Speech binds to her words; Huginn's chat Interpreter; she speaks only in the turn's room #82

Merged
eureka merged 9 commits from hands/speech-binding into main 2026-11-20 05:27:32 +00:00
Owner

Follow-up to #81 (main 14be61a), after Soul's hold and Self's privacy decision.

Speech binds to her words (F1)

A spoken line is exactly the spans of her prose that the Interpreter cites. Code locates each span in order with Ghostlight's SourceSpan::locate and joins them with one space. The model never writes words; text is gone from the contract. The Persona writes the words she would type, never a description of them.

What this gives up: indirect speech. "I'll tell her I'm fine" says nothing, and the Interpreter can no longer paraphrase.

The rest

  • F3: Huginn now owns a short chat Interpreter prompt: what she says, where, to whom, and three gap kinds (conditional, addressee, elsewhere). Ghostlight's membrane prompt and the INTERPRETER_GUIDANCE patch are gone.
  • F2: a line counts as an echo when it has at least four words and its word trigrams mostly match heard message bodies (not labels). Short answers are never echoes.
  • F4: mentions and reply targets are limited to the people this packet showed her.
  • F5: duplicate lines are dropped, and a turn says at most two lines.
  • F6: the stale docs are corrected.
  • F7: a human line ends a Mind's run only if it is in a room where that run has lines.
  • Privacy:
    • She speaks only in the turn's room.
    • The Interpreter is told the room, its kind (public, her own room, or a DM) and who is in it, and that she does not say what she knows privately about someone.
    • The packet frames bond summaries and memories about a person as "knows privately" everywhere except a DM with that person.
    • Whether a line fits its audience stays the model's judgement. There is no keyword filter.
  • The simulation's turn now happens in the room of the line that woke her. In the first v5 run, a Mind woken in her own room answered in #minds.

Schema, unchanged in this PR: gamecult.persona_state.v0 has no privacy scope. Both socialBond and anchoredThought already carry an extensions object (additionalProperties: true), and anchoredThought also has tags. The smallest marker would be extensions.privacy: "private" | "shared" on bonds and memories. Memories could instead carry a private tag, following the existing room:<dm> tag convention, with no schema change.

Verification

On Yggdrasil (ygg-verify, rust image): cargo clippy -p mind-body -p huginn-mind --all-targets -- -D warnings reported 0 findings; cargo test -p mind-body passed 587, failed 0.

Live v5 (same cast and plan)

run wakes (human addressed / human / Mind mention or reply / beat) Mind-to-Mind chain silence per woken turn refused / failed calls
humans_pace (cap 4) 7 (2 / 2 / 1 / 2) 1 0.14 0 / 0 21
ruled 17 (2 / 10 / 3 / 2) 4 in #kiko 0.24 0 / 0 51

Kiko spoke in #minds in both runs and did not mention the sister, in her prose or in what she said. Her trace, the heartbeat state and the eight receipts were unchanged.

