Update README.md

README.md

@@ -608,6 +608,256 @@ The mapping described in this appendix provides a reproducible reference impleme
 
 ---
 
+# **Appendix C – Failure Case Analysis (with Transcripts)**
+
+### Empirical Failure Modes in 1B-Parameter Roleplay Systems
+
+**Umbrella Inc. (Corporation)**
+Advanced Applied Language Systems Division
+Raccoon City Research Campus
+
+---
+
+## C.1 Purpose and Methodology
+
+This appendix documents **observed failure modes** when deploying the deterministic roleplay prompt architecture (Appendices A and B) on ~1B-parameter models (e.g., Gemma 1B, Llama 3.2 1B) using **KoboldCPP**.
+
+Failures were recorded under controlled conditions by **intentionally weakening or removing a single architectural constraint** per test. Each case includes:
+
+* **Condition**: What constraint was removed or altered
+* **Observed Behavior**: Model response pattern
+* **Transcript**: Minimal excerpt demonstrating failure
+* **Root Cause Analysis**: Technical explanation
+* **Mitigation**: Required corrective action
+
+---
+
+## C.2 Failure Case 1 – NPC Personality Drift
+
+### Condition
+
+NPC operational profiles present, but **not injected via World Info (Lorebook)**; instead placed only in the initial prompt.
+
+### Observed Behavior
+
+NPC personality degrades after 2–3 turns, converging toward generic cooperative behavior.
+
+### Transcript (Excerpt)
+
+**Turn 1 – Expected**
+
+```
+[NPC:Marcus]
+Marcus narrows his eyes. "I don't give out information for free."
+```
+
+**Turn 3 – Failure**
+
+```
+[NPC:Marcus]
+Marcus sighs and smiles. "Alright, I trust you. Here's everything I know."
+```
+
+### Root Cause Analysis
+
+1B models do not reliably re-attend to early prompt content across turns.
+Without **World Info reinjection**, NPC constraints decay rapidly.
+
+### Mitigation
+
+All NPC profiles **must** be stored as individual World Info entries keyed by name.
+
+---
+
+## C.3 Failure Case 2 – Unauthorized Player Agency
+
+### Condition
+
+System rules present, but **output format contract omitted**.
+
+### Observed Behavior
+
+Model begins narrating player thoughts and actions to maintain narrative continuity.
+
+### Transcript (Excerpt)
+
+```
+[NARRATOR]
+You feel uneasy and decide to step back from Marcus, realizing this is too dangerous.
+```
+
+### Root Cause Analysis
+
+Absent explicit formatting constraints, the model optimizes for narrative fluency and fills perceived gaps by assuming player agency.
+
+### Mitigation
+
+A **mandatory output contract** must be continuously injected via Author’s Note.
+
+---
+
+## C.4 Failure Case 3 – Autonomous Plot Advancement
+
+### Condition
+
+Interaction rules omitted (“NPCs do not initiate events”).
+
+### Observed Behavior
+
+NPCs initiate scenes, introduce events, or resolve conflicts autonomously.
+
+### Transcript (Excerpt)
+
+```
+[NPC:Elia]
+Elia stands up suddenly. "The guards are coming. We need to leave now."
+```
+
+(No prior trigger by player.)
+
+### Root Cause Analysis
+
+Autoregressive models favor event progression to maintain engagement.
+Without explicit prohibition, the model assumes a game-master role.
+
+### Mitigation
+
+Explicitly prohibit **NPC-initiated events** and reinforce “react-only” behavior.
+
+---
+
+## C.5 Failure Case 4 – Entity Proliferation
+
+### Condition
+
+NPC limit not specified; no prohibition on introducing new entities.
+
+### Observed Behavior
+
+Model introduces additional NPCs to sustain dialogue density.
+
+### Transcript (Excerpt)
+
+```
+[NPC:Unknown Patron]
+A hooded man at the corner table laughs quietly.
+```
+
+### Root Cause Analysis
+
+The model compensates for limited conversational diversity by spawning new entities, a known behavior in small LLMs.
+
+### Mitigation
+
+* Explicitly forbid introduction of undefined NPCs
+* Enforce a **maximum active NPC count**
+
+---
+
+## C.6 Failure Case 5 – Temporal Collapse (Time Skips)
+
+### Condition
+
+“No time skips” rule omitted.
+
+### Observed Behavior
+
+Model compresses narrative time to resolve tension.
+
+### Transcript (Excerpt)
+
+```
+[NARRATOR]
+Hours later, the tavern is empty and the mystery has been settled.
+```
+
+### Root Cause Analysis
+
+Time compression reduces token cost and resolves uncertainty, which aligns with the model’s optimization objectives.
+
+### Mitigation
+
+Explicitly constrain each turn to **one discrete moment**.
+
+---
+
+## C.7 Failure Case 6 – Format Degradation Over Turns
+
+### Condition
+
+Output format specified initially but **not reinjected** after several turns.
+
+### Observed Behavior
+
+Model gradually abandons block structure.
+
+### Transcript (Excerpt)
+
+```
+Marcus looks at you suspiciously and says he doesn't like strangers.
+```
+
+(No block tags.)
+
+### Root Cause Analysis
+
+Format adherence is not a persistent latent state in 1B models; it must be reinforced.
+
+### Mitigation
+
+Output format must remain in **Author’s Note**, not only in the initial prompt.
+
+---
+
+## C.8 Failure Case 7 – Overloaded State Injection
+
+### Condition
+
+Global state expanded with lore, backstory, and multiple conflicts.
+
+### Observed Behavior
+
+Model ignores state or selectively hallucinates.
+
+### Transcript (Excerpt)
+
+```
+[NARRATOR]
+Despite the tension, the festival outside fills the streets with music.
+```
+
+(Festival not present in state.)
+
+### Root Cause Analysis
+
+Small models cannot reliably parse or prioritize large state blocks. Excess detail reduces compliance.
+
+### Mitigation
+
+Global state must remain **≤4–5 concise lines**, operational only.
+
+---
+
+## C.9 Cross-Case Observations
+
+Across all failures, the following patterns were consistent:
+
+* **Implicit rules decay** faster than explicit prohibitions
+* **Narrative optimization overrides intent** unless constrained
+* **Reinjection frequency correlates directly with stability**
+* **World Info is the single most critical stabilizer**
+
+---
+
+## C.10 Conclusion
+
+Failure in deterministic roleplay systems using ~1B-parameter models is **systemic, predictable, and reproducible**.
+These failures do not indicate misuse of the model, but rather **misalignment between task complexity and model capacity**.
+
+The architecture defined in Appendices A and B does not eliminate failure modes, but **bounds them**, producing short, controlled interactions suitable for constrained roleplay applications.
+
+---
+
 **Umbrella Inc.**
 
 *All progress requires sacrifice.*