How the run runs

Foresight runs a computed map forward: many worlds, not one guess, and the spread across them is the output. This is the engine as specified, the four reads it would emit, the rules printed on its face, and where the specification stands against the vendors who sell simulated audiences today.

Spec'd, not built

01 . What it takes in

The map, and a move. Nothing else.

The map is the one the studio computes: personas carrying a quant vector, and the signed edges between them. The run reads the vector as the agent and the edges as the topology, so the map is the model and there is no second one. The move is what the tenant wants to test: a frame, a plan with a budget and a burn curve, a lever set. The run returns the spread of outcomes across an ensemble of worlds and the levers that change the spread.

02 . The four layers

A numeric complex-contagion core carries every claim; nothing is role-played. That is why the whole thing is specified to run in a browser rather than on a cluster.

fig 1 . the engine, layer by layer

microArchetype agents on the quant vector. Thresholds are sampled per world from a spread, not fixed at a point, because outcomes live in the tails. Adoption needs reinforcement from more than one source type before the gate opens; a single weak tie carries information, not behaviour. State carries activation, hostility and fatigue, and activation decays without re-exposure. Adopters churn.

mesoThe map's edge matrix is the topology. Affinity edges are the wide bridges a complex contagion needs to cross a community boundary; tension edges are hostile contact, on which backlash propagates as its own signed channel.

macroThe environment. Finite attention, taxed by rival spikes per world. A media budget and its schedule, so exposure scales with remaining spend and a paid burst is a runnable counterfactual. The recommender layer explicit: algorithmic exposure multiplies with current engagement, adoption and hostility both counting, which is where cumulative advantage and trapping emerge rather than being scripted. A skeptic compartment for organised opposition.

the readsFour coupled ensemble reads, below. A single trajectory renders only as a ghost line behind a band.

The meso layer is free: the map's edges are literally the interaction matrix, which is why the map and the run are one system and not two products.

03 . What it puts out

Four reads, all ensemble-first. Never a point forecast, never a magnitude, never a dated promise.

fig 2 . the ensemble fan, drawn as a specimen and not from a run

share of the map, 0 at the base, 1 at the topt0 on the left, the horizon on the right

adoption, p10 to p90, median drawn backlash, p10 to p90, median drawn one world, a ghost

A drawing of the read's shape. Nothing here was run; the run is not built. The three pigments are the house's, in their fixed order, and the ghost is a neutral.

fig 3 . the four reads

Read	What it shows	The fact it renders
the fan	Adoption and backlash as p10, p50, p90 bands over the horizon.	the spread is the product; a single trajectory is an artifact
the outcome histogram	Final outcomes across worlds, with fork detection.	bimodality is a path-dependent future, and a mean of two futures describes neither
the diffusion map	The persona graph animated over a replayed world: where the cascade compounds, where it traps in a cluster, where it fractures.	adoption and hostility can win different regions of the same map
the lever read	Which parameter classes move the median: threshold spread, recommender gain, budget, backlash amplification.	the lever list is the deliverable

04 . The rules on the face

Inherited from the regrounded discipline and carried on the run's surface rather than in a footnote. Each is a constraint the engine enforces, not a promise.

Ensembles only.
Every read is a distribution across worlds. A single world renders as a ghost line and never as the answer.
Directional until calibrated.
Every output carries the measurement-ladder rung. No output graduates past directional until its parameters survive a backtest with a held-out episode. Every run this house could produce today would be directional, and the face says so.
Relative outcomes only.
Which frame, which plan, which lever. Never a point forecast, a magnitude, or a date.
The synthetic floor, disclosed.
The ICC/ESOMAR International Code 2025 requires disclosure that synthetic data was used, with method and minimum data thresholds, and sets an augmentation floor: synthetic segments at or under 15% of sample, with at least 300 real respondents in the base. The run prints the disclosure with every read.
No raw accuracy number.
The one figure this page quotes is normalized and says so: the generative-agent study most cited for this class of engine reports 0.85 normalized accuracy, which is a raw 65.7% over the participants' own 79.5% two-week test-retest consistency. The claim is that the agent is 85% as consistent with a person as the person is with themselves. Vendors drop the normalization; the house does not.

05 . How it would earn a rung

A rung is earned by a program, and the program has three directions.

Backward. Foresight is the forward mode of the same engine that Hindsight runs backward, so every Positioning Backtest episode calibrates it. That loop is the only route from directional to calibrated, and Hindsight is spec'd, not built, which is the fact that fixes this page's rung.

Forward. Every engagement's intervention card is a dated, resolvable prediction. It is logged, scored when it resolves, and published as a register with its resolution method stated. A register with a stated resolution method has precedent in the field for simulated humans only, and executed loosely; nobody runs one for a cultural read.

Market-priced. Where a scenario touches a listed culture market, the house's directional call is logged beside the market price at freeze time, so the two can be compared when the market resolves.

06 . Against the field

The tier that competes with this instrument is the LLM audience-simulation tier, and it is where the money and the published validation sit. Every figure in the table is the vendor's own and none has been independently audited.

fig 4 . the simulation tier, and the distinction

Vendor	Published posture, self-reported	Where this differs
Artificial Societies	300 to 5,000+ AI personas from real profiles; models peer influence. An evaluation page states 86% distribution accuracy while the homepage states 95%, with no limitations section.	Peer influence is the meso layer here too, but from the tenant's own map. No accuracy percentage is marketed at all.
Subconscious.ai	Causal framing on discrete choice theory; lift with a confidence interval; a public leaderboard with a cleanly defined metric.	The nearest analogue to the register in § 05. The house adds the rung and the synthetic-floor disclosure on the face.
Aaru	Behaviour simulation; a 0.90 median correlation in one six-month study.	One case is not a suite. The register is a suite by construction.
Synthetic Users	Synthetic interviews on personality-trait personas; from $12,500 a year, $2 to $60 per interview.	Interviews an isolated persona. The run models propagation across a map, with the levers as the output.
Roundtable, now Proof of Human	Pivoted from selling synthetic audiences to selling detection of AI respondents.	The same buyers purchase simulated humans and proof that respondents are not simulated. The disclosure rule in § 04 is the house's answer to that tell.

Nobody in the tier sells the mechanism: how this specific move propagates through a modelled population, who adopts, who attacks, where it forks, and which lever changes that. None reads meaning; all predict response; none hands the read to the people it describes. That is the lane, and the sweep of July 2026 found nobody in it.

07 . Its state

Spec'd, not built. Ruled 2026-09-02, in the same words on every surface of the house.

A v0 prototype landed in July 2026 in a sibling repository and is not in this one. The engine described above is a specification with a research base, a scenario set and a validation programme, and no code the house serves. The run cannot exist before the map it reads, and the map waits on the calls its own page lists; two of the run's own calls are below.

S5Ruled: the run reads the quant block, not an activation-space sense of the persona. Reading 2 is closed.
the nameOpen. The engine is not named Signal; that word survives only as data-layer vocabulary. What it is called is a call the record holds.
the studyOpen. Nobody has re-run the clustered-versus- random topology experiment with agents of this kind. A small replication is both validation and the first public result.

The map this run reads is described at how the map is computed. The instrument's place among the four is on the front door, in the same words.