Evidence as of October 11, 2026. Every page cited here was retained on October 11, 2026, and is quoted with attribution to the organisation that published it. The evidence record behind this piece states what it does and does not establish: source identity and the located excerpts are checked, while truth, rights clearance and editorial judgement are separate steps still pending. No publication times were reconstructed beyond the calendar day.

A simulation interface makes three promises, and only one of them is easy to keep. It promises you something to turn. It promises that turning it changes something for a reason. And it promises a result you can later hold up against the world. Most screens deliver the first, gesture at the second, and quietly skip the third.

Watch the video record for the full comparison. This article is the written companion, with the sourced evidence attached.

Part one: the dials you can turn

Start with what is on the screen. Climate Interactive's En-ROADS user guide, in its tutorial page retained October 11, walks a reader through 19 sliders and tells them to "share your unique scenario link" — so a setting is not just a setting, it is an address someone else can open.

Energy Innovation's Energy Policy Simulator goes a step further in its online model tutorial: alongside its sliders and implementation schedules, each policy carries a section headed "How the model handles this policy". That is a small design decision with a large consequence. The dial is annotated with its own mechanism.

Units matter as much as range. The GDP Dynamics Shock Simulator (version 1.12), published by Ty Keynes, states on its own page that one of its inputs is "not a percent of GDP" — printed next to the control, not buried in a manual. A reader who mistakes the unit does not get a slightly wrong answer; they get a different question answered.

And someone has to build these. isee systems lists Stella Architect, one of the commercial tools for making such interfaces, at 3,999 on its retained store page — a four-figure list price. Minsky, by contrast, is open source; the System Dynamics Society's "useful open-source tools" page describes it as the most "math-like" interface in system dynamics. (System dynamics is the practice of modelling a system as accumulating quantities — stocks — connected by rates of change — flows.) Steve Keen, writing on his Substack, describes that approach as "explicitly mathematical", and in the same post calls Ty Keynes the "premier Minsky modeller". Ravelation Pty Ltd's product page lists RavelSim, Ravel Data Analysis and Ravel Pro, and describes the "Minsky engine" underneath.

Part two: the loops behind them

A dial without a visible mechanism is a wish. What the better screens expose is the feedback loop — a chain where an output feeds back to change its own input, so effects compound or damp instead of adding up neatly.

The GDP Dynamics simulator names this directly: its page carries a "Dominant Feedback Loops" element, which puts the question which loop is currently driving the result? on the screen rather than in a footnote.

Bret Victor's 2011 essay "Up and Down the Ladder of Abstraction" is the design argument for why this works: his subject is "moving between levels of abstraction" — letting a reader step from one specific run to the whole space of runs and back. Climate Interactive's "simulators and science" page shows the production version of the same idea, offering an assumptions view and an equations viewer, with the model compiled from Vensim to WebAssembly so it runs inside a browser. (WebAssembly is a format that lets compiled code run in a web page at near-native speed.)

Exposure has costs, and the people who built these tools say so. In a 2014 Dice interview, Minsky co-creator Russell Standish discussed the project's "rough edges", and the same piece uses the word "spaghetti" — the familiar complaint about diagrams that show everything and therefore show nothing. A screen that reveals the machinery still has to stay readable.

Part three: the number you can afterwards check

This is where most interfaces stop, and where the sourced evidence is most consistent.

Climate Interactive states plainly in its En-ROADS baseline documentation that the baseline scenario is "not a forecast of what is most likely to happen". The Federal Reserve Bank of Atlanta's GDPNow page, as retained, says the estimate carries "no subjective adjustments" and is not an official forecast of the Bank or its president. Both tell you what kind of number you are looking at before you cite it.

Then there is the question of which run produced the number. NetLogo's BehaviorSpace documentation describes a tool that "runs a model many times" across varied settings — the standard answer to a single suggestive line on a chart. Minsky's repository README documents a headless batch mode (headless means running without the graphical interface) and a Python programming interface, which is what makes a run something a script can repeat rather than something a person remembers doing. En-ROADS' shareable scenario link is the consumer-grade version of the same property: a result with an address.

What the critics say a screen hides

The scepticism is older than the current generation of tools. Paul Starr's 1994 essay "Seductions of Sim: Policy as a Simulation Game", in The American Prospect, warns about the "black-box nature of the models" and about the risk that users "cede power to those who define the models". Three decades later, Climate Interactive's "simulators and science" page states that its "Model files are proprietary" — which is exactly the tension Starr identified, and which the assumptions view and equations viewer are an attempt to answer.

Two further limits are structural, not fixable by better design. The Lucas critique — summarised in the Wikipedia article retained here, which is an encyclopedia entry rather than the primary paper — holds that policy changes "systematically alter the structure of econometric models", so a relationship fitted under one policy regime may not survive the next. And NBER working paper 21097, "The Use and Misuse of Models for Climate Policy", uses the word "illusory" in its retained abstract page about what such models appear to deliver.

To Climate Interactive's credit, it publishes its own critiques: a support page titled "What are your biggest critiques of the model and possible future areas of improvement?", whose retained text names permafrost.

The disclosure rulebooks already exist

HM Treasury's Aqua Book (second edition, 30 July 2025) specifies an "Assumptions log" and addresses "assuring black box systems" — government guidance on how analysis should be produced and reviewed. In machine learning, Mitchell et al.'s 2018 paper "Model Cards for Model Reporting" (arXiv:1810.03993) built the companion idea for models: a short document stating, among other things, the model's "intended use".

Neither rulebook mentions sliders. Both describe what a simulation screen would need to print next to its answer.

Comuvia's judgement

This paragraph is Comuvia's judgement, not sourced evidence. Across the retained material, the three parts are not equally hard. Building a dial is a weekend. Making the loop behind it legible is a design problem some teams solve well. Printing a number that can be checked later — with its provenance, its width, and the identity of the run that produced it — is the part almost nobody ships, and it is the only one that creates accountability. Our judgement is that an honest simulation screen owes a reader four things beside every number: where each setting came from, how wide the answer is, which run produced it, and what the number is not.

Comuvia's own public simulation work, the ForeGlass Engine, is in the same comparison table as the tools above rather than outside it, and the ForeGlass forecast ledger is scored in the open on October 15, 2026. Where it falls short of these four requirements, the ledger is where that will show. Not investment advice.

What this establishes, and what it does not

Established: that each organisation named above published, as retained on October 11, 2026, the statement attributed to it. Not established: whether any of these models is accurate, whether the proprietary components match the published descriptions, or whether any tool's results have been independently verified. Those are separate exercises against separate evidence.

The accompanying video record uses AI-generated narration and a synthetic presenter.