Provenance

Inspiration

Important business decisions are often based on information that has passed through several people, documents, and teams. A customer's original complaint may be shortened by a support agent, summarised by a manager, and converted into a product requirement. These sumaries are needed to reduce the information overload (millions of customer calls/feedbacks). Although nobody intentionally changes the meaning, important context can gradually disappear.

We were inspired by the idea that organisations should be able to verify whether a final decision still reflects the original source. Provenance was created to make this loss of meaning visible before it leads to the wrong action.

What It Does

Provenance tracks how information changes as it moves from an original source to a final business decision.

Users can provide the original customer message, the intermediate summaries or transformations, and the final decision taken by product owner/manager. Provenance then compares each stage and identifies:

Important facts or concerns that were lost
Meaning that was added without evidence
Changes in tone, urgency, or intent
The point in the communication chain where the largest distortion occurred

The system presents a clear provenance trail and an overall alignment assessment. It also links identified issues back to the original source, allowing decision-makers to understand what changed and correct the final decision.

How We Built It

We built Provenance as a lightweight web application with a React frontend connected to backend API endpoints.

The frontend allows users to enter or select the original message, intermediate versions, and final decision. The backend processes each stage using Mistral's language models to extract important claims, concerns, requested outcomes, tone, and contextual details.

The system then compares consecutive stages in the communication chain. It evaluates whether important information was preserved, removed, weakened, exaggerated, or introduced without support.

The backend returns structured results containing alignment scores, detected changes, explanations, and references to the relevant source text. These results are displayed in a simple interface that helps users inspect the full decision trail without reading every document manually.

Challenges We Ran Into

One of our biggest challenges was defining the difference between acceptable summarisation and harmful information loss. A summary does not need to preserve every word, but it must preserve the details that materially affect the final decision.

We also had to present complex language-model analysis in a way that was understandable and actionable. Instead of only displaying a generic similarity score, we focused on explaining what changed, where it changed, and why the change mattered.

Another challenge was integrating the frontend and backend within a limited hackathon timeframe. We reduced the product scope to the most important workflow: tracing one communication chain from the original message to the final decision.

Accomplishments That We're Proud Of

We are proud that Provenance turns an abstract problem, context loss across teams, into something users can directly inspect.

Rather than treating each summary as an isolated piece of text, Provenance evaluates the entire chain of communication. This allows users to identify the exact stage where the original meaning began to drift.

We are also proud that the system does more than detect differences. It explains the business significance of those differences and directs users back to the original evidence. This makes the output useful for real decision-making rather than simply producing another AI-generated score.

What We Learned

We learned that information does not need to be factually incorrect to become misleading. Meaning can change through omission, simplification, altered emphasis, or the removal of uncertainty.

We also learned that language models are most valuable in this problem when they provide structured, evidence-based explanations rather than replacing human judgement. Provenance is designed to help people review decisions more effectively, not automatically decide whether a business action is right or wrong.

Finally, we learned the importance of reducing scope during rapid development. By focusing on one clear end-to-end workflow, we were able to demonstrate the core value of the product.

What's Next for Provenance

The next step is to integrate Provenance directly with the tools organisations already use, such as customer-support platforms, email, Slack, Jira, Confluence, and product-management systems.

Future versions could automatically construct provenance chains across these platforms and alert teams when an important requirement, customer concern, or risk is lost.

We would also like to add organisation-specific evaluation rules, collaborative review workflows, and analytics showing recurring sources of communication drift.

Ultimately, Provenance could become a verification layer for organisational decision-making, ensuring that important decisions remain connected to the people, evidence, and original context that produced them.