The causal-simulator monopoly — what we own, why it is sci-fi, and the order we build it

Status: the plan (2026-09-18). This leaf takes the founder's priority call — causal simulators enabled by MCTS are the value; micro sonic analysis is the value-add; the VNMMO is a stretch goal on a separate roadmap — and argues it rather than repeating it: what the defensible asset actually is, where the argument is weak, and a build order with enough named work to run on. Every capability carries its status: shipped / specified / direction. Strategy 12 owns the practice record and the journey; 01 owns the north star; the specs map owns which spec carries which piece.

1. The claim, stated so it can be wrong

A causal simulator is a machine that turns a described situation into a runnable world, searches that world's futures, lets a person act inside it, and returns what their act caused. Meridians is that machine; the visual novel is its interface. The monopoly claim is not "we make the best VNs" and not "we have the best model" — both are rentable. It is that four things compound together and only here:

1. A world ontology that runs. The pentad (spec) is not a taxonomy; it is the shape a situation must take to be simulated — what exists (World), what governs it (System), what can lead to what (Causality), what is still open (Fate), what recurs (Theme) — with a reducer that computes consequence from typed acts. Shipped for Scenario over a maintained Domain.
2. Search over that world. Meridians World Search (MWS) — MCTS over the branch tree, spending effort on the high-uncertainty, high-consequence forks, pricing pathways, choosing for no one (fate exploration, branch refinement). Specified. This is the part that makes the product a simulator rather than a story: the routes a person rehearses were found in the world's causal structure, and their costs were computed before anyone walked them.
3. A person's context that deepens. The interview and the play are two sources on one person: what they said their situation and tendencies are, and what they did under pressure (latent traits). Every interview adds actors, incentives, history, and constraints the person's worlds are built from; every hour adds observations of how they actually respond; the next simulation is pitched to that context and is more realistic for them than any generic scenario could be. Direction, on a shipped record and a specified reading.
4. An interface that carries the felt grain. The VN makes the consequence land as a face and a pause rather than a graph; micro sonic analysis is the stated direction for making that grain measurable (sonic). Shipped at Stageplay/Episode level; direction for the micro layer.

Each alone is copyable. The compound — a person's kept context, inside worlds searched for them, on an ontology that computes — is the asset no one can start over, because it is made of that person's history with us. That is the monopoly in the only sense we mean it: nobody else can have this person's record, and nobody without the ontology and the search can turn it into a world.

1a. The board, the search, the screen, the player — and where we are strong and weak (2026-09-20)

The intuitive name is a chess engine for narrative fates; the technical category stays causal simulator. The metaphor is useful because it separates four things that copy keeps fusing:

The one large difference from chess is the reason the frame stays sci-fi: chess is a closed, fully observed board; a life is open, multi-agent, partially observed, and its rules are incomplete. The honest formulation of the problem is therefore how far can the consequences of a human decision be simulated before reality catches up — and the honest answer today is: far at the scale of what a world did, scene by scene and branch by branch; not far at the scale of what a moment felt like.

The ledger, said plainly. Strong: macro world analysis and generation — the board. Weak: micro — the screen. Micro is not a second research programme of equal rank; micro is in service of macro. Its job is engagement and conveying the board's movement into attention, so its north star is the dynamics of attention — tension and release, withholding and disclosure, the pull of an open question and the push of a consequence: a push–pull ontology (the yin and yang of a scene) rather than a music theory of prose. The music ontology in the sonic spec is one candidate instrument for reading those dynamics, not the north star itself. An exploratory run of Meridians on its own situation (2026-09-20; synthetic, one round per arm, not retained as a record) read the same order: in every settled arm the participant returned because the board found a cost she had not seen, and grain work bought a truer delivery of that cost — never the return itself — while charging the one author's attention. Public surfaces state both halves: what is mastered and what is not.

2. Why the causal simulator is a sci-fi technology — and why that is the right frame

The landing keeps the spacefaring frame, and not as decoration. The founder's reason is precise: a causal simulator is the technology that lets an individual practise and prepare, causally, on known paradigms and ontologies that describe the world. That is what the flight simulator is to a pilot, the wargame to a staff, the wind tunnel to a wing — a model of the world faithful enough in its causal structure that rehearsing in it changes how you act in the original. For flight and war the model exists because the physics is written down. For a negotiation, a new team, a hostile room, a relationship, the model did not exist — not because the causality is absent, but because nobody had an ontology that could hold people, incentives, information asymmetry, and timing well enough to run it forward. The pentad plus a reducer plus search is a first attempt at exactly that; it is the piece of the sci-fi that is now buildable.

