KCAP — Kemet Core Algorithm Protocol

1. Charter

Kemet Protocol Governance establishes KCAP as the central ranking orchestrator of the Network. We do not curate speech; we orchestrate relevance under cryptographic identity, published policy constraints, and integrity signals from companion protocols. KCAP is the only protocol that composes the full retrieval stack in a single request lifecycle.

2. Surfaces Served

• Following — corpus from followed Nodes; graph proximity dominates.
• For You — personalized blend of graph, semantic, engagement, and momentum signals.
• Explore — outward discovery; semantic and momentum dominate; graph bias intentionally reduced.

3. Request Pipeline

Feed Request
  → KCCP cache check (hit → return)
  → KCAP-SI  viewer context ingest
  → KCAP-CG  candidate pool assembly
  → KCAP-FE  per-candidate features
  → KCAP-MS  multi-factor scoring (coefficients withheld)
  → KCAP-TF  trust, diversity, exposure budgets
  → KCAP-FR  final order + impression log
  → KCCP write + response

4. Sub-Instruments

• KCAP-SI — identity, interest vector, follow graph slice.
• KCAP-CG — graph candidates, semantic k-NN, policy-eligible corpus.
• KCAP-FE — feature materialization under latency budget.
• KCAP-MS — scoring composition (surface-specific weights withheld).
• KCAP-TF — KTSP, KVEP, diversity caps.
• KCAP-FR — slice, log impressions, respond.
• KCAP-CS — cold start when history sparse.
• KCAP-FS — follow suggestions (separate RPC lifecycle).

5. Post-Create Integration

Before content enters ranking corpora, post-create pipelines invoke KLDP, KEVP, KMSP, and KSLP. Only policy-clean entries become candidates for KCAP-CG.

6. Impression Transparency

We log ranking impressions — viewer, post, surface, rank band, and engagement outcomes — for governance research and future model training under privacy constraints. Feature snapshots attached to logs do not include withheld coefficients.

7. Evolution

Phase 1 operates published heuristic governance. Later phases may introduce learned ranking instruments trained off impression logs. Model artifacts, serving regions, and training pipelines are not published.