Skip to main content

Story Signals

Story signals are a set of deterministic, structured summaries extracted from a finished simulation. They turn the raw chat log and entity graph into a small, human-readable picture of what happened — who took which side, how the run unfolded over time, and which moments mattered. The same signals feed both the Story result view and the grounding context the report is written from, so the narrative and the report describe the same underlying run.

Story signals are computed by pure functions — no LLM calls and no external I/O — so they are reproducible from the simulation artifacts.

What the signals measure​

  • Stakeholder positions — agents are clustered by their dominant stance across all their posts (opposed, supportive, neutral, or split). Each cluster lists its members, its size, and the keywords that characterize its rationale.
  • Named coalitions — stakeholder clusters with two or more members are promoted to named groups, so blocs that actually formed get a label rather than just a stance.
  • Disagreement axis — the single line of contention the run was organized around, expressed as the top theme on the supportive side versus the top theme on the opposed side (stance words themselves are filtered out so the axis reflects substance, not just "support vs oppose").
  • Phase boundaries — the run is split into early / mid / late phases (or a single "full horizon" for very short runs), each tagged with its dominant topic and the calendar week range it maps to on your forecast horizon.
  • Quotable posts — the highest-engagement post per phase per stance, deduplicated so the same agent isn't quoted twice — the lines that carried the conversation.
  • Simulation scale — honest aggregates about the run: how many participants took part, the horizon in days, the number of blocs, and whether any market stress was actually observed (rather than implied).

How to read them​

Read the signals as the shape of the simulation. Start with the disagreement axis and named coalitions to see what the population split over and how it organized; walk the phase boundaries to see how the dominant topic moved early to late; and use the quotable posts as concrete evidence behind each phase. The simulation-scale figures tell you how much weight to give the result — for instance, whether market stress was genuinely present or simply not observed.

Because every signal is derived directly from what the agents did, nothing here is invented: a coalition, axis, or quote always traces back to real posts in the run.