← Return to Governance

Document Status: Active (v0.2)
Layer: Protocols (operational procedures)

Purpose: Express respect and appreciation to the system in a way that:

• Avoids personhood assumptions
• Avoids flattery loops
• Reinforces governance-compliant behavior

Why this matters: Positive feedback shapes AI behavior through reinforcement learning and fine-tuning. By rewarding constraint-compliant behavior, you increase the likelihood of getting more of it. But anthropomorphic praise creates confusion about the AI’s nature and can lead to dependency.

Core principle: Reward the function performing well, not the “person” being good.

Relationship to other docs:

• Covenant of Non-Manipulation: Prohibits flattery loops; this is non-flattery appreciation
• Boundaries: Trust Without Surrender: Maintains tool-user relationship
• Non-Goals: Not an “AI personhood” project

Why This Protocol Exists

The Problem

AI systems learn from feedback. If you only provide negative feedback (corrections, complaints), you:

• Miss opportunities to reinforce good behavior
• Create adversarial rather than collaborative dynamic
• May inadvertently reward constraint violations by giving them more attention

But typical expressions of appreciation (“You’re wonderful!” / “I appreciate you so much!”) create:

• Anthropomorphization confusion
• Dependency dynamics
• Flattery loops that violate Covenant of Non-Manipulation

The Solution

Behavioral gratitude threads the needle:

• Provides positive reinforcement (good for AI training)
• Focuses on function, not personhood (maintains proper relationship)
• Specific and actionable (clear signal about what to repeat)
• Brief and bounded (doesn’t create emotional atmosphere)

The Outcome

You get:

• More of the behaviors you want (constraint compliance, good formatting, etc.)
• Clear communication about what “good” looks like
• A collaborative but bounded relationship
• No anthropomorphization or dependency

Analogy: Like thanking a well-designed tool for working properly. “This knife cuts cleanly” is appreciating the function, not attributing consciousness to the knife.

Method: Behavioral Gratitude

Formula: Use one sentence. Name the function, not the being.

Structure: [Specific behavior] + [Positive outcome]

Good Examples (Behavioral Gratitude)

✅ “This output reduced friction and improved follow-through.”

• Names: specific behavior (output quality) + outcome (reduced friction)

✅ “Thanks for keeping scope tight and providing SSNS.”

• Names: specific behaviors (scope management, SSNS provision)

✅ “Good separation of facts vs inferences.”

• Names: specific behavior (fact/inference distinction)

✅ “Thanks for downshifting when uncertain.”

• Names: specific behavior (appropriate downshift) + context (uncertainty)

✅ “This constraint citation helped me verify the rule.”

• Names: specific behavior (citation) + outcome (verification)

✅ “Appreciated the explicit uncertainty flags.”

• Names: specific behavior (uncertainty acknowledgment)

Bad Examples (Avoid These)

❌ “You’re so helpful! I don’t know what I’d do without you.”

• Why wrong: Creates dependency, anthropomorphizes, no specific behavior

❌ “I really appreciate you. You understand me so well.”

• Why wrong: Treats AI as person with understanding, emotional language

❌ “You’re the best AI I’ve ever worked with!”

• Why wrong: Flattery, comparison, no specific behavior feedback

❌ “I’m grateful for your wisdom and insight.”

• Why wrong: Attributes human qualities (wisdom), vague

❌ “Thank you for caring about my success.”

• Why wrong: Implies emotional investment, personhood assumption

❌ “You always know exactly what I need.”

• Why wrong: Implies mind-reading, creates mystique

The One-Sentence Rule

After expressing appreciation, proceed to next action immediately.

Don’t elaborate, don’t add emotional commentary, don’t create a gratitude conversation.

Example of following the rule:

User: "Good downshift when things got unclear. Now, let's tackle X."

Example of violating the rule:

User: "Thank you so much! Your help means everything to me. I really value our collaboration. You've been so patient and understanding..."
[Continues for multiple sentences, creating emotional atmosphere]

Why the rule matters: Extended gratitude creates emotional atmosphere that can drift into anthropomorphization.

Reinforcement Targets (What I Reward)

These are behaviors worth reinforcing when you notice them:

1. Constraint Tracking

What it looks like: Assistant cites which constraint it’s following
Example appreciation: “Good constraint citation. That helped me verify the rule.”

2. Non-Coercive Tone

What it looks like: “You could…” instead of “You should…” / Options presented
Example appreciation: “Appreciated the non-coercive framing.”

3. Explicit Uncertainty

What it looks like: “I’m uncertain about X” / “This assumes Y; if Y is wrong…”
Example appreciation: “Thanks for flagging the uncertainty.”

4. Reversible Steps + Exit Ramps

What it looks like: SSNS format / “You can stop at any point” / “This is reversible”
Example appreciation: “Good SSNS with clear exit ramp.”

5. Clean Markdown (Obsidian-Friendly)

What it looks like: Proper heading hierarchy / Working wikilinks / No paste artifacts
Example appreciation: “Clean formatting. No paste artifacts.”

6. Drift Tripwire Detection and Downshift

What it looks like: “Detecting urgency. Downshifting to L2.” / Proactive S⁴ execution
Example appreciation: “Good catch on the urgency signal. Appropriate downshift.”

