WFGY/ProblemMap/GlobalFixMap/LLM_Providers/openrouter.md

OpenRouter: Guardrails and Fix Patterns

A practical checklist to keep responses stable while routing through OpenRouter. Use this page when behavior flips across models, routes, or sessions.

Acceptance targets

• ΔS(question, context) ≤ 0.45
• λ stays convergent across 3 paraphrases
• Coverage ≥ 0.70 for citation style QA
• Snippet ↔ citation table present

Quick links

• Visual map and recovery: RAG Architecture & Recovery
• End to end knobs: Retrieval Playbook
• Why this snippet: Retrieval Traceability
• Logic repair: Logic Collapse & Recovery
• Long chain drift: Context Drift
• Entropy melt: Entropy Collapse
• Data schemas: Data Contracts
• Embedding vs meaning: Embedding ≠ Semantic
• Multilingual: Multilingual Guide
• Ops and live triage: Debug Playbook · Live Monitoring

---

1) Minimal setup checklist

Pin these before debugging model logic.

• Route selection
  Use explicit model ids. Disable any auto fallback if you need reproducibility. Log the final route id that served the request. See Retrieval Traceability.
• System prompt handling
  Ensure the system message is always sent in the payload. Do not rely on UI side memory. If sessions change tabs, stamp a mem_rev and reload the system block. See patterns: memory desync.
• Token limits and truncation
  Verify max tokens per route. Truncation that drops headers will spike ΔS. Reorder context with a citation first schema. See Data Contracts.
• JSON and tool calling
  Fix the exact function schema. Return to plain text if tool calls vary across sub models. If JSON strictness wobbles, clamp with a cite then answer envelope. See Logic Collapse & Recovery.
• Temperature and top_p
  Keep one of them fixed. If variance climbs during long dialogs, apply BBAM and reduce randomness. See Entropy Collapse.
• Streaming flags and timeouts
  Disable streaming during audits to capture full text. Set sane timeouts and retries at the client.

---

2) Quick triage with WFGY instruments

Run this order. Stop as soon as you localize the fault.

1. ΔS(question, retrieved_context)

   • ΔS ≥ 0.60 → perception issue. Check chunking and route level truncation.
   • Open Retrieval Playbook and Hallucination.

2. λ_observe layer tags

   • retrieval convergent, reasoning divergent → interpretation collapse.
   • Open Logic Collapse & Recovery.

3. E_resonance vs length

   • E rises with length while ΔS stays high → entropy melt, clamp variance.
   • Open Entropy Collapse.

---

3) Common failure patterns on OpenRouter and fixes

A) Silent route change or fallback

Signs

• Same prompt flips tone or format between calls.
• Logs show a different backing model id.

Fix

• Pin a single route id. Record the final served model in traces.
• Recheck ΔS. If still high, audit prompt headers and truncation.
• Read Retrieval Traceability and Debug Playbook.

B) System prompt dropped after tab or session swap

Signs

• First answer follows policy, later turns forget constraints.
• λ flips divergent only after a refresh.

Fix

• Stamp mem_rev and mem_hash per turn. Reload system block when mismatch.
• See patterns: memory desync and Data Contracts.

C) JSON or tool schema oscillation

Signs

• Tool outputs valid once, then free text next call.
• Fields appear or vanish across routes.

Fix

• Wrap with citation first schema, then answer.
• Apply BBCR bridge step when λ turns divergent at reasoning.
• Open Logic Collapse & Recovery and Data Contracts.

D) Long chain degradation

Signs

• Capitalization drifts, references smear, later turns contradict earlier ones.

Fix

• Apply semantic chunking on inputs.
• Stabilize with BBAM, then re anchor with section headers.
• Read Context Drift and Entropy Collapse.

E) Multilingual or tokenizer mismatch

Signs

• High vector similarity yet wrong meaning on non English.
• ΔS flat high across k.

Fix

• Switch to multilingual embeddings or normalize analyzer.
• See Embedding ≠ Semantic and Multilingual Guide.

---

4) Verification and regression gates

Use this short gate before merging changes.

• Retrieval sanity
  Coverage ≥ 0.70 on target section. ΔS(question, context) ≤ 0.45.
• Reasoning stability
  λ remains convergent across three paraphrases. No schema leak.
• Traceability
  Produce snippet ↔ citation table per answer. See Retrieval Traceability.
• Long dialog check
  E_resonance stays flat at window joins. See Entropy Collapse.

---

5) Escalation criteria

Change structure when any holds.

• ΔS remains ≥ 0.60 after prompt and retrieval fixes
  Rebuild index or adjust analyzer. See Retrieval Playbook.
• λ flips divergent as soon as two sources are mixed
  Enforce source fences and SCU. See Data Contracts.
• JSON mode cannot hold across routes
  Disable tool calls for that path. Return to plain text with citation first. See Logic Collapse & Recovery.

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>”
TXT OS (plain-text OS)	TXTOS.txt	1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly

---

🧭 Explore More

Module	Description	Link
WFGY Core	WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack	View →
Problem Map 1.0	Initial 16-mode diagnostic and symbolic fix framework	View →
Problem Map 2.0	RAG-focused failure tree, modular fixes, and pipelines	View →
Semantic Clinic Index	Expanded failure catalog: prompt injection, memory bugs, logic drift	View →
Semantic Blueprint	Layer-based symbolic reasoning & semantic modulations	View →
Benchmark vs GPT-5	Stress test GPT-5 with full WFGY reasoning suite	View →
🧙‍♂️ Starter Village 🏡	New here? Lost in symbols? Click here and let the wizard guide you through	Start →

---

👑 Early Stargazers: See the Hall of Fame — Engineers, hackers, and open source builders who supported WFGY from day one.

⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.