Tribunal Evidence Collaboration - Retrospective Review v1.0

Purpose

This review considers the reconstructed Tribunal Evidence Collaboration Framework against the way the project actually developed. It is not an attempt to rewrite the original framework with hindsight. Its purpose is to identify what worked, what the work itself taught us, and why defining the working environment at the beginning of an AI-assisted project matters.

What we did well

The collaboration developed a strong evidence-first approach. Facts, allegations, interpretation and legal significance were kept distinguishable; chronology was used to test competing narratives; employer knowledge was not automatically equated with the date of formal diagnosis; and general medical or legal research was not treated as proof of the individual case. Evidence supplied by the Respondent was examined for material helping either party, not simply mined for favourable points. Missing evidence remained a gap rather than becoming an assumed fact, and conclusions were revised when later documents changed the picture. The claimant remained the source of personal evidence, while AI was used primarily to organise, compare, question and cross-reference a large body of material.

The project also developed useful controls organically. Large document sets were converted into indexes, chronology, evidence maps, contradiction analysis and structured rebuttal. This reduced reliance on memory and made relationships between documents easier to examine. Importantly, the process became increasingly willing to expose weaknesses, uncertainty and adverse material rather than attempting to construct an artificially perfect claimant narrative.

Improvements discovered through the work

The principal weakness was not the absence of useful rules, but that those rules were fragmented across conversations instead of being established as a project environment at the outset. Because the framework existed implicitly rather than explicitly, the same boundaries sometimes had to be rediscovered or restated. The AI could also expand a task beyond the immediate need, produce more analysis than was useful, or begin designing an improved methodology when the actual requirement was simply to document an existing one. These behaviours were normally corrected through conversation, but each correction consumed time and increased the amount of material that later had to be reviewed.

The later Chat Project made this particularly visible. Reconstructing the history showed that many principles eventually formalised as collaboration rules had already emerged during the tribunal work. Had they been written down when first agreed, they could have acted as persistent constraints throughout the project rather than relying upon conversational memory. The retrospective review therefore identifies project governance itself as one of the most important improvements.

Summary - Build the environment before building the output

AI-assisted work benefits from an explicit project environment: a small set of rules defining purpose, authority, evidence standards, boundaries, responsibilities and what the AI should do when uncertain. These rules should constrain the process without predetermining its conclusions.

Without that environment, an AI can remain helpful while gradually accumulating unnecessary analysis, repeating decisions already made, expanding the scope, or following an interesting line of reasoning away from the actual task. The result is bloat: more material to manage without a corresponding increase in value.

A project framework therefore does more than describe good practice. It gives the collaboration a boundary. It preserves the user's intent across long-running work, allows new findings to improve the process without silently rewriting its history, and gives both human and AI a stable reference point for deciding what belongs in the project and what does not.

The lesson from this project is simple: establish the environment rules early, then allow the evidence and the work to develop inside them. The framework should keep the project on course; it should never decide in advance where the evidence must lead.