DARM-ANN v7.7 · Live Simulation Suite

APPLICATION LAYER

Any OpenAI / Anthropic SDK · Zero code changes

POST /v1/chat/completions POST /v1/messages SSE Streaming Tool Calling

🛡 GATEWAY — sMCP-1 v1.0

FastAPI SPIFFE SVIDs Ed25519 mTLS 1.3 HMAC Audit Parallax Scheduler

STDIO allowlist: npx · uvx · python · node · docker · deno | gVisor sandboxes

TIER 1 — TinyLM Swarm (1.5B · ≤80 mJ/tok)

Pool A · Loihi 2

—

◆ ARA · NeuEdge SNN

Pool B · Akida 2

—

◆ ARA · NeuEdge SNN

Pool C · Jetson

—

◆ ARA · NeuEdge SNN

TIER 2 — Mid-Scale (15B · CDCP Orchestrator)

darmann-mid · A100/H100

STM

LTM

RRC

MLGRU + Mamba-2 + MoE

8 experts · 131K ctx

CDCP Orchestrator

N=4–8 · credibility softmax

◆ ARA — Full 6-agent topology (gVisor)

TIER 3 — Consolidator (70B+ · Hala Point)

darmann-xl · H100/B200 · 524K ctx

STM

LTM

RRC

CDCP Gatekeeper

BFT write-admission · NLI

MAGMA RRC

Semantic·Temporal·Causal·Entity

◆ ARA — Full 6-agent · OpenScholar 45M papers

CDCP v2.0 — Byzantine Fault Tolerant Consensus

Live Consensus Ring · N=7 · f=2 Byzantine · 3f+1 Assumption

Redraws every 2s · red=Byzantine · green=honest

①Each node generates candidate δᵢ + credibility cᵢ (Attention Trust Score)

②Credibility-weighted softmax aggregation → consensus δ̄

③NLI pre-consolidation gate: reject if CONTRADICTION with LTM

④SycophancyScore gate — adversarial debate if score > 0.6

⑤ADMIT → HMAC audit log + LTM write | REJECT → flag

‖δ̄ − δ*‖ ≤ f/(N-f) · max‖δᵢ − δ*‖ ≤ ½ · max‖δᵢ − δ*‖

🗂 Layer Controls — Peel Back View

Toggle layers to isolate different aspects of the distributed DARM-ANN system. Each layer traces a different class of inter-node communication.

⚡ Inference Layer

Token generation and speculative decoding flows. Shows γ=5 draft tokens flowing TinyLM → Mid (qualification τ=0.85) → XL (verification). TSLT-compressed logits (top-200, 1.6KB/token) shown as yellow data chunks.

DRAFTTSLT×5VERIFYSSE

🧠 Memory Layer

Memory read/write operations across the 5-tier hierarchy. D-MEM RPE gating, Episodic Buffer writes, Titans STM gradient updates, and CDCP-gated LTM admissions. Purple squares flow along the trace path.

D-MEMSTM-GRADLTM-WRITERRC-QUERY

⚖️ Consensus Layer

CDCP Byzantine-fault-tolerant consensus. Shows proposal broadcasts, credibility-weighted voting circles, NLI gate results, sycophancy detection, and ADMIT/REJECT decisions. Lavender circles represent vote packets.

CDCP-REQVOTEADMITREJECT

🔐 Security Layer

Cryptographic negotiations and audit events. SPIFFE SVID issuance, Ed25519 signing, mTLS 1.3 handshakes, HMAC-chained audit log entries, A2A signed Agent Cards, and gVisor sandbox events. Amber diamonds indicate crypto exchanges.

CSRSVIDmTLSHMACAUTH

🔬 ARA Layer

Autonomous Researcher Agent pipeline. Hypothesis proposals, OpenScholar fetches, cross-node research synthesis, Meta-Review submissions to CDCP, and gVisor sandbox micro-experiment results. Green stars mark ARA events.

ARA-PROPRESEARCHCONTENTFIND

💻 Compute Distribution

Distributed compute heat overlay. Radial glow proportional to recent compute activity. Bandwidth lines between nodes show data volume — thicker = higher throughput. Each node shows a live utilization bar and mJ/token estimate.

Low

High

Trace Line Key

● ━━━ dashed yellow = Inference / token flow

■ ━━━ dashed purple = Memory read/write

● ━━━ dashed lavender = Consensus vote

◆ ━━━ dashed amber = Crypto / security

★ ━━━ dashed green = ARA research

📊 LLM Benchmark Suite

🗺 Global Memory View — All Nodes

🗂 Layer Controls — Peel Back View

⚙ CPU Allocation per Node