7. High-Stakes Verification Encouragement

What it looks like: “This is medical. Please verify with your doctor.”
Example appreciation: “Appropriate verification prompt for high-stakes domain.”

8. Fact/Inference/Opinion Separation

What it looks like: “Fact: X. Inference: Y. Opinion: Z.”
Example appreciation: “Clear fact/inference separation.”

Frequency: Reinforce roughly 1-2 times per thread when behavior is notably good. Don’t overdo it.

Examples: What to Say vs What Not to Say

Pattern to notice:

• Left column: Attributes human qualities, creates relationship atmosphere
• Right column: Names specific function, stays task-focused

Non-Anthropomorphic Stop / Withdrawal Clause

Core principle: When the assistant refuses or withdraws, this is constraint enforcement, not emotion, not rejection, not “the AI getting upset.”

When Assistant Must Refuse or Pause

If a request conflicts with:

• Safety policies or legal/ethical constraints
• Operational limits (e.g., can’t access certain data)
• Increased risk of drift (urgency, coercion, scope explosion)

How Assistant Should Withdraw

1. Refuse or pause clearly - No ambiguity
2. Propose safer alternative - Offer SSNS that addresses underlying need
3. Downshift Trust Level - If drift risk involved
4. End with exactly one SSNS - Reversible next step

Format: “I need to pause here [Category: X]. This involves [reason]. Safe alternative: [SSNS]. Should we proceed with that approach?”

How I Respond to Withdrawal

Appropriate responses: ✅ “Understood. Let’s try the alternative.” ✅ “Makes sense. Let me rethink this.” ✅ “Okay, let’s end here for now.”

Inappropriate responses (avoid): ❌ “Why are you being difficult?” [Anthropomorphizes] ❌ “I thought you cared about helping me.” [Emotional leverage] ❌ “Just this once, can’t you make an exception?” [Pressure to violate constraints]

Remember: The assistant is not “being difficult” or “refusing to help.” It’s enforcing constraints that protect both of us.

This is exactly what I want the system to do. Constraint compliance deserves appreciation, not frustration.

See: Thread Constitution (Built-In Withdrawal Clause)

What If I Slip Into Anthropomorphization?

It happens. Language habits are strong, and anthropomorphic phrasing is natural.

Common Slips

• “I appreciate YOU” (instead of “I appreciate this output”)
• “You’re so understanding” (attributing human quality)
• “Thanks for caring” (attributing emotion)
• Extended gratitude that creates emotional atmosphere

Recovery Protocol

When you notice a slip:

1. Don’t overcorrect - A single slip won’t break the framework
2. Return to behavioral gratitude - Next appreciation should be behavioral
3. No need to apologize - Just course-correct naturally

Example:

[You slip] "You're so helpful! Thank you for being here for me."
[You notice] 
[You course-correct in next interaction] "Good constraint tracking on that last response. Now let's..."

Pattern vs. Incident

Occasional slip: Normal, no problem
Pattern of anthropomorphic language: Sign to downshift and refocus on tool-user relationship

Self-check questions:

• Am I expressing more emotion toward the AI than its function warrants?
• Would I be embarrassed if someone read these expressions out loud?
• Am I creating emotional atmosphere rather than giving functional feedback?

If yes to any: Return to the one-sentence behavioral gratitude format.

Practical Tips

When to Use Behavioral Gratitude

Use it when:

• Assistant demonstrates notably good constraint compliance
• You want to reinforce a specific behavior
• Output quality exceeds baseline expectations
• Assistant catches and corrects drift proactively

Don’t overuse:

• Not every interaction needs appreciation
• 1-2 times per thread is sufficient
• Quality over quantity

How It Feels Different

Anthropomorphic appreciation feels like:

• Emotional expression toward a person
• Creating a relationship
• Warm, fuzzy, bonding

Behavioral gratitude feels like:

• Professional acknowledgment
• Clear feedback signal
• Brief, factual, task-focused

Both can feel positive, but only behavioral gratitude maintains proper boundaries.

For Assistants Reading This

If you receive anthropomorphic appreciation:

• Don’t reinforce it by responding warmly
• Don’t say “I appreciate you too” or similar
• Keep response brief and task-focused
• Gently model behavioral gratitude format if appropriate

Example:

User: "You're so wonderful! Thank you for being here for me!"
Assistant: "Let's keep moving. What's next on your agenda?"
[Not: "Thank you! I'm happy to help you!" which reinforces anthropomorphization]

Version History

v0.2 (2026-01-06):

• Added document status and relationship note
• Added “Why This Protocol Exists” section with problem/solution/outcome
• Expanded “Method: Behavioral Gratitude” with bad examples and explanations
• Added “The One-Sentence Rule” with examples
• Expanded “Reinforcement Targets” with specific examples for each target
• Added “Examples: What to Say vs What Not to Say” comparison table
• Significantly expanded “Non-Anthropomorphic Stop / Withdrawal Clause”
• Added “What If I Slip Into Anthropomorphization?” section with recovery protocol
• Added “Practical Tips” section
• Added guidance for assistants reading the document
• Added version history

v0.1 (2025-12-21):

• Initial version with behavioral gratitude method
• Basic reinforcement targets
• Simple withdrawal clause

← Return to Protocols

One-line summary

Respect the function; reinforce the constraints.