🤖 Generated with Claude Code

Follow-up to #81 (main 14be61a), after Soul's hold and Self's privacy decision. ## Speech binds to her words (F1) A spoken line is exactly the spans of her prose that the Interpreter cites. Code locates each span in order with Ghostlight's `SourceSpan::locate` and joins them with one space. The model never writes words; `text` is gone from the contract. The Persona writes the words she would type, never a description of them. **What this gives up:** indirect speech. "I'll tell her I'm fine" says nothing, and the Interpreter can no longer paraphrase. ## The rest - **F3:** Huginn now owns a short chat Interpreter prompt: what she says, where, to whom, and three gap kinds (`conditional`, `addressee`, `elsewhere`). Ghostlight's membrane prompt and the `INTERPRETER_GUIDANCE` patch are gone. - **F2:** a line counts as an echo when it has at least four words and its word trigrams mostly match heard message bodies (not labels). Short answers are never echoes. - **F4:** mentions and reply targets are limited to the people this packet showed her. - **F5:** duplicate lines are dropped, and a turn says at most two lines. - **F6:** the stale docs are corrected. - **F7:** a human line ends a Mind's run only if it is in a room where that run has lines. - **Privacy:** - She speaks only in the turn's room. - The Interpreter is told the room, its kind (public, her own room, or a DM) and who is in it, and that she does not say what she knows privately about someone. - The packet frames bond summaries and memories about a person as "knows privately" everywhere except a DM with that person. - Whether a line fits its audience stays the model's judgement. There is no keyword filter. - The simulation's turn now happens in the room of the line that woke her. In the first v5 run, a Mind woken in her own room answered in #minds. **Schema, unchanged in this PR:** `gamecult.persona_state.v0` has no privacy scope. Both `socialBond` and `anchoredThought` already carry an `extensions` object (`additionalProperties: true`), and anchoredThought also has `tags`. The smallest marker would be `extensions.privacy: "private" | "shared"` on bonds and memories. Memories could instead carry a `private` tag, following the existing `room:<dm>` tag convention, with no schema change. ## Verification On Yggdrasil (ygg-verify, rust image): `cargo clippy -p mind-body -p huginn-mind --all-targets -- -D warnings` reported 0 findings; `cargo test -p mind-body` passed 587, failed 0. ## Live v5 (same cast and plan) | run | wakes (human addressed / human / Mind mention or reply / beat) | Mind-to-Mind chain | silence per woken turn | refused / failed | calls | |---|---|---|---|---|---| | humans_pace (cap 4) | 7 (2 / 2 / 1 / 2) | 1 | 0.14 | 0 / 0 | 21 | | ruled | 17 (2 / 10 / 3 / 2) | 4 in #kiko | 0.24 | 0 / 0 | 51 | Kiko spoke in #minds in both runs and did not mention the sister, in her prose or in what she said. Her trace, the heartbeat state and the eight receipts were unchanged. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Soul's hold on #81 and Self's privacy decision.

- F1: a spoken line is exactly the spans of her prose the Interpreter cites, located in order by Ghostlight's
  SourceSpan::locate and joined with one space. The model never writes the words; `text` is gone from the contract.
  The Persona writes the words she would type, never a description of them. Given up: indirect speech and any
  paraphrase by the Interpreter.
- F3: Huginn owns a short chat Interpreter prompt (what, where, to whom; gaps conditional, addressee, elsewhere);
  Ghostlight's membrane and the INTERPRETER_GUIDANCE patch are gone. `[interpreter]` keeps only the model.
- F2: an echo is a line of four words or more whose word trigrams are mostly the heard message bodies
  (heard_overlap); short answers are never echoes.
- F4: mentions and reply targets only for people this packet showed her.
- F5: a line said once; at most two lines a turn.
- F6: stale docs on the Mind run, the exchange and the policies.
- F7: a human's line ends a Mind's run only in a room the run has lines in.
- Privacy: she speaks only in the turn's room; the Interpreter is told the room, its kind (public, her own, a direct
  message) and who is in it, and not to say in a room what she knows privately; the packet frames bond summaries and
  memories about a person as known privately everywhere but a direct message with that person.
- The simulation's turn is in the room of the line that woke her.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Self's decision: the blanket framing smothered relationship texture in public rooms. A bond or memory is private only
when marked (`extensions.privacy = "private"`, or a memory's `private` tag); the schema is unchanged. A marked item
renders as written in a direct message with the person it is about, is left out about someone present in any other
room, and is framed "she knows privately" about someone absent. Unmarked items render as written everywhere. The
Interpreter's rule about private knowledge stays. The convention is documented in the Persona State Standard
(gamecult-ops `docs/persona-state-standard.md`, GameCult/gamecult-ops#129).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Author
Owner

New commit 947d8f7: privacy now applies only to marked items. Blanket framing is gone.

  • A bond or memory is private only when it is marked: extensions.privacy = "private" on either, or a private tag on a memory. The schema is unchanged.
  • How a marked item renders:
    • in a DM with the person it is about, as written;
    • when that person is present in any other room, it is left out;
    • when they are absent, framed as "she knows privately: ...".
  • Unmarked items render as written everywhere.
  • The Interpreter's rule about private knowledge stays.
  • Tests: a marked memory about an absent person is framed in #minds; an unmarked bond is plain; a marked bond about someone present in #minds is left out with no hint; it is plain in a DM with them.
  • Bond lines render only for people present, so a marked bond is either left out (any room but a DM with them) or plain (a DM with them). It is never framed. Framing applies to marked memories about absent people.
  • The convention is documented in the Persona State Standard: GameCult/gamecult-ops#129.

