The practice record — the journey, the measure, the proof

# The practice record — the journey, the measure, the proof

> **Status: the plan.** Meridians grows people through repeated, generated, emotionally consequential hours
> in worlds made for them, measures the change on a calibrated scale, and proves it transfers. It never
> passes a verdict on anyone. Every claim below carries its evidence status: *designed / measured in
> fiction / shown to transfer*.

## What we monopolise (2026-09-18)

**The causal simulator, enabled by Meridians World Search (MWS).** Not a library of visual novels and not a persistent
universe yet: a consequential situation turned into a kept world whose pathways MCTS has grown and priced,
rehearsed through the VN, remembered, and retried. That is the value, and value today lies in **causally
deploying simulations from a voice interview** — the person tells us what they are preparing for; the
world is built; the pathways are searched; they rehearse before reality does it for them
([01](01-the-north-star.md#meridians-is-a-causal-simulator-the-visual-novel-is-its-interface)). Micro
sonic analysis is the value-add on the felt grain of those hours. The primary roadmap is **Causal MCTS VNs**;
the VNMMO is a stretch goal on a separate roadmap: a persistent universe that relies on the same simulator,
more automated, kept on the side until the
simulator has earned it ([specs map](../../../specs/README.md#priorities-2026-09-21)). The record below is
what the simulator leaves behind in the person, and it is the part no one else can regenerate. The full
argument — the compound asset, why the simulator is a sci-fi technology in earnest, the weak points, and
the build-ordered roadmap — is [14](14-the-causal-simulator-monopoly.md).

The durable thing a training world leaves behind is not the world. It is **the person's record of having
been in it**: what they committed to under what they knew, what they expected and what happened, which
seats they re-entered, what they carried across runs — and, across a journey, how far they moved.

## The journey — a concierge live service

The hour ([training worlds](../../../specs/training-worlds/README.md)) is one session in a journey:

```text
journey = trait × person × [hour₁ … hourₙ] × weekly cadence × debrief
hourₖ   = world_k(question_k, forks aimed at θ̂_{k−1}) → reading_k → θ̂_k
anchor  = a frozen hour, unchanged for everyone, at baseline · mid · end — what pins the scale
```

Six to ten hours over as many weeks, aimed at one trait the person named (courage under loss, romantic
judgment, perspective-taking, calibration, promise-keeping). Week one baselines; the middle weeks grow, each
new world generated from the last reading — new setting and cast, same trait, forks pitched at the edge of
what the person can see; a **transfer world** near the end that resembles nothing before it; a delayed
**retention** return; then the journey's reading, in the world's words, with its uncertainty. The next
hour is prepared *between* sessions by the Factory pipeline from the record; a concierge voice frames each
hour, receives the reading, and proposes the next. Spacing and the debrief are what make repeated hours
compound rather than blur. Sold as a monthly subscription on trait change and emotional experience; the
operator subscription ([04](04-the-market.md)) stays as the floor for people who use the substrate.

## The measure — Rasch

```text
ln[ P(right move at fork i) / (1 − P) ] = θₙ − δᵢ
```

A fork in the flag plan is an item; the decision on the canonical good route is the keyed response; δᵢ is
how hard the right move is to see (Fate's belief gives the prior, players give the estimate); θₙ is the
trait as this person expresses it, in logits. Information `P(1−P)` peaks at θ = δ, so the fork that
measures best is the one that teaches best; `SE(θ̂) ≈ 1/√ΣI` is the honest uncertainty on every reading; fit
finds bad forks, keys that are really values disagreements, and worlds that measure two things. Rasch
rather than a tally or a portrait because of **specific objectivity**: players take different routes and
meet different forks, and only a calibrated, anchored scale puts them on one θ regardless.

