MeridiansMeridians

Training worlds — requirements for an hour that is remembered

From the Meridians Wiki · Public · Maintained · human-contract

Status: rough spec (direction). Semantic requirements and intent for the next roadmap, not an implementation contract and not a new ontology. It composes the shipped substrate and the existing visual-novel specs (branching, decisions, flags, fate exploration, latent model) into one target: a training world is a short, immersive, original, emotionally true branched world that a person can play through fast, that rewards seeing the right decision, and that they come back to.

The picture to hold is a spy inserted into an unfamiliar country: no briefing covers what they will meet, so they read the room, work out who is telling the truth, judge what a move will cost, act before they are sure, and live with it — then are pulled out, debriefed, and sent in again, better. A training world is that insertion made safe, short, replayable, and worth being in.

The goal is simple and is stated without hedging: release effective training worlds. Effective means the hour changed how the player weighs a choice; immersive means they forgot they were being trained; original means the world is ours and could not be found elsewhere; remembered means they tell someone. Everything below is a requirement on the world, its creation, or its measurement, in service of those four words. What follows is vibes and semantics on purpose — the specifics belong to the owning specs.

1. What a training world is

one world · one hour · one question the player must get right
a canonical good route, earned by right decisions
divert  →  rejoin, or diverge for good if the decisions carried enough sway
consequence felt inside the fiction; the reading of it afterwards
  • A bounded hour. A training world is consumed in under an hour at the player's pace, and in far less at speed. It is a Scenario-and-Experience passage over one Domain, prepared before play, never a persistent world. It ends. The ending is part of the training.
  • One question. Each world is built around a decision the player has to see — not a puzzle with a hidden key, but a situation in which the right move is available to someone paying the right kind of attention. The world is a lesson in what to attend to.
  • Original. Worlds are Meridians' own — original settings, casts, and stakes that carry an emotional charge. Adaptation is a production shortcut for proving the pipe; the released catalogue is ours.
  • Or brought. A training world is a causal simulator, and the situation it simulates may be the player's own: a negotiation, a hard conversation, a new team, a new country. The intended intake is an interview — a voice agent that builds the situation model, the actors and their incentives, the history, constraints, and what the person may be missing — from which the world is constructed and Meridians World Search (MWS) prices its pathways before play. Today that step is done by hand through MCP (timeline-wisdom); the voice intake and automated construction are direction, and are the first-priority direction (specs map). A brought world is held to every requirement below; its cast are people the player will meet, so the fiction stays a fiction and the reading stays about the play.
  • Emotional and immersive first. The player must care about the people in the room before a decision is asked of them. Sprites, plates, staging, voice, sound, and the cadence of a scene exist so that the fork lands on someone who is already inside the world. A fork that arrives before immersion is a menu.

2. The canonical good route

Every training world has a canonical good route: the path the world would take if the player made the right decisions at each fork. It is not the only ending, and it is not the popular one; it is the one the world's own rules and people vindicate.

  • The good route is earned, not signposted. The right decision is visible to a player who anticipates consequence, reads the room, and holds a stance under pressure. The world encourages the player toward it — through characters who tell the truth as they see it, information that is there to be noticed, and stakes made legible — but never labels it. No option is marked correct; no meter tilts.
  • Diverting is real. A wrong or different decision moves the world. The player leaves the good route and the world continues coherently from where they actually are, with the people and state they actually have. Nothing snaps back.
  • Rejoining is allowed and costs something. Many diversions can rejoin the canonical route at a later bottleneck — the same event happening in every route, read differently in each (branching §7). A player who rejoins arrives with less: a lost ally, a spent resource, a truth learned late. The rejoin is where the lesson is felt.
  • Diverging for good is allowed when the sway is enough. If accumulated decisions carry enough weight, the world commits to a different path and does not come back. Divergent endings are as finished and as cared-for as the canonical one. A bad end teaches by showing what would have made it go otherwise; a strange end is a real place the player chose to go.
  • Sway is world state, not a score. Whether a diversion rejoins or diverges is decided by what the decisions actually changed — typed deltas, closed threads, moved relationships — not by a hidden counter. Flags carry the player's half of that record (flags); the world's half is always state or nothing.

