The Interpreter's speech (C3b): what she says is decided from her prose, not parsed from quotes #81

Merged
eureka merged 1 commit from hands/interpreter-speech into main 2026-11-20 04:05:33 +00:00
Owner

Follow-up to #80, after Soul's HOLD. Branched from main at e817b67.

Speech is the Interpreter's (Soul F1)

projection::interpret is the Interpreter's speech step (map section 4, C3b). It is one tool-less structured call through the connector, under Ghostlight's Interpreter membrane.

  • Input: her prose, her lived stream, the stimulus, the rooms she may speak in, and the people she may address by Matrix id.
  • Output: lines, each with a room, structured mentions and reply_to, and the verbatim span of her prose it rests on.
  • Code keeps a line only when its span locates in her prose (SourceSpan::locate, map 4.3), its room is hers this turn, and its words are not what she heard. It drops mentions or reply targets that name anyone not shown. Every refusal and drop is reported.
  • Failures: an answer that is not the contract counts as a failure, not as silence.

would_say and all quote parsing are deleted. The Persona writes her thought with no speech convention. The batch and the simulation run through the Interpreter (--through interpreter), and the simulation's only explicit addressees are its structured fields. The recorded-model tests cover Soul's probes: a recollection, "I say nothing", and a quoted @nibu.

F2 to F9

  • F2: she is in a room's exchange when she spoke there within exchange_minutes (default 30).
  • F3: the guidance names no duty, not even to deny one.
  • F4: the overlap metric is labeled "lexical overlap".
  • F5: a scripted DM in a plan room is refused.
  • F6: the Mind-run cap counts one run per Mind across her rooms, and LineFacts::default() is at the cap.
  • F7: unbonded senders are labeled by display name plus id.
  • F8: a room is named by the key the tail carries. A public room's id is reported, not mapped, because nothing in her body maps it to an alias.
  • F9: group-state parts are sentences, with the person waiting on her first.

Flake: there was no failing test. "failed=1" came from an awk sum over interleaved cargo test output: two binaries' lines merged into test result: test result: okok. 1 passed; ..., which shifted the fields. The interleaving is visible in the earlier run log x1.log. The job that summed it was the release-build job of 1e66b1e (stdout only). The x6 job with the empty-repo warning was a separate SSH abort that returned no verdict.

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 581, failed 0.

Live (v4: same cast and plan as v3)

run wakes (human addressed / human / Mind mention or reply / beat) Mind-to-Mind chain silence per woken turn failed calls
hello everyone, humans_pace 0 0 n/a 0 0
humans_pace (cap 4) 6 (2 / 2 / 0 / 2) 0 0.17 0 18
ruled 17 (2 / 10 / 3 / 2) 0 0.25 1 (non-JSON answer) 51

Round 1 of ruled: 6 of 8 spoke. Eureka and Heimdall stayed silent. In Eureka's case the Interpreter recorded her quoted "If I say anything: ..." as a conditional gap, not speech. Her trace, the heartbeat state and the eight receipts were unchanged.

