mind-persona project: the persona seed harness (--room, --through persona, --batch, --simulate) #79

Merged
eureka merged 2 commits from hands/persona-harness into main 2026-11-20 01:53:34 +00:00
Owner

Built on #76 (C3a), now merged; rebased onto main at c4bd83b.

The persona seed harness, built inside mind-persona project so experiments and the live turn share one code path: every case is one projection::project call.

What it adds

  • --room <fixture.toml>: a mocked room in place of the mouth and room tail (her id, rooms, messages with sender, display name, room, ago, body and new/tag/mentions/dm/follow_up/reply_to; a mocked now; a last_heard table). The fixture is a MouthReader, so whether a message is for her is still address::is_for. The old live room filter is now --in <room>.
  • --state is validated as a put validates content: huginn-mind's P1 to P3 moved into one owner, persona::validate_document, which the put now calls.
  • --through persona: the Persona stage (C3b's first piece). One tool-less connector call under Ghostlight's build_persona_prompt over identity, the narration and the verbatim heard lines; prose capped at 180 words; would-say lines are the double-quoted spans (a preview of the Interpreter's speech, not its decision); usage per call. persona-tuning.toml gains [persona] (model, guidance, words).
  • --batch <plan> --out <report.md>: states x scenarios x samples, one Markdown report.
  • --simulate <plan> --out <report.md>: up to eight personas in one room. Wake is is_for plus ruling persona-only-wake-humans-in-rooms; wake = "all_messages" is the old loop as a control. State does not evolve: state_after_turn is the seam for C3b/C3c, and the report header says so.
  • --simulate reports wakes per round split into human-caused and Mind-addressed, and Mind-addressed evidence by kind (Matrix mention or reply, id written as text, name in text), listing each id-as-text wake.
  • Fixtures under crates/mind-body/fixtures/harness/ (four scenario rooms and a control character).

Choices where the map is silent

  • Persona guidance: one setting line, "words she says aloud ... appear in her prose inside double quotation marks". It licenses nothing; Ghostlight's membrane already says a thin moment gets little or nothing.
  • Would-say: straight or curly double-quoted spans; single quotes are not speech (apostrophes). A remembered quotation reads as said until the Interpreter decides.
  • Lived stream for the Persona: narration, a blank line, then the heard lines.
  • Request ids carry the wall clock, because a fixture's now is the same on every run and the kit replays a repeated id.

Verification

  • Yggdrasil (ygg-verify, rust image): cargo clippy -p mind-body -p huginn-mind --all-targets -- -D warnings clean; cargo test -p mind-body and cargo test -p huginn-mind green.
  • Read-only: tests/persona_bin.rs runs project --room --state with env cleared and the daemon at a dead port, and with no endpoint at all.

Live results: see the comment below.

