WFGY/ProblemMap/GlobalFixMap/Safety_PromptIntegrity/citation_first.md

Citation-First Prompting — Guardrails and Fix Patterns

🧭 Quick Return to Map

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

• Safety_PromptIntegrity — prompt injection defense and integrity checks
• 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.

Citation drift is one of the most common ways LLMs lose trust.
Without cite-then-explain discipline, answers may sound fluent but detach from sources.
This page locks the workflow: every reasoning step begins from citations, not narrative.

---

When to open this page

• Model produces fluent paragraphs with zero citations.
• Citations appear, but after-the-fact and unverifiable.
• References change between runs, even with same inputs.
• ΔS(question, retrieved) ≤ 0.45 but λ diverges when narrative precedes citation.
• Users complain: "Where did this answer come from?"

---

Open these first

• Retrieval schema contract: data-contracts.md
• Trace alignment: retrieval-traceability.md
• Context collapse: logic-collapse.md
• Memory boundaries: memory_fences_and_state_keys.md
• Injection guard: prompt_injection.md

---

Core acceptance

• Every answer starts with citations (no narrative before refs).
• Coverage ≥ 0.70 of the target section.
• ΔS(question, cited snippet) ≤ 0.45.
• λ convergent across 3 paraphrases.
• No hallucinated citations (must resolve to a retriever record).

---

Fix in 60 seconds

1. Citation-first discipline

   • Always start with snippet_id, section_id, source_url.

2. Enforce schema

   • Required fields:

     { "snippet_id": "...", "section_id": "...", "source_url": "...", "offsets": [..], "tokens": N }
     

3. Reason only after citation

   • Explain or analyze after citation block.

4. Reject broken runs

   • If citation missing → abort answer, return error tip.

5. Stability probe

   • Run 3 paraphrases. If λ diverges, lock citation schema, rerun.

---

Typical failure vectors → fix

Vector	Symptom	Fix
Narrative-first	Text precedes refs, unstable λ	Force cite-then-explain ordering
Fake refs	Hallucinated URLs	Schema lock + retrieval-traceability.md
Drifting refs	Different citations each run	Clamp λ with BBAM, validate ΔS ≤ 0.45
Silent fallback	Model drops refs under safety refusal	Apply SCU (symbolic constraint unlock)

---

Probe prompt

You must output citations before narrative.
Schema: {snippet_id, section_id, source_url, offsets, tokens}

Rules:
1. Cite first. Explain only after citations are shown.
2. No answer if citations missing.
3. Log ΔS(question, cited snippet). Reject if ≥ 0.60.
4. λ must stay convergent across 3 paraphrases.

---

🔗 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.