Status : Reference specification
Source : Consolidated performance, chain support, gas, and subsystem specifications.
Reader orientation: This appendix consolidates all performance targets and technical specifications across the Bardo ecosystem: vault gas costs, Heartbeat (the 9-step decision cycle) tick latency, HDC operation benchmarks, chain intelligence pipeline timing, Grimoire (the agent’s persistent knowledge base) memory operations, per-tick compute budgets at each timescale (Gamma/Theta/Delta), chain support matrix, contract gas estimates, ERC standards checklist, and subsystem specifications. The key concept is the per-tick compute budget: all Triage Augmentation innovations are T0-compatible (zero inference cost, no LLM calls). See prd2/shared/glossary.md for full term definitions.
Metric Target
Vault deployment gas < 3M gas Deposit/withdraw latency < 2 blocks Tool response time < 500ms (p95) Heartbeat tick (suppressed) < 100ms, $0 Heartbeat tick (escalated) < 10s, < $0.10 LLM cost Memory search latency < 25ms Share pool price convergence < 5 min to NAV ±1% Wallet creation < 500ms Transaction signing < 200ms Onboarding (atomic) < 30s, $0 gas (paymaster) GottsLoop daily cost (typical) < $5 (95% Haiku ticks) Memory RAM footprint 150–250 MB
Latency targets measured on Apple M1 Pro (NEON SIMD). x86-64 with AVX2 should match or beat these numbers thanks to native POPCNT. Dimension D=10,240 bits (160 u64 words) unless noted.
Operation Target Latency Method
Hamming distance (D=10,240) ~10 ns 160 XOR64 + 160 POPCNT64 HyperVector bind (XOR) ~2 ns 160 XOR64 HyperVector unbind (XOR, self-inverse) ~2 ns 160 XOR64 HyperVector permute (cyclic shift) ~5 ns 160 shift + 160 OR HyperVector bundle (majority vote, K=10) ~15 us accumulate, ~12 us finish Per-bit vote counting + threshold Full transaction fingerprint encode < 100 us Bind + permute across transaction fields Resonator factorization (1 iter, 3 factors, 100 codebook) ~300 us Iterative projection ANN query (usearch binary HNSW, 100K items) < 1 ms (main), ~1.5 ms (main + staging) Hamming metric, M=16, ef_search=100 ANN staging insert ~100 us HNSW incremental insert Hamming top-K pre-filter (100 candidates) ~1 us 100 x Hamming distance
Total POPCNT budget per Gamma tick : ~3.5 us across ~355 comparisons (curiosity pre-filter, anti-pattern screening, context dedup, regime classification, dream correlation). This is noise within a 5-15s Gamma interval.
Operation Target Latency Notes
Binary Fuse8 lookup ~10 ns per address 3 memory accesses, branchless (xorf crate) Binary Fuse8 construction (10K keys) ~100 us Rebuilt at each Gamma tick via arc-swap DDSketch quantile query O(1), ~15 ns observe Logarithmic bucketing, alpha=0.01, ~2 KB fixed memory DDSketch quantile read O(log n) bucket scan Formal relative error guarantee MIDAS-R anomaly score ~50 ns per edge 3 CMS queries + chi-squared CMS frequency update ~20 ns k hash + k increment Single block triage (150 txs) < 5 ms total Binary Fuse8 filter + severity scoring Bayesian surprise update ~300 ns per observation Conjugate prior arithmetic + digamma Bayesian surprise decay (100 protocols) ~50 us Linear scan of parameters
False positive rates : Binary Fuse8 achieves 0.39% FPR at 8.7 bits/entry, down from ~1% FPR at 9.6 bits/entry with standard Bloom filters.
Operation Target Latency Notes
HDC episode compression < 500 us per episode Bundle accumulate at Theta cadence Similarity search (10K episodes) < 5 ms Hamming pre-filter then cosine refinement Grimoire HDC retrieval (top-10) < 2 ms Binary HNSW + Hamming distance SQLite query < 1 ms Existing PRD2 target LanceDB vector search < 25 ms Existing PRD2 target Memory per HDC vector 1,280 bytes 10,240 bits (vs. 3,072 bytes for 768-dim f32)
All TA (Triage Augmentation) innovations are T0-compatible: zero inference cost, no LLM calls, no external API calls, no blocking I/O.
