Paper Trail Methodology

Paper Trail is an unprecedented collection of data about Canadian companies and organizations involved in lobbying and government contracting. It allows any interested Canadian to look up an organization and see its activity, under many different names, in one place.

When these organizations make their mandatory public disclosures, they record their names however they like. This means Rogers could go by “Rogers,” “Rogers Communications”, or “Rogers Telecom” – or even a typo, like “Rogers Communicaiton,” of which there are many. The IJF has identified 40 such spellings of Rogers in their disclosures about lobbying and government contracting; in government contracting, for the top 10 companies we found 692 different names.

For Paper Trail, the IJF built a new AI pipeline to help surface possible names for each organization. Then, we carried out a large-scale human review of these AI suggestions, following a rigorous and conservative methodology. As of September 2026, more than two dozen reviewers have reviewed a total of 176,000 different names.

The original data

Paper Trail is built on top of existing IJF data, from our lobbying data — Lobbying Registrations, Lobbying Communications, Revolving Door, Government Funding — and notices of awards in our Procurement data. These data sources describe activities by organizations. These sources comprise records by which a particular organization is named, either as one lobbying the government or receiving a government contract. It is in these records that the variations in names appear: when submitting a lobbyist registration, for example, the lobbyist may record the name of their client in many ways. The IJF already applies some data-cleaning to these records, but not enough to handle the variation in names.

What is covered

Because we use manual, human review, our pace is limited and not all records in the IJF’s data will be covered. We set minimum thresholds to ensure the records that matter most always appear. Any record associated with an organization that meets one of these criteria will be included in Paper Trail:

1. They have registered to lobby since September 1, 2025
2. They have won a government contract (federally or in British Columbia) worth at least $500,000 since September 1, 2024

These are guaranteed minimums. As the IJF continues to collect new lobbying and procurement data every day, we will maintain manual review so that these two thresholds are satisfied at a delay of at most one week. 

We cover many contracts and registrations beyond these minimum thresholds. Here are three other qualities of the reviewed dataset:

1. Of all dollars awarded by the government to organizations in our data, more than 80% of them (approximately $750 billion) are covered
2. Every company listed in the TSX 100 index is covered
3. Every company we have reported on since 2025 is covered

Finding names with AI

The IJF built an Entity Resolution (ER) system to search more than one million records we have across our lobbying and procurement data. Its job is to suggest groupings of similar names, all of which our staff later approve or reject.

A system like this must necessarily compare every name against every other, so the problem consists of evaluating these comparisons as cost-effectively as possible while remaining accurate, since the number of all combinations is in the many millions. 

The first step is typically called “blocking” in ER literature. It is a simple, cheap pass to remove obvious negatives. We use a k-Nearest Neighbours approach, the distance calculation being the cosine similarity between the SentenceBERT embedding of two names. We found a k of 100 to work best for our data. 

The second step is to filter out more pairs with an LLM as a classifier. Though we tried to avoid it, we found an LLM was necessary to make the many semantically complex associations we required. For example, because an LLM has general world-knowledge, it would link “Rogers Comms” with “Rogers Wireless”, but not “Pattison Group” with a man named “Harold Pattison”. 

We tried using DITTO, a popular entity matching system published in 2020, but found its customary use, with small language models like ROBERTA, was not sophisticated enough. In addition, the data that organizations submit to the government is regularly incomplete or even incorrect. This deprived us of other context clues that would help match up organizations. With such incomplete information, we found LLM suggestion with human review to be the viable path.

Manual review

Every name appearing on the Paper Trail platform has been reviewed by IJF staff. The names they do review are produced as suggestions by the ER system described above. However, the final say, with total latitude, falls to our reviewers. 

Our reviewers are a mixture of our reporters, software developers and research and development associates, totalling more than two dozen people. They followed a methodology designed to be cautious, so as to prevent false positives – putting a name to an organization that shouldn’t be there.

The core principle of this methodology is as follows: to link together any two names, their spellings should be equivalent, and additional data about the records to which they correspond should also match. A summary of the full methodology is below.

Matching names

Most name differences can be accounted for by the following rules:

1. Missing spaces: SYSCO CANADA vs SYSCOCANADA
2. Repetitions: ONTARIO NATURE vs ONTARIO NATURE ONTARIO
3. Typos: AMERICAN SCIENCE ENGINEERING vs AMERICAN SCIENCE ENGINEEERING
4. Affiliations: ESSAR STEEL ALGOMA FORMERLY ALGOMA STEEL vs ALGOMA STEEL
5. English + French: THE CANADIAN RED CROSS SOCIETY vs THE CANADIAN RED CROSS SOCIETY / LA SOCIÉTÉ CANADIENNE DE LA CROIX-ROUGE

Many other idiosyncratic differences which do not change the meaning of the name can exist.

Additional evidence

Not only the name must match, but also some other evidence about the records corresponding to the names. We often find matches by address, for instance by an identical postal code, forward sortation area, or same civic number on the same street. We also consider the content of the record. For example, an awarded contract should be awarded by a buyer that is in the sector in which the company or organization operates: Lockheed Martin is less likely to be procuring for Fisheries and Oceans than for the Department of Defence.