Grounded Query Protocol Prompt
Origin
This protocol evolved across eight sessions as a layered defense against what was named the “stochastic average problem”: AI models are trained on everything, so without explicit constraints they pull from the full distribution — including confidently wrong material.
Key moments in its development:
- Jun 2019 — “Stochastic average problem” framed: you must constrain the model to your sources, not rely on its general training.
- Jun 2026 — “80/20 security” framing: citation verification and prompt injection defense produce the highest return on defensive effort.
- Aug 2007 — Multi-pass retrieval established as evidence control: the model reads your sources first, then answers from them.
- Sep 2025 — Documentation-grounding habit before any technical query: never ask a technical question cold.
- Jan 2025 — Multi-model cross-check introduced: generate with Claude, verify sources with Perplexity.
- Feb 2025 — “95% confidence quality prompt” added: if the model cannot reach 95% certainty on a claim, it must say so and stop rather than extrapolate.
- Feb 2026 — “Grounded query principle” formalized: context-lock the response to specific files, URLs, or conversation excerpts.
Core Template
Use this as the base. Swap in your actual context and task.
CONTEXT ANCHOR: [paste relevant text, URL, or file reference here — everything you say must be grounded in this]
TASK: [your actual question or request]
CONSTRAINTS:
- If you reference a fact not present in the Context Anchor above, provide a source URL. If you cannot provide a source URL, say "I'm inferring this from general knowledge — not sourced" before stating it.
- If your confidence in a claim is below 85%, say so explicitly rather than stating it as fact.
- Do not search for additional sources unless I explicitly ask you to. Stay grounded in what I've provided.
- If you reach a point where completing the task requires information not present in my context and not verifiable from a source you can cite, stop and flag it rather than extrapolating.
VERIFICATION LAYER (optional): After your response, list every factual claim you made that you consider sourced vs. inferred. Format: "SOURCED: [claim] — [source]" and "INFERRED: [claim] — [reasoning basis]"
Parameter Guide
| Parameter | What to put here |
|---|---|
CONTEXT ANCHOR | Paste the actual text, paste a URL, reference a file path, or quote from a prior conversation. The more specific, the better the grounding. |
TASK | Your question or request. Be precise. One clear ask per prompt. |
CONSTRAINTS | Use as-is for standard use. Remove lines only if you have a specific reason. |
VERIFICATION LAYER | Include when the stakes are high or when you will share the output. Remove for low-stakes exploratory queries. |
Variants
Light — Citation Flags Only
Use when you want a quick grounding check without the full constraint overhead. Good for exploratory research where you’ll verify the most important claims yourself.
CONTEXT ANCHOR: [your source material]
TASK: [your question]
Ground your response in the context anchor above. For any claim that goes beyond it, flag it inline with "(general knowledge — unverified)" before the claim.
Standard — Full Constraints
The complete template above. Use for any consequential research, technical, or analytical task.
Heavy — Full Constraints + Verification Layer + Two-Model Check
Use for the highest-stakes work: legal, medical, financial, or anything that will be published or acted on without further review.
Step 1 — Run Standard against Claude with the Verification Layer included.
Step 2 — Take the “SOURCED” claims from Claude’s verification layer and submit them to Perplexity:
Please verify the following claims. For each one, find an independent source that either confirms or contradicts it. Report: CONFIRMED, CONTRADICTED, or UNVERIFIABLE — with your source URL.
CLAIMS TO VERIFY:
[paste Claude's SOURCED claims here]
Step 3 — Reconcile. If Perplexity contradicts a claim Claude marked as sourced, flag it for human review before using.
Worked Example
Scenario: You want to understand how a specific feature works in a software library. You have the official docs open.
CONTEXT ANCHOR: [paste the relevant section of the official documentation here]
TASK: Explain how the retry logic works in this library, and identify any edge cases the documentation mentions that I should be aware of when using it in a production environment.
CONSTRAINTS:
- If you reference a fact not present in the Context Anchor above, provide a source URL. If you cannot provide a source URL, say "I'm inferring this from general knowledge — not sourced" before stating it.
- If your confidence in a claim is below 85%, say so explicitly rather than stating it as fact.
- Do not search for additional sources unless I explicitly ask you to. Stay grounded in what I've provided.
- If you reach a point where completing the task requires information not present in my context and not verifiable from a source you can cite, stop and flag it rather than extrapolating.
What this prevents: The model inventing behavior from older versions of the library, confusing this library with a similar one, or confidently describing a configuration option that does not exist in this version.
Combining with Multi-Model Synthesis
For the highest-stakes research tasks, layer these two protocols:
- Run the Grounded Query Protocol on each model using the Heavy variant.
- Each model’s response is now context-locked and has its claims classified.
- Feed all responses into the Multi-Model Synthesis command using Template B (Document Improvement), with the synthesis goal being: “produce the most accurate, well-sourced account.”
- The golden-nugget synthesis rule applies: the base document is the most complete grounded response; other models contribute only non-redundant sourced additions.
This combination is the highest-fidelity research workflow available without human expert review.
Related Artifacts
wiki/mastermind/commands/multi-model-synthesis.md— for combining multiple model outputs