Three constraints make that scale real rather than nominal
([training worlds §5a](../../../specs/training-worlds/README.md#5a-the-measure--rasch)):

- **Pinned by frozen anchor hours.** Generated hours are new items every week, and a bottleneck re-staged
  in a new setting is a new item whose difficulty drifts; with only the person linking two hours, "they
  grew" and "the world got easier" are the same data. The scale is pinned by a small library of worlds
  that never change, played by everyone at baseline, mid-journey, and end, plus counterbalanced ordering of
  generated hours across the cohort so drift averages out.
- **Hours are testlets.** Forks on one route depend on each other, so an hour's information is deflated
  for within-hour dependence (≈ ±1.3 logits per hour, not ±1).
- **Per person an interval, per cohort a claim.** `Δθ = θ̂_end − θ̂_start` carries `SE(Δθ) ≈ 0.6–0.7` after a
  full journey — wider than the change a journey can plausibly produce — so the person's reading shows the
  interval and the movement inside it, never a growth pass/fail. "Changed" is asserted for a cohort, when
  |Δθ| > 2·SE(Δθ) on anchored items *and* holds on the transfer world *and* at retention; otherwise "no clear
  change yet".

Partial credit and many-facet extensions (world, seat, occasion, generating model) when needed. The unit
already fits: Fate runs in logits.

## What the record is

```text
commitment at a stated vantage → consequence → reading back → next insertion, harder or from another seat
```

- **Yours.** Attributed to one person against one kept world. Another studio can make a better world; it
  cannot make your history in this one.
- **In fiction, about the play.** Anticipations met, stances held under cost, perspectives taken,
  promises kept — described in the world's own words, from a declared seat, over a declared interval.
- **Read back, never shown live.** The ending card and the replay compare are where it lives. A gauge
  during play turns training into optimisation and breaks immersion.
- **Steers the next hour.** Every VN decision is useful information about the person; the record's first
  consumer is generation, which pitches the next world at the edge of what they can see and presses where
  they surprised the instrument ([latent-traits spec](../../../specs/latent-traits/README.md)).

## How what we sell feeds it

- **The training world** starts it cheaply: one hour, one world, one reading.
- **The pocket dimension** deepens it: what another seat knew is the sharpest thing a debrief can say, and
  only other wills can supply it. Entry is earned in training worlds; the reading stays the person's own
  inside the crowd ([01](01-the-north-star.md#pocket-dimensions-with-care)).
- **Agents** get both as twins: the same kept world, the same log, a policy in the seat ([13](13-consequential-test-worlds.md)).

## What it is not

- Not a score, rank, diagnosis, or prediction of a person. The reading is a mirror the player holds up to
  their own run ([promise](../positioning/promise.md)).
- Not clinical. Trait labels are the world's — *held the line at the hearing* — not inventory scales; the
  [latent-model boundary](../../../specs/latent-model/README.md#2-what-psychometrics-contributesand-what-it-does-not)
  applies to the player as to a character.
- Not sold to anyone else. Nothing about a person leaves the person.

## The proof programme

We claim trait change, so we owe the evidence, rung by rung, each named on the tin:

| Rung | Status | Evidence |
| --- | --- | --- |
| **Designed** | spec exists | every hour built to a fork plan aimed at a declared trait |
| **Measured in fiction** | not started | θ̂ trajectories with SE, pinned by frozen anchor hours, testlet-deflated, fit-checked, keys retired where players split on values; cohort read first, per-person intervals always |
| **Shown to transfer** | not started | pre-registered journey vs. wait-list study, external measure of the trait's real expression before and after, retention at three months, run or audited independently, published either way |

Copy never runs ahead of the rung: *built to change how you decide* → *measured to change how you decide in
these worlds* → *shown to change how you decide*. The old gate was a reason for silence; this is a
programme with dates. Ethics, minimal: informed consent that the service is designed to change behaviour;
the person owns and can delete their record; traits are opt-in and named by the person; no
protected-characteristic inference; no third-party disclosure; a referral path if a journey surfaces
distress.

Still true: **the simulation cannot certify its own transfer**
([WORLD_ENGINE](../../architecture/WORLD_ENGINE.md#training-environments-preserve-the-learning-conditions)).
Rung three is how *we* certify it, externally.

See [05 — The moat](05-the-moat-ordered-by-what-survives-the-worst-case.md) for the world-side twin,
[human practice](../positioning/human-practice.md#from-remembered-world-to-player-progression) for the loop
that produces the record, and the [training worlds spec](../../../specs/training-worlds/README.md) for the
unit that starts it.

Source: [2026-09-10 concierge live service / Rasch handover](../../../sources/founder/2026-09-10-2026-09-10-concierge-live-service-rasch-handover.md).

Raw Markdown source