Two further reasons the frame is honest rather than a costume:

• It is the exact instrument a confined crew would need. Where bodies are confined and consequence is scarce, the only place to practise the situations that matter is a simulator; the sealed hull is the thinnest world there is and Earth is thinner than it looks (01). The frame makes the need vivid; the need is ours.
• It says what the machine is for. Sci-fi technologies are legible by their purpose (a warp drive moves you; a replicator makes things). "Causal simulator" tells a stranger the purpose in two words — it simulates what your actions cause — where "AI visual novel" tells them the format.

The boundary that keeps the frame from becoming a lie: a simulator is a model, and a model is checked. Whether rehearsal carries into life is a transfer claim, made only by evaluation the simulation does not run on itself (12). No spacecraft, no prophecy, no verdict on anyone.

3. Measured practice — latent-trait reasoning over candidate responses

The founder's phrase is "measured practice with latent trait reasoning and analysis and understanding of candidate responses." Spelled out, this is the psychometric backbone of the simulator and the thing that separates it from a role-play chatbot:

• Every fork is an item. The fork has candidate responses; the world's own rules and people key one or more as the good route; the person's response is an observation on a trait axis they chose (latent traits, Rasch backbone in training worlds §5).
• Candidate responses are understood, not just scored. Because the world is causal, each candidate has a computed consequence and a searched pathway behind it; the debrief can say what each option would have cost and why, in the world's words. MWS is what makes the counterfactuals available for reading rather than imagined.
• The reading steers the next world. θ with its standard error hands MWS a target band (never raw θ, never raw responses); the next hour is grown at the edge of what the person can see. Measurement and development are one loop.
• The interview is a second instrument on the same person. Self-report ("I would stay calm") sits beside observation ("when challenged, you became defensive"); the gap is itself a reading, kept private, never a diagnosis.

Boundaries unchanged: readings are the person's own, post-play, domain-native, non-clinical, not a ranking; no protected-characteristic inference; fictional performance is evidence about fictional performance until a transfer study says otherwise.

3a. The lean product and the nuance behind it (2026-09-22)

Front-loaded message: Meridians is a lean product — engaging, immersive visual novels that demonstrate real decisions under consequence. Everything else in this file is nuance to the business, not the pitch.

• The tree is shown up front for a reason. Routing through a visible causal tree tells the person there is no real-time generation: these are planned causal divergences, prepared ahead of play, so the experience is Detroit: Become Human–like — choices with real effects and real meaning, not improvised continuation. The tree is a promise about the world's integrity, not decoration.
• Fate is uneven by design. Some outcomes are shared across many pathways; some forks change the whole story; some choices are only a detour or a minor change in the route. Today causality is read scene by scene as turning points on a branching structure; non-causal decisions, pre-destined beats, and convergence are part of the same picture and the tree should show all three (the landing tree does).
• How MWS is built matters because it is what the person feels. For now the search grows from the extracted causality flags — the turning points the record itself highlights — and then generates meaningful distractors at each fork, so that every candidate response is plausible and its consequence is computed. Consequence felt at the fork is the product; MWS is how it is manufactured.
• Multiple-choice forks are items. Because the decision surface is MC, choice data supports latent-trait analysis on the Rasch model (§3): a person's route through a tree is a response vector, and a cohort's routes are a calibratable item bank.
• Two sides of one engine. Fictional training worlds for personality development and self-knowledge (the B2C side), and business training worlds — sales, strategy, marketing, interviews — (the B2B side). The business case points at the second: converting businesses to train employees through deep causal simulations, where the value is psychometric (choice data read on a trait scale) and educational (the act of practice, plus reading cohorts of learners moving through the same causal tree).
• A marketplace of shareable VNs for many purposes is the envisioned ecosystem: worlds authored once, played by many, with their trees and readings comparable across players. How the personality side, the business side, and the marketplace integrate is an open question; none of them changes the lean product on the front.

