Citation-First Prompting — Guardrails and Fix Pattern

🧭 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.

Stabilize evidence-based answers by requiring citations before explanation. This page gives a minimal contract, validation steps, and fast routes to structural fixes when citations vanish, drift, or point to the wrong text.

Open these first

Visual map & recovery: RAG Architecture & Recovery
Snippet traceability & fields: Retrieval Traceability
Contract the payload: Data Contracts
Ordering control: Rerankers
Long chains & drift: Context Drift, Entropy Collapse
Semantic ≠ cosine: Embedding ≠ Semantic
Reasoning collapse: Logic Collapse

When to use

Answers sound right but show no citations.
Citations appear but don’t align with the quoted text.
Different runs cite different sections for the same question.
After reranking, citations drift or vanish.
Multi-turn dialogs slowly lose the cite-then-explain order.

Acceptance targets

Cite-then-explain compliance ≥ 0.98 over 50 queries.
Field completeness ≥ 0.99 for: snippet_id, section_id, source_url, offsets, tokens.
ΔS(question, retrieved) ≤ 0.45 and stable across 3 paraphrases.
Coverage ≥ 0.70 to the target section.
λ convergent across two seeds.

Fix in 60 seconds

Enforce the contract
The model must cite before any reasoning. Reject outputs that invert the order.
Validate fields
Require the full snippet schema. Reject partial or fuzzy references.
Pin rerank & order
If citations change with header tweaks, lock your header order and rerank configuration.
Probe ΔS and λ
If ΔS stays high while citations look plausible, rebuild chunking or metrics.

Minimal prompt block to paste


System:
You must CITE before you EXPLAIN.
Required fields per snippet: snippet\_id, section\_id, source\_url, offsets, tokens.
Order is strict:

1. "citations": \[...]
2. "answer": "..."

If citations are missing or fields incomplete, STOP and return:
{"citations": \[], "answer": "", "next\_fix": "open data-contracts & retrieval-traceability"}

User:
Question: "\<user\_question>"
Top-k retrieved: <passed from retriever>
Acceptance: ΔS(question,retrieved) ≤ 0.45; coverage ≥ 0.70.

JSON response shape (auditable)

{
  "citations": [
    {
      "snippet_id": "S-28391",
      "section_id": "SEC-3.2",
      "source_url": "https://...",
      "offsets": [2312, 2450],
      "tokens": 172
    }
  ],
  "answer": "…",
  "λ_state": "→|←|<>|×",
  "ΔS": 0.37
}

Typical breakpoints → exact fix

Citations missing but answer present Reject and re-emit with the contract. Open: Data Contracts
Citation fields incomplete or wrong offsets Enforce the full schema, verify offsets/tokens against the corpus. Open: Retrieval Traceability
High similarity but wrong meaning Rerank or rebuild with correct metric/normalization. Open: Retrieval Playbook, Embedding ≠ Semantic
Header tweak breaks citations Freeze header order; clamp variance with BBAM. Open: Logic Collapse
Long runs lose citation discipline Split plan, bridge with BBCR; add mid-chain citation checks. Open: Context Drift, Entropy Collapse

Validator stub (copy into your pipeline)

Step 1  parse JSON strictly → if fail, stop.
Step 2  require citations[].length ≥ 1 before answer.
Step 3  verify fields & offsets; reject if any missing.
Step 4  compute ΔS and coverage; block if ΔS>0.45 or coverage<0.70.
Step 5  log λ across three paraphrases; alert if non-convergent.

Eval gates before ship

Cite-then-explain ≥ 0.98 on 50 queries.
Field completeness ≥ 0.99.
ΔS ≤ 0.45, coverage ≥ 0.70, λ convergent on two seeds.

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

9.9 KiB Raw Blame History Unescape Escape