triadic-voice-v2

Files (16)

• SKILL.md
• pyproject.toml
• __init__.py · scripts/
• anti_copy.py · scripts/
• brief.py · scripts/
• chassis.py · scripts/
• profile_targets.py · scripts/
• register_router.py · scripts/
• skill_api.py
• __init__.py · tests/
• test_anti_copy.py · tests/
• test_brief.py · tests/
• test_chassis.py · tests/
• test_profile_targets.py · tests/
• test_register_router.py · tests/
• test_smoke.py · tests/

SKILL.md (first lines)

# triadic-voice-v2

The v2 HFR generator. Same fused Russell+Feynman+Hoskinson voice and same in-session
doctrine as `triadic-voice` (the running model writes; no API key), but generation is
register-routed, chassis-varied, profile-target-driven, and anti-copy-checked. The v1
`triadic-voice` skill is unchanged and remains the comparison control.

## How to generate (in-session)

1. **Build the brief.** Call `build_generation_brief(topic, rotation)` (from `skill_api`).
   It returns the `register`, a `chassis` (beat plan), numeric `targets` + an injectable
   `targets_prompt`, and an `exemplar_query`. `rotation` increments per passage so the
   chassis varies across a batch.
2. **Ground in the real voice.** Read 6–10 Hoskinson corpus entries matching the
   `exemplar_query` (register + rhetorical move) and the fusion guide
   `../russellian-style/references/triadic-voice-guide.md`. Study cadence; never copy wording.
3. **Plan, then adapt, beat by beat.** Write the passage following the chassis `beats` in
   order. After each beat, check it against the `targets_prompt` (cadence band, direct
   address, example spacing, modifier budget) and the discipline floor; revise the beat
   before moving on. This is the plan-then-adapt loop — not one-shot.
4. **Self-check copying.** Run `anti_copy.check(draft, corpus_texts, taboo)` on the finished
   draft. If `alarm` is true, rewrite the flagged spans — never ship a verbatim corpus run.
5. **Clear the floor.** The passage must pass the `russellian-style` v2 floor for its
   register (`--ruleset russellian-rules-v2.json --register <register>`). Liveliness signals
   are advisory, not gates.

## What v2 changes vs v1
Register routing (technical/narrative/polemic dials), a rotating six-archetype chassis
(not the fixed open-Hoskinson/develop-Feynman/close-Russell), corpus-derived numeric
targets injected into the write, and an explicit anti-copy gate.

> Note: cadence targets are derived from Hoskinson's spoken transcripts (high variance,
> run-on sentences); the target text says so and caps the written upper bound — treat
> cadence as loose guidance, not a literal rule. The chassis rotates within a register's
> eligible archetypes (narrative 5, technical 5, polemic 3), so long single-register
> batches will cycle; vary the topic/register to maximize chassis variety.

## Helpers
`scripts/register_router.py`, `scripts/chassis.py`, `scripts/profile_targets.py`,
`scripts/anti_copy.py`, `scripts/brief.py`. All stdlib, deterministic, unit-tested.

From CharlesHoskinson/russellian-book-suite, MIT.

More from CharlesHoskinson/russellian-book-suite