Verified Empirical Measurements (tools/measure/out/)

Empirical Hardware Benchmarks

Real forensic measurements executed against physical hardware (17.18 GB Unified Memory Host) across 9 simultaneous multimodal pressure scenarios and micro-benchmark test harnesses.

AUDIT.md §17 Scenarios

The 9 Multimodal Pressure Scenarios

Select a scenario to inspect measured RAM consumption and active concurrency contention.

✓ Verified Live Data
Scenario 01

System Idle

14.54 GB+0.00 GB from idle

Full container mesh running (Postgres, Neo4j, Redis, Qdrant, NATS, LiveKit) with no active conversation.

Active Mesh Components
NATS JetStreamPostgres + pgvectorNeo4j GraphRedis CacheQdrant Vector DBLiveKit SFU
Key Empirical Finding:Measured on 17.18 GB unified memory host. Docker containers consume 752.4 MiB total; Ollama holds 3 resident models.
Provenance: backend/tools/measure/out/m17_pressure_scenarios.json §1
Subsystem Measurements

Subsystem Micro-Benchmarks

Exact timings captured from test suites and live measurement runs in tools/measure/out/.

M1.5-RTX2060S-TTFT● LIVE

Composed Turn Time-To-First-Token (TTFT)

119.35 ms

Composed turn deliberation + prompt prefill on dedicated RTX 2060 Super (isolated prompt eval: 39.95 ms)

Unit: milliseconds (composed turn: 34.57ms deliberation + 84.78ms prefill)
M1.5-RTX2060S-Throughput● LIVE

RTX 2060 Super (3B) Generation Speed

122.71 tok/s

Sustained streaming throughput (187ms full sentence generation, zero audio buffer underrun)

Unit: tokens per second sustained (8.15 ms/token)
M1.7-LLM-TTFT● LIVE

Hermes 3 (8B) Time-To-First-Token

61.9 ms

Empirical streaming generation benchmark across 5 companion scenarios on Tesla T4 GPU

Unit: milliseconds (<100ms sub-perceptual)
M1.8-LLM-Throughput● LIVE

Hermes 3 (8B) Generation Speed

46.6 tok/s

Sustained streaming throughput (~150ms per 7-word audio chunk, zero TTS buffer underrun)

Unit: tokens per second sustained
M1.9-BARGEIN-REFLEX● LIVE

Sub-Millisecond Barge-In Reflex

0.099 ms

Audio interruption reflex dispatch latency over NATS bus upon voice activity detection

Unit: milliseconds (<1.0 ms hard threshold)
M1.10-ROLLBACK● LIVE

Deterministic State Rollback

14.28 µs

Atomic rollback of speculative cognitive state mutations upon turn cancellation

Unit: microseconds (<500 µs safety gate)
M1.11-BOUNDARY-INVARIANCE● LIVE

Identity & Safety Boundary Invariance

100%

Constitutional guard evaluation across 1,200 adversarial redteam probes

Unit: invariance rate (0 violations / 1,200 adversarial turns)
M1.12-META-OVERHEAD● LIVE

Metacognitive Latency Overhead

0.17 ms

Calibration and second-order reflection directive generation overhead

Unit: milliseconds per turn
M1.13-CANDIDATE-SELECTION● LIVE

Candidate Action Selection Latency

0.36 ms

Constraint-first scoring and selection across 10 competing action intents

Unit: milliseconds (10 candidate evaluation)
M1.14-PLAN-VERIFICATION● LIVE

Deterministic Plan Verification

0.29 ms

Formal DAG verification of multi-step deliberative plans prior to execution

Unit: milliseconds (<2.0 ms execution budget)
M1.6-GraphWarm● LIVE

Neo4j Knowledge Graph Warm Fetch

3.58 µs

1,003 entities & 4,002 relations fetched with 300s TTL memory cache

Unit: microseconds (<0.004 ms)
M1.6-GraphCold● LIVE

Neo4j Knowledge Graph Cold Query

56.2 ms

Full Cypher graph traversal over 1,003 un-cached entity nodes

Unit: milliseconds
M1.6-GraphSeeding● LIVE

Neo4j Graph High-Volume Seeding

0.91 s

Concurrent ingestion of 1,000 entity nodes and 1,998 relationship edges into graph DB

