Signal tiers

Scienter’s engine grew in numbered “tiers,” each answering a progressively harder question. This page describes what’s actually shipped per tier, distinct from what’s still a roadmap entry. If a capability isn’t listed as shipped, don’t build against it — it isn’t there yet.

​

Tiers 1–5 — the foundation

“Is there a trade here, and who’s actually worth copying?”

• Copy-trading engine. Scans ~40,000+ Hyperliquid traders, HL vaults, OKX lead traders, and Polymarket, judges candidates with the Deflated Sharpe Ratio against the entire scanned population, and watches the survivors live for entries.
• The edge engine — four independent signal kinds, each optional and keyless where possible:

  Signal	Source	Keyless?
  Funding divergence	Hyperliquid metaAndAssetCtxs	Yes
  Liquidation clusters	Hyperliquid clearinghouseState (real positions, not modeled leverage)	Yes
  Exchange netflow	CryptoQuant / Glassnode	Needs a key
  Whale wallet buys	Helius + Kolscan/Cielo	Needs a key

• Conviction scoring. A 0–100 score built from independent groups, not raw signal count — two reads from the same group (e.g. funding + liquidations, both reading Hyperliquid positioning) count as one corroboration, not two, so conviction can’t be manufactured by counting correlated inputs twice.
• Free / pro tiering. tier=free delays evidence by 4 hours. This is a rendering choice today, not an enforced paywall — see API Reference for what that actually means.

​

Tiers 6–10 — the quant layer

“Has this ever been worth anything, and can we execute it well?” Nothing in this tier places, sizes, or closes an order — by construction, not configuration.

​

Tiers 11–12 — institutional risk

“How much can this book lose, how big should each leg be, and does anything independent agree?”

• Risk (11d). Three VaR models always run together (historical, Gaussian, Cornish-Fisher) — when they disagree, the disagreement is the finding. Expected shortfall, correlation-regime detection, and a circuit breaker that survives process restarts.
• Portfolio sizing (11e). Quarter-Kelly sizing off the ensemble’s own closed, marked-to-market history — never a default or an assumed win rate.
• LLM council (11c). Five frontier models (Claude, GPT, Gemini, Grok, Llama) independently judge each high-conviction signal; disagreement is surfaced, not averaged away.
• Forecasting (11a). 24h return distributions, not point estimates, with a running calibration check against actual coverage.
• Signal weights (11b) & wallet clustering (12c). Held-out A/B testing before any weight change is even proposed; a hypergeometric test for coordinated wallets, correcting for the number of pairs tested.
• Alt-data (12a/12b). On-chain (Bitcoin mempool, keyless) is live. SEC Form 4 and additional chain feeds are wired but currently dormant pending configuration.

This tier works with zero configuration for the core risk math (it runs off public Hyperliquid candles). Optional API keys unlock the LLM council and a foundation-model forecaster — anything without a key reports itself dormant, with the reason, not a silent degrade.

​

Tiers 13–15 — specced, not built

These exist only as roadmap entries with honest cost/benefit writeups, not code. In order of what the roadmap itself rates as worth doing first: coordinated-wallet (“airdrop farming”) detection, MEV-density as a contrarian topping signal, retail-spending proxies, and satellite-imagery macro signals (explicitly not recommended by the roadmap’s own analysis — weak, unstable correlation to the asset class at Scienter’s holding horizon). None of this is live. Don’t document or build against it as if it were.

​

Tier 16 — premium UI

A wallet-connect dashboard scaffold (React, wagmi, RainbowKit) exists as a skeleton with no built views yet. The dashboard you’ll actually use today is the self-contained HTML UI served directly by ghostcopy/server.py (/, /edge, /edge/risk, /edge/ledger, /execute).