An open-source workspace for investigating recurring coordination in public social-media activity.
Early research release · Version 0.3.3
I'm releasing TopicStream, an open-source investigation tool from Argelius Labs. The idea is straightforward: start with a topic, collect or import relevant posts, and look for accounts that repeatedly act together. Then inspect the evidence behind those patterns.
A burst of similar posts can be worth examining. It can also be people reacting to the same news, sharing a press release, or participating in an openly organized campaign. A useful investigation needs enough context to tell those explanations apart. TopicStream gives you a place to start that work.
This is an early research release. It runs locally, keeps cases on your computer, and makes the detection rules available for inspection. The goal is to help investigators develop and examine coordination leads while keeping the underlying evidence accessible.
You begin by entering a topic or situation and reviewing a proposed search plan. You can adjust the phrases, hashtags, dates, sources, and collection budget before making requests. That review matters: the current planner creates an editable draft. It doesn't automatically understand every meaning of a topic or discover all of its aliases and translations.
Once data is collected or imported, TopicStream looks for repeated text, shared links, and amplification of the same posts within short time windows. The default recurring-group rules require at least three accounts, with qualifying account pairs appearing in at least three separate episodes across two UTC days. Findings retain the posts and events that support them, so you can examine why accounts were connected.
The workspace also includes coverage information, graphs, timelines, analyst reviews, and exportable reports. Reports show the scope and activity triggers up front, with collapsible evidence, a light theme, and print support. You can save an investigation and return to it later. If you already have a suitable dataset, you can import a CSV, including account profiles, instead of making new collection requests. The project includes adapters for X, Bluesky, and Mastodon, with different access requirements and visibility limits. The first live acceptance test described here used X.
Coordinated inauthentic behavior involves more than accounts posting together. Establishing deception, common control, or sponsorship requires additional evidence. TopicStream keeps the coordination finding separate from an analyst's assessment of inauthenticity. Its ranking scores are not probabilities that an account is fake, and legitimate organizations can produce recurring coordination patterns.
For the first live pilot, I chose the current discussion around Singapore's haze situation. The National Environment Agency was publishing updates, giving us a timely topic with a concrete public reference. We searched X over a six-day interval from October 30 through November 5, 2026, using the word haze alongside Singapore, Singapura, or the Chinese spelling of Singapore.
The search retrieved 285 distinct posts. After requesting full-length text, 284 posts from 238 accounts were included in the analysis. The detector found four short-window repost clusters, but no recurring account group met the default thresholds. The resulting assessment was: no strong recurring coordination detected in the observed data.
That result applies to this query, interval, dataset, and configuration. It doesn't establish that the wider topic is free of coordinated inauthentic behavior. The query still required the English word haze, and it covered one platform. X's counts endpoint also reported 320 matches while search pagination ended at 285 retrieved posts. That difference remains unresolved; the report doesn't assign it a cause.
The pilot was useful because it exercised the whole collection and analysis workflow and exposed things that needed fixing. Some long posts arrived as shortened previews, hiding the search term from our local filter. Requesting the full text recovered 17 relevant posts without adding new post IDs. We also fixed a chart that omitted the partial first day. Follow-up feedback helped improve the billing model and make the reports easier to inspect.
Costs deserve the same care as the findings. TopicStream's local estimate originally put this pilot at US$4.02 initially and US$7.98 with verification. That model overcounted author expansions and repeat reads. The operator-reported account activity instead supports an inferred pilot allocation of about US$1.44, not a separately verified debit; a half-cent discrepancy remains in the supplied whole-account arithmetic. X has since notified us that expanded objects become billable at their standard resource rates from November 22, 2026. TopicStream currently requests only author-profile expansions, priced at US$0.010 per user resource under the new policy. Version 0.3.3 automatically includes that charge from the announced date and reserves for it in advance. The November 5 pilot remains a pre-change example. Daily resource deduplication still applies, and balance observations remain separate from local estimates because delayed charges and other account activity can affect them. A case budget is local to its database.
Version 0.3.3 is now on the repository's main branch. The current suite has 98 passing tests and one optional test skipped, with checks passing on Windows and Linux. That gives us evidence about software behavior. Measuring detection accuracy requires a separate evaluation using independently labeled cases, legitimate campaigns, other topics and languages, and examples with incomplete data.
If you want to try it, clone the repository and run python -m topicstream serve from the project folder with Python 3.9 or later. Open the local address printed in the terminal. Start with the synthetic example to explore the workflow without API credentials or collection charges, then read the source and budget guidance before a live run.
I'm particularly interested in feedback from investigators who can test it against cases they understand. Reports of missed patterns, ordinary activity that gets flagged, confusing coverage information, or discrepancies in cost accounting will help improve the next release. Please use the GitHub issue tracker for reproducible feedback and keep credentials and sensitive case material out of public reports.
TopicStream is available now through Argelius Labs. Try the example, inspect the method, and follow a finding back to its evidence.
Explore TopicStream on GitHub · Read the detection methodology