> For the complete documentation index, see [llms.txt](https://docs.cournot.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.cournot.ai/core-mechanism/verifiable-reasoning.md).
# Verifiable Reasoning
Cournot’s Proof of Reasoning (PoR) separates oracle resolution into two distinct trust problems:
1. **Evidence Trust:** whether the protocol collected valid evidence from allowed sources;
2. **Reasoning Trust**: whether the final verdict is correctly derived from the agreed evidence.
PoR already addresses Evidence Trust through **PromptSpec**, **DataRequirements**, and **EvidenceReceipts / EvidenceBundles**. Once evidence has been collected and consensus has been reached on the evidence bundle, Cournot then treats the reasoning step as a separate, bounded computation problem.
#### **Key Design Principle**
Cournot does not initially attempt to verify the entire web retrieval process with opML. Instead, Cournot first freezes the inputs to the reasoning stage:
* semantic specification (PromptSpec)
* finalized evidence bundle
* verdict schema
* reasoning policy (model/runtime constraints)
By fixing these inputs, the protocol creates a **replayable reasoning target**. This makes reasoning verification significantly cheaper and more practical.
#### **Verification Tiers After Evidence Consensus**
After evidence has been collected and consensus has been reached on the evidence bundle, Cournot treats the remaining reasoning step as a bounded execution problem with frozen inputs (semantic specification + finalized evidence + verdict schema + reasoning policy). This is where Cournot can introduce a stronger verification lane ("opML" in our terminology) for deterministic replay and adjudication in disputed or high-risk cases.
Recent research supports this design direction from two complementary angles. First, [LLM-42](https://arxiv.org/abs/2601.17768) shows that inference nondeterminism is often a systems-level issue (e.g., dynamic batching, reduction order variability), and proposes a verify-rollback style approach to enforce deterministic behavior more efficiently than fully disabling batching. Second, work on numerical sources of nondeterminism in LLM inference shows that even greedy decoding can diverge due to floating-point precision effects and hardware/runtime configuration differences, especially for reasoning-heavy models. ([paper](https://arxiv.org/abs/2506.09501))
Together, these findings reinforce Cournot’s architecture choice: rather than assuming "temperature=0" is enough, Cournot defines a replayable reasoning target after evidence consensus and applies stronger deterministic verification only when needed.
---
# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.
## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.
Perform an HTTP GET request on the current page URL with the `ask` query parameter:
```
GET https://docs.cournot.ai/core-mechanism/verifiable-reasoning.md?ask=
```
The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.
Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.