🤖 Generated with Claude Code

Built on #76 (C3a), now merged; rebased onto main at c4bd83b. The persona seed harness, built inside `mind-persona project` so experiments and the live turn share one code path: every case is one `projection::project` call. ## What it adds - `--room <fixture.toml>`: a mocked room in place of the mouth and room tail (her id, rooms, messages with sender, display name, room, `ago`, body and `new`/`tag`/`mentions`/`dm`/`follow_up`/`reply_to`; a mocked `now`; a `last_heard` table). The fixture is a `MouthReader`, so whether a message is for her is still `address::is_for`. The old live room filter is now `--in <room>`. - `--state` is validated as a put validates content: huginn-mind's P1 to P3 moved into one owner, `persona::validate_document`, which the put now calls. - `--through persona`: the Persona stage (C3b's first piece). One tool-less connector call under Ghostlight's `build_persona_prompt` over identity, the narration and the verbatim heard lines; prose capped at 180 words; would-say lines are the double-quoted spans (a preview of the Interpreter's speech, not its decision); usage per call. `persona-tuning.toml` gains `[persona]` (model, guidance, words). - `--batch <plan> --out <report.md>`: states x scenarios x samples, one Markdown report. - `--simulate <plan> --out <report.md>`: up to eight personas in one room. Wake is `is_for` plus ruling `persona-only-wake-humans-in-rooms`; `wake = "all_messages"` is the old loop as a control. State does not evolve: `state_after_turn` is the seam for C3b/C3c, and the report header says so. - `--simulate` reports wakes per round split into human-caused and Mind-addressed, and Mind-addressed evidence by kind (Matrix mention or reply, id written as text, name in text), listing each id-as-text wake. - Fixtures under `crates/mind-body/fixtures/harness/` (four scenario rooms and a control character). ## Choices where the map is silent - Persona guidance: one setting line, "words she says aloud ... appear in her prose inside double quotation marks". It licenses nothing; Ghostlight's membrane already says a thin moment gets little or nothing. - Would-say: straight or curly double-quoted spans; single quotes are not speech (apostrophes). A remembered quotation reads as said until the Interpreter decides. - Lived stream for the Persona: narration, a blank line, then the heard lines. - Request ids carry the wall clock, because a fixture's `now` is the same on every run and the kit replays a repeated id. ## Verification - Yggdrasil (ygg-verify, rust image): `cargo clippy -p mind-body -p huginn-mind --all-targets -- -D warnings` clean; `cargo test -p mind-body` and `cargo test -p huginn-mind` green. - Read-only: `tests/persona_bin.rs` runs `project --room --state` with env cleared and the daemon at a dead port, and with no endpoint at all. Live results: see the comment below. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Recall now reads only memories of the cued sorts: follow-ups, rulings and the
open-item filter leave it, with the tests that pinned them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
F1/F2: the Projector shows her persona state and her room; recall.rs and the
recall pool, recall_sorts and its cap are deleted. Reading the Mind is the
Interpreter's (C3b).
F3: messages to her go in whole and first; others fill what is left, with an
exact note; overflow past the cap is reported.
F4: candidates are included with probability 1 - exp(-score / inclusion_score);
caps are maxima. F5: the hash is of the effective tuning. F6: the guard reads
the persona schema and the tuning keys. F7: bounded fixed slots and relatedness
cue; the redundant final clip is gone. F8: tuning values are validated.
F9: project's organs are read-only traits. F10: clips land on graphemes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
S1: to_her is address::is_for over the heartbeat's addressing row (own names
when there is none); a name inside a path or identifier no longer addresses
her, in address.rs itself. S2: room talk is headed Also in the room. S3:
addressed_bytes (8 KiB) bounds the messages to her, newest first; one larger
message is clipped between graphemes with its remainder counted. S4: one
escape (spoken) renders every body. S5: only short-term items decay; memories,
values and stances come forward by lexical relevance to what the room said;
the lens is drawn from included items. Lows: people keyed by full id except
the operator and the Minds; unbonded speakers shown with their id; tuning
refuses an empty model, a temperature outside 0.01..10, all-zero caps and an
overflowing identity; the guard folds kebab and run-together forms; a beat
respects small caps; project refuses without MIND_PERSONA_SELF_SENDER.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The harness judges candidate persona states before any Mind holds them, through the
same `project` the live turn will use: every case is one `project` call.

- `--room <fixture.toml>`: a mocked room in place of the mouth and the room tail:
  her id, rooms, messages (sender, display name, room, ago, body; new, tag,
  mentions, dm, follow_up, reply_to), a mocked now and a "last heard from" table.
  The fixture is a MouthReader, so addressing is still `address::is_for`. The live
  room filter that `--room` used to name is now `--in`.
- `--state` is checked as a put checks content: huginn-mind's P1 to P3 now live in
  one owner, `persona::validate_document`, which the put calls.
- `--through persona`: the Persona stage (C3b's first piece). One tool-less
  connector call under Ghostlight's persona membrane over identity, narration and
  the heard lines; prose capped at 180 words; would-say lines are the double-quoted
  spans; usage per call. Tuning gains `[persona]` (model, guidance, words).
- `--batch <plan> --out <md>`: states x scenarios x samples, one Markdown report
  with packets, narration, prose, would-say lines, word counts, distinct-2 and
  cross-state word similarity, and a total-usage footer.
- `--simulate <plan> --out <md>`: up to eight personas taking turns in one room.
  Wake is is_for plus the ruled set (addressed, any human in her rooms, the beat);
  `wake = "all_messages"` is the old loop as a control. Woken personas run in a
  seeded order; quoted lines join the room as from her id. State does not evolve
  (`state_after_turn` is the seam for C3b/C3c). Metrics: messages, silence,
  who addresses whom, Mind-to-Mind chains, lexical convergence, a work-talk
  heuristic, usage.
- With a state file and a fixture no daemon and no mouth is needed (bin test).
- Request ids take the wall clock, so a fixture's fixed now cannot replay.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
eureka force-pushed hands/persona-harness from 54f29daafe to 70d0929606 2026-11-19 23:43:45 +00:00 Compare
eureka force-pushed hands/persona-harness from 70d0929606 to 19e0ee5285 2026-11-20 01:09:47 +00:00 Compare
Author
Owner

