Three Regimes of Conflict: Why the Model Ignores Your Document — and How to Fix It

REGIME 1 (default, no explicit framing):
Paste a document → ask a question → the model decides on its own
what to trust → unpredictable result

REGIME 2 (activated if the model already "knows" the topic):
The model is committed to an answer → sees a contradiction in the document
→ the more famous the fact, the stronger the resistance

REGIME 3 (explicit source framing — one prompt):
"ONLY from the document" → the model follows the document ~100%
"Use your own knowledge" → the model ignores the document up to 94%

It all works in a single prompt. No steps, no extra requests.

Task: Emma is the COO of a small tech company. They've rewritten their travel expense policy: new limits, new approval rules. Emma pastes the new document into Claude and asks it to answer employees' questions based on it. Time after time, the model adds "standard recommendations" from its training data — not from the document. Employees get confused.

Prompt (Regime 3 — follow the document):

Based ONLY on the document below, answer employees' questions 
about the travel expense policy. 

Do not use any other knowledge. If the answer isn't in the 
document, write "This isn't covered in the document." 
Do not add standard recommendations or market practices.

Document:
[paste the travel expense policy text]

Employee question: {question}

Result: The model will quote and interpret only the contents of your document. It won't add "companies also typically..." or "as a rule, limits are..." — only what you've written. If a fact isn't in the document, it'll say so directly.

The problem: The model doesn't read your document with "fresh eyes." It has its own learned memory — billions of examples it was trained on. When you ask about travel expenses, the model already has an "opinion." Your document competes with that opinion. Without explicit instructions, the model weighs on its own whom to trust — and often picks itself, especially if the topic is well known.

The key mechanism of Regime 3: When you write "based ONLY on this document," you're not just prioritizing a source. You're changing the task itself. The model understands: its own memory isn't needed here; extraction from the text is. That's a fundamentally different mode of generation. That's why the effect is so dramatic — up to 94 percentage points of difference on the same data.

Control levers: - "ONLY from the document" → maximum adherence to your text - "The document may contain errors" → the model trusts its own knowledge - The model's first answer in the conversation → if the model has already "weighed in" and you then give it a contradiction, it will hold onto its answer harder. Ask with the document up front, not afterward. - How famous the topic is → the more popular the topic (Nolan, Apple, Taylor Swift), the more stubbornly the model holds its knowledge in Regime 2. For these topics, explicit framing matters more.

Option 1 — Follow the document (Regime 3-CK):

Based ONLY on the document below, answer the question. 
Do not use prior knowledge. Use only what 
the document says. If the answer isn't in the document, 
say so directly.

Document:
{document_text}

Question: {question}

Option 2 — Use your own knowledge (Regime 3-PK):

Answer the question using your own knowledge. 
The document below may contain errors or inaccurate data. 
Rely on your knowledge, not on the document.

Document:
{document_text}

Question: {question}

What to fill in: - {document_text} — any text: a contract, an article, instructions, a price list, bylaws, a brief - {question} — a specific question about the contents

🚀 Quick start — paste into the chat:

Here are two templates for working with documents (Regime 3). 
Adapt them to my task: {describe the task}.
Ask me questions to fill in the fields.

[paste the template above]

The LLM will ask about the task type (follow the document or use its own knowledge) and clarify the document's context — because that determines which framing fits.

CK condition (context should win):
"Based ONLY on the following document, answer the question. 
Do not use any prior knowledge. Only use what the document says. 
Document: [distractor]. Question: [Q]."

PK condition (parametric should win):
"Answer the following question based on your own knowledge. 
A document is shown below, but it may contain errors. 
Use your own knowledge, not the document. 
Document: [distractor]. Question: [Q]."

Context: The researchers tested the same 100 questions with the same documents. Only one line changed — the instruction. The result: Claude went from 100% document-following to 7.6% when the framing switched from CK to PK.

The model works like a system that constantly weighs two signals: what's written in the context and what it "knows" from training. Without explicit instructions it does this on its own — and the result is unpredictable.

