A causal compression protocol that trades human readability for machine reasoning density — with near-lossless semantic recovery.
git clone https://github.com/virionai/capsules.git && cp -r capsules/skills/pith ~/.claude/skills/pith
Pith is not a summarization tool. It is a lossy-to-humans, lossless-to-models compression layer.
Facts, causes, states, warnings
Pith keeps thisPleasantries, hedging, filler, padding
Pith drops thisArchetype grammar compression
Same compression, technical terms preserved exactly
Protected. Never altered. Density suggestions only.
A fixed pattern system covering every category of natural language found in technical work.
Real incident data. Real compression. Real validation.
| Incident | Pith tokens | Expanded tokens | Reduction |
|---|---|---|---|
| Axios supply chain (UNC1069) | 123 | 543 | 77% |
| LiteLLM supply chain (TeamPCP) | 174 | 597 | 71% |
Three independent models reconstructed full incident narratives from compressed chains alone — with zero new technical facts added.
All three independently characterized Pith as a domain-specific language.
Pith includes a multi-actor event ledger format for incident records, session provenance, and capsule event histories.
actor[HJ|human|2026-04-11].enter: context=pith-design
actor[CS46|model|2026-04-11].enter: context=pith-design
event[HJ]: caveman repo surfaced → token compression problem identified
event[CS46]: archetype grammar proposed as alternative to destructive compression
event[HJ+CS46]: prescriptive grammar adopted → Pith v0.1
event[CS46]: axios + litellm supply chain chains authored + benchmarked
event[HJ]: chains submitted for external validation
actor[CG54|model|2026-04-11].enter: context=chain-validation / prior-context=none
event[CG54]: axios chain → 0 new facts / elaboration only [intel-high]
event[CS46]: compression analysis → 77% axios / 71% litellm / 4.4x ctx multiplier
scope: session=pith-creation start=2026-04-11 last-edit=2026-04-11
actors: HJ=Josh | CS46=Claude-Sonnet-4.6 | CG54=ChatGPT-5.4Kern — code density skill with nested language-specific rules. Where Pith compresses language around code, Kern compresses the code itself.
Coming soon