Canada and France signed four years of AI cooperation. The project ledger starts empty

THE SIGNAL IN ONE SENTENCE

International cooperation is easy to photograph. Two research leaders stand together. Flags behave. The announcement contains the right nouns: safety, science, health, climate, open source, multilingual AI, students, startups and the common good. Then everybody goes home and the verbs have to begin. On October 8, Montreal's Mila and France's Inria announced a four-year agreement to deepen artificial-intelligence and digital-science cooperation between Canada and France. Mila is a nonprofit AI research institute rooted in Quebec's university ecosystem. Inria is France's national institute for digital science and technology. These are not ceremonial organizations borrowing scientific language for an afternoon. They have researchers, graduate students, software projects, public mandates and established links to industry. A durable bridge between them could be useful. The agreement covers foundational machine learning, agentic AI, evaluation and safety, AI for science and health, energy and environmental sustainability, language, computer vision, robotics, advanced computing infrastructure, and the social and economic effects of AI. The institutions say they plan to organize joint projects, co-supervise master's and doctoral students, support researcher and student mobility, pursue publicly funded initiatives and launch innovation projects. Applied work in quantum and AI, open source and multilingual AI receives particular attention. Companies and startups may join. Other Canadian and French organizations may be added. A joint steering committee will guide the partnership and track progress. That is a sturdy frame. It is not yet a project ledger. The public announcement does not identify an initial research project, principal investigator, participating laboratory, committed budget, named funding award, delivery milestone, repository, dataset, license, mobility place or outcome measure. It does not publish the steering committee's membership, meeting schedule, decision rules or reporting cadence. The absence of those details in one announcement does not mean the institutions have no internal plan. It means the public cannot yet inspect one. The plain signal is this: a cooperation agreement becomes public value when readers can follow resources into work and work into evidence. Four years is long enough to build something consequential and short enough to disappear into a sequence of conferences. The distinction matters because the announcement mixes several kinds of cooperation. Scientific research is not the same as student mobility. Open-source software is not the same as startup creation. A publicly funded initiative is not the same as an industry partnership. Multilingual AI may require datasets, community governance and evaluation practices that look nothing like a robotics collaboration. Put every promise into one paragraph and the partnership sounds enormous. Separate the promises into accountable workstreams and the practical questions arrive. For joint research: What problem is each project trying to solve? Who leads it? Which laboratories participate? What result would count as progress? When should another scientist expect a paper, dataset, model, benchmark or negative result? For student supervision: How many students receive support? Who pays tuition, salary, travel, housing and insurance? How are candidates selected? Can a student spend meaningful time at the other institution, or does bilateral supervision mean two names on a document and twelve more video calls? For mobility: Are opportunities open through a published application? Do they include technical staff and research engineers, or only senior scientists with existing networks? What assistance exists for visas, disability access, caregiving and relocation costs? An exchange route is not genuinely open if only people with spare money can use it. For open source: Which repository? Which license? Which artifacts? Who maintains the project after the paper? How are security reports handled? Can researchers outside the two institutions reproduce the work and propose changes? For multilingual AI: Which languages and communities? Who decides what good performance means? Are evaluations limited to translated English tasks, or do they test local law, culture, professional language and ordinary speech? Who can inspect or contest the dataset? For startups: Which intellectual-property rules apply when work was publicly funded, jointly supervised and built across borders? What remains open? Who receives equity? Can a company obtain an exclusive right to infrastructure or research that the public already helped finance? None of these questions argues against the agreement. They are the maintenance manual for taking it seriously. Mila and Inria already point toward a governance mechanism. Their joint steering committee is supposed to guide development, track progress and identify new areas of interest. That committee can turn the announcement into a living public record. Start with a simple page. Give every project a stable name and identifier. List the responsible people and institutions. Record the start date, target end date, funding source, committed resources and compute access. State whether the expected output is a paper, model, dataset, software package, clinical study, policy method, startup or something else. Then add the fields that partnership announcements usually avoid. What slipped? What failed? Which result could not be reproduced? Which dataset cannot be released, and why? Which license changed? Which student position went unfilled? Which safety concern stopped a deployment? Which community objected to a multilingual dataset or evaluation? A useful ledger is not a trophy shelf. It is a memory system. That memory has special value in a four-year collaboration because personnel will change. Students graduate. Postdoctoral researchers move. Funding calls open and close. A model that looks important in year one becomes ordinary in year three. Without a public project record, the partnership can keep announcing motion while losing the chain of decisions that explains where it went. The research agenda itself is broad enough to create that risk. Foundational machine