Athean Council — BTC $120k by 2026-12-31
Identical +17pp edge signal — clean portfolio, no correlated BTC exposure already on the books. Eleven agents deliberate, Themis resizes from raw half-Kelly 10.5% to a category-capped 5% NAV, Areopagus approves. Watch the four rounds and the final size land at the constitutional cap.
Sign one tx. Your wallet lands on chain.
Optional — read-only. The demo never transacts. Connect to see the same SIWE flow the production API uses, switch to Mantle Sepolia so the on-chain Proof of Restraint records render properly, and sign a verification message in your wallet.
No injected wallet detected. Install MetaMask, Rabby, Coinbase Wallet, or similar and reload.
What this proves: the Mantle Sepolia pipeline works for visitors end-to-end (wallet → contract → arcscan → dashboard feed). What it does NOT prove: that the council picks profitable trades. See the empirical backtest panel below for that. Cost: fractions of a cent in testnet USDC.
No injected wallet detected. Install MetaMask, Rabby, or Coinbase Wallet and reload.
Bybit Testnet
Bybit Testnet provides free USDT for paper trading. Create a testnet account, fund it from the faucet, and the pipeline uses your Bybit paper balance.
Bybit Testnet FaucetWatch the council deliberate.
Twin scenarios · same signal, different portfolio
Approval and Restraint scenarios share the identical +17pp edge signal (0.59 oracle vs 0.42 market). What differs is the portfolio state: clean book → APPROVED, already-correlated book → VETOED. An edge alone does not justify a trade.
Source · Dialogue is captured from prior live Gemini runs. Signal scoring, Areopagus verdict, half-Kelly sizing, and paper trade all compute through production code paths (apollo.scorer, areopagus.court, strategos.paper).
Real BTC bars, naïve estimator.
⚠ This panel lost money — by design
The strategy is naïve momentum (long if last bar up, short if down) — a deliberate stand-in for the council. The point is to prove the plumbing survives real market data: 4% round-trip fees mathematically kill momentum. The council's job is to refuse those trades.
Run: scripts/live_paper_trade_coingecko.py on 7 days of BTC/USD hourly bars. Fills route through the real strategos.paper.PaperBook with half-spread, slippage, and 2% taker fees.
Does the council help?
200 resolved Manifold binary markets, run through the 5-role council via scripts/backtest_sources_xml.py. Brier decomposition, per-source verdicts, $0.12 cost on Gemini flash-lite.
Honest framing: the council does not beat free human consensus on this 200-market Manifold sample (0.149 vs 0.126 Brier). What it does do is close 80% of the gap between a single-shot LLM and the human prior. Aggregation across 5 distinct roles is worth ~0.11 Brier — an order of magnitude more than any individual data source. Total cost $0.120 on Gemini flash-lite (council, 5 roles × 8 batches).
| Forecaster | Brier | Reliability | Resolution |
|---|---|---|---|
Manifold consensus (baseline) free play-money human consensus — the wall to beat | 0.126 | 0.012 | 0.123 |
Single-shot Gemini raw LLM alone — worse than play-money humans | 0.260 | 0.046 | 0.049 |
5-role council aggregated closes 80% of the gap to Manifold | 0.149 | 0.019 | 0.088 |
| Source | Δ Brier | Applicability | Verdict |
|---|---|---|---|
attention (Wikipedia) | -0.0040 | 33.5% | ADOPT |
crowd_sentiment (Nitter) | -0.0040 | 36.0% | ADOPT |
geopolitical_risk (GDELT) | -0.0013 | 8.0% | HOLD |
lead_lag (TradingView) | -0.0006 | 14.0% | HOLD |
perps_signal (Binance) | -0.0001 | 6.3% | HOLD |
orderbook_imbalance | -0.0001 | 8.0% | HOLD |
onchain_tvl (DeFiLlama) | +0.0004 | 7.5% | HOLD |
basis_arb | +0.0003 | 6.0% | HOLD |
cot_positioning (CFTC) | +0.0010 | 5.5% | HOLD |
consensus_delta | — | 0.0% | UNTESTABLE |
macro_basis (FRED) | — | 4.0% | UNTESTABLE |
macro_release_consensus | — | 4.0% | UNTESTABLE |
What this means: council aggregation alone is worth ~0.11 Brier vs single-shot, while individual source signals are worth ~0.003–0.004 each. Aggregation > sources by an order of magnitude. The two ADOPT sources (Wikipedia attention + Nitter crowd_sentiment) are already wired in apollo.scorer at ±0.05 contribution caps. The HOLD sources may shine on a Polymarket-flavoured corpus once the geo-block proxy is deployed.
Where the prior comes from.
12 sources are plumbed in. Only 2 have survived empirical Brier-delta testing — Wikipedia attention + Nitter crowd_sentiment. The other 10 are HOLD or UNTESTABLE pending a Polymarket-flavoured corpus. A source only graduates to ADOPT when its negative Brier-delta beats noise.
| Source | Apollo feature | Mechanism | Auth | Tier |
|---|---|---|---|---|
| Odds API | basis_arb | Sportsbook consensus vs Polymarket — vig-stripped, fee-net basis | free key | high |
| Kalshi REST | basis_arb | US-regulated event-contract prices; cross-listing arb with Polymarket | none | high |
| Manifold Markets | consensus_delta | Free human-consensus prior; wide gap = wider council, smaller size | none | high |
| Binance perps | perps_signal | Funding-rate z + OI delta; contrarian on extreme over-leverage | none | high |
| Polymarket L2 WS | orderbook_imbalance | Live CLOB book depth; per-side imbalance into oracle prior | operator key | high |
| Source | Apollo feature | Mechanism | Auth | Tier |
|---|---|---|---|---|
| CFTC CoT | cot_positioning | Weekly speculator-net z vs 26-week history; contrarian on crowded longs | none | medium |
| GDELT 2.0 | geopolitical_risk | Real-time global news volume + tone, 100+ languages, 15-min cadence | none | medium |
| FRED | macro_basis | 816k US macro time series; Fed / CPI / NFP gaps vs threshold | none | medium |
| Wikipedia pageviews | attention | Per-article hourly pageview velocity z-score = investor-attention proxy | none | medium |
| DeFiLlama | onchain_tvl | Chain TVL + stablecoin mints + protocol yields — flow indicator | none | medium |
| Source | Apollo feature | Mechanism | Auth | Tier |
|---|---|---|---|---|
| Nitter (X/Twitter) | crowd_sentiment | Free RSS scrape; in-tree VADER-style scorer for retail-flow sentiment | none | experimental |
| TradingView screener | lead_lag | Cross-asset technical screens; macro lead/lag for derived markets | none | experimental |
Falsification protocol · all 12 sources plumb into apollo.scorer.MarketSnapshot. A source graduates HOLD → ADOPT only when its paired Brier test beats the council baseline by >0.002. Non-adopted sources stay live for telemetry; their oracle delta is set to zero. No source ships unfalsified.