When you generate a scenario, your selected drivers are passed to a large language model along with your chosen time horizon, tone and domain, and, if you set them, the active company profile and its research brief. The model returns five sequential stages, each with a description, a one-word theme, 3–4 investable category archetypes and the one to three selected drivers that shaped it most. It also scores where each stage sits on the Future Cone: its distance from the projected path (the business-as-usual continuation of current trends), which places it in the Probable, Plausible or Possible ring, and its desirability, which places it above the line (preferred) or below it (less desirable). Desirability is judged relative to the projected path on three tests: value for business and the economy, human welfare, and the planet. Each driver's push toward one side or the other is shown in the Driver Influence lanes under the cone, and a separate outlook rates how favourable the projected path is in its own right. Tone sets the conditions explored, not the outcome: optimistic surfaces emerging plays, baseline the consensus thesis, challenging the defensive moats and incumbents-who-survive, and any tone can end on either side of the line.

The 100 macro drivers in this workspace were synthesized from publicly available frameworks published by the World Economic Forum and McKinsey & Company, then de-duplicated and re-categorized across the STEEP framework:

Your team can add its own drivers with Create Driver; they join the library alongside the 100. Each driver carries a source attribution surfaced on its card:

Drivers that appear in both WEF and McKinsey carry both labels (de-duplicated overlap).

The Foresight OS driver library doesn't weight its drivers — each carries equal standing. Teams can rank drivers against their own priority lists (for example, investment theses); those rankings only change the order of the Driver Cards view. When a user selects active drivers and clicks Generate Scenario, those drivers are passed to Claude (Anthropic's LLM), which synthesizes the scenario narrative and its position on the Future Cone based on how it reasons the selected drivers would interact.

This is a more powerful approach than a numeric scoring scheme. A fixed formula can only capture interactions its designer anticipated in advance, and it treats every combination of drivers the same way regardless of context. An LLM instead reasons fresh about each combination — picking up second-order effects, historical analogues, and domain nuance a static weight could never encode. The same drivers can yield meaningfully different, well-reasoned scenarios depending on time horizon, tone, and lens, because the model is reasoning about interactions rather than running arithmetic.

The tradeoff is less determinism and auditability than a scoring model would offer — identical inputs can produce somewhat different framing across runs, and the reasoning lives in the model rather than an inspectable rule set. Scenarios should be treated as exploratory, AI-informed outputs meant to provoke strategic thinking, not validated quantitative forecasts.

Each of the 100 drivers carries a title, a one-sentence descriptor, a STEEP category, and a source attribution. The descriptor is a concise signal statement — not an extended analysis. When you generate a scenario, the platform passes the selected driver titles to the language model as equal-weight inputs. When you generate a Ripple Effects flow, it passes both titles and descriptors, giving the model slightly richer signal to reason about interactions.

There is no numeric weighting applied to drivers. All selected drivers are treated as equal inputs, and the model synthesizes their interactions, tensions, and compounding effects autonomously. Effective weighting is entirely user-driven:

• Selection — only the drivers you select are passed to the model. Unselected drivers are invisible to generation.
• STEEP balance — if your selection contains five Technology drivers and one Society driver, the model's outputs will naturally lean toward the Technology cluster. Varying the STEEP mix produces substantively different scenarios even from the same topic area.
• Exclusion toggle — the Active Drivers bar in the Scenario view lets you temporarily remove individual drivers from a generation without deselecting them, useful for testing how sensitive a scenario is to any single driver.
• Driver count — a small, focused selection (3–5 drivers) produces tight, coherent scenarios; a larger selection (10+) produces richer, more complex outputs with broader scope.

Future releases will introduce optional driver signal-strength ratings to allow explicit up- or down-weighting of individual inputs before generation.

• Global Risks Report (annual)
• Future of Jobs Report
• Top 10 Emerging Technologies
• Strategic Intelligence Platform — driver and transformation taxonomies

• Technology Trends Outlook (annual)
• McKinsey Global Institute reports — productivity, AI, energy transition, demographics
• McKinsey Quarterly — business model, talent, capital markets coverage

Of the 100 drivers, attribution breaks down as: 22 from both WEF and McKinsey (de-duplicated overlap), 29 from WEF only, 22 from McKinsey only, and 27 synthesized sub-drivers. This roadmap will broaden over time to incorporate additional public sources — likely Gartner, IEA, IMF, IPCC, and government foresight units — to reduce single-source dependency and increase methodological diversity.