🤓 Latest Submissions
ClearSpread
ClearSpread is a transparent, human-approved OSINT options agent built on Alpaca's official CLI. "Transparent, human-approved, defined-risk" is table stakes here, so we built what that claim usually leaves out. REAL MULTI-AGENT DISAGREEMENT. Three separate Gemini calls (News, SEC Filings, Price/Momentum) each see only their own slice of the data, never each other's output. A fourth Critic agent gets all three opinions and must name any disagreement explicitly as a risk flag instead of averaging it away. On a real NVDA run, News came back bullish 88/100 on acquisition headlines, Filings neutral 65/100 after finding three executive Form 4 insider sales, and Price bullish 70/100 on a 5% rally that happened on 0.81x average volume. The Critic caught the conflict, named it in the risk flags, and synthesized a conservative bullish 72/100. ACTOR-CRITIC VALIDATION. Every agent's output passes a deterministic, non-LLM Reviewer before it is accepted: schema, enum and range checks. Invalid output is re-prompted with the exact errors, up to two revisions, then the pipeline fails loudly. A HASH-CHAINED LEDGER, NOT AN "AUDIT LOG" IN NAME. Every decision, human action and execution appends to a SQLite ledger where each row's SHA-256 hash covers the previous row's hash plus its own payload. One click recomputes the chain and reports where it breaks if anything was edited out of band. DEFINED RISK ONLY, PROVEN LIVE. The engine proposes debit spreads exclusively, so max loss is known up front and no naked option selling exists anywhere. Real multi-leg orders went out on a fresh, hackathon-dedicated paper account with options level 3, confirmed on the exchange with real order IDs. APPROVAL IS A HARD GATE. trading/executor.py, the only code path that can reach the order book, is never called except from the Approve button. A MAX_NOTIONAL_PER_TRADE guard blocks oversized orders even after approval, and a blocked approval is logged into the same chain. Nothing quietly disappears.
4 Sep 2026
