System
How the machine is built: the eight-plane Tier-1 architecture, the live operational surfaces, and the fleet of autonomous agents that survive a reboot.
Agent fleet · live
source: offline
Total
—
Healthy
—
Warning
—
Critical
—
Status reads from the live agent heartbeat (agent_health, hourly). When offline, fields show "—".
Eight-plane architecture (Tier-1)
Each plane is an independent, deterministic safety layer. The cycle is fail-safe: if any plane fails, no trade goes through.
Validation
Deflated Sharpe, PSR, min track length on every result
Risk
Deterministic RiskPolicy gate — no LLM, cannot be overridden
Audit
Append-only audit trail of every cycle and decision
Execution readiness
Pre-trade checks before any live capital deployment
Governance
Owner review gates, ADR-tracked policy changes
SSOT
Single source of truth mirrored to the site
Disaster recovery
NAV reconciliation, state recovery after host restart
Anomaly detection
APY-spike, threat-reactor, kill-switch on critical threats
Live operational surfaces
Live dashboard
Canonical command center — equity, positions, go-live
Agent fleet
The autonomous launchd agents that run the machine (dashboard · System tab)
Strategy tournament
Backtest → paper → live promotion pipeline (dashboard · Tournament tab)
Risk-tier packages
Unified Preserve / Core / Max-Yield taxonomy
Verify us
Don't trust us — re-derive every published hash yourself
Promotion pipeline · Strategy Lab
source: offline
The research decision layer: which research sleeves graduate backtest → paper → live, and why. Promotion is a deterministic risk gate, not a sales funnel. We publish what FAILS, not just what passes.
Paper research · advisory · not live. The honest finding:
Only the stable sleeves reach PAPER_CANDIDATE. The crypto (ETH/BTC) sleeves are REJECTED — they do not beat the ~3.4% risk-free floor on a risk-adjusted basis. The gate is deterministic: pure arithmetic over the criteria, no LLM, cannot be overridden.
The promotion ladder
Stages past backtest are only reached by clearing the gate. The current track phase is paper; no sleeve goes live until the track is confirmed. RWA floor everything is measured against: —
Paper candidate
—
Backtest pass
—
Rejected
—
Loading promotion verdicts… When offline, the criteria are shown but not live values.
Criteria (deterministic, from the SSOT config): beats floor · drawdown within band · not killed by real risk · walk-forward consistency · capacity · positive net APY. A data-gap kill (a missing feed value) is tolerated as a fail-closed safe-hold, not a risk failure. Source: spa_core/strategy_lab/promotion.py → /api/strategy-lab/promotion.
Risk refusal · what we won't hold
source: offline
The assets we REFUSE to hold — and why
Paper research · a measurement / advisory engine · not financial advice
Most yield products advertise APY. We publish the assets we REFUSE and why — a risk engine that flagged toxic restaking tokens (ezETH-type) before they collapsed, validated on 2 years of data including real depeg events. We'd rather miss yield than hold a book that blows up.
This is a measurement / advisory engine, not a guarantee. It scores the tail risk of each yield-bearing asset and emits a verdict of SAFE / WATCH / REFUSE / UNKNOWN. Source: /api/refusal.
Loading refusal-engine verdicts… When offline, live values are unavailable.
tail_score (0–1) is the normalised tail-risk score: depeg, downside volatility, funding regime. REFUSE/WATCH are listed first. Verdicts are advisory; the live deterministic execution gate is the RiskPolicy (LLM forbidden). Source: GET /api/refusal.
Proof of Reserves
source: static fallback
Verifiable NAV — every dollar accounted for, cryptographically verifiable
loading…NAV (paper)
Delta
—
Verification hash
—
NAV = deployed + cash + accrued, reconciled against reported equity. Paper proof-of-reserves (methodology on the virtual $100k book, not real custody). Full panel → /track-record
The daily cycle
→ multi-strategy runner (S0–S77, tournament)
→ allocator → target allocation with caps
→ RiskPolicy gate (deterministic, LLM forbidden)
→ virtual rebalance → daily yield accrual
→ GoLiveChecker → 29 criteria
Stack: Python 3, stdlib-only at runtime. Atomic writes to all state files. Agents are installed via launchd and survive reboots.