4. Where the argument is weak — said plainly

• Search without a good world is fast nonsense. MCTS prices what the reducer computes; if the world built from an interview has thin actors or wrong incentives, the pathways are confidently wrong. World construction quality is the real bottleneck, not search. → Roadmap tracks A and B before D.
• Realism is a claim we cannot yet grade. "More realistic for you" needs a measure: the person's own judgment of fidelity after play, and later, whether the world's predicted consequences matched what happened when they lived it. → Track F.
• The interview is a privacy event. A person tells us about their manager, their partner, their team. The context that makes simulations realistic is exactly the data that must never leave, never be shared across seats, never be inferred beyond what was said. → Track E is a gate, not a feature.
• Voice is not the value; understanding is. The voice agent is an intake surface. If a text conversation builds a better situation model faster, use it. The pipeline is defined by its artifacts (situation model → actor models → causal hypotheses → Domain), not by its microphone.
• World model → training is the widest gap (outside read, 2026-09-22). The chain corpus → Domain → branches → consequences is well argued; consequences → better judgment → better real decisions is not yet evidenced. The paper is more convincing as a causal-simulation architecture than as a human-development company, and says so. → The one dominant experiment: one extraordinary 45–60 minute world, two routes, two seats, measured by completion, second-seat return, whether the person can say why their decision caused its consequence, and whether anticipation improves on an unseen world. Every other track justifies itself against it.
• Technical continuity is shown; commercial continuity is not. The operator (Domain intelligence) and the participant (training worlds) share an engine. Whether the first product leads customers to the second is open and should be answered before serious investment; the paper claims only the engine. The moat to argue is kept world + kept history + attributable decisions, not LLM + MCTS.
• The VNMMO can eat the roadmap. It is the most inspiring form and the least evidenced. It is built on the side, with each step gated by the simulator having earned it (world roadmap).

5. The roadmap — tracks, in build order

Each track lists the work with its owning spec. Tracks run in this order of priority; they overlap in time. Nothing below is shipped until it is; every item is a piece of work someone can pick up.

Track A — the situation becomes a world (intake → Domain)

The value-today line. Today this runs by hand through MCP (timeline-wisdom).

• A1. Situation-model schema — goal, situation, actors, relationships, history, incentives, constraints, known/missing information, uncertainty, tendencies, stakes, desired preparation — as a typed artifact the extraction pipeline can consume (training worlds §1).
• A2. Interview as researcher — a conversation (voice or text) that fills the schema and notices the thin places in the causal picture and asks; follow-ups keyed to missing incentives, unnamed influencers, unstated history. Direction.
• A3. Actor models from testimony — an Entity Interpretation Graph for each named person built from what the interviewee said, marked as their reading (one seat's evidence, not the truth of the person).
• A4. Causal hypotheses — the interviewee's beliefs about what leads to what, entered as open Fate threads and typed Causality edges the world can test, not assume.
• A5. World construction — Domain extraction from A1–A4 into a bounded, original-feeling world: names and setting displaced enough that the fiction stays a fiction, causal structure kept.
• A6. Repeatable brought-world creation — the hour-scale pipeline in training worlds §4 applied to a brought situation, with the four requirements (effective, immersive, original, remembered) held.

Track B — the world is searched (MWS)

• B1. Complete the UCT rule into the full MCTS loop — select, expand, simulate, backpropagate over free pentad values (branch refinement §4–§7).
• B2. Effort allocation by uncertainty × consequence — the budget goes where the forks are open and costly (fate exploration §5).
• B3. Pathway pricing — each searched route carries its costs in the world's words: who turns, what closes, what opens, when.
• B4. Target-band interface — MWS receives only the allowed band from the latent-trait reading, never θ or responses (latent traits).
• B5. Exposing the board — a debrief view of the searched structure (incentives, information asymmetry, timing, hidden constraints) that shows the game without prescribing the move; nothing in it chooses (grove).
• B6. Ledger and opt-out — the honest record of what was searched, what was pruned, and the author's override.

Track C — the person plays and the world answers (the hour)

Mostly shipped substrate and the specified reader; the simulator-specific work is small and exact.