3. Requirements on the world

  • Consequence has weight. Every fork that stops the player changes something that can be pointed at later. Forks whose branches differ only in prose are not offered.
  • The right decision is discoverable. Before a fork, the world has placed what a careful player needs — a line, a look, an artifact, a rumour with a source — and placed it in the fiction, not in a hint. Maximal support before the fork; none at it.
  • Perspective genuinely withholds. The player knows what their seat could know. Part of the training is deciding without omniscience; part of replay is returning with what another seat knew.
  • Bottlenecks carry the spine. A small number of route-invariant events hold the world together so it can be established once and re-entered cheaply, and so divergence reads as this world going differently rather than a different story.
  • Endings are finished. Canonical, rejoined, diverged, and bad endings each get an ending card that reads the player's route back — what they committed to, what it cost, what would have gone otherwise.
  • The world is warm. Stakes are personal before they are strategic. The people the player meets are consistent across seats and routes because their behaviour is grounded in evidence (latent model), and they remember what the player did.

4. Requirements on creation

The foundations exist: Domain extraction, branches with priced futures, Scenario resolution, Stageplay composition, Factory assets, fate exploration, and the flag ontology. What is required is that they compose into fast, repeatable branched-world creation aimed at §1–§3.

  • Scout for the question, then the route. Exploration finds the fork worth building the world around and the canonical route that vindicates the right answer; it then scouts the diversions worth producing — the ones a plausible player would actually take — and where they rejoin or break away. The standing form of that scout is MWS: the branch tree grown between readings toward forks whose futures diverge, inside the difficulty band the flag plan declares.
  • Flags are the design grammar. Divergent, conditional, and personal flags (flags §1) are how a world is planned at the reader's grain: where the route splits, what small commitments the world later reads, what is asked only to be read back. A training world is designed as its flag plan before it is written.
  • Generation fills, structure holds. Generation writes prose, scenes, variants, and gap continuations between the structural points the scout and the flag plan fixed. It does not decide where the forks are, what they cost, or where the good route runs.
  • Prepared, then served. A training world is produced before anyone plays it, so play is fast and stalls on nothing. Speed of play is a property of preparation, not of a smaller world.
  • Cast and place are reusable. Factory lineage keeps a world's people and rooms the same across routes, seats, and revisions, so a second world in the same setting is cheaper and a replay is recognisably the same place.
  • Short worlds, many of them. The unit of release is small enough to make often. A catalogue of remembered hours beats one large world nobody finishes.

5. Trait measuring, inside the boundary

Training worlds measure. They measure what the player did — anticipations stated and met, stances held under cost, perspectives taken, promises kept — and read it back afterwards as a portrait of how they played. This is the flag ontology's personal nature and the calibration and coherence readings already described as hypotheses (flags §7).

  • Measured in fiction, about the play. A reading describes decisions in a declared world, from a declared seat, over a declared interval. It is a mirror the player holds up to their own run.
  • Read back afterwards, never shown live. No meter during play; the ending card and the replay compare are where the reading lives. A live gauge turns practice into optimisation and breaks immersion.
  • Domain-native, not clinical. Trait labels are the world's — held the line at the hearing, trusted the wrong messenger twice — not inventory scales. No clinical validity, no population norms, no protected-characteristic inference, no cross-person ranking. The latent model boundary applies to the player exactly as it applies to a character.
  • Encouragement, not judgment. The reading's job is to make the next run better: what to notice, whose word to weigh, when to hold. It never grades the person, never predicts their life, and never leaves the fiction as a claim about them (promise).
  • Improvement is the purpose; its evidence status is stated. Every claim about change is designed, measured in fiction, or shown to transfer, and copy never runs ahead of the rung (strategy 12).

5a. The measure — Rasch

Each fork in the flag plan is an item; the good-route decision is the keyed response.

ln[ P(right move at fork i) / (1 − P) ] = θₙ − δᵢ
  • δᵢ is how hard the right move is to see. Fate's belief about the fork gives the prior; players give the estimate; where they disagree, the world hid the key.
  • θₙ is the trait as this person expresses it, in logits, with SE ≈ 1/√Σ P(1−P) always attached.
  • The fork that measures best teaches best: information peaks where δ ≈ θ, so hours are pitched at the edge of what the person can see — a few easy forks for confidence, most at the edge, one or two stretch.
  • Fit flags a bad fork, guessing, or a world that is really measuring two things.
  • The key is calibrated, not assumed. The canonical good route supplies the provisional key; a fork where well-measured players split near 50/50 and misfit is a values disagreement, not a hard item, and is retired from the scale (it may stay in the world). θ measures the trait only to the extent the key survives that test.

How each decision is written to the person's record, how the reading is understood, and how it steers the next generated hour is the latent-traits spec.

One hour of 6–10 forks reads a person to roughly ±1 logit — nearer ±1.3 once the testlet correction below is applied; six to eight linked hours to ±0.4–0.5. That is about the size of change a journey can produce, which is why the journey is as long as it is. Items derive from Flags and canonical routes; there is no second source of truth.