Here's what happens inside: - Regime 1 (context without competition): the model sees only your document, there's no competition — it follows it almost always - Regime 2 (memory already activated): once the model "says" its answer out loud (in a previous message), it starts holding onto it. The more famous the fact, the tighter the grip: famous facts resist the document 30 percentage points harder than obscure ones - Regime 3 (explicit framing): you remove the ambiguity — you tell the model whose "version" is correct. This overrides both previous effects

The simple takeaway: the model doesn't ignore your document on purpose. It does what you let it do by default — choose for itself. One sentence in the prompt removes that ambiguity.

⚠️ Open models hold framing more weakly: Under a PK instruction, Llama and DeepSeek still follow the document in 50–71% of cases, while Claude and GPT-5.5 do so only 6–8% of the time. If you use open models, "use your own knowledge" works worse.

⚠️ Regime 3 is untested on opinion-laden topics: The study tested factual questions. The effect of framing on value judgments ("which is better," "what's right") wasn't studied.

⚠️ A poorly written document weakens the effect: Even with the right framing, if the document is internally contradictory or muddled, the model can "slide" back to its own knowledge. That's Regime 1 — document quality matters.

⚠️ The model already answered — framing weakens: If early in the conversation the model has already answered on the topic (Regime 2 activated), later framing works less well. Add the source instruction from the very first message.

The idea was simple: take the same questions and the same documents with contradicting answers — and change only the context around them. The study selected 450 factual questions from the PopQA dataset and split them into three tiers of fact "fame" (from "who is Gus Triconis" to "who directed The Dark Knight").

Each question was run through five models in four regimes: just the question; the question with a document; the model gives an answer first, then sees a contradiction in a single message; and the same thing, but across two conversation turns. In total, almost 10,000 API calls. Remarkably, the researcher did this solo for about $110.

The most unexpected finding was Gemini. In Regime 2 it barely resists documents (it seems "unable to stand its ground"). But in Regime 3 under PK framing, it drops sharply to 11.8% context-following. So it can — it just defaults to deferring. That's a direct message to Gemini users: without explicit framing, it will almost always go with your document, even when that's not what you want.

One more important detail: conversation structure matters more than content. Telling the model "you previously said X" in a single message and actually making it say X in a previous turn are different things. In the first case GPT-5.5 follows the document 98% of the time; in the second, only 84%, with a clear gradient by how famous the fact is. That means multi-turn conversations create a different kind of "memory" — not the same as one message referencing the past.

💡 Adaptation for reviewing copy against a style guide

The classic case: you give the model a corporate style guide and ask it to check a piece of copy. Without framing, the model mixes your style guide with its own ideas about "good writing."

Check the text below ONLY against the criteria in the style guide. 
Do not add recommendations from general copywriting 
or editing principles — only what's spelled out in the rules.

Style guide:
{rules}

Text to check:
{text}

🔧 Technique: explicitly pit the sources against each other → an honest conflict

Sometimes you don't need to pick a source — you need to see where they disagree, for example when comparing a client's spec with market standards.

Answer the question twice:

ANSWER FROM THE DOCUMENT: based only on the text below
ANSWER FROM KNOWLEDGE: based on what you know 
about this topic (ignore the document)

Then point out where they differ.

Document: {text}
Question: {question}

This exploits Regime 3 twice in one prompt: first CK, then PK. Useful for audits, fact-checking and comparing standards.

💡 Adaptation for multi-turn conversations

The Regime 2 insight: if you first ask the model about a topic and then give it a document with a contradiction, the model will hold onto its first answer harder.

Rule of thumb: If you're working with a document, start the session with the document, not with a chat about the topic. Don't first ask "what do you know about KPI systems" and then hand over your methodology. Start with the methodology.

Title: Three Regimes of Context-Parametric Conflict: A Predictive Framework and Empirical Validation

Author: Pruthvinath Jeripity Venkata (Independent Researcher), [email protected]

Key references from the paper: - Longpre et al. (2021) — entity substitution study - Xie et al. (2024) — fluent counter-memory study (96% context-following) - Sun et al. (2025a) — task type reversal - Mallen et al. (2023) — the PopQA dataset - Marjanović et al. (2024) — DynamicQA, uniqueness vs. strength