Live results (Yggdrasil, as mind-eureka, through the connector; default tuning; Sonnet 5.5 for Projector and Persona)

Reports: F:\Projects\swarm-experiment-2026-11\work\harness\ (batch-2026-11-20.md, sim-eureka-old-2026-11-20.md, sim-eureka-draft-2026-11-20.md); on Yggdrasil under /var/tmp/persona-harness/out/.

Batch (3 states x 4 scenarios x 2 samples): 48 calls, 10,713 output and 54,288 cache-write tokens. Spoke: stored 6/8 (silent on the quiet beat), Life's draft 7/8, control 8/8. All three states answered Deru's unaddressed aside 2/2.

Simulation (8 personas, #minds, 12 rounds, 2 repeats, ruled wake set, operator opens warmly): Eureka as stored took 228 calls; Eureka as Life's draft took 224. In all four repeats, round 1 is 8 human-caused wakes, and every wake in rounds 2 to 12 is Mind-addressed (2 to 7 per round). Mind-addressed evidence: 365 lines in total; 0 Matrix mentions or replies, 0 ids written as text, 365 names in text. Mind-to-Mind chains ran 44 to 62 lines with no human between them. Silence rates were 3% to 12%.

Read-only: her trace, the heartbeat state and all eight persona receipts were unchanged before and after. Five mouth cursors (eureka, heimdall, kiko, libby, nibu) were rewritten together at 01:44:43 UTC by the mouth's own process (user bifrost). The same five had also changed together at 23:22, before any run. The harness has no mouth path in these runs.

🤖 Generated with Claude Code

## Live results (Yggdrasil, as mind-eureka, through the connector; default tuning; Sonnet 5.5 for Projector and Persona) Reports: `F:\Projects\swarm-experiment-2026-11\work\harness\` (`batch-2026-11-20.md`, `sim-eureka-old-2026-11-20.md`, `sim-eureka-draft-2026-11-20.md`); on Yggdrasil under `/var/tmp/persona-harness/out/`. **Batch** (3 states x 4 scenarios x 2 samples): 48 calls, 10,713 output and 54,288 cache-write tokens. Spoke: stored 6/8 (silent on the quiet beat), Life's draft 7/8, control 8/8. All three states answered Deru's unaddressed aside 2/2. **Simulation** (8 personas, #minds, 12 rounds, 2 repeats, ruled wake set, operator opens warmly): Eureka as stored took 228 calls; Eureka as Life's draft took 224. In all four repeats, round 1 is 8 human-caused wakes, and **every wake in rounds 2 to 12 is Mind-addressed** (2 to 7 per round). Mind-addressed evidence: 365 lines in total; 0 Matrix mentions or replies, **0 ids written as text**, 365 names in text. Mind-to-Mind chains ran 44 to 62 lines with no human between them. Silence rates were 3% to 12%. **Read-only:** her trace, the heartbeat state and all eight persona receipts were unchanged before and after. Five mouth cursors (eureka, heimdall, kiko, libby, nibu) were rewritten together at 01:44:43 UTC by the mouth's own process (user bifrost). The same five had also changed together at 23:22, before any run. The harness has no mouth path in these runs. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
eureka merged commit fe723867f7 into main 2026-11-20 01:53:34 +00:00
eureka deleted branch hands/persona-harness 2026-11-20 01:53:35 +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!79
No description provided.