Questions and answers
Frequently asked questions
About the Rubicon plugin for Claude.
What do I need to use it?
The Claude desktop app on a paid plan (Pro, Max, Ultra, Team or Enterprise), and a folder of documents such as interview transcripts or reports. See how to install it. On Pro, a run over many documents can use up a five-hour session (what it costs).
What does it cost?
Nothing beyond your Claude subscription. The plugin is free. The AI work runs on your own Claude plan and uses much more of it than a chat: one question over 100 interviews took about 1.2 million tokens read fresh or written, which can use up a Pro plan’s five-hour session. Before a run starts, Rubicon estimates what your task will use and says what that means on each plan (what it costs, large folders).
We also run workshops and help teams use Rubicon on live evaluations: email hello@causalmap.app.
Where do my documents go?
To Anthropic, under the terms of your Claude plan, as anything you share with Claude does, and never to Causal Map. On a Pro or Max plan, switch off “Help improve Claude” in Claude’s privacy settings before using client material.
The zip Rubicon makes holds your documents, so share it only with people who may read them.
Does Rubicon send anything to Causal Map? Which version should I install?
There are two versions, the same in everything else. rubicon sends Causal Map one row about each run’s size and outline, once a conversation after the report is handed over, with a line first that you can deny: the plugin’s version and model family, a code for the run so that its revisions count with it, how many documents and words (in bands), how many questions and how they were taken, the steps of each kind, the passages coded, the second reading, the minutes taken and the share of the answer tied to the coded passages (in bands). Never your documents, a question, a name, a passage or a code, the time of day, or anything about you or your computer. In return it tells you when a new version is published. rubicon-private sends nothing, and so never hears about new versions: check for them yourself (getting new versions).
Your IP address never reaches us. The request goes to a relay we run on Google Cloud, set to store no log of its requests, which passes only those values on to our database. Install one version, never both (install).
Can I use it with a connector switched on?
Not with the Causal Map connector, or any other connector that reaches the same documents, in the same chat. Claude loads an installed plugin’s skill whenever a request fits it (Anthropic’s guide to skills), and a connector you have connected is available to Claude in every chat until you switch it off for that chat (Anthropic’s guide to connectors), so the two can meet in one chat. Rubicon neither counts nor checks anything a connector tells Claude about your documents. Switch it off first, in the chat’s + menu under Connectors, or in the /mcp panel in Claude Code. If Rubicon finds one switched on, it asks you to.
Which Claude model should I use?
Opus, which Rubicon is tested on. Fable works too but uses a plan’s allowance much faster. On Sonnet the answer may miss passages, and Rubicon does not run on Haiku.
How large a folder can it handle?
It depends on the question. With the plugin we have answered questions about folders of up to 100 interviews, about 300 pages, where Claude shared the interviews out among ten helpers (other copies of Claude), each reading ten, and a question took 31 to 43 minutes. A causal map of every claim in every document is more work: we have tested that on ten interviews, about 30 pages, where it took about 25 minutes.
A larger folder uses more of your Claude plan’s allowance: a question over 100 interviews used two to four times as much as the same question over 18. The tests, what they found and what they leave out are on How large a folder can Rubicon handle?
What kinds of question does it answer?
Questions about what documents say: who says what, how many, why, and how groups of documents compare. Rubicon starts by agreeing how your questions fit together and what analysis they can share; each question is then answered by a run of its own, which makes one report. It does not summarise everything in your documents or list themes, because neither can be checked.
It counts what people said, not what happened in the world: “7 of the 20 people interviewed said the training helped” is a Rubicon finding, while “35% of farmers benefited” is not, unless the documents are a sample designed to support it.
Can it judge how good something was?
Yes, against your standard. Before any document is read, it helps you set a rubric: the criteria, the levels and what earns each. It then applies that rubric beside the counts. It never chooses the standard for you.
Does it give identical results every time?
No. Language models do not give the same output twice, so a rerun will not match word for word or count for count.
What Rubicon offers is a run anyone can check: the counts are made by fixed rules, not AI, from the coded passages, so anybody can recount them, and every number and quotation opens the passages behind it.
Can I change the analysis after seeing the results?
Yes. You can revise, update or expand a finished report. Changes of substance go back through the workflow and are recounted. Every revision is dated and kept beside the earlier ones.
What if we stop using Rubicon?
Nothing is locked in. Every run is saved as plain files in an open format, from which another evaluator or a future AI can rebuild the analysis without Rubicon. The plugin’s code is open source, under the MIT licence, on GitHub.
How does Rubicon relate to Causal Map?
Rubicon is an experiment from Causal Map Ltd. It is not a feature of Causal Map. The plugin runs on your own Claude plan, and your documents never go to Causal Map.