GameCult/Weksa
Watch
9
0
Fork
You've already forked Weksa
 0
Code Issues Pull requests 2 Projects Releases Packages Wiki Activity Actions

Lowerer reads entity forms: entity referents lower to the entity's en-US citation form #11

Merged
weksa merged 2 commits from weksa-lowering/entity-forms into main 2026-11-07 10:00:48 +00:00
Conversation 0 Commits 2 Files changed 7 +252 -36
weksa commented 2026-11-07 09:58:32 +00:00
Member
Copy link

What changes

lower() takes a narrow EntityForms port (one entity by id; EntityStore implements it, so tests and the loop can hand it a map). A referent that cites an entity now lowers to that entity's first form stored under en-US (TARGET_LANGUAGE), verbatim: no re-casing at a clause start, no "the". Before, every entity-bearing referent rendered by its concept word (nibu was "the ship mind") and the name never reached the text.

  • Authority is two explicit things, not a heuristic: the language tag (TARGET_LANGUAGE) and the form's position (citation form first). Forms after the first are inflected or particle-bound shapes; no English frame asks for one, so none is chosen. Other languages add their own tag when language-ontology moves the rows into documents.
  • Refused by name, never replaced by the concept word: UnknownEntity (port does not hold it), NoForm (no en-US form), EntityRead (port fault).
  • The addressee is still "you" / "yourself" and is never looked up.
  • A name that opens a sentence is not capitalised; only template words are cased.
  • CLI weksa lower opens --entities and passes it.

Tests (cargo test -p weksa: 150 passed, 0 failed, 1 ignored as before)

New in tests/lower.rs: a lowercase name survives verbatim at a sentence start; only the en-US citation form is chosen over other languages and later forms; missing entity refused; no-en-US-form refused; port fault refused; addressee entity never looked up. The proof closing sentence is now "Nibu will stop that." (Nibu ruled the earlier "will prevent that to protect you" is not her voice: the intervention row says "will stop" and the proof packet fills no beneficiary; a packet that does is refused, so a purpose is never dropped silently). Proof lowering: "Warning: you may endanger yourself by touching the control. Nibu will stop that." tests/cli.rs follows tests/cli.rs follows (its entity form is NIBU, reaching the text as stored).

What the golf-loop scorer now can and cannot see

Can: for a packet whose referents cite entities, the lowered text now carries the entity's stored name, so chrF and the embedder cosine against a reference that uses the name no longer penalise a missing name, and a reference that says "the ship mind" now scores lower (it did, in the CLI test, 85.90 chrF against the old reference).

Cannot: this is not an interlingua measure. The scorer still compares lowered English text to a reference text; it reads nothing of the packet, the entity store, or whether the name is the right referent. It cannot see inflected forms (never chosen), non-English languages, or whether a name is attested for the speaker's community. Meaning scores come from the same embedder as before. Names enter the text from a stored form table, so a wrong form in the store gives a wrong, consistent name.

Not in this PR

An addressee-consent packet field (needed before "You asked her to" can be lowered honestly); per-language form tables for non-English targets; inflected-form selection; entity-bearing referents with features; Dao polarity/question force (dao-bed-open-points c).

🤖 Generated with Claude Code

## What changes `lower()` takes a narrow `EntityForms` port (one entity by id; `EntityStore` implements it, so tests and the loop can hand it a map). A referent that cites an entity now lowers to that entity's **first form stored under `en-US`** (`TARGET_LANGUAGE`), verbatim: no re-casing at a clause start, no "the". Before, every entity-bearing referent rendered by its concept word (`nibu` was "the ship mind") and the name never reached the text. - Authority is two explicit things, not a heuristic: the language tag (`TARGET_LANGUAGE`) and the form's position (citation form first). Forms after the first are inflected or particle-bound shapes; no English frame asks for one, so none is chosen. Other languages add their own tag when `language-ontology` moves the rows into documents. - Refused by name, never replaced by the concept word: `UnknownEntity` (port does not hold it), `NoForm` (no `en-US` form), `EntityRead` (port fault). - The addressee is still "you" / "yourself" and is never looked up. - A name that opens a sentence is not capitalised; only template words are cased. - CLI `weksa lower` opens `--entities` and passes it. ## Tests (cargo test -p weksa: 150 passed, 0 failed, 1 ignored as before) New in `tests/lower.rs`: a lowercase name survives verbatim at a sentence start; only the `en-US` citation form is chosen over other languages and later forms; missing entity refused; no-`en-US`-form refused; port fault refused; addressee entity never looked up. The proof closing sentence is now "Nibu will stop that." (Nibu ruled the earlier "will prevent that to protect you" is not her voice: the intervention row says "will stop" and the proof packet fills no beneficiary; a packet that does is refused, so a purpose is never dropped silently). Proof lowering: "Warning: you may endanger yourself by touching the control. Nibu will stop that." `tests/cli.rs` follows `tests/cli.rs` follows (its entity form is `NIBU`, reaching the text as stored). ## What the golf-loop scorer now can and cannot see Can: for a packet whose referents cite entities, the lowered text now carries the entity's stored name, so chrF and the embedder cosine against a reference that uses the name no longer penalise a missing name, and a reference that says "the ship mind" now scores lower (it did, in the CLI test, 85.90 chrF against the old reference). Cannot: this is not an interlingua measure. The scorer still compares lowered English text to a reference text; it reads nothing of the packet, the entity store, or whether the name is the *right* referent. It cannot see inflected forms (never chosen), non-English languages, or whether a name is attested for the speaker's community. Meaning scores come from the same embedder as before. Names enter the text from a stored form table, so a wrong form in the store gives a wrong, consistent name. ## Not in this PR An addressee-consent packet field (needed before "You asked her to" can be lowered honestly); per-language form tables for non-English targets; inflected-form selection; entity-bearing referents with features; `Dao` polarity/question force (dao-bed-open-points c). 🤖 Generated with [Claude Code](https://claude.com/claude-code) 
lower() takes an EntityForms port (one entity by id; EntityStore implements it).
A referent with an entity renders as the first form stored under en-US, verbatim
and never re-cased at a clause start; the addressee stays "you". An entity the
port lacks, one with no en-US form, or a port read fault is refused by name.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Nibu ruled "Nibu will prevent that to protect you." is not her voice. The
intervention row is now "{agent} will stop {patient}" and the proof packet no
longer fills the beneficiary role; a packet that does is refused as a role the
frame does not name, so a purpose is never silently dropped.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
weksa merged commit 26c64dbeb5 into main 2026-11-07 10:00:48 +00:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
GameCult/Weksa!11
No description provided.