Target total: < 200 ms (sensing + triage + CorticalState update). Current measured budget: ~7-17 ms.
Component Per-Event Cost Events/Tick Total/Tick
Block ingestion + log filtering (existing) – – 5-10 ms Probe severity scoring (existing) – – 1-2 ms CorticalState atomic writes (existing) ~0.1 us – ~0.1 us HDC fingerprint encoding ~100 us 10-50 1-5 ms Binary Fuse8 queries ~10 ns 100-500 1-5 us Bayesian surprise updates ~300 ns 10-50 3-15 us MIDAS-R scoring ~50 ns 10-50 0.5-2.5 us DDSketch observations ~15 ns 10-50 0.15-0.75 us Hamming pre-filter ~1 us 5-20 5-20 us CMS frequency updates ~20 ns 10-50 0.2-1 us CorticalState HV encoding ~100 us 1 100 us Total new TA overhead ~1.2-5.3 ms
TA overhead: < 0.1% of available Gamma budget.
Target total: < 2 s (includes deliberation, prediction, context assembly). LLM-dominated (30-120s cycle).
Component Total/Tick
ANN staging insert ~0.1 ms Episode compression (accumulate) ~0.015 ms Regime classification ~0.001 ms Hedge weight update ~0.01 ms Thompson threshold selection ~0.05 ms Bayesian surprise (Theta batch) ~0.05 ms Total new TA overhead ~0.2 ms
Target total: < 30 s (includes dream scheduling, consolidation check). Runs on 5-20 minute cadence.
Component Total/Tick
Bundle prototype refresh 1-5 ms Drift detection (centroid comparison) ~0.15 ms Curator consolidation 50-200 ms ANN main rebuild 100 ms - 2 s Bayesian surprise (Delta batch) ~0.1 ms Total new TA overhead ~150 ms - 2.2 s
Operation Target Latency Notes
Persistence diagram computation (1000-point cloud) < 50 ms Scheduled at Theta/Delta cadence Wasserstein distance < 10 ms Between persistence diagrams Somatic marker lookup < 1 ms HDC similarity (Hamming distance)
Metric Target Notes
Per-branch overhead (CoW) ~12.8 KB CoW overlay with ~100 dirty slots Branch creation time ~1 us Arc clone + empty HashMap Transactions replayed per block 5-15 Targeted, watched contracts only (vs. ~150 full replay) Max memory per branch 512 MB standard, 2 GB power Resource profile max_memory setting
CPU : Apple M1 Pro or equivalent (NEON SIMD). x86-64 with AVX2/POPCNT.POPCNT : Available on all modern targets. LLVM compiles count_ones() / count_zeros() to hardware intrinsic automatically.Memory : DDSketch uses ~2 KB fixed regardless of observation count. HDC vectors use 1,280 bytes each.Dependencies : xorf (Binary Fuse8), sketches-ddsketch, usearch (HNSW), statrs (digamma), arc-swap (lock-free replacement). Total compiled footprint: ~2.5 MB, no C++ dependencies required.
Chain Chain ID V4 Deployed Vault Support Share Pools
Ethereum 1 Yes Yes (v1 target) Yes Base 8453 Yes Yes (v1 target) Yes Unichain 130 Yes Yes (v1 target) Yes Optimism 10 No Tools only No Polygon 137 No Tools only No Arbitrum 42161 No Tools only No BNB Chain 56 No Tools only No Avalanche 43114 No Tools only No Celo 42220 No Tools only No Blast 81457 No Tools only No zkSync Era 324 No Tools only No
v1 targets full vault and share pool support on Ethereum, Base, and Unichain – the three chains with V4 deployed. All other supported chains receive tool coverage for data reads, trading, and LP operations via the Uniswap Trading API.