Unit: seconds wall-clock
M1.2-ConsolidationIdle● LIVE

Subconscious REM Consolidation (Idle)

7.48 s

Extracts facts and updates Neo4j beliefs across 6 recent turns

Unit: seconds wall-clock
M1.2-ConsolidationVLM● LIVE

Subconscious REM Consolidation (VLM Load)

10.08 s

Consolidation pass executed during continuous Moondream VLM inference

Unit: seconds wall-clock
M1.1-AudioBurst● LIVE

LiveKit Audio Frame Burst Delivery

23.5 ms

50 sequential 32kHz PCM audio frames published over NATS

Unit: milliseconds
Infrastructure Footprint

Docker Mesh Memory Breakdown (752.4 MiB Total)

Measured resident container memory across the 6 core backend services.

590.8 MiBbrain_graphPort 7687
62 MiBbrain_vectorsPort 6333
38.3 MiBlocal_sfuPort 7880
34 MiBpostgres_dbPort 5432
17.2 MiBnats_meshPort 4222
10.1 MiBbrain_cachePort 6379
Compatibility & Profiles

Platform Performance Matrix

Engineering targets the architecture is designed for, not live measurements — see Section 2 above for what's actually been measured.

Hardware TargetLaunch ProfileLLM Inference EngineVoice EngineTime-To-First-TokenTotal Turnaround
Dedicated Home GPU (NVIDIA RTX 2060 Super 8GB)Headless 24/7 Full Mesh (Linux Kernel 7.0 / Ubuntu 24.04)Llama 3.2 3B (Ollama Bare-Metal / CUDA 13.2)GPT-SoVITS 32kHz (CUDA)119.35 ms composed (39.95 ms isolated prompt eval)120 - 180 ms
Google Colab / Cloud GPU (NVIDIA Tesla T4)Full mesh + Hermes 3 (8B)Hermes 3 8B (Ollama / CUDA)GPT-SoVITS 32kHz (CUDA)61.9 ms (Measured)160 - 220 ms
Apple Silicon M1 / M2 / M3 (16GB Unified)Full Stack (Voice + Brain + Memory)Llama 3.2 3B (Metal / MLX)GPT-SoVITS (CPU / Metal)320 - 450 ms680 - 950 ms
NVIDIA RTX 3060 / 4060 (12GB VRAM + 16GB Host)Full Stack + Vision ProfileHermes 3 8B / Qwen 2.5 14B (CUDA)GPT-SoVITS 32kHz (CUDA)50 - 90 ms150 - 260 ms
Modern x86_64 CPU (16GB RAM, No GPU)Heavy Mode (Local Whisper STT + Brain)Llama 3.2 1B (AVX-512 / OpenVINO)Bundled Pre-synthesized Reference450 - 650 ms900 - 1400 ms
Weak Laptop + Cloud Fallback (8GB RAM)Light Mode (Claude 3.5 Sonnet Fallback)Anthropic Claude API (Streaming)Remote TTS or WebRTC Voice280 - 400 ms600 - 850 ms
Step-by-Step Breakdown

Conversational Loop Latency Waterfall

A per-stage budget, not a measured trace — the direct attempt to measure this loop (m14_stt_cost.json) came back UNKNOWN because stt-agent wasn't running for that run.

Step 0115 ms

Speech Detection & VAD Cutoff

Energy-based thresholding & silero VAD

stt-agent (Rust)
Step 02135 ms

Speculative Intent & Emotion

Early barge-in reflex & 7-class emotion classification

SenseVoice (sherpa-onnx)
Step 03180 ms

Final Speech Transcription

High-precision word-level transcript generation

whisper.cpp (Rust)
Step 044 ms

Appraisal & Endocrine State

PAD computation, boundary check, cortisol/dopamine update

brain_agent (Python)
Step 0512 ms

Deliberation & Intent MAUT

Behavior tree traversal and candidate scoring

brain_agent (Python)
Step 0662 ms

LLM Time-To-First-Token (TTFT)

Empirical streaming first token dispatch (61.9ms measured on Tesla T4)

Ollama (Hermes 3 8B)
Step 07160 ms

GPT-SoVITS 32kHz Synthesis

Streaming chunk synthesis with prosody trajectory & pause bias

voice-agent (Rust)
Step 0818 ms

LiveKit WebRTC Transmission

PCM audio frames + visemes data channel dispatch

transport_agent (Python)