What the Huginai stack actually runs.
Nine layers, from raw social noise to the paper ledger. The pipeline works — ingestion, clustering and scoring all still run. What it produced is what closed the directional book, and the one construction that survived measurement sits at the bottom of the list.
Live price streams
Persistent Binance WebSocket connections carry tick-level price and volume for every asset on the watchlist, plus the funding rates the carry work is built on.
- Binance spot and futures streams
- WebSocket subscriptions, not polling
- Volume and order-book depth
- Funding rate monitoring
Signal clustering and scoring
A model router reads the ingested sources — Gemini for the high-volume extraction pass, Claude for the conviction pass — clusters related mentions, weighs source credibility and emits a 1–10 score with the inputs that produced it. That score measured anti-predictive across all three cohorts, so it gates nothing and sizes nothing today.
- Conviction scoring 1–10 · retained, not trusted
- Per-signal audit chain, recorded internally
- Source quality weighting
- Duplicate de-clustering
On-chain monitoring
The chain-watcher reads wallet movement from Etherscan and DEX liquidity events from DexScreener, both public endpoints, and feeds them into the same stream as everything else.
- Labelled wallet tracking (Etherscan)
- DEX liquidity events (DexScreener)
- Large transfer detection
- Hyperliquid public perp positioning
News and social ingestion
RSS feeds, Twitter/X keyword monitors and Telegram channel readers run continuously. Items are filtered, de-duplicated against what has already been ingested, and stored with their source.
- RSS feed aggregation
- Twitter/X keyword monitoring
- Telegram channel reading
- Public sources only
Paper trading
Everything trades on paper against live prices, with taker fees and slippage in the model. This is the layer that measured the directional book and returned the verdict that closed it: 40% win rate against the 57.4% its payoff shape needed.
- Automatic paper execution of every signal
- Taker fees and slippage modelled per leg
- Equity, win rate and drawdown per cohort
- No real capital, ever
Operator notifications
One Telegram bot, reporting to the operator who runs the desk. There are no user accounts to subscribe and no alert product — the bot exists so a paper position opening at 04:00 is not discovered the next morning.
- Single operator channel
- Position, cohort and kill-rule events
- Muteable, and the mute is enforced at the sender
- No subscriber alerts
The desk dashboard
The desk interface: signal queue, paper portfolio, source health, and the audit trail from raw mention to final score. It is published as a demo build on synthetic sample data, behind a permanent banner saying so — the backend it would read from is not reachable from the internet.
- Signal queue with scores
- Paper portfolio and source health
- Per-signal audit drill-down
- Published at /app/ · synthetic data
REST API and webhooks
Bearer-token endpoints and HMAC-signed webhooks are written and covered by tests. Nothing is exposed publicly: there are no keys to issue, because there are no accounts to issue them to.
- Bearer-token API keys
- HMAC-SHA256 signed webhooks
- Signals, regime and IV endpoints
- Not reachable from the internet
Funding-carry forward-test
Short the Hyperliquid perpetual, hold the spot leg, collect the funding difference. Delta-neutral, so the part that failed — predicting direction — is not involved. It runs as a paper forward-test with an automatic kill rule.
- HYPE 5.6% · NEAR 5.2% · LDO 3.9% net APR
- Window 2026-05-10 → 2026-07-08, re-costed 2026-08-13
- Taker fees + 5bps assumed slippage, per leg, per side
- Kill rule: trailing 7-day funding sum under 7bps
The carry figures close a 60-day window that ended in July, so they are already old. They were 6.4–7.7% before slippage entered the cost model; that earlier version was an upper bound. LDO at 3.9% now sits below the 4.5% net APR the desk requires before a position is considered for capital, so on its own expectation it no longer qualifies.
Built, and not built
This page describes a research system, so it is worth being exact about which parts a visitor can actually reach. There are no plans and no accounts — the tier table that used to sit here compared three products that were never sold.
| Capability | State | Note |
|---|---|---|
| Signal ingestion from social, Telegram and on-chain sources | Runs | Public endpoints only |
| Clustering, source weighting and conviction scoring | Runs | The score measured anti-predictive, so it no longer gates anything |
| Paper trading with taker fees and slippage modelled | Runs | Against live prices |
| Hyperliquid funding-carry forward-test | Runs | On paper, with a kill rule |
| Directional signal book | Closed | 2026-07-08, after three unprofitable cohorts |
| Demo dashboard on synthetic data | Runs | Published at /app/, behind a permanent demo banner |
| Accounts, signup, billing | Absent | Never built |
| Public API and webhooks | Absent | Written, never exposed |
| Live trade execution with real capital | Never | Paper and testnet only, by standing rule |