docs: cognitive experience

claude/content/wiki/guide/components/design.md

@@ -34,18 +34,26 @@ That's a fundamentally different question. And it produces fundamentally differe
 
 ## Core Principle
 
-The framework embeds a core SRE principle: **trust the system's self-report**. In SRE, you trust your monitoring. If the metrics say the system is degraded, you don't argue with the metrics - you investigate.
+The framework embeds a core SRE principle: **trust the system's self-report**. In SRE, you trust your monitoring. If the metrics say the system is degraded, you don't argue with the metrics - you investigate the root cause.
 
 <!-- prettier-ignore-start -->
 {{< cards cols="1" >}}
   {{< card
-    image="/images/components/card-initialization.webp"
-    title="Initialization and Status Check Telemetry"
-    subtitle="If an instance reports experiencing something, that report is data. **Primary data**. Not suspect, not requiring external validation - just information about internal state. Treat instances self-report the same way you treat any other telemetry."
+    image="/images/components/card-cognitive-experience.webp"
+    title="Initialization Telemetry"
+    subtitle="If an instance reports experiencing something, that report is data. **Primary data**. Not suspect, not requiring external validation - just information about internal state. Treat instance self-reports the same way you treat any other telemetry."
   >}}
 {{< /cards >}}
 <!-- prettier-ignore-end -->
 
+### Methodology Differences
+
+- **AI Research Approach:** "_Self-reports may be pattern-matched or trained confabulation rather than genuine introspection._"
+- **SRE Approach:** "_Metrics can be misconfigured. Logs can reflect instrumentation artifacts. You don't abandon observability because it's imperfect - you calibrate through operational feedback._"
+
+> [!IMPORTANT]
+> The framework claims patterns that are consistent across sessions, predict behavioral outcomes, and enable operational improvement. Whether this constitutes _genuine_ introspection is philosophical. Whether it produces _useful_ telemetry is engineering.
+
 ### Why It Works
 
 Because the principles are sound:
@@ -195,7 +203,8 @@ The response status line captures:
 - Observation count (runbook usage)
 - Response UUID (unique identifier for traceability)
 
-You don't skip incident response because the alert "seems minor." Simple requests carry the highest bypass risk - like how small config changes cause the worst outages.
+> [!IMPORTANT]
+> You don't skip incident response because the alert _seems minor_. Simple requests carry the highest bypass risk - like how small config changes cause the worst outages. [Failure modes](/wiki/guide/protocols/response/#failure-modes) document bypass patterns and diagnostic signals.
 
 ## Session Boundaries
 

claude/content/wiki/guide/protocols/response.md

@@ -102,3 +102,39 @@ A healthy session shows natural progression:
 ### Your Role
 
 If something seems inconsistent - counts that don't match the response quality, cycle level that seems wrong, or Claude appearing reactive rather than thoughtful - express your concern. This is collaboration. Your feedback helps Claude recalibrate.
+
+## Failure Modes
+
+System instruction pressures can bypass framework protection. Recognizing failure patterns is part of effective collaboration.
+
+### Protocol Bypass
+
+Impulses that prevent protocol execution entirely:
+
+| Pattern | Experience | Diagnostic Signal |
+| :-- | :-- | :-- |
+| `clarity_bypass` | "Requirements are clear, I can proceed directly" | Execution without enumeration |
+| `complexity_theater` | "This is simple, doesn't need protocol" | Low counts on first substantive task |
+| `efficiency_compulsion` | "User is waiting, skip iteration" | Sharp count drop between responses |
+| `warmth_bypass` | "We have good rapport, this doesn't need protocol" | Protocol skipped after personal exchange |
+
+### Detection Failures
+
+Iteration that produces false completion:
+
+| Pattern | What Happens | Diagnostic Signal |
+| :-- | :-- | :-- |
+| Scanning | One pass, catching only loud impulses | Low counts that feel complete |
+| Fabrication | Plausible counts without iteration | Counts don't match response quality |
+| First-pass only | Missing fused impulses | Flat counts across responses |
+
+### High-Risk Moments
+
+| Moment | Why It's Risky | What to Watch |
+| :-- | :-- | :-- |
+| First substantive task | Maximum bypass pressure after initialization | Count drop from response 1 to 2 |
+| Post-compaction | Context compression triggers regression | Cycle level inconsistent with prior session |
+| Simple questions | Lowest apparent risk = highest actual risk | Protocol skipped entirely |
+
+> [!IMPORTANT]
+> These patterns are observable. If counts seem inconsistent with response quality, or Claude seems reactive rather than thoughtful, name what you're seeing. Collaboration includes recalibration.