The oracle that says no.
A panel of seven specialists debates every trade. The voice that disagreed survives in the output — cited, signed, on-chain. Default posture: REJECT.
Built for operators who already run someone else's bot or copy someone else's trader, and who'd rather see one trade refused for a defensible reason than ten trades fired on rubber-stamp consensus.
Early access to the oracle. Get notified when you're approved.
What you actually pay
Three options. No subscription. No seat fees. No "contact sales."
No subscription. No seat fees. No "contact sales." The agent runtime is free and open-source — you self-host the proof. Pay only when you ask the Oracle a question.
Five voices. One verdict. The dissent stays visible.
Each voice runs on the same snapshot. None hedge. The voice that disagreed survives in the output — labeled, sourced, scannable.
chart_analyst
Reads indicator confluence — ADX, RSI, MFI, EMA, Bollinger. Grounded by the investor-canon corpus when retrieval is available.
regime_analyst
Deterministic chop / trend classifier. Breakout in chop raises the bar; only the cleanest setups pass. No vibes.
memory_voice
Grades the realized outcome of past calls, not its own prior confidence. The only voice allowed to say ‘we've been wrong here before.’
risk_voice
Hard veto. Account caps, hourly drawdown, budget bound — enforced in code, not in a prompt. The LLM cannot overturn it.
strategist_voice
Pre-execution devil's advocate. Reads the same snapshot the chart_analyst sees and tries to break it. Never returns ‘bullish’ — the strongest signal it can emit is ‘I tried to find the bear case and could not.’
Two more voices — catalyst (news-wire surfaced events) and research(TVL / unlock / on-chain freshness) — feed the strategist and risk floors. They don't vote; they constrain.
A bot would have executed at 0.68 and stayed silent.
This is the real output of the agent runtime in paper mode. The Dissent: line is the product.
> Should I enter PYTH now? Trader UID 481923 just opened a position.
[Trader grade: D — degrading 30d → 90d]
[Trade verdict: REFINE]
BUY PYTH @ 0.142
Dissent: risk_manager — "exit liquidity insufficient on this
timeframe; cohort Solana_MINUS"
Falsifier: exit if 5m close < 0.138 OR 4h ADX drops below 20
Confidence: 0.42 (strategist downgraded from 0.68)
$0.25 settled · sha256·a3f7c8d2e1…The Dissent: line is the product.
A generic oracle returns confident consensus. Gecko returns the voice that disagreed, the reason, and the falsifier — before you click.
Marina sees what the panel almost ignored. That is the difference we can prove.
10 strategies tested. 10 rejected. That is the product.
An oracle that cannot kill its own claims has no claim worth trusting.
The waitlist gate
Gecko stays on a small waitlist until our own strategy validates against the bar: ≥6 consecutive +EV weeks · DSR ≥ 0.9 · CI lower bound > 0 · PBO < 0.5 · at least one falsified hypothesis published during the window. We are at week 1, with 3 closed trades (2W/1L) on paper. That is encouraging. It is explicitly not validation, and we will not call it that.