+27.46 bps aggregate mean signed test return.
Alpaca AI Trading Agents Hackathon · paper trading only
24/7 crypto context.
Bounded U.S. options execution.
ClockCross measures BTC-to-COIN repricing gaps, falsifies the signal chronologically, asks AI for a tightly bounded thesis, then lets deterministic risk gates decide whether a defined-risk Alpaca options spread is allowed to exist.
ResearchMUTATE
Six-month replay36 AI trades
Live candidateCOIN only
Competition statusEpisode verified
Evidence, not promises.
The research was allowed to say “no” — and the backtest was allowed to find a bug.
The first Alpaca research run returned MUTATE, not GO. After the first autonomous competition episode, a literal chronological replay then exposed a structural spread-construction defect. ClockCross preserved both results, fixed only the structural defect, and left the signal, timing, AI policy, and risk budget frozen.
01 · Initial falsification
What survived the first research pass
Historical window: 2025-01-02 → 2026-08-28. Equity history used consolidated SIP data. Numbers below are underlying-return diagnostics, not an options-fill backtest.
The recent COIN regime survived while the broader thesis was narrowed.
MSTR deteriorated and remains intentionally execution-ineligible.
The broad-market control remained materially weaker than the crypto-sensitive surface.
The failure stays in the record
Both COIN and MSTR failed the original 100 bps underlying-friction promotion check. ClockCross did not erase that result. The approved mutation narrowed execution to COIN and asked whether a current signal could be expressed through an actual 7-21 DTE defined-risk spread.
02 · Literal chronological replay
Six months, day by day — including the bad regimes.
March 2 → September 1, 2026 · 128 actual Alpaca market days. No historical order endpoint was used. Directional returns are diagnostics, not compounded account returns.
58.3% directional hit rate and +50.7 bps mean chosen-direction 60-minute COIN return.
+15.9 bps mean directional 60-minute return. The AI beat it over six months, not in every regime.
June: -38.3 bps
July: -57.1 bps
23 quarter signal contexts × 3 calls; 23/23 unanimous actions and 0 action-flip contexts.
The recent 5/5 is not the headline
From August 3 through September 1, the current AI chose the correct 60-minute underlying direction on all five signal days while frozen continuation was correct on one. The six-month replay is shown above precisely so that favorable recent month is not mistaken for a universal regime.
03 · Backtest → defect → bounded fix
The backtest found a production-selection pathology.
The fix was structural, predeclared, symmetric for calls and puts, and did not retune the residual threshold, beta, AI prompt/model, risk budget, entry time, or exit horizon.
01
Defect exposed
FOUNDThe old net-delta/debit objective could prefer extremely far-OTM shorts with ~0.00-0.02 absolute delta, creating 50-60 point-wide spreads with poor historical short-leg print coverage.
02
0.10 short-delta floor
FIXEDThe single promoted change requires the short leg to retain at least 0.10 absolute delta. It kept a candidate on all 5 recent signal days.
03
Structure improved
BETTERMean width fell from 56.9 to 29.4 points; historical entry-print coverage improved 40% → 80% and exit-print coverage 20% → 40%, while mean net delta remained ~0.370.
04
Post-fix stress
PASS12,000 live indicative-surface constructor trials remained 100% constructible in both directions with zero risk-invariant violations.
04 · External preflight
Five live surfaces checked. Five passed.
Encrypted development environment · 2026-08-30. The preflight is read-only and cannot create a trading episode.
01
Paper account state
PASSActive and not trading-blocked.
02
Options permissions
PASSApproved and current trading capability satisfy Level 3 spread requirements.
03
COIN option-chain coverage
PASS416 parseable contracts found in the 7-21 DTE window on the indicative feed.
04
Alpaca MCP
PASSRead-only get_clock invocation succeeded through the official MCP surface.
05
Bounded AI gateway
PASSThe deployed model returned a schema-valid decision under the exact ClockCross decision contract.
competition episode · 2026-09-01
$ clockcross competition-session --date 2026-09-01
preflight: PASS
opening MLeg: filled
research-horizon close: filled
terminal_state: CLOSED05 · Autonomous competition proof
The full paper lifecycle executed.
On 2026-09-01, the dedicated competition account passed pristine preflight, the opening Alpaca MLeg filled, and the deterministic 10:55 ET exact-contract closing MLeg filled on its first close attempt. The persisted lifecycle reached terminal state CLOSED.
No unresolved opening or closing order remained and no manual trade-selection intervention was used. This proves the lifecycle, not future profitability; final event account equity/P&L is reported separately from research diagnostics.
06 · Autonomous decision path
AI has authority. It does not have the keys.
The AI can choose a thesis only after evidence gates pass. Contract selection, position sizing, freshness, buying power, and execution remain deterministic.
- BTC
24/7 market move
- β
rolling BTC/COIN beta
- Δ
09:25 ET premarket residual
- OOS
chronological evidence gate
- 09:40
opening confirmation frozen
- 09:55 ET
earliest decision boundary
- MCP
read-only Alpaca context
- AI
bounded thesis
- RISK
deterministic governor
- MLeg
paper limit order
- DB
durable decision ledger
07 · Deterministic safety envelope
No thesis can override these gates.
COIN onlyApproved mutation surface
7-21 DTENo 0DTE contracts
0.45-0.65 ΔLong-leg absolute-delta window
≥ 0.10 ΔShort-leg absolute-delta floor
≥ 0.30 ΔMinimum net directional delta
1:1 verticalsSame expiration, defined debit
Limit MLegNo naked option execution
1%Max defined loss per position
5%Max aggregate defined loss
Fresh quotesStale or incomplete chains fail closed
Idempotent ordersAmbiguous outcomes reconcile; never blind retry
development safety proof · 2026-08-30
$ clockcross smoke-mleg
{"ok": false, "error": "U.S. market is not open; refusing MLeg smoke"}08 · Fail-closed order proof
Unsafe conditions stop before submission.
Development-order opt-in was deliberately enabled while the U.S. market was closed. ClockCross consulted the market clock and refused before option-chain construction or order submission. No smoke order was created.
Development and competition paper accounts remain separated; the competition account is never used for smoke testing.
09 · Alpaca implementation
Sponsor technology is inside the decision path, not added as decoration.
BTC and equity history drive the residual research, six-month replay, and live signal reconstruction.
Actual COIN contracts drive deterministic feasibility and spread construction; historical option prints bound what can honestly be reconstructed.
Read-only account/market context is captured as an auditable AI evidence surface.
Atomic paper MLeg limit orders use deterministic identities, persisted recovery, and exact-contract exits.
Current competition state
Research replayed.
Autonomous lifecycle proven.
The first dedicated-account competition episode is externally verified and the post-episode constructor correction is merged with full CI. This remains a static, zero-secret judge surface: no trading controls, credentials, account connection, or mutable application state are exposed. Final event account equity/P&L is captured separately before submission rather than inferred from research or proxy returns.