The three pillars, in depth.

MD — Markets Directional

MD is the directional perp-trading pillar on Hyperliquid. It identifies directional setups (long or short), executes them with venue-side SL and TP, and broadcasts each executed trade to Telegram.

The signal flow: every 90 minutes (at Tier 0; faster at higher tiers) the engine pulls the top 30 Hyperliquid perps — crypto and HIP-3 TradFi markets alike — by a composite score that combines log-volume, momentum quality (a bell curve peaking at 4–10% 24h move — this is the "pre-breakout sweet spot"), and ATR (real-time volatility). The composite is deliberately tuned to reject already-extended chase candidates and surface setups with room to extend.

Each candidate is then evaluated by a dual-triage: Gemini Flash + Claude Haiku in parallel, looking for a pattern match against a known taxonomy (TREND_CONTINUATION, BREAKDOWN_SHORT, FUNDING_FADE, MEAN_REVERSION_FROM_EXTREME, VOLATILITY_EXPANSION). Triage is cheap, fast, and intentionally permissive — its job is to surface candidates, not gate them.

Survivors of triage go to the 3-vendor judge committee: Claude Sonnet, xAI Grok-4.3, and Google Gemini-Flash judge each setup independently, with full asset context (news, regime, funding state) injected into the prompt. Two of three must PASS at confidence ≥ 0.65 for the trade to fire. The vote is egalitarian — no single model can gate or override.

ORACLE — Polymarket prediction markets

ORACLE trades binary YES/NO contracts on Polymarket. The pillar's edge is in reading world events better and faster than the market consensus, then betting against any implied probability that's drifted away from fair value.

The universe filter is strict: markets must resolve at least 48 hours from now (to eliminate same-day anomaly trades), have at least $10,000 of liquidity, and trade at neither extreme (yes-price between 0.05 and 0.95). Markets that pass the universe filter feed into the same dual-triage / 3-vendor-judge stack as MD.

ORACLE also has a deterministic POST-EVENT GUARD: a hard-coded set of rules that blocks any market within 24 hours of resolution if the price is at or near extreme (≥ 0.95 or ≤ 0.05). This is the safety net for the "the market knows the answer already" failure mode that lost real money on an early ORACLE trade.

The other ORACLE strategy is PCE_v2 (Patient Capital Edge) — an older mean-reversion-vs-FRED-model strategy that fires high-conviction low-edge trades like "Fed will hold rates at the June meeting" at near-1.0 prices. Asymmetric R:R, but high hit rate. These currently fire as manual-approve while we evaluate folding them into the AUTO_VETO pipeline.

LP Agile — concentrated liquidity income

LP is the income engine. The pillar provides liquidity to concentrated-liquidity pools on Aerodrome (Base) and prjx (HyperEVM), earning trading fees plus emissions when the asset trades within your chosen price range.

The hard problem in concentrated LP: range selection. Too tight and the asset exits range quickly, you stop earning, and you eat impermanent loss. Too wide and your capital efficiency is low. LP Agile computes a target range based on live ATR (so the range adapts to volatility regime), the pool's historical price distribution, and the operator's reinvest-vs-take-profit ratio.

Every 30 minutes the LP scanner re-evaluates open positions. If a position's price exits the range, the scanner suggests a rebalance (close + open with new range) or a wait-and-see (if the price excursion is small and likely temporary). Operator approves via Telegram; engine executes.

The harvest engine runs hourly: when accrued fees + emissions cross a threshold (default $1.50 USD worth per position), the planner generates a harvest transaction that claims, swaps to USDC, and reinvests 80% / takes 20% as cash profit. This converts compounding into income on a continuous basis.

Funding-rate strategies — inside MD

Perp funding rates are themselves a signal. When funding on a Hyperliquid market reaches an extreme, the crowd is leaning hard in one direction — and paying for the privilege. The FUNDING_FADE pattern in MD's triage taxonomy trades exactly this: enter against the crowded side, collect the funding tailwind, and exit when funding normalizes.

Historically this ran as a standalone cross-venue funding-arbitrage pillar. That pillar has been retired and absorbed into MD: with all perp trading consolidated on Hyperliquid, funding extremes are now just one more trigger in the MD pipeline — same dual-triage, same judge committee, same Telegram delivery as every other MD trade.

How the AI judge committee works

The committee is a deliberate trustless construction:

• Three different vendors — Anthropic, xAI, Google. One vendor going down doesn't take the whole committee down.
• Three different models — Claude Sonnet, Grok-4.3, Gemini-Flash. Each is biased differently; consensus is more meaningful than any single verdict.
• Egalitarian voting: each voter has equal weight. The incumbent Claude does not get veto power; the challengers do not get veto power. Two of three must PASS for execution.
• Confidence floor: a PASS at 0.50 doesn't count. The floor is 0.65, calibrated against historical signal performance.
• Direction-aware prompting: the judges see the live market regime (BTC trend, funding state) and explicitly evaluate both long and short setups symmetrically. The system isn't biased toward longs.

When a judge returns UNAVAILABLE or PARSE_ERROR, the vote falls back to the remaining live voters. If only one voter is alive, no signal fires — we never trade on a lone opinion. This sometimes means a slow news day fires zero signals; that's correct behavior.

The graduation tier system

The engine starts at Tier 0 (UNPROVEN): 90-minute scan cadence, $1/day combined AI spend cap. The faster cadence and higher spend only unlock after the engine actually closes winning trades:

• Tier 0 → 1 after 1 winning closed trade with +PnL.
• Tier 1 → 2 after 5 winners + 7-day cumulative positive PnL. Cadence drops to 60 min, cap to $4/day.
• Tier 2 → 3 after 28-day cumulative positive PnL. Cadence drops to 30 min, cap to $6/day.

This is deliberate: an unproven engine doesn't earn the right to spend more on AI calls. The cost scales with proven edge.