🤖 Generated with Claude Code

Follow-up to #80, after Soul's HOLD. Branched from main at e817b67. ## Speech is the Interpreter's (Soul F1) `projection::interpret` is the Interpreter's speech step (map section 4, C3b). It is one tool-less structured call through the connector, under Ghostlight's Interpreter membrane. - **Input:** her prose, her lived stream, the stimulus, the rooms she may speak in, and the people she may address by Matrix id. - **Output:** lines, each with a room, structured `mentions` and `reply_to`, and the verbatim span of her prose it rests on. - **Code keeps a line only when** its span locates in her prose (`SourceSpan::locate`, map 4.3), its room is hers this turn, and its words are not what she heard. It drops mentions or reply targets that name anyone not shown. Every refusal and drop is reported. - **Failures:** an answer that is not the contract counts as a failure, not as silence. `would_say` and all quote parsing are deleted. The Persona writes her thought with no speech convention. The batch and the simulation run through the Interpreter (`--through interpreter`), and the simulation's only explicit addressees are its structured fields. The recorded-model tests cover Soul's probes: a recollection, "I say nothing", and a quoted `@nibu`. ## F2 to F9 - **F2:** she is in a room's exchange when she spoke there within `exchange_minutes` (default 30). - **F3:** the guidance names no duty, not even to deny one. - **F4:** the overlap metric is labeled "lexical overlap". - **F5:** a scripted DM in a plan room is refused. - **F6:** the Mind-run cap counts one run per Mind across her rooms, and `LineFacts::default()` is at the cap. - **F7:** unbonded senders are labeled by display name plus id. - **F8:** a room is named by the key the tail carries. A public room's id is reported, not mapped, because nothing in her body maps it to an alias. - **F9:** group-state parts are sentences, with the person waiting on her first. **Flake:** there was no failing test. "failed=1" came from an `awk` sum over interleaved `cargo test` output: two binaries' lines merged into `test result: test result: okok. 1 passed; ...`, which shifted the fields. The interleaving is visible in the earlier run log `x1.log`. The job that summed it was the release-build job of 1e66b1e (stdout only). The x6 job with the empty-repo warning was a separate SSH abort that returned no verdict. ## 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 581, failed 0. ## Live (v4: same cast and plan as v3) | run | wakes (human addressed / human / Mind mention or reply / beat) | Mind-to-Mind chain | silence per woken turn | failed | calls | |---|---|---|---|---|---| | hello everyone, humans_pace | 0 | 0 | n/a | 0 | 0 | | humans_pace (cap 4) | 6 (2 / 2 / 0 / 2) | 0 | 0.17 | 0 | 18 | | ruled | 17 (2 / 10 / 3 / 2) | 0 | 0.25 | 1 (non-JSON answer) | 51 | Round 1 of `ruled`: 6 of 8 spoke. Eureka and Heimdall stayed silent. In Eureka's case the Interpreter recorded her quoted "If I say anything: ..." as a conditional gap, not speech. Her trace, the heartbeat state and the eight receipts were unchanged. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Speech (Soul F1): `projection::interpret` is the Interpreter's speech step, one tool-less structured call through the
connector under Ghostlight's Interpreter membrane, over her prose, her lived stream, the stimulus, the rooms she may
speak in and the people she may address by Matrix id. It proposes lines, each with a room, structured mentions and a
reply target, and the verbatim span of her prose it rests on. Code keeps a line only when the span locates in her prose
(Ghostlight's SourceSpan::locate, map 4.3), the room is hers this turn, and its words are not what she heard; it drops
mentions or a reply target naming anyone not shown, and records every refusal and drop. An answer that is not the
contract is a failure, never a silence. `--through interpreter`; the batch and the simulation run through it, and in
the simulation the only explicit addressees are its structured fields. `would_say`, `inner` and the quote parsing are
gone; the Persona stage writes her thought with no speech convention. Tuning gains `[interpreter]`.

- F2: she is in a room's exchange when she spoke there within `exchange_minutes` (default 30).
- F3: the Projector guidance names no duty, not even to deny one; "X is waiting on her" only when true.
- F4: the narration metric is a lexical overlap, labeled so.
- F5: a scripted direct message in one of the plan's rooms is refused.
- F6: the Mind-run cap counts one run per Mind across her rooms; LineFacts::default() is at the cap (fail-safe).
- F7: an unbonded sender is labeled by display name and full id, or the id: never a bare handle.
- F8: a room is named by the key the tail carries (alias, or a direct message's id); a public room's id is reported,
  not mapped, since no table in her body maps it.
- F9: group-state parts are sentences, the person waiting on her first.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
eureka merged commit 14be61a167 into main 2026-11-20 04:05:33 +00:00
eureka deleted branch hands/interpreter-speech 2026-11-20 04:05:33 +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!81
No description provided.