learning, agents, health, energy, robotics, language, computing infrastructure and economic impacts could each support a separate institutional partnership. The range may be a strength if it allows teams to form around real opportunities. It may become fog if every future project can be described as part of the agreement after the fact. The steering committee should therefore publish selection criteria before selecting the portfolio. One criterion can be scientific complementarity. What can a Mila-Inria team do together that either organization would struggle to do alone? Another can be public value. Is the work likely to produce reusable knowledge, open infrastructure, safer practice or a measurable benefit beyond the participating institutions? A third can be additionality. Did the agreement create a new project, fund a new researcher, expand access to compute or remove a real barrier? Counting an existing collaboration under a new banner may be administratively convenient. It does not show what the agreement added. A fourth can be access. Can students, smaller laboratories, outside evaluators and affected communities participate, or does the partnership mainly connect people who were already connected? A fifth can be evidence. Is there a baseline, a target, a method and a date for checking the result? That last one is where international research announcements often get soft. Publication counts are easy. They are also incomplete. Ten papers do not tell readers whether software is maintained, students gained durable opportunities, multilingual systems work for real communities, safety findings changed practice or a climate project reduced any measurable burden. The scorecard should match the work. For research, report papers, artifacts, replications, negative results and outside use. For students, report positions, completion, exchange duration, support, access and career outcomes. For open source, report repositories, licenses, contributors, releases, unresolved issues, security response and maintenance funding. For multilingual AI, report languages, community partners, dataset governance, evaluation design and error patterns. For innovation, report companies involved, public and private funding, intellectual-property terms, pilots, survival and what stayed publicly accessible. Do not mash those into one partnership score. Different work deserves different evidence. There is also an opportunity here that deserves more than skepticism. Canada and France sit inside overlapping but distinct research, language and policy environments. Quebec brings its own institutional and linguistic context. France brings a national research system and European regulatory setting. Mila and Inria can test whether responsible AI principles survive when systems cross borders, funding structures and legal regimes. That could make multilingual evaluation more honest. It could make open-source governance more durable. It could expose where a safety method depends on one language, one institution or one definition of acceptable risk. It could give students access to mentors and infrastructure on both sides of the Atlantic. It could also make research harder in productive ways. Data cannot always move. Licenses conflict. Ethical review processes differ. Export controls, privacy rules, procurement requirements and intellectual-property policies can slow a collaboration that looked frictionless in a press release. Those obstacles should not be edited out of the record. They are part of the result. If the partnership develops a reliable method for cross-border data governance, publishes a reusable agreement for joint open-source maintenance or shows how a multilingual evaluation failed before it improved, that institutional knowledge may travel farther than one model. The student layer deserves particular care. Co-supervision and mobility can turn an agreement from institutional branding into a human network. They can also create invisible administrative labor and unequal access. Publish the routes. A student should be able to see available projects, supervisors, eligibility, funding, location, expected travel, language requirements, intellectual-property terms and support. Count applicants and participants in ways that reveal whether access reaches beyond established laboratories. Ask participants what the exchange made possible and what paperwork nearly killed it. The dry administrative details are where opportunity becomes real. The same principle applies to public funding. The announcement says the institutions will pursue publicly funded initiatives. It does not name an award or commit a sum. Until a funding body approves one, "pursue" means intent. Future coverage should preserve that verb. When funding arrives, the ledger should identify the program, amount, dates, cost-sharing, recipient, public deliverables and reporting obligations. If private partners contribute compute, data or cash, record that too, along with the access or rights they receive. Transparency does not require publishing every contract or sensitive research detail. It requires enough structure to distinguish a live project from an attractive category. The sequence can be practical: First, publish the agreement's governance and selection criteria. Second, register each project with its owner, resources, dates and expected outputs. Third, link the papers, repositories, datasets, student opportunities and funding awards as they appear. Fourth, publish a short progress note on a fixed schedule. Fifth, record failures and corrections beside successes. Sixth, invite outside use and evaluation. Seventh, close each project with a result, not a vanishing web page. That sequence creates accountability without demanding certainty before research begins. Research is allowed to fail. In fact, failure can be evidence that the question was real. What institutions should not do is let uncertain research become untraceable administration. Mila and Inria have made a promising commitment: four years, two serious research communities and a deliberate connection between science, students, open infrastructure and innovation. The announcement gives them room to find projects rather than pretending every answer was known at signing. Now the empty space needs names, dates and links. The photograph can stay. So can the broad ambition. Beside it, put the ledger.