Narrative Governance

Two ways coherence breaks, both of them familiar

The most useful recent research does not promise coherence. It names precisely how coherence breaks, which is far more actionable. Two failure modes of long-horizon multi-agent narrative have been given names. Social memory stacking is what happens when memory only appends: contradictory relational states pile up until nobody can say what the relationship currently is. Narrative-spatial dissonance is a character sitting in a place the plot has already contradicted.

Both walk straight out of fiction and into a company. Social memory stacking is four incompatible versions of the story about one product, held by four teams, each true when written and none retired. Narrative-spatial dissonance is the press release describing a position the company left two quarters ago. Neither is a writing problem. Both are governance problems, and both are invisible to anyone reading one piece at a time.

A thirty year old problem with a birth certificate

This is not a new management fashion. Interactive narrative research has called it strong story against strong autonomy since the Oz Project, led by Joseph Bates at Carnegie Mellon in the early nineties. Give the director enough power and the world becomes coherent and dead. Give the actors enough freedom and the world becomes alive and incoherent. Three decades in, no dominant solution has emerged, which is worth knowing before buying one.

In 2008, David Roberts and Charles Isbell wrote down ten requirements a system of this kind has to satisfy. Four of them show the shape of the difficulty on their own: authorial control, player autonomy, invisibility, measurability. Every item is achievable. They pull against each other.

Both answers were built, and both cost something

The top-down answer is drama management: a layer above the agents that orchestrates them so the story stays globally coherent. Facade, built by Michael Mateas and Andrew Stern and released in 2003, held twenty seven story beats and around two and a half thousand joint dialogue behaviours, sequenced by preconditions, weights, priorities and their effect on story tension. It worked. It also cost so much authoring that Mateas described writing interactive drama in that architecture as possible, but not necessarily easy.

The bottom-up answer is emergent narrative: let autonomous agents interact and let the story arise from what they do. Comme il Faut, the first social physics engine, made social relationships playable in Prom Week and ran on roughly five thousand hand-authored social rules and about a thousand dialogue templates. Versu inverted the premise entirely. The author configures the initial state and not the plot, and relationships nobody wrote appear between characters.

Riedl's Automated Story Director sat between the two, letting characters act freely until an experience manager intervened, with contingency narratives precomputed by a planner. Read as a set, the lesson is consistent. Every architecture in this lineage buys one property and pays for it with another, and the payment is usually made in authoring labour.

Language models paid one bill and opened another

Then fluency arrived and the authoring bill collapsed. Nobody has to hand-write five thousand social rules to get plausible dialogue any more. What arrived with the fluency was a different debt: hallucination, inconsistency over long horizons, and no causal guarantee that what happens in chapter nine follows from chapter two.

The problem moved. It used to be writing enough rules. It is now trusting what got written. A better sentence does not help with that, so the field converged on architecture instead of on a bigger model. A director separated from the actors. A shared narrative graph. A game master that is a configurable component instead of hidden plumbing. Memory split into working, episodic and consolidated tiers.

Govern with reasons the world already contains

The most useful recent finding for anyone designing this is about the shape of intervention. In work presented at ACM IUI in 2026, a system running specialized agents over a shared narrative graph tested how participants respond to being redirected. Flat refusal reads as railroading and gets rejected. Redirection grounded in the internal logic of the world, delivered through more information, visible consequences and pressure from other characters, gets accepted, and it did not cost participants the sense of acting freely.

That is one study in a fictional setting, and it points at a design principle without proving one about organizations. It is still the sharpest evidence available, and it says: govern by giving reasons the world already contains.

Governance is not a bigger veto. It is a better one. A company that answers a team with because the guidelines say so has picked the version the evidence found weakest.

The four questions to answer in your own company

Strip the fiction away and four questions remain. Most organizations can answer one of them.

1. What does coherence mean here, concretely enough that two people judge the same piece the same way?
2. Who is allowed to break it, and through which door?
3. What evidence proves it held, given that the measurement problem is unsolved and every proxy has to be declared as a proxy?
4. What happens when it does not hold?

The fourth is the one brand systems almost never answer, because they were written as documents and a document has no runtime. There is also an unsolved problem underneath all four: computational metrics still do not capture what a human means by a good story, and the field lists standardized measurement of narrative quality as open. Treat any governance layer sold as having solved measurement with suspicion.

Reading which stories are forming before you write anything is real-time narrative intelligence. Running the whole thing as a continuous function is storytelling management.

If your answer to question four is that nothing happens, that is the place to start, and the conversation begins here.

Questions