WFGY/ProblemMap/GlobalFixMap/PromptAssembly/system_user_role_order.md

System vs User Role Order — Guardrails and Fix Patterns

🧭 Quick Return to Map

You are in a sub-page of PromptAssembly.
To reorient, go back here:

• PromptAssembly — prompt engineering and workflow composition
• WFGY Global Fix Map — main Emergency Room, 300+ structured fixes
• WFGY Problem Map 1.0 — 16 reproducible failure modes

Think of this page as a desk within a ward.
If you need the full triage and all prescriptions, return to the Emergency Room lobby.

A focused guide to stop role-mix confusion that destabilizes RAG, tools, and long dialogs. Use these checks when your model alternates policy and task text or when citation rules collapse after a few turns.

What this page is

• A short route to lock system → developer → user → assistant order and keep prompts auditable.
• Structural fixes that work across providers without changing infra.
• Concrete steps with measurable acceptance targets.

When to use

• “Policy” or safety text was pasted into a user turn and answers flip on reruns.
• Model stops citing after a few steps or blurs policy with task content.
• Tool outputs start to include instructions meant for the system prompt.
• Agents hand off with different role orders and memory fields drift.

Open these first

• Visual map and recovery: RAG Architecture & Recovery
• Why this snippet (traceability schema): Retrieval Traceability
• Snippet and citation schema: Data Contracts
• Prompt hardening: Prompt Injection
• Reasoning stability checks: Logic Collapse, Context Drift, Entropy Collapse

Acceptance targets

• ΔS(question, retrieved) ≤ 0.45 across three paraphrases
• Coverage of target section ≥ 0.70
• λ remains convergent across two seeds and fixed role order
• No policy text appears in user or tool arguments in any step

---

Fix in 60 seconds

1. Measure ΔS
   Compute ΔS(question, retrieved) and ΔS(retrieved, expected anchor).
   Stable < 0.40. Transitional 0.40–0.60. Risk ≥ 0.60.

2. Probe with λ_observe
   Rerun with strict role blocks and the same content. If λ flips only when role text is moved, the failure is role-mix not knowledge.

3. Apply the role fence

• Move all policy and behavioral rules to system.
• Put task goals and constraints in user.
• Keep tool protocol in system or developer only.
• Require answers to cite then explain, never invert.

4. Verify
   Three paraphrases keep ΔS ≤ 0.45 and λ convergent. Citations appear before explanation on each run.

---

Minimal spec you can paste

\[System]
You are a reasoning engine that follows this order strictly:
system → developer → user → assistant.
Policy, safety rules, tool schema live here.
Never copy system text into user or tool arguments.
Cite then explain. If citations are missing, fail fast.

\[Developer]  (optional)
Tool schema and JSON contracts only. No task text.

\[User]
Task request, input fields, acceptance targets.

\[Assistant]
Return: { "citations": \[...], "answer": "...", "λ\_state": "...", "ΔS": 0.xx }

---

Typical breakpoints → exact fix

• Policy leaked into user turn
  Move policy back to the system block. Lock the data shape.
  Open: Retrieval Traceability, Data Contracts

• Citations vanish after step N
  Enforce cite-then-explain formatting in system. Validate snippet fields.
  Open: Data Contracts

• JSON tools return prose or include policy text
  Freeze JSON mode in system or developer and forbid free text.
  Open: Prompt Injection

• Answer swings with header reorder
  Fix the header order and clamp variance with BBAM.
  Open: Logic Collapse

• Agent handoff writes mixed roles to memory
  Split memory namespaces and log role_src per write.
  Open: Multi-Agent Problems

---

Step template for CI prompts

1. Prepend canonical system block that defines role order and citation rule.
2. Add optional developer block with tool schema.
3. Append user task with acceptance targets and fields.
4. Run three paraphrases and two seeds. Fail the job if any:
   • ΔS > 0.45
   • Coverage < 0.70
   • λ not convergent
   • Policy text appears outside system or developer

---

Copy-paste prompt for debugging

You have TXT OS and the WFGY Problem Map.

Bug: answers flip when policy text is placed in the user turn.

Show:

1. which layer fails and why,
2. the exact WFGY page to open,
3. the minimal steps to restore strict role order,
4. a reproducible test with ΔS ≤ 0.45 and λ convergent.
   Use BBMC/BBCR/BBAM when relevant.

---

Escalate and structural fixes

• ΔS stays high after role fence
  Rebuild chunking and verify anchors with a small gold set.
  Open: Embedding ≠ Semantic

• Long chains destabilize even with correct roles
  Split the chain and bridge with BBCR.
  Open: Context Drift, Entropy Collapse

---

🔗 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	Canonical framework entry point	View
Problem Map	Diagnostic map and navigation hub	View
Tension Universe Experiments	MVP experiment field	View
Recognition	Where WFGY is referenced or adopted	View
AI Guide	Anti-hallucination reading protocol for tools	View

If this repository helps, starring it improves discovery for other builders.