episteme-compaction git:20261018.b2b1b02A

Immutable. This exact content never changes and is served at /api/v1/blob/ef2e1c389ed521eb.

---
name: episteme-compaction
description: |
  Lossless compaction of epistemes: prose ↔ compact typed-slot notation (F4-F5;
  ~65% median token savings, measured on the instruction-file corpus) with a
  source-claims vs [bracketed] split and reconstruction back to source. Use for
  AI-agent or FPF-literate-human output needing token economy and auditability.
  NOT narrativization. Depends on fpf-core.
---

# Episteme Compaction — Dispatcher

**Depends on:** `fpf-core`
**Bounded context:** lossless compaction of epistemes — prose ↔ compact typed-slot notation (F4-F5), governed by the FPF A.6.3 (Epistemic Viewing) family.

**Source of truth:** this skill is a **compacted runtime projection** — a representation carrying method + trigger + router only, not the source. The canonical carrier is `skills/episteme-compaction/` in the Episteme Compaction repo; it owns the pattern graph, provenance, full E.8 pattern bodies, and acceptance cases.

**Readiness (collective):** `source-faithful` as a compacted projection — runtime allow-list complete, episteme split LEAK/DROP-free, routing complete, FPF grounding in-block, canonical source cited. Not `case-validated`; per-card status and acceptance-case results live in the canonical source.

## Routing table

| Situation | Load |
|---|---|
| Decide formality level (F0 vs F4-F5 vs F3 hybrid) for an output | `references/ECPF.1.md` |
| Render prose claims as typed slots (the core move) | `references/ECPF.2.md` |
| Render Γ aggregation / F-G-R-CL trust metrics | `references/ECPF.3.md` |
| Render evidence provenance / ADI reasoning | `references/ECPF.4.md` |
| Check strict distinctions / category errors | `references/ECPF.5.md` |
| Evaluate a rendering before admission | `references/ECPF.6.md` |
| Reconstruct prose from a compact episteme (reverse render) | `references/ECPF.7.md` |
| Govern the episteme↔prose round-trip / drift | `references/ECPF.TG.md` |

## Navigation rule

The dominant chain is a rendering pipeline: start at `references/ECPF.1.md` (formality gates everything), then `references/ECPF.2.md` (the core typed-slot render), then `references/ECPF.3.md` / `references/ECPF.4.md` for aggregation/trust and evidence/reasoning, then `references/ECPF.5.md` for distinction compliance, closing with `references/ECPF.6.md` for admission. For the reverse direction (notation → prose), load `references/ECPF.7.md`; load `references/ECPF.TG.md` when the round-trip across passes must stay drift-free.

## Lookup rules

- **Routing table first, then "no".** Before looking for a pattern, check the routing table and the `Continues` lines of the card you already hold. If it is not routed, the pattern does not exist here — answer "no"; do not open files searching for it.
- **Read once.** Do not reopen a file you have already read — reuse its already-read content from your context.

## Single surface

`references/` is the consumer surface of this projection. Each card is one `episteme` block: no card frontmatter (identity = filename stem + body heading; FPF grounding = in-block `Grounding (FPF)` slot; local links = `Continues`/`Reopen`). No `INDEX.md`, `relations.md`, graph generator, or evals are shipped — the dependency graph and provenance live in the canonical source. `SKILL.md` is routing-only. Prose reconstruction of this projection is out of scope (a skill is read, not reconstructed).