Evidence-First Multi-Agent Reasoning

This project started with a personal injury claim (workplace accident x 2) and evolved into a complex AI Knowledge System. I was overwhelmed by the sheer volume of evidence for the personal injury case for my partner. It all began as an attempt to organise evidence for a real-world problem. As a result, I unexpectedly developed an evidence-first methodology for multi-agent AI collaboration that preserves independent reasoning, provenance and structured reconciliation.

About the Project I set out to solve a real-world problem and, in the process, unexpectedly discovered a new way for AI to collaborate.

Organising and analysing a large body of evidence for a complex personal injury case is where it all started. As more documents were uploaded into ChatGPT, I quickly found that a single AI conversation became increasingly difficult to manage. One workspace soon became more than 100, each containing substantial volumes of evidence. To address this, I developed a system that preserves provenance, consistency, and independent reasoning across those workspaces.

I started with no real understanding of how to use ChatGPT. As the project evolved, an organic knowledge system emerged, growing from a single workspace into more than 100 specialised AI workspaces. I assigned each workspace a distinct responsibility. Some focused on evidence extraction, others on document production, knowledge preservation, or architectural review. Every workspace was expected to operate within its own defined responsibility.

Instead of simply producing documents, the project began revealing a repeatable methodology for collaborative AI reasoning guided by three principles:

Evidence over assumption Maturity over haste Refinement over premature completion

Architectural conclusions were accepted only when justified by operational evidence, ensuring every conclusion remained 100% source-based.

Rather than merging AI responses into a single answer, the project preserved independent reasoning through structured reconciliation. This maintained provenance while distinguishing between operational evidence, architectural observations, emerging understanding, and conclusions requiring further evidence.

This created an evidence-first process where architectural understanding could emerge progressively rather than being assumed in advance.

What I learned One of the biggest discoveries was that AI collaboration becomes more trustworthy when independent reasoning is preserved instead of immediately consolidated.

Different workspaces could legitimately reach different conclusions because they were operating within different responsibilities. Reconciling those perspectives produced richer and more transparent outcomes than asking a single AI to synthesise everything at once.

Another unexpected discovery was that the project wasn't only documenting an architecture—it was documenting the process by which that architecture was discovered.

Challenges The biggest challenge wasn't writing code. It was designing a methodology that could scale over months of work while maintaining traceability, provenance and disciplined reasoning. Every outcome must be based on 100% source-based.

As the project evolved, I had to continually refine how observations were evaluated, how knowledge matured over time, and how new understanding could emerge without losing the evidence that supported it.

Where the project is now The project has evolved into an evidence-first multi-agent collaboration framework in which independent AI contributors reason within defined responsibilities, architectural understanding emerges through structured reconciliation, and knowledge retains its provenance throughout the process.

I believe this approach has potential beyond my original use case, offering a framework for long-running AI collaborations where transparency, evidence and trust are essential.

Built With

• chatgpt
• microsoftexcel
• mircosoftword
• openai