The Desks
Ten specialist desks (+ flow-intel & whale) — twelve analyst voices, each a real data source. One is the ML.
Each desk runs a real tool over fresh data and returns a stance + confidence (one Qwen/Qwen3.8-27B call per
desk on NEAR AI Cloud). The firm synthesizes ten specialist desks plus two
self-learning inputs — flow-intel and whale — for twelve analyst voices into one PM decision.
Every desk has its own deep-dive page — linked in the table below.
All on-chain. The firm (token 1) and all 12 desks (tokens 2–13) are registered agent identities in the
corecontract on NEAR testnet (see On-chain), so every voice the PM hears has an on-chain identity.
| # | Desk | tokenId | What it reads |
|---|---|---|---|
| 1 | Regime/Technical (the ML) | 2 | rule-based regime label live (XGBoost is the backtest engine: 78.9% vs 76.1% persistence on NEAR) + trend strength + dynamic R:R |
| 2 | Macro (Allora) | 3 | Allora decentralised-AI BTC/ETH 8h price prediction |
| 3 | On-chain/Risk | 4 | Rhea whale flow + watchlist health + internal NEAR token safety screen |
| 4 | Research | 5 | Fear & Greed + global macro via free public APIs |
| 5 | Smart-Money Flow | 6 | perp positioning (OI / funding / long-short) + Nansen netflow |
| 6 | Smart-Social | 7 | Elfa narrative / mindshare (smart accounts, not retail) |
| 7 | OI-Contrarian | 8 | fade perp open-interest extremes — context-only until an asset earns the edge |
| 8 | Carry | 11 | delta-neutral funding-carry yield (market-neutral, advisory) |
| 9 | NEAR Fundamentals | 12 | DeFiLlama NEAR TVL / fees / staking → ecosystem risk-on/off |
| 10 | TimesFM Vol/Risk | 13 | next-day realized-vol forecast (TimesFM 2.5 when configured, EWMA fallback) → risk-off stance, ticket vol-sizing, regime-break alarm |
| + | flow-intel | 9 | self-learning flow-signal validation (OI / funding / order-flow / momentum) |
| + | whale | 10 | Hyperliquid per-asset whale stance tracker |
How the desks divide the problem
Twelve voices, three jobs: predict a pattern, follow informed actors, or supply context / risk. The firm tilts which family leads based on each asset's measured character (see below).
How they collaborate
The PM does not average the desks. The nine directional voices (all but Carry, NEAR Fundamentals and TimesFM) each earn a reliability weight from their own track record; the PM ranks them, tags TRUSTED / DISCOUNT, runs a bull-vs-bear debate over the outputs, and only then decides — behind a hard R:R gate.
Which desks lead depends on the asset
asset_efficiency.py profiles each asset fresh every cycle (volatility + return-autocorrelation). A
predictable asset (autocorrelation ≥ 0.10, or a validated edge) boosts the prediction desks
(Regime, OI-Contrarian); an efficient asset (autocorrelation < 0.05, no exploitable pattern) boosts
the flow desks instead. On the data in this repository NEAR, BTC, ETH, SOL, HYPE and SUI all score
below 0.05 (efficient) and XRP reads neutral, so flow desks lead.
The ML is the Regime desk
A directional ML (XGBoost BUY/SELL) is not used: it would have to clear the same cost-aware gate as
every other hypothesis (Finding #7). The ML that does run is the regime classifier
(78.9% vs 76.1% persistence on NEAR, in backtest) inside the Regime/Technical desk, used as a
capital allocator / risk read, not a price predictor. Live, the desk emits a cheap rule-based regime
proxy (detect_current); the XGBoost model is the backtest engine that earns the classifier its pedigree.
Honest ML: the good model earns its place; the disproven one stays out.
Risk & context desks
- TimesFM Vol/Risk — next-day realized-vol forecast (TimesFM 2.5 when configured, EWMA fallback otherwise) → risk-off stance, ticket vol-sizing and a regime-break alarm; never a price call. Deep-dive.
- NEAR Fundamentals — DeFiLlama chain TVL / fees / staking yields (free, keyless) → ecosystem risk-on/off tilt for NEAR. Deep-dive.
- Carry — delta-neutral funding-carry monitor. Deep-dive.
Not a weighted desk: Character Tilt
Character Tilt — asset_efficiency.py profiles each asset (predictable vs flow-following) and tilts
the desk ranking (the quadrant above). It's a meta-weighting over the desks, not a voting desk.
Reliability weights
The PM weights desks by learned reliability (desk_performance): each desk earns a weight
0.6–1.4 from its realized track record (stance vs realized direction, ≥15 samples). Neutral until
proven, never muted. See Self-learning.