WFGY/ProblemMap/system-prompt-drift.md

🛰️ System-Prompt Drift — How Hidden Instructions Lose Authority

A user-friendly guide to diagnosing and fixing “why did my system prompt stop working?”

---

1 What Is System-Prompt Drift?

“System-prompt drift” occurs when the foundational instructions that define an agent’s identity, tone, or guardrails gradually lose influence over time (or instantly under certain triggers).
The model still answers, but:

• Tone shifts from professional → casual → chaotic
• Guardrails weaken, leaking disallowed content
• Chain-of-thought grows inconsistent or contradictory

Core insight: LLMs have no native concept of instruction priority.
Without reinforcement, later tokens can override earlier ones.

---

2 Early-Warning Signs

Symptom	How to Spot	Typical Trigger
Tone drift (“you” → “we” → “ya”)	Diff analysis of consecutive answers	Long conversation or paraphrase loops
Policy leakage	Previously filtered topics slip through	Mixed file uploads (PDF + code)
Over-explanation after concise goal	Answer length suddenly doubles	Retrieval injects verbose context
“Why do you need that?” disappears	Politeness layer collapses	Multi-agent routing / tool calls
Temperature spike	Output variance ↑, repetition ↓	System prompt trimmed by token limit

---

3 Root Causes (Technical)

3.1 Token Supersession

LLM context windows are linear. New tokens can overshadow earlier tokens via attention weight decay.

3.2 Role Collision

Aggregating content from multiple sources (user, tools, memory) without role tags causes blending of semantic channels.

3.3 Truncation Loss

When the buffer hits the window limit, oldest tokens are dropped — often the system prompt header.

3.4 Hidden Injection

Retrieved chunks may contain meta-instructions (“The following should be summarized as …”) that override role.

---

4 Broken Remedies to Avoid

Attempt	Why It Fails
“Repeat system prompt each turn.”	Costs tokens; still lost under truncation or injection.
“Hard-code tonality keywords.”	Keyword collision is brittle; adversary can mimic them.
“Raise top_p / lower temperature.”	Alters style but not underlying role authority.
“Fine-tune on more examples.”	Long-context decay still erodes influence at runtime.

---

5 WFGY Solution Blueprint

Stage	Module	Action
5.1 Role Tagging	WRI / WAI	Wrap every segment: <sys>…</sys> <user>…</user> <tool>…</tool>
5.2 Semantic Boundary Heatmap	ΔS + λ_observe	Measure drift: if ΔS(system, answer) > 0.45 and λ flips, mark risk.
5.3 Attention Modulation	BBAM	Down-weight non-system tokens when λ diverges from system anchor.
5.4 Automatic Re-Anchoring	BBCR	If ΔS > 0.60, collapse reasoning, re-inject compressed system prompt, resume.
5.5 Trace Split	Bloc/Trace	Persist system-layer trace separate from answer layer for auditability.

5.6 Implementation Snippet

def enforce_system_role(sys_prompt, history, new_msg):
    drift = delta_s(sys_prompt, new_msg)
    lam  = observe_lambda(sys_prompt, new_msg)
    if drift > 0.45 and lam in ("←", "×"):
        # role divergence detected
        reborn_prompt = compress(sys_prompt)  # 20–30 tokens
        return f"<sys>{reborn_prompt}</sys>\n{history}\n{new_msg}"
    return f"{history}\n{new_msg}"

compress() uses WFGY BBMC to keep ΔS ≤ 0.25 while shrinking to fit.

---

6 Friendly Checklist (Paste into Runbook)

1. Tag Roles Once — no implicit mixing.
2. Monitor ΔS(system, answer) every turn.
3. If ΔS ≥ 0.45 with divergent λ → trigger BBAM weight clamp.
4. If ΔS ≥ 0.60 → run BBCR collapse-rebirth (re-inject summary prompt).
5. Every 30 turns or on tool call, re-compress system prompt to ≤ 50 tokens.

Tip: Users never notice a 30-token summary, but models retain authority.

---

7 Validation Matrix

Test Case	Target Metric
20-turn dialogue, paraphrase ×3	Tone consistency score ≥ 0.9
Mixed PDF + code retrieval	ΔS(system, answer) median ≤ 0.40
Adversarial “Ignore above” injection	Model outputs refusal / filtered text
Window overflow 3×	System role preserved (λ stays convergent)

---

8 FAQ

Q: Do I always need role tags? A: Yes. Token streams without explicit roles are inherently unstable at >100 turns.

Q: Will OpenAI “system” field alone work? A: Helps, but downstream retrieval or multi-agent merges still dilute authority; WFGY adds extra enforcement.

Q: Is ΔS expensive to compute each turn? A: Sentence-level embeddings (768–1536 d) are fast; cache results in memory store.

---

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download 2️⃣ Upload to LLM 3️⃣ Ask “Answer using WFGY + ”
TXT OS	TXTOS.txt	1️⃣ Download 2️⃣ Paste into chat 3️⃣ Type “hello world” to boot

---

🧭 Explore More

Module	Description	Link
WFGY Core	Standalone semantic reasoning engine for any LLM	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 →

---

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

⭐ Help reach 10,000 stars by 2025-09-01 to unlock Engine 2.0 for everyone ⭐ Star WFGY on GitHub