ALPACA PAPER · AUTONOMOUS OPTIONS RESEARCH
Three strategy agents challenge the same market. A critic questions every thesis. Deterministic risk gets the final veto—and every refusal leaves evidence.
LIVE BROKER OBSERVATIONS
Source-labelled Alpaca option snapshots. Stale observations remain visible for evidence but cannot authorize execution.
The backend returned no current contract observations. ThesisCircuit will not substitute prototype prices.
DETERMINISTIC FEATURE ENGINE
A rule-based classification recomputed each cycle. Missing or stale inputs resolve to uncertainty, not confidence theater.
Regime state will appear only from a recorded backend cycle.
MULTI-AGENT OPTIONS ENGINE
Three independent theses compete on the same recorded state. Capital allocation remains zero unless every hurdle clears.
No strategy state is fabricated while the research ledger is unavailable.
ADVERSARIAL GOVERNANCE
Agent proposals flow through an independent critic, allocator, and deterministic risk officer.
No debate or verdict is synthesized without backend evidence.
ACTUAL ALPACA PAPER POSITION
Broker-reported valuation and source-labelled option metrics. Recommendations are research only; closing authority is disabled.
Only an actual broker-reported position can populate this panel.
PRESERVED PHASE 1 EVIDENCE
One controlled opening order, reconciled from Alpaca and persisted in Supabase. No additional or closing order was sent.
Readiness records are never presented as executed orders.
DETERMINISTIC CAPITAL GOVERNOR
The agent layer cannot override these checks. Any blocked or unknown condition prevents broker dispatch.
An absent risk decision never implies approval.
COUNTERFACTUAL · NEVER SENT TO ALPACA
Rejected ideas are marked against later observed bids to measure decision quality without claiming fills, fees, or executable returns.
The system does not invent counterfactuals when no eligible rejected proposal exists.
EVIDENCE-WEIGHTED ALLOCATION
Scores shrink toward a neutral prior with sparse samples. Phase 1 performance is not attributed to these research agents.
Leaderboard values appear only from recorded evidence.
FINITE DRY-RUN DISPATCH
Recorded research cycles are idempotent and non-overlapping. Autonomous broker execution remains disabled.
No autonomous cycle values are synthesized.
IMMUTABLE RESEARCH LEDGER
A replayable stage trail from underlying selection through risk, shadow research, and position review.
Audit stages appear only from a persisted cycle.
BUILT TO FAIL CLOSED
Execution is a separately authorized, durable transaction—not a side effect of agent output.
One active research cycle; stale locks expire safely.
Every broker action is recorded before network contact.
Database coordination prevents overlapping submission ownership.
Unknown broker responses are queried by client order ID.
Opening, closing, and total order budgets are atomic.
Completion, expiry, kill, or failure forces execution off.