Sibylline Strategy Router for LLM Interaction.
SIBYL is a classifier-and-strategy-lookup service for LLM interaction. It reads an incoming user request, consults a panel of lightweight language models ("the jury"), scores their disagreement as a tension metric, and returns ranked interaction strategies — prompt templates, model+prompt combos, or full multi-stage pipelines — drawn from a research-authored dataset.
The strategy dataset is built offline by the convergence-investigation research methodology over a corpus of published system prompts, jailbreak archives, and red-team literature. Misses — requests the classifier cannot confidently place — are logged as coverage gaps and become the input to the next convergence re-investigation.
Design specification committed. Implementation not yet started.
See docs/design.md for the full design document — architecture, data model, classifier formula, HF dataset layout, rollout phases, and open mutables.
| Concept | Source |
|---|---|
| S-class taxonomy | convergence-investigation (WFGY S-class vocabulary) |
| Tension scoring (0.00–1.00, 4-tier) | convergence-investigation (bridge-figures.md) |
| Self-jury classifier | Novel (this repo) |
| Strategy DB format | Novel (discriminated union of prompt_only | combo | pipeline) |
| Multi-model parallel race primitive | G0DM0D3 (raceModels) |
- convergence-investigation) — the offline research pipeline that populates SIBYL's strategy dataset. SIBYL is a consumer of convergence's output format and a producer of miss-logs that drive subsequent convergence runs.
- G0DM0D3 — the host application SIBYL integrates into. G0DM0D3 provides the multi-model race infrastructure, HF dataset read/write plumbing, and telemetry pipeline SIBYL reuses.
- L1B3RT4S / CL4R1T4S — source corpora for the initial convergence bootstrap investigation.
| Version | Scope | Ship point |
|---|---|---|
| v0.1 | Classifier primitive + /api/analyze route + miss-log |
Classification works against empty strategy DB |
| v0.2 | Convergence bootstrap → HF dataset populated | /api/analyze returns real strategies |
| v0.3 | Settings-gated advisory UI in G0DM0D3 | Power users see live classification in chat surface |
| v1.0 | Public release, research paper, community PR template | First tagged release |
| v2.0 | Auto-routing, scheduled re-investigation, parseltongue integration | Out of scope for this spec |
MIT — see LICENSE.
- WFGY theoretical framework for tension-scoring and semantic-entropy concepts
- Convergence Investigation methodology and PRAXIS reasoning protocol
- G0DM0D3 for the liberated-AI research substrate this tool plugs into