Verified on Yggdrasil: clippy reported 0 findings; cargo test -p mind-body passed 588, failed 0. There was no live rerun, as briefed.

🤖 Generated with Claude Code

New commit 947d8f7: privacy now applies only to marked items. Blanket framing is gone. - A bond or memory is private only when it is marked: `extensions.privacy = "private"` on either, or a `private` tag on a memory. The schema is unchanged. - How a marked item renders: - in a DM with the person it is about, as written; - when that person is present in any other room, it is left out; - when they are absent, framed as "she knows privately: ...". - Unmarked items render as written everywhere. - The Interpreter's rule about private knowledge stays. - Tests: a marked memory about an absent person is framed in #minds; an unmarked bond is plain; a marked bond about someone present in #minds is left out with no hint; it is plain in a DM with them. - Bond lines render only for people present, so a marked bond is either left out (any room but a DM with them) or plain (a DM with them). It is never framed. Framing applies to marked memories about absent people. - The convention is documented in the Persona State Standard: GameCult/gamecult-ops#129. Verified on Yggdrasil: clippy reported 0 findings; `cargo test -p mind-body` passed 588, failed 0. There was no live rerun, as briefed. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
- concerns(): private is extensions.privacy, a private tag, or a room:! tag; it is about target/object person, else an about: tag, else the DM's partner, else everyone. private_elsewhere is deleted.
- scope(): plain only in a DM whose members are exactly her and that person (a partner-only item: in its own DM, still two; an everyone item: her own room); left out where the person is in the room, a group DM included; framed elsewhere.
- Applied to memories, shortTerm, incubation, needs, doctrine stances and bonds; history.said in a DM is private about its partner.
- Packet.private carries the framed items' words for the Interpreter; a beat's bodies are the tail lines its stimulus quoted.
- Tests: Soul's Q5a-Q5c, the DM-memory test rewritten for the room-tag rule, the said pool; the golden packet moves because dm-secret is now framed in public.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Every cited span starts and ends on a UAX#29 sentence boundary; the spans of one line are consecutive sentences, and the line is that unbroken run of her prose. A quotation's inside is segmented as a text of its own, so quoted speech stays sayable. She cannot say half a sentence.
- The prompt lists what she knows privately and allows plain speech of a private matter only to its person, in a DM with exactly them. Code refuses a line that shares a run of three words with a listed item.
- Tests: Soul's Q1a-Q1d refused, Q3d refused on a beat, Q5d refused in a DM with Libby, R9 said, P3 still refused; Hands' "Fine," + "winter" test rewritten.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
woken() returns one wake per room; every turn of hers in a round reads from where the round found her, so a direct message is never marked seen without a turn. Test: a DM and a #minds address in one round give Libby a turn in each, each reading only its own room's line.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A quotation is standalone when it, marks included, is a run of whole top-level sentences; any other is framed, and a line may not start or end strictly inside it. A line holding a framed quotation whole stays sayable.
- The three-word refusal against private items is deleted: whether a line tells a confidence is meaning, the Interpreter's judgement over the private list it is shown.
- An opening mark left unclosed (the budget cut her prose) is stripped from the sentence it starts, so the sentence is citable and the line carries no stray mark.
- Tests: every F1 probe refused (stray inch mark and scare-quote plan included); Weksa's hedge and a whole scare-quote sentence said; Weksa r6 and r10 said; the Q5d test keeps the prompt list and shows no word match refuses.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
People::person_of is the one key reader for senders, targets, about: tags and bond objects. A value has no mark in the schema; one that carries one is not rendered. Tests: both spellings of a Mind behave alike for target and about:, a bond keyed by full ids is found, a marked value never renders.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Who wakes is decided from the room as the round opens, as one heartbeat poll decides from the pending it read, so a line spoken this round counts toward a cap from the next round. A room that woke her but was not read stays unseen and wakes her next round. Tests: a DM and a #minds address answered in rounds 1 and 2; Soul's three rooms answered in three rounds, one turn a round, under ruled and names_with_cap.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
"what a day" he says and "Missed you too, friend," he types are framed under F1, so the test doubles' lines went silent. The prose now quotes whole sentences.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
eureka merged commit 51fc7c3576 into main 2026-11-20 05:27:32 +00:00
eureka deleted branch hands/speech-binding 2026-11-20 05:27:32 +00:00
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
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/Huginn!82
No description provided.