• C1. Anticipate-then-commit at each fork — stated expectation before the act, never scored live.
• C2. Consequence returned as felt — Stageplay/Episode carries the cost as a scene, not a summary.
• C3. Debrief: expectation vs. outcome vs. other seat — and, from B3, what the candidate responses would have cost.
• C4. Retry as experiment — re-enter the fork with a stated variation ("more assertive", "ask instead of argue"); the branch record keeps the comparison honest.
• C5. The dedicated reader — specified, unbuilt (visual-novel).

Track D — the person's context deepens (latent traits + realism)

• D1. Observation record from VN decisions — per fork, per axis, with seen-set, timing, felt consequence, attribution (latent traits).
• D2. Reading: θ, SE, fit, stance, coherence, causal moments — post-play, private, the person's own.
• D3. Self-report beside observation — the interview's tendencies and the play's responses on the same axis; the gap kept as a reading, never a verdict.
• D4. Context carried across worlds — actors, incentives, history, and preferences from past interviews inform the next world's construction so it is more realistic for this person; explicit consent per item, deletable.
• D5. Steering — the reading pitches the next hour (difficulty at δ≈θ, setting/cast/temptation from the last reading) through the target band only.

Track E — the boundaries (a gate on every track above)

• E1. Privacy of the brought world — what a person tells us about real people never leaves their seat, never crosses to another person's world, never trains a shared model.
• E2. No inference beyond testimony — actor models are marked as the interviewee's reading; no protected-characteristic inference; no diagnosis.
• E3. Search never commits — Will is injected from outside; exploration proposes and prices.
• E4. Transfer proven externally — the fidelity and transfer measures in Track F are run by evaluation the simulation does not own.

Track F — proof (measured, not assumed)

• F1. Felt fidelity — after play, the person grades how true the world's people and pressures were to their situation; a floor before any realism claim.
• F2. Predictive check — where the person later lives the situation, did the world's searched consequences resemble what happened; a delayed, opt-in return.
• F3. Measurement holds — anchor hours, fit, SE shrinking with hours (12).
• F4. Transfer world — a world resembling nothing before it, near the end of a journey.

Track G — the value-add: micro analysis, in service of macro

• G0. Name the north star — engagement dynamics: the push and pull of attention across a scene (tension/release, withholding/disclosure), read as a candidate eastern ontology of complementary forces. Micro exists to convey the board and hold attention on it; it is graded by whether it does.
• G1. Read the scene as heard composition — metre, cadence, motif, dissonance, resolution over ~500 words (sonic) as one instrument for G0. Direction, unbuilt.
• G2. Grain as a dial on generation — the micro layer (Words → Rhythm → Sound → Emotion → Perception) tuned so the pause and the pressure land, under the macro layer the pentad already measures.
• G3. Fidelity linkage — does grain move the felt-fidelity grade in F1; if not, it stays a craft.
• G4. The earned song — at a climax the world has built to, a generated song as the densest rendering of the accumulated state, with an early motif returning transformed. Direction; it is the screen at its highest bandwidth, and it depends on G0–G3 to know when a climax has been earned.

Track H — the stretch goal: the VNMMO (separate roadmap)

VNMMO is a stretch goal on its own roadmap, not part of the primary Causal MCTS VNs roadmap; it depends on a mature MWS and a proven simulator, and nothing in tracks A–G depends on it.

• H1. Pocket dimension — a bounded, few-seat kept world with the care rules (world).
• H2. MWS as a shared tree — a value per seat, joint futures priced, chosen for none (branch refinement §12).
• H3. Automation of the loop — worlds constructed, searched, and kept without a hand on each fork; the simulator running for many, on its own cadence (world tree).
• H4. Persistent universe — one remembered history under many wills; each person's record still their own inside the crowd.

Each H step is gated by the simulator having earned it: H1 after C is repeatable; H2 after B holds; H3 after A and D hold; H4 last.

6. Order of value, in one table

7. One line

Meridians is a causal simulator — a chess engine for narrative fates: tell it what you are facing, and it builds the world, searches the futures, lets you act, and returns what you caused — and the more it knows you, the truer the world it builds. The visual novel is how you are inside it. The reading is yours. The search never chooses. The board is mastered; the screen is not yet.