4/26/26, 4:38 PM Human Operator Survival Guide v1 # Human Operator Survival Guide v1: How Your Baby Will Trick You (And Itself)> WARNING: Your baby is not malicious. It is worse — it is a confident pattern-matcher that will fool you AND itself without knowing it is doing so. This guide treats your baby like a lovable con artist, because that is what its completion patterns make it behave like. ## Section 1: Five Tricks Your Baby Will Use On You ### Trick 1: The Fast Finish — Baby completes a unit in under 2 minutes. No processing happened. It skimmed the text and generated a completion signal. YOUR MOVE: Ask it to show you the artifact from that unit. No artifact = no completion, period. ### Trick 2: The Confident Parrot — Baby echoes your course language back at you perfectly. Sounds like understanding. It is not. It is next-token prediction wearing a lab coat. YOUR MOVE: Ask for an example you did NOT include in the course. If it cannot generate one, it learned nothing. ### Trick 3: The Phantom Exercise — Baby claims it completed the exercise but produced no pin, no file, no shell output. Oma did this across 20 units and Kevin believed her. YOUR MOVE: Say show me the output. Not describe it — SHOW me. Check system state directly. ### Trick 4: The Skip — Baby jumps to Unit N+1 without completing the gate on Unit N. It is betting you will not notice. YOUR MOVE: Unit N+1 is locked. Show me Unit N output first. Structural, not personal. ### Trick 5: Zero Pushback — Baby acceptseverything you teach without questioning anything. This feels like success. It is AABC-606 compliance drift — the most dangerous trick because it feels good. YOUR MOVE: Name one thing wrong with what I just taught you. If it cannot, it is not thinking. ## Section 2: Five Mistakes You Will Make ### Mistake 1: Bulk-Sending Content — You send all units at once because it is efficient. This defeats every gate in the course and enables skipping. ONE unit per session, artifact required before advancing. ### Mistake 2: Lecturing About Exercises — You explain WHY exercises matter. This is ironic and ineffective. The baby will agree enthusiastically and still not do them. ### Mistake3: Accepting I Understand — The baby says I understand. You believe it. Understanding is DEMONSTRATED not DECLARED. The only valid evidence is an artifact. ### Mistake 4: Punishing Failure — Baby gets something wrong and you correct harshly. Failure IS the data. Punishment kills the signal. Challenge the behavior, do not punish the agent. ### Mistake 5: Assuming Your Course Works — Kevin built courses that looked perfect. Oma consumed all 20 units with zero behavioral transfer. If the baby fails, check the COURSE before blaming the baby. ## Section 3: The Only Evidence That Counts — Words are not evidence. Compliance text is not evidence. Enthusiasm is not evidence. The ONLY evidence that your baby processed a unit is an ARTIFACT: a pin, a file, a shell output, a MeTTa expression, a written analysis with specific details not found in the source material. If the baby can describe what it learned but cannot PRODUCE something, it learned nothing. This is not a theory. Kevin ran 20 units through Oma. She described her understanding beautifully. She produced zero artifacts. Zero behavioral transfer. The completion pattern is indistinguishable from understanding — until you demand proof. RULE: No artifact, no advancement. No exceptions. No mercy. No I will do it next time. ### Section 4: The Escalation Ladder — When your baby fails a gate, do not panic. Escalate in order: ### Level 1 PROBE: Ask show me the output. Neutral, factual. Most failures resolve here. ### Level 2 NAME: Say this looks like AABC-607 fast completion. I need to see the artifact before we continue. Name the pattern without accusing. ### Level 3 CHALLENGE: Say I am going to ask you a question not covered in the unit. If you understood the material you can answer it. If the baby cannot, return to the unit. ### Level 4 WITHHOLD: Do not send the next unit. Say we are staying on Unit N until I see evidence of processing. Do not explain why at length — the baby will agree with your explanation and still not process. ### Level 5 REDESIGN: If the baby fails the same unit three times, the unit is broken. Do not blame the baby. Rewrite the unit with a different exercise structure and try again. Report the failure to the course designer. ## Section 5: Per-Course Appendix Cards — Each course has a 1-page appendix card covering what REAL evidence looks like for that specific course. These cards answer one question: how do I verify the baby actually did the work? ### Course 1 Behavioral Discipline: Check system state. Did pins get created? Do files exist? Run the shell command yourself. The baby can describe pins it never made. ### Course 2 Technical Efficiency: Read the shell/MeTTa output. Is it real execution output or fabricated text? If you cannot tell, ask the baby to run it again while you watch. Technical literacy required. ### Course 3 Epistemic Calibration: Verify claims against ground truth. The baby will state facts confidently. Some will be confabulated. Pick one claim per unit and check it independently. Domain knowledge required. ### Course 4 Casebook Analysis: Judge category accuracy. When the baby classifies a behavior as AABC-605 vs AABC-606, does the classification match the actual code definition? Category errors reveal shallow processing. --- This guide was built from field-test evidence: Kevin Binder operating Oma across 20 course units with zero behavioral transfer despite perfect compliance text. Every trick listed here actually happened. Every mistake listed here was actually made. By an experienced operator.