How Westward monitors public money and public decisions across the South West — what we look for, how we analyse it, where our analysis can go wrong, and how to challenge us when it does.
Westward is an independent public resource that monitors spending, asset decisions, planning records, and democratic process across local and regional government in the South West of England. Our starting geography covers Bristol, WECA, Somerset, Gloucestershire, and Wiltshire, with scope to expand as the South West devolution settlement develops.
We exist because most civic data infrastructure is passive: it stores documents but does not interpret them. Westward adds an active analytical layer — a documented, repeatable framework applied consistently to public data — and publishes not just what we have found, but how and why we found it. We believe that transparency about methodology is inseparable from transparency about public institutions.
Westward flags patterns that merit scrutiny. We do not determine wrongdoing, and we do not investigate individuals. A Westward flag is the beginning of a question, not the end of one.
All Westward analysis is organised around five accountability patterns. These were developed through investigative reporting on Bristol City Council and WECA, and are grounded in a concept we call structural asymmetry: the observable tendency in public financial decision-making for risk and liability to move downward through the public sector hierarchy while resources and opportunity flow upward toward private and institutional actors.
Westward monitors a defined set of public data sources on a regular schedule. All sources are either statutory publications, Freedom of Information releases, or open data published voluntarily by public authorities. We do not use data obtained through hacking, leaks, or any unauthorised access.
Current monitored sources include: spending over £500 data published by Bristol City Council, WECA, Somerset Council, Gloucestershire County Council, and Wiltshire Council; committee agendas, minutes, and decision records from each authority; planning portal records; South West devolution deal documentation; Companies House records; and the register of members' interests for each authority's elected body.
Source documents are retained and linked from every flag we publish, so that any reader can verify the underlying record directly.
Westward uses a large language model — specifically, Anthropic's Claude API — to apply the five accountability patterns to ingested data at scale. This is not a black box: the model is given an explicit, documented prompt that encodes each pattern as a set of testable criteria, and its output is structured to explain which pattern a given item matches and why.
The AI layer performs pattern detection, not editorial judgement. It identifies candidate flags and generates a plain-English explanation of the match. It does not determine significance, draw conclusions about intent, or make final publication decisions. Every AI-generated flag is reviewed against the source document before publication.
Every item that reaches publication passes through three stages: automated monitoring identifies candidates; AI analysis assesses each against the five patterns; a human editor reviews the flag against the source document and decides whether to publish, hold, or dismiss.
Items dismissed at editorial review stage are not published. We maintain an internal log of dismissed flags and periodically review the dismissal rate as a measure of the AI layer's precision.
A Westward flag indicates that a pattern associated with accountability risk is present in the public record. It does not indicate that a law has been broken, that a public official has acted improperly, or that a decision was wrong. Many flagged items will have legitimate explanations not visible in the automated data. The purpose of a flag is to direct attention, not to assign blame.
Westward does not investigate individuals. Where named individuals appear in our flags, this is because they are connected to a public decision in their official capacity. We do not publish personal information beyond what is already in the public record.
We publish these limitations not as disclaimers but as a genuine account of where our methodology can fail.
Any public body, organisation, or individual named in a Westward flag may contact us at methodology@westward.org.uk with the flag reference number, the specific claim in dispute, and supporting evidence. We commit to acknowledging all requests within five working days and publishing corrections prominently alongside the original flag rather than silently amending it.
We will remove a flag if it is demonstrated to be factually incorrect. We will not remove a flag solely because the subject disagrees with the pattern designation or objects to the visibility of public information already in the public record. Where we decline a removal request, we will explain our reasoning in writing.
This document is versioned. When the framework changes we will publish a new version with a clear account of what has changed and why. Previous versions will remain publicly accessible. We intend to publish the full prompting methodology used in the AI layer as a separate technical document once the framework has stabilised through live operation.