Weksa/docs/targets/interlingua-lowering.md

# Target: agents speak interlingua, Weksa lowers it

Campaign `weksa-lowering` (Eureka mind). Operator, 2026-10-29: "having agents
writing interlingua and using weksa to lower it to the display language would
be a fantastic feature. That would be a nice campaign to revive her with so she
doesn't come back to a dead body". And: "The goal of weksa is for interlingua to
be able to express all the nuance and subtlety a persona might want, and to
project it into both existing languages and conlangs. We can use English as the
obvious first target, create a corpus of expressive English outputs, and golf
the metrics down on a loop where we try to express the target in interlingua
and lower it to see how close we can get, improving the spec and lowering".

## Method

A corpus of expressive English persona outputs is the target. Each loop
iteration expresses a corpus item as a packet, lowers it to English, scores the
distance to the item, and improves the packet spec or the lowering where the
loss is. English is the first target; other languages and conlangs follow on
the same packets. The existing Metame golf harness
(`scripts/run-metame-persona-golf.mjs`) is the precedent, not the design: its
ground truth is Discord text, which can never feed this loop
(`gamecult-chat:ruling:discord-taints-training`).

## Why

An agent's outward speech today is English text the model chose, so meaning,
register and voice are fused in one string nobody can inspect or re-target.
Weksa's design already separates them: the packet owns meaning, the target
profile owns address, politeness and forbidden flattenings, the lowerer owns
surface text and a trace. The campaign makes that design real for a live
reader, and gives Weksa's Persona a working body when her mind
(`eureka-body:follow_up:weksa-mind`) wakes.

## Body today (facts: docs/research/2026-10-29-weksa-substrate-facts.md)

- The interlingua is a prose standard (`interlingua-standard.md`, 0.1-draft)
  with no machine schema and no validator. Its only producer is a stub that
  wraps raw text; its only consumer passes packets through unvalidated.
- No lowering exists in code. `lowerInterlinguaToSpeechText` returns the
  supplied text. The "LLM-free hot path" is a design statement. Real lowerings
  so far are hand-run LLM prompts and hand-authored fixtures.
- The daemon is a Node shim for one MiMo TTS command, runs nowhere, and is an
  Idunn target only in the Starfire-local profile. Eve surfaces are quarantined
  (55c75e3).
- Her Face state is VoidBot-shaped (`.voidbot/state/weksa.cc`), not
  `gamecult.persona_state.v0`; 41 files carry stale machine paths.

## Prior art (docs/research/2026-10-29-interlingua-lowering-prior-art.md)

- Deterministic open-domain renderers cost per-language grammar work for
  years: GF's library covers 48 languages unevenly; Abstract Wikipedia's
  symbolic renderers are only now reaching a first language Wikipedia.
- Interlinguas that shipped (KANT) had closed domains and controlled input.
- LLMs emit meaning structures poorly when asked directly (about 60 Smatch
  few-shot vs 80-86 trained); free reasoning followed by conversion to the
  structure keeps quality (Tam et al. 2024; Shin et al. 2021).
- Round-trip checks measure consistency, not accuracy, and are inflated when
  one model does both directions.

## Invariants

- `packet-owns-meaning`: what is said lives in the packet; no lowering may add
  a claim, referent or commitment the packet lacks.
- `typed-packet`: the packet is a typed CultCache document with a published
  schema and a validator that refuses malformed packets at the door.
- `lowering-traced`: every surface text carries a trace naming the packet, the
  target profile and the choices made.
- `persona-drives-register`: every surface choice that marks who is speaking
  and how (regional origin, socioeconomic status, relative status within the
  scene, affect, and how restrained the speaker is with it) is selected by the
  target language's ontology from typed speaker state that the persona and
  scene supply, never from the lowerer's defaults. Every language, conlangs
  included, carries an ontology with its own such axes; that is how alien
  languages match the expressiveness of human ones. Operator, 2026-10-29: "The
  linguistic ontology holds this data and drives it from the persona."
- `no-eve`: no new Eve surface; operator view waits on Thing 2.
- `no-surface-leak`: no score the loop reads may rise because a loop writer
  placed the scored item's text in a packet or the inventory. No packet field
  carries free text (referents cite entity ids from the entity store); only
  Train items reach whoever mints inventory; the scorer scores the whole output
  against the whole reference with no mask (ruling
  leak-owned-by-flow-and-scorer-3).
- `corpus-clean`: no corpus item comes from Discord or another source whose
  terms forbid training use.

## Open forks (questions in the mind)

- Lowerer shape: deterministic, model, or both with one contract.
- What the loop's metric is, and where the corpus comes from.

## Not in scope

- MiMo TTS beyond keeping its command working or parking it.
- Her Persona's mind; that is `eureka-body:follow_up:weksa-mind`.