Identifiability — the scale must be pinned by something that does not change. Every generated hour is a new world, so every fork in it is a new item; a bottleneck re-staged in a new setting is a new item too, and its δ is expected to drift (differential functioning by world is the default, not the exception). If the only thing two hours share is the person, then "the person moved" and "this week's world was easier" produce identical data — Δθ and Δδ are not separately identifiable. Two designs pin the scale, and a journey uses both:

  • Frozen anchor hours. A small library of worlds that never change — same prose, same forks, same key — played unchanged by every person at baseline, mid-journey, and end. Their items are calibrated once on the cohort and then fixed; every generated hour is linked to the scale through them, never through a fork that merely resembles one elsewhere. Frozen anchors are the only forks the spec calls anchor forks.
  • Counterbalanced cohorts. Generated hours are assigned in counterbalanced order across people, so an hour's difficulty is estimated from players who met it at different points in their journey and world drift averages out of the cohort's Δθ. Item estimates are always cohort estimates; a fork met by one person has no δ.

Local dependence — forks on one route are a testlet. A player only reaches fork 6 because of fork 3, so responses within an hour are conditionally dependent and Σ P(1−P) overstates the information an hour carries. Treat each hour as a testlet: estimate the within-hour dependence from the cohort and deflate the hour's information accordingly (the ±1.3 above), or fit the hour as a single partial-credit item over its route when dependence is strong. Never sum raw fork information across a route as if the forks were independent.

Per person, an interval; the claim, a cohort. With SE ≈ 0.4–0.5 at each end, SE(Δθ) ≈ 0.6–0.7 and the 2·SE rule needs |Δθ| ≳ 1.2–1.4 logits — above what one journey can plausibly produce (≈ 0.5–0.8). So the per-person reading shows the interval and where the route moved inside it, never a pass/fail on growth; "you changed" is a cohort statement, decidable because n makes it so, and the person is told which cohort read their journey sits in and how wide their own interval is. The 2·SE rule, the transfer world, and retention gate the cohort claim.

5b. The journey and the concierge

The hour is one session in a journey: 6–10 hours over 6–10 weeks, one trait the person named, one world per week generated for them from the last reading. Baseline first; growth in the middle, new setting and cast each time; a frozen anchor hour at baseline, mid-journey, and end (§5a); a transfer world that resembles nothing before it; a delayed retention return; then the journey's reading. A concierge voice — in-world or as the service — frames each hour, receives the reading, proposes the next, and never reports a score. Emotional design is first-class: attachment, loss, temptation, dread, relief — consequence that is not felt does not change anyone, and the reading says what moved them. The plan and its proof programme are strategy 12.

6. Replayability and being remembered

  • The second run is a different story. Another seat, another route, or the same route with what the player now knows. Skip-over-seen makes the common route bearable; withheld perspective makes the second seat worth it.
  • The world remembers. Endings, flags, and the reading persist for the player; a replay is compared against the first run, not started cold.
  • Memorable means one scene. Each world is built so that one moment — the fork, the rejoin, the ending — is the thing a player describes to a friend. If the scout cannot name that scene, the world is not ready to produce.
  • Fast is a feature. A player who can finish in twenty minutes at speed and an hour with care will replay; one who needs an evening will not.
  • Never the same world twice inside a journey, unless replay from another seat is the point.

7. What effective looks like

Not numbers yet. The world is effective when a player who finishes can say what the right decision was and why they did or did not see it; when second runs happen unprompted; when the reading is kept rather than dismissed; and when someone describes the world to another person by its scene rather than its mechanics. Numbers arrive when there is a played world to take them from.

8. What this is not

  • Not a new ontology or a fourth medium. It is Scenario decided and Experience read, over one Domain, prepared and bounded.
  • Not a persistent or populated World; not an agent benchmark. Those remain their own directions.
  • Not a test of the person. It is a practice ground with a mirror, and the mirror stays in the fiction.
  • Not authored-by-hand branching. Routes are found by exploration and fixed by the flag plan; the world computes consequence.

9. Open

  • how strongly the world may encourage the good route before encouragement becomes a label;
  • how much sway is "enough" to diverge for good, and whether the player should ever be told;
  • how many diversions a one-hour world should produce versus leave as cliffs;
  • the first original setting, and how many worlds it should carry;
  • what the reading shows first — calibration, coherence, or the single decisive moment;
  • how many frozen anchor hours a journey can carry before they read as tests, and how often a frozen hour may be retired and re-calibrated; when a fork is partial-credit rather than keyed; how the generating model becomes a facet so drift is measured, not suffered;
  • the within-hour dependence actually observed, and whether an hour reads better as a testlet or as one partial-credit item;
  • the minimum cohort size at which Δθ is decidable for one trait, and what the person is shown before it is.