TESTNET
Build Notes
Transparency for review: what the live system carries on top of the code submitted to GitHub.
Frozen submissiongithub.com/Eras256/Aurex @ 9eb95a4
The repository code submitted for judging is frozen at the deadline commit. The public deployment kept iterating after that date; every change over the frozen submission is listed, fully and honestly, below.
The goal is that reviewers can reconcile the two: what was judged is exactly the frozen commit, and what is live is that commit plus the extensions detailed on this page.
The 17 finalists voted for Option B: an official extension window through Sunday 12 Jul 2026 to expand the project (final-phase branch). Everything listed under "Final Phase" below is that official work, already merged to main and deployed.
Final Phase · Extension Window
FEED-SYNC FIX
Binance L2 sync rewritten to the official sequence rules (resync-storm fix)
- ▸The Binance diff-depth client treated stale overlapping events (normal replays for a few hundred ms after every snapshot) as sequence gaps, causing an infinite resync/reconnect storm that could keep the venue offline for hours while hammering the REST snapshot endpoint. Events are now classified per the official Spot rules — stale replays are dropped silently, snapshot-straddling events apply idempotently, and only true forward gaps resync. Covered by new deterministic unit tests.
- ▸A failed snapshot fetch (e.g. a REST rate limit) previously left the client buffering diffs forever with no recovery path; it now recycles the socket so the backoff reconnect cycle owns recovery, and the sync buffer is capped so a stalled fetch can never grow memory unboundedly.
RISK ENFORCEMENT
Every advertised risk knob is now actually enforced
- ▸maxTradesPerMinute was exposed in the schema and Risk page but never enforced; the RiskManager now applies it as a rolling 60-second booked-trade limiter (triangular fills consume the same budget), with deterministic unit tests.
- ▸maxPositionQuotePerExchange was likewise unenforced; a trade whose quote notional would commit more than the cap to a single venue is now rejected with an explicit reason.
TELEMETRY HONESTY
Telemetry and ledger honesty fixes
- ▸Trade provenance is now explicit end-to-end: every ledger row records whether it was a real testnet fill (signed IOC orders matched on an exchange test environment) or a simulator fill. TESTNET rows carry a badge in the Trade Ledger, the CSV export gained an execution column and was renamed from simulated_trades_*.csv to aurex_trades_*.csv (with testnet mode active, calling every row "simulated" was inaccurate), and the tag survives restarts via the persistence layer.
- ▸The Trade Ledger net-P&L cell hardcoded a green "+$" prefix, so a realised loss would have rendered as a glowing green "+$-3.21"; it now renders sign-aware amounts in green or red.
- ▸evals/sec now counts every evaluation cycle (including coalesced re-runs) instead of only the first book of a burst; trade events report the actually-filled volume (a 0.001 BTC testnet clamp is no longer narrated as the full theoretical size); the triangular loop honours per-exchange taker-fee overrides; sizing volumes are float-rounded so ledger rows never show 0.44999999999999996-style artifacts; and shutdown now disconnects all five venues.
PARAMETRIZATION
Deep runtime parametrization
- ▸Every previously-hardcoded engine constant is now a live, schema-validated (Zod), hot-applied knob in EngineConfig: sizing step, per-pair execution cooldown, circuit-breaker multiple, leg-fill failure probability, volatility-breaker %, consecutive-loss limit, loss/volatility cooldown durations, inventory rebalancing thresholds, and the statistical-arbitrage z-score gate — all editable from the Risk page with no restart.
- ▸Per-exchange taker-fee overrides with one-click Retail/VIP presets, so the operator can make the fee assumption explicit and adjustable instead of hidden.
AI · LIVE MODEL
Copilot upgraded to a real, live LLM — grounded on engine state
- ▸The AI Copilot — and every in-app AI widget (trade critiques, opportunity explanations, system-health diagnostics and risk calibration) — is now backed by a real model (OpenAI gpt-4o-mini) instead of the deterministic mock. It is reached through two secure server-side Next.js routes; the API key is read server-side only and never reaches the client bundle.
- ▸Answers are grounded on the live engine state (GET /state): the model reasons over the operator’s real P&L, risk config, recent fills and rejected windows, so it cites actual numbers rather than templated boilerplate.
- ▸Resilience is preserved and made honest: if the key is unconfigured or a request fails, each call falls back transparently to the deterministic mock so the demo never breaks — and the UI labels which mode produced the answer (live model vs. offline fallback) so the distinction is never overstated.
STRATEGY+
Statistical-arbitrage gate wired to execution
- ▸The rolling per-pair z-score (already computed for ranking) now optionally gates execution: when enabled, only windows more anomalous than a configurable threshold are executed, prioritising mean-reverting dislocations over merely-positive spreads.
ROBUSTNESS+
Deterministic robustness coverage
- ▸New unit tests exercise the configurable circuit breakers deterministically (volatility breaker trips at the configured %, consecutive-loss breaker trips at the configured limit and not before, and the loss streak resets after a win) — proving robustness rather than only asserting it.
REBALANCE+
Rebalancing visibility
- ▸The Wallets page now surfaces the active (operator-configurable) rebalancing thresholds per asset plus recent REBALANCE settlement activity pulled straight from the engine event stream.