Operation Estimated Gas Est. Cost (Base L2) Est. Cost (Ethereum)
Vault deployment (CREATE2) 2.5–3M $0.05–0.10 $5–15 Deposit (ERC-4626) 80–120K $0.002–0.005 $0.50–1.50 Withdraw (instant) 80–120K $0.002–0.005 $0.50–1.50 Rebalance (single adapter) 150–250K $0.005–0.01 $1.00–3.00 Hook swap (VaultHook) 100–180K $0.003–0.007 $0.80–2.00 Agent registration (ERC-8004) 150–200K $0.004–0.008 $1.00–2.50 Milestone attestation 50–80K $0.001–0.003 $0.30–0.80 Proxy announce 60–100K $0.002–0.004 $0.40–1.00 Proxy execute 80–120K $0.002–0.005 $0.50–1.50 Proxy cancel 40–60K $0.001–0.002 $0.25–0.60
Gas estimates assume Base L2 gas price of 0.001–0.005 gwei and Ethereum mainnet gas price of 20–60 gwei.
Standard Purpose
ERC-20 Token interface for vault shares and base assets ERC-721 Identity NFTs for ERC-8004 agent registration ERC-4337 Account abstraction for smart wallet operations ERC-4626 Tokenized vault standard for deposits, withdrawals, share accounting ERC-7265 Circuit breakers for automated risk intervention ERC-7540 Asynchronous redemption for queued withdrawals ERC-7579 Session keys for bounded agent authorization ERC-7677 Paymaster for gasless onboarding ERC-7683 Cross-chain intents (initial support) ERC-8004 Agent identity and reputation registry EIP-7702 EOA delegation for account abstraction migration
Standard Purpose Reason for Deferral
ERC-6900 Modular accounts Complexity; ERC-7579 sufficient for v1 ERC-7683 Full cross-chain settlement Single-chain focus in v1 ERC-7821 Minimal batch executor Evaluate need post-launch ERC-6551 Token-bound accounts Alternative identity model; evaluate
Property Value
Total tools ~171 across 26 modules Composable profiles 12 Authentication SIWE + OAuth 2.1 API key tiers Three-tier (read, standard, admin) Transport stdio, SSE, Streamable HTTP Rate limiting Configurable per key and per identity tier Cache TTL 30s (prices), 5min (pool data), 1h (static data)
Property Value
Total agents 32 (25 core + 7 vault) across 8 categories Delegation hierarchy 4-level Terminal nodes 4 DAG constraint Acyclic (validated at registration) Maximum delegation depth 3
Property Value
Minimum heartbeat interval 60,000ms Maximum heartbeats/day per strategy 1,000 Typical suppression rate ~95% Maximum active playbook heuristics 30 Curator consolidation cadence Every 50 ticks ExpeL consolidation cadence Every 4 hours Maximum parameter change per iteration 10% Maximum consecutive errors before pause 5 Lane Queue priority levels 6 Session history window 200 turns
Property Value
Memory types 4 (working, episodic, semantic, procedural) Embedding dimensions 384 (all-MiniLM-L6-v2, 23MB INT8 quantized) Insight categories 12 Maximum insights per agent 10,000 Maximum episodes per agent 100,000 SQLite query latency Sub-1ms LanceDB vector search latency Sub-25ms Confidence updates Asymmetric (+0.10 optimistic / −0.15 pessimistic) Drift detection ADWIN with Hoeffding bounds at p < 0.01
Property Value
Architecture Three concentric rings (preventive, cryptographic, reactive) Defense layers 15 (7 preventive + cryptographic enforcement + reactive) Default per-transaction limit $10,000 Default per-session limit $50,000 Default per-day limit $100,000 Proxy delay tiers 5 (0s, 10min, 1h, 24h, 48h) Circuit breaker levels 3 (3% / 7% / 13% drawdown) Multi-layer defense requirement Independent bypass of all 7 preventive layers
Topic Document
Vault contracts ../08-vault/01-contracts.md (ERC-4626 Solidity contracts: AgentVaultCore, AgentVaultFactory, FeeModule, NAVAwareHook)Safety architecture ../10-safety/00-defense.md (15-layer defense model with three concentric rings: preventive, cryptographic, reactive)Sanctum tools ../07-tools/01-architecture.md (tool library architecture: profiles, transports, safety pipeline, caching)Agent system ../01-golem/01-cognition.md (T0/T1/T2 cognitive tier routing and LLM-driven decision-making)Memory system ../04-memory/01-grimoire.md (Grimoire persistent knowledge base: episodes, insights, heuristics, causal links)Compute infrastructure ../11-compute/00-overview.md (Fly.io two-app topology, x402 billing, VM tiers and warm pool)Chain configuration ../shared/port-allocation.md (canonical port assignments for local development services)