The gap between draft and publish
Most content pipelines measure after the fact. You publish, then you watch the
analytics. By then the message has already done its work — or its damage. The flywheel I
run is gap → brief → draft → publish → measure → re-seed, and the honest weakness was
always that "measure" looked backward.
The fix is a stage that didn't used to exist: simulate. Between draft and publish, run the message through a synthetic audience — a set of AI personas generated from the same knowledge graph that powers the intelligence work — and read the predicted reception before the real world supplies it.
``` gap → brief → draft → [SIMULATE] → revise → publish → measure → re-seed ```
This is the second stop on the learning path — **Multi-Agent Systems → Simulation → Predictive Intelligence*. The first pathway mapped a multi-agent simulation engine* as a codebase. This one runs it.
What "simulate" actually means here
Take a strategy document — for this worked example, a real engagement with a century-old health authority weighing whether to break its institutional silence with a "confidence-graded second voice." (Client anonymized; the method is the point.)
1. Ontology from the document. The engine reads the brief and proposes the entity types that matter — not generic "people," but the actual stakeholder segments the strategy will touch: the institution itself, the expert bench, peer organizations, the everyday consumer, the health journalist, even the AI chat surfaces the strategy worries about. 2. A persona graph. Each segment becomes one or more agents with a distinct voice, stance, and activity pattern — generated from the document's own knowledge graph, so the audience is grounded in the material, not invented. 3. A simulation. The agents post, reply, reshare, and shift over several rounds on stand-in social platforms. You watch the narrative propagate. 4. The analysis. The reactions get analyzed back into a structured intelligence report — and, because the discourse is written back into the knowledge graph, the report is grounded in what the synthetic audience actually said, not just the source document.
The whole loop runs locally — local knowledge graph, local open-weights model, no data leaving the building. That "capability without custody" property is what makes it usable on confidential material.
What the synthetic audience caught
The instructive part is what the simulation surfaced that the brief didn't emphasize. Across every segment, the recurring objection wasn't "this is too fast." It was "accuracy and trust will need to be addressed." And the strategy's actual differentiator — grading how confident the institution is — barely showed up in how the agents re-voiced the message. The audience heard the riskier half of the idea.
That reframes the whole strategy: it succeeds or fails on whether the confidence grade itself is the visible hero of the message, not a footnote. You want to learn that before you publish, not after. A synthetic audience hands it to you in an afternoon.
The catch worth naming
A simulation is a modelled prediction of reception, not fielded research. Persona realism is bounded by the model you run; small runs echo themselves; and a synthetic audience is a rehearsal, not an oracle. Used honestly — to shape a message and find its weak seam before it ships — it earns its place in the pipeline. Sold as truth, it's theater.
Try it
The pattern generalizes past content: pressure-test a pitch deck against synthetic investors, or a report's thesis against synthetic skeptics. Anywhere you'd want a dress rehearsal with a stakeholder audience, a knowledge-graph-grounded simulation gives you one.
Part of the Developer Workbench series. Companion: the codebase map of the simulation engine itself (Pathway 1), and the integration plan that makes "simulate" a standing stage of the flywheel.