ANALYTICS+
Max-drawdown metric
- ▸A max-drawdown figure (largest peak-to-trough equity decline) is computed client-side from the same equity history driving the P&L chart, alongside the existing win rate and honest Sharpe.
REAL EXECUTION
Real order execution on exchange test/demo environments
- ▸An optional executionMode: "testnet" routes arbitrage legs to real signed IOC orders on venue test environments — Binance Spot Testnet, OKX Demo and Bybit Testnet, all three verified live with real filled orders (six real-executable directed routes) — fake balances, no real funds, real matching engines. Falls back to the internal simulator per-trade if a leg is unconfigured or fails, so the default demo experience never depends on it.
- ▸The execution layer also gained a Coinbase Exchange sandbox real-fill leg (active when sandbox credentials are configured) and an honest Kraken dry-run: Kraken offers no public spot testnet, so its leg calls the real order-validation endpoint (validate-only) and the fill is transparently booked by the simulator — never presented as a real fill.
UI & METRICS FIXES
Dashboard coherence and Copilot accuracy fixes
- ▸Portfolio Equity is now starting allocation + realised net P&L; it previously summed the seeded per-venue wallet reserves, showing an incoherent multi-million figure that contradicted the initial-allocation label and the Live Net P&L card. KPI cards also gained responsive font sizing.
- ▸The cumulative trade count reports the true executed total instead of freezing at the 50-row display slice (which also inflated average profit per trade); the AI Copilot’s execution-cost figures were corrected for a scale mismatch against the engine’s cost model; and a Risk-page bug where sliders reset while being edited was fixed.
INFRASTRUCTURE
Redeployed backend with durable cloud persistence
- ▸The bot now runs on a new Fly.io app with Supabase Postgres persistence (trades, opportunities, balances, config and P&L survive machine restarts/redeploys — the previous state was ephemeral local storage), still in the Frankfurt region required to reach all 5 exchange feeds, on a single machine (the engine is a stateful singleton; a second machine would split state).
Pre-Final-Phase · Public Deploy Delta
REALISM
Execution realism & honest metrics
- ▸Stochastic two-sided fill drift: realised fills are drawn around the modeled adverse cost (Box–Muller), so the win rate is no longer a perfect 100%.
- ▸Cross-venue leg-execution risk: a fraction of approved trades fill one leg and miss the other, unwinding at a realised loss — the dominant real-world loss source.
- ▸Per-pair execution cooldown (60s): once a dislocation is captured the pair is not re-fired every tick, so cumulative returns are realistic instead of compounding an artificial spread.
- ▸SKIPPED window surfacing: sub-threshold venue pairs are recorded as transparently-rejected even on executing cycles, and the feed is status-balanced so SKIPPED is never empty.
STRATEGY
Strategy intelligence
- ▸Single-venue triangular arbitrage on Binance (USDT → BTC → ETH cycle), net of three taker fees.
- ▸Statistical-arbitrage ranking: a rolling z-score per venue pair prioritises the most anomalous, mean-reverting dislocations when multiple windows are profitable.
EXECUTION
Execution & latency modelling
- ▸Adverse price-movement model during the fill window, scaled by volatility × √time.
- ▸Pure compute latency separated from wire latency, with a p99 detection-latency metric in the header ticker.
AI COPILOT
AI Quant Copilot layer
- ▸AI Copilot workspace and an engine settings modal with provider/model selection (advisory only; simulated by default).
- ▸Spread explainability and execution-cost attribution surfaced from live telemetry.
RELIABILITY
Reliability & operations
- ▸Always-on engine guard and liveness watchdog with self-heal recovery counting.
- ▸Honest Sharpe ratio: withheld until at least 20 trades exist, rendered as "Calculating n/20" rather than a flattering placeholder.
- ▸Binance resync-storm fix (reconnect on persistent desync) and persistence honesty (report missing cloud state instead of masquerading defaults).
- ▸Real Supabase audit trail, dynamic risk-calibration API, and a live WebSocket telemetry stream.
UI / I18N
Interface & localisation
- ▸Full English/Spanish localisation across the terminal, plus responsive fixes (mobile bottom-sheet modal, header ticker, overflow tables).
- ▸Markets grid and spread matrix across the 5 venues, wallet/CEX coverage view, logo-to-home links and repositioned language selectors.
TESTS & QA
Tests & quality
- ▸Integration test suite for the Express bot API (HTTP routes, secure-guard auth, calibration and audit endpoints) on top of the existing unit tests — 41 tests passing.
DOCS & LEGAL
Documentation & legal
- ▸Bilingual README rewrite, technical paper, operational runbook and AI Copilot phase roadmaps; absolute claims softened for technical honesty.
- ▸Aurex-specific Terms & Privacy plus the official Coding Challenge Mexico policy pages, routed from the footer.
TOOLING / CI
Tooling, CI & refactors
- ▸CI workflow YAML fix, removal of dead i18n scaffolding, elimination of `any` types for stricter typing, and assorted build/config cleanups.
Simulated environment. Not financial advice. This page does not modify the submitted repository.