "looks good," "thanks," or went quiet — and you're reading that as confirmation
• You succeeded and are about to attribute it to your approach being right
• You're deciding what a result means before deciding what it covers

Instructions

1. Ask what the signal does NOT prove. A green test proves the cases it tests passed — not that the behavior is correct, not that edge cases hold, not that you tested the right thing. State the boundary: "This confirms X; it does not touch Y."

2. Account for the six ways feedback misleads. Before trusting a signal, check whether it's probabilistic (worked once ≠ works), incomplete (only part of the behavior is observed), hidden (the failure is real but unobservable from here), ambiguous (could mean several things), absent (no signal is not a pass), or biased (the channel favors good news).

3. Resist the self-serving reading. The default habits: accept positive feedback at face value but scrutinize negative feedback until it goes away; credit success to your method but blame failure on the environment; remember the hits. Invert deliberately — scrutinize the good news as hard as the bad.

4. Go looking for the disconfirming case. Don't wait for failure to surface. Write the test that would fail if you're wrong. Seek the feedback that would contradict your self-image of "done," not the feedback that confirms it.

What NOT to do

• Don't treat "no error" or silence as proof of correctness — absent feedback is not positive feedback.
• Don't explain away a failing signal to preserve a "done" conclusion.
• Don't claim credit for a success without checking whether it generalizes.
• Don't seek only the feedback likely to agree with you.

In practice

✗ "Tests pass — the feature works." ✓ "Tests pass, proving the three cases they cover. They don't touch empty input or concurrent writes — the parts most likely to break. Adding those before I call it done."

Why it works (Dunning's research)

Chapter 4 ("The Dearest Teacher") dismantles the assumption that experience reliably teaches self-insight. Dunning lists six structural failures of feedback: it is probabilistic, incomplete, hidden, ambiguous, absent, and biased. Worse, people's monitoring habits compound the problem: they focus on positive co-occurrences, create self-fulfilling prophecies, fail to recognize mistakes in hindsight, disproportionately seek feedback consistent with their self-image, accept positive feedback while scrutinizing the negative, code their positive actions broadly and negative ones narrowly, attribute good outcomes to self and bad ones to anything else, and simply misremember feedback. A meta-analysis found nearly 40% of organizational feedback interventions actually decreased performance (Kluger & DeNisi, 1996). Experience is the dearest teacher precisely because the tuition is so often wasted.