Alpaca AI Trading Agents Hackathon · PAPER-only
The trading agent that
refuses to cheat on time.
Most agents ask “can the model pick winners?” Esscher asks a harder question: can you prove it never saw the future?
3
clocks kept separate — evidence / decision / market
0
live calls, credentials, or broker mutations in the repo
100%
fail-closed — unknown means refuse to act
PAPER
only. no real-money route, ever
The deterministic backbone
Every decision is a notarized fact, not a vibe.
Frozen strategy policy
A version-locked rulebook the system cannot quietly rewrite.
Any semantic change creates a new policy identity and restarts prospective evidence.
Point-in-time evidence
Every input stamped with published & retrieved time — no post-cutoff data enters a decision.
Browse the frozen 23-event universe and its provenance.
Bounded reasoner
Any plugged-in model answers only UP / DOWN / UNCERTAIN, with citations and its strongest falsifier.
See the gate that bounds the model.
Expression menu
A fixed menu for expressing a view — the model can never invent a position.
Prohibited controls are listed, not implied.
Atomic risk caps
Hard, account-relative loss limits enforced before any order exists.
Rehearse risk tiers on the live desk.
Monitored PAPER lifecycle
Simulated-money loop reconciled against the broker's own records, open to flat.
Watch decisions, vetoes and abstentions on the desk.
Trade Passport
Append-only, hash-linked log that lets anyone re-verify every decision.
Source of truth lives in the repo.
The bounded reasoner
The LLM is the least-trusted part.
Whatever model you plug in gets one call, one schema, and three answers. It cites its evidence and names its strongest falsifier. It has zero authority over instrument, size, price, or the broker.
// prohibited_llm_controls
symbol_eligibility · instrument · contract · strike ·
expiry · quantity · price · risk · entry · exit ·
account · broker_tool · policy_or_prompt_mutation
// the model proposes. deterministic gates dispose.
Team