Overview
Dual-Phase Holographic Cache
Separates stable prefill context from a rolling decode frontier weighted by causal surprise. Novel tokens write harder; redundant tokens are suppressed.
Provenance-Weighted Resonant Memory
Cryptographically isolated multi-tenant memory using SOURCE/TRUST/TENANT role vectors. Prevents resonance bomb attacks via trust-weighted accumulation.
Capability-Scoped HMKs
Stateless computing tokens with HMAC integrity, hop limits, expiry, and resonant shard consensus for adversarial perturbation detection.
Mathematical Formalization
Dual-Phase Cache
Prefill Cache: 𝒫 = Σ exp(i · φ(kt))
Decode Frontier: ℱt = ℱt-1 + wt · exp(i · (φ(kt) + φ(pt)))
Causal Surprise: wt = max(0, 1 - R(kt, 𝒫))
Provenance Memory
Secure Entry: ℰ = (k ⊕ v) ⊗ (ρTENANT ⊕ encode(z)) ⊗ (ρSOURCE ⊕ encode(s))
Trust Weighted: ℰweighted = (τ · R(SRC, A)) · ℰ
Capability HMK
Integrity: ι = HMACSECRET(serialize(s⃗) ‖ σ ‖ h ‖ ε)
Shard Consensus: reject if |R(s⃗i, α) - μ| > κ · σ
API Reference
/api/v1/dual-phase/prefill/api/v1/dual-phase/decode/api/v1/provenance/store/api/v1/provenance/query/api/v1/hmk/create/api/v1/hmk/verifyDeployment
Cloudflare Worker
export CLOUDFLARE_API_TOKEN=your_token
npm install
npm run deploy
Docs (Cloudflare Pages)
npm run docs:build
# Deploy docs/dist to Cloudflare Pages
Local Development
~/.kimi/catalyst-brain/.venv/bin/python -m pytest tests/ -v
~/.kimi/catalyst-brain/.venv/bin/python visuals/generate_charts.py