Latent traits — the practice record as an instrument that steers generation

Status: rough spec (direction). Semantic requirements for measuring a person's latent traits from what they decided in visual novels, and for using that reading to generate the next hour for them. Not an implementation contract and not a new ontology. It composes pieces that already exist — the fork as an item and the Rasch measure (training worlds §5a), the personal flag and the mirror harvest (flags §1, §4, §5), the journey and its proof programme (strategy 12), and the psychometric boundary (latent model §2, §10b) — into one target: every decision a person makes in a training world is useful information about them, kept in one record, read to understand their latent traits, and used to make the next world fit them.

The thesis it serves: the world around us is causally thin, so a training world builds causality and consequence from nothing; with enough play the same hours measure a trait and develop it (strategy 01). Measurement and development are one loop. This spec is the loop. Inside the causal simulator it is also the compounding part (strategy 14 §1, §3, track D): what a person tells the interview (their situation, actors, tendencies) and what they do in play are two sources on one person, and together they are the context from which each next world is built more realistic for them.

Nothing here is built. The Rasch fit, the record, the steering, and the analysis are all unimplemented.

---

1. The loop

hour played → decisions written to the record → reading (θ per trait, SE, fit, stance)
           → next hour generated from the reading → hour played …

Three requirements hold the loop together:

• Every fork writes. A decision is never consumed by the story alone; it lands in the person's record with what it was a decision about. Items derive from Flags and canonical routes — there is no second elicitation and no questionnaire.
• The reading is for the next hour first, the person second, the cohort third. Its first consumer is generation; its second is the mirror shown after play; its third is the cohort claim that a journey changed anyone. Each consumer sees a different projection of the same record.
• The person chose the trait. Nothing is measured that the person did not name as what they came to work on, and a trait axis is always the world's word for it, never an inventory label.

---

2. The record — what one decision writes

One decision at one fork writes one observation, and the record is the ordered set of them across every world the person has played. The observation carries enough for all three consumers and nothing more:

Everything else — stance, coherence, attention, conviction, attribution, register — is derived from these, exactly as flags §4 already requires. The design test is the same: if it can be derived from the observation plus world state, it is not a new field.

The record is the person's. It is private to them and to the generation that serves them; nothing in it is disclosed to another player, a seat, or a third party; the person may read all of it and delete it (flags §3 — losing it loses a portrait, never content).

---

3. The reading — understanding latent traits

The reading is what the record says about the person right now. It has three layers, and each is honest about what it is.

• The measure. Per trait axis, θ in logits with its SE, fitted over the keyed items as training worlds §5a specifies: frozen anchor hours pin the scale, hours are testlets, a fork that splits well-measured players near 50/50 and misfits is a values disagreement and leaves the scale. Fit is read per person too: a person who misfits an axis is not "bad at it" — the axis may be measuring two things in them, and that is information for steering.
• The stance. The distribution of personal flags and the coherence of that distribution across a route: who the person is being in this world, whether they held it, where they drifted. Not a score.
• The moments. The few forks where this person's route diverged from what the reading predicted — the places they surprised the instrument. These are their causal moments: where a decision changed what followed, and where the next hour should press.

Multi-trait is the default. A person names one trait for a journey, but a fork exercises more than one axis, and the record accumulates evidence on axes the person did not name. That evidence may steer generation and may appear in the mirror as the world's observation; it may not become a claim, a label, or a suggestion that the person work on it unless they ask.

What "understand their latent traits" may conclude, and may not:

---

4. Steering — the reading generates the next hour

This is the centre of the spec. The reading exists so the next world fits the person, and "fits" has a precise meaning:

• Pitched at the edge. Forks are planted where δ ≈ θ on the named axis: a few easy for confidence, most at the edge of what the person can currently see, one or two stretch. Information and learning peak in the same place (training worlds §5a).
• Pressed at the moments. The next hour re-stages the kind of fork the person surprised the instrument at — a new setting, a new cast, the same shape of decision — so a causal moment becomes a faculty rather than a one-off.
• Shaped by what moved them. The record's felt field tells generation which consequences landed — attachment, loss, temptation, dread, relief — and which passed unnoticed. Consequence that is not felt does not change anyone; the next hour spends its emotional budget where this person feels it.
• Varied by what they never look at. Seen-set omissions become the next world's density: what the person consistently does not attend to is where the next key is hidden.
• Never repeated. A new world every hour inside a journey, with the anchor hours frozen and unchanged. Steering chooses shape, difficulty, and emotional register; it does not recycle content.
• Bounded by the person's ask. Steering optimises the axis the person named. Evidence on other axes may vary a world but may not redirect the journey without the person's say.
• Built from their context. (Direction.) Actors, incentives, history, and constraints the person gave by interview, and the self-report of how they usually respond, inform the next world's construction so its people and pressures are true to their situation — with the self-report kept beside the observed response on the same axis, the gap a private reading and never a verdict. Each item of context is consented per item, deletable, never crosses to another person's world, and is marked as the person's own reading of the people in it, not the truth of those people.

Steering is a projection of the reading into the flag plan of the next world (training worlds §4); it is not a second generator and does not write canon. Concretely it writes one thing the search can read: a target band — a difficulty range on the named axis, about ±1 logit wide — and Meridians World Search (MWS) (branch refinement §3) grows the tree toward forks inside that band. The search never sees θ, the response, or the record; the band is the whole of the hand-off, so the world is selected and the person is measured, never the reverse. The concierge voice frames what steering chose and why, in the world's words, and never reports the number (training worlds §5b).

---

5. The mirror — what the person is shown

After play, never during. The mirror shows the person their own record for this hour, in this world's words: the stance they played, the fork that was decisive, what they anticipated and what they never looked at, the interval their route sits in and where it moved inside that interval. It shows the cohort read only when n makes it decidable, and says so. It never shows a live gauge, a grade, a comparison, or the raw θ (flags §5–6).

The mirror's job is the next hour: what to notice, whose word to weigh, when to hold.

---

6. Refusals

The boundary in latent model §2 and VALUES applies without exception, and the loop adds three of its own:

• No measurement the person did not come for. The named trait is the contract; everything else is the world's observation, kept for steering, never turned into a claim.
• No steering that hides the good route to keep the number moving. The canonical good route is earned by right decisions; a world may be harder, never rigged.
• No record leaves the fiction as a fact about the person. Copy never runs ahead of the proof rung (built → measured → shown, strategy 12).

---

7. What effective looks like

Not numbers yet. The loop is effective when the next hour is visibly for this person — a fork they nearly saw last week now sits at the centre; a consequence lands where the last one passed them by — and when the person recognises themselves in the mirror without being told a number. When there is a played journey, the numbers are the ones §5a already names: SE per hour, SE across a journey, the cohort Δθ.

8. Open

• the observation's exact shape, and whether felt is witnessed by the record or reported by the person;
• how many axes one fork may be planted for before it measures nothing well;
• how strongly steering may follow un-named axes before the journey is no longer the one the person chose;
• what the mirror shows first — the decisive moment, the stance, or the interval;
• the minimum record before steering is better than a good default flag plan;
• whether an agent in the seat has a record of the same shape, and what steering means for it.