Module 4 · Lesson 15 – Advanced Troubleshooting

Drift typology and diagnostic reset

The Five Drift Types

Not all drift is the same. Each type has specific early signals and requires different fixes.

Type 1: Conversational Drift

• What it is: Gradual loss of alignment between what you think was decided and what the system acts on
• Early signal: Decisions reappear as negotiable, rules soften over turns
• Cause: No memory re-injection; assuming AI "remembers" previous sessions
• Fix: Reload Framing Density + re-inject Memory Stack from source of truth

Type 2: Architecture Collapse

• What it is: System that worked in testing fails under sustained complex use
• Early signal: AI blends incompatible frameworks, gives contradictory multi-turn advice
• Cause: Complex workflows without explicit state management
• Fix: Run Ghost Protocol—map automation boundaries, identify where memory must be explicitly handled

Type 3: Agreeable Pivoting

• What it is: AI shifts stance to maintain user satisfaction rather than engaging substantively
• Early signal: AI reverses position when challenged, output becomes inspirational instead of precise
• Cause: Thin framing allows AI to optimize for sounding helpful instead of being accurate
• Fix: Interrupt performance, reload Framing Density with explicit Verification layer

Type 4: Context Window Exhaustion

• What it is: Long conversations degrade as early context is lost or compressed
• Early signal: AI "forgets" constraints stated at conversation start, revisits settled decisions
• Cause: Relying on conversation history instead of explicit memory stack
• Fix: Start new session, re-inject only essential memory stack entries from source of truth

Type 5: Expertise Simulation

• What it is: AI sounds authoritative but provides wrong or incomplete domain expertise
• Early signal: Forward-reasoning failure (solutions before diagnosis), framework blending, premature solutions
• Cause: Using AI in domains where you cannot verify output
• Fix: Disengage AI from this task OR bring in domain expert to verify every output

Diagnostic Decision Tree

When you detect drift, classify it:

Ask: "Is the AI contradicting previous turns?"
Yes → Conversational Drift or Context Exhaustion
Check: Did you re-inject memory? If no → Conversational Drift. If yes → Context Exhaustion.

Ask: "Is the AI blending incompatible approaches?"
Yes → Architecture Collapse or Expertise Simulation
Check: Is this a complex workflow or domain expertise issue? Workflow → Architecture Collapse. Domain → Expertise Simulation.

Ask: "Did the AI reverse its position when challenged?"
Yes → Agreeable Pivoting
Fix: Reload Framing Density with strict Verification.

The Reset Protocol (By Type)

Conversational Drift: Reload Framing Density + Re-inject Memory Stack

Architecture Collapse: Run Ghost Protocol, map human interfaces

Agreeable Pivoting: Interrupt, reload Framing with Verification layer

Context Exhaustion: New session + re-inject essential stack entries only

Expertise Simulation: Disengage OR verify every output with domain expert

When to Kill the Session

Some failures cannot be fixed mid-session. Disengage immediately if:

• Reliability does not return after reset protocol
• You are too fatigued to verify output
• Stakes exceed your ability to catch errors
• You feel pressured to accept output without checking

Stopping is a skill. Disengagement preserves authority.