trdrbot — an options-trading agent that knows the difference between right and lucky

One sentence: trdrbot senses markets and news, forms a falsifiable thesis, simulates several
option structures expressing it, sizes the winner by Kelly gated on its own earned track
record, executes and manages it through Alpaca, then scores afterwards whether its view was
right, its structure was right, or it just got lucky — and only the first two ever move its
confidence.

The loop

A scheduler wakes the agent every 60s. Cheap deterministic work — analytics, broker
reconciliation, exit-rule checks — runs every tick; the one LLM decision cycle runs every
~15 min. Three independent sources propose theses into one inbox: research (daily top-down:
regime + news + prediction-market odds → opportunities), discovery (bottom-up: news nominates
companies, a deterministic gauntlet of technicals/fundamentals/liquidity filters before an LLM
ever writes one up), and the muse (creative collision of random wiki concepts × news,
argued into chains, adversarially gated). The decide cycle reads the inbox and very often
declines — a no-op is a logged, legitimate answer, not a failure.

AI logic — what the model decides, what the code decides

The LLM's job is narrow: form a thesis, propose ≥2 structurally different expressions of it,
state a probability, and state a vol view if the thesis is about volatility. Everything after
that is arithmetic. simulate_experiments prices every candidate under one declared
measure — the thesis's own drift and (optionally) its own realized-vol forecast — so directional
and volatility theses are judged on the terms they actually claim, with the market's own pricing
shown alongside as the contrast. size_position computes Kelly from the structure's
conditional payoff (E[win|win]/E[loss|loss] — the naive max/max ratio is measurably biased
toward buying premium) and a probability shrunk toward the agent's own measured calibration: a
90%-caller who's right half the time gets cut down; a 70%-caller whose 70%s land 70% of the time
earns the full fraction. Facts (payoff at expiry, breakevens — contract arithmetic) and models
(P(profit), EV — a distributional assumption) are shown under separate headings, deliberately, so
neither gets mistaken for the other.

Risk gates — deterministic, no LLM veto layer

There is no separate approval step that vetoes an LLM decision after the fact. What replaces it
is stricter: the agent cannot execute a mistake the math itself refuses to compute. Kelly
refuses any structure with unbounded loss outright (unbounded profit, like a long call, is
fine — only the loss side is the danger). Size is capped three ways at once — per-position,
per-underlying, whole-book, all in dollars of defined max loss — and a payoff the sizing tool
can't verify against a real simulated structure refuses to size at all rather than falling
back to an optimistic estimate. Once open, a position's own agent-authored exit rules run against
the underlying every 60 seconds — debounced against quote artifacts, but decisive the instant the
underlying itself corroborates a real move. Two rules the agent cannot override: a
competition-deadline sweep, and a rule that closes a position outright if a leg vanishes at the
broker (early assignment) — the survivor of a broken spread can be an unbounded naked position,
worse than the one it replaced. A "no trade" verdict from any gate is a correct answer, not a
fallback.

Alpaca infrastructure

The Alpaca MCP server is the sole broker interface — one stdio session shared for a whole tick
rather than re-spawned per call (−78% wall-clock) — for order placement, the options chain,
quotes, and account reads. Every order's client_order_id is derived deterministically from the
decision batch and enforced at the tool-call boundary, so a crash-and-retry resumes the same
intended order rather than opening a second one. A reconciler runs first every tick, diffing
Alpaca's real holdings against trdrbot's own records before the exit-rule evaluator sees them,
so a fill, assignment, or external close is never evaluated against stale local state.
trdrbot doctor probes every configured LLM and the live MCP connection before anything trades;
trdrbot health separately asks, of every subsystem, "did it run and produce, or run and
silently do nothing" — the failure mode that has actually cost this project more than any
exception.

The differentiator: scored honesty, not just P&L

A genuinely 60%-edge agent only beats a coin flip 69% of the time over 20 trades — measured, not
assumed. So position size is earned, on a four-tier ladder gated on resolved theses, calibration
reliability, and — the distinctive part — attribution rate: whether the agent could actually
explain why it was right, not just whether it made money. A profit on a wrong thesis is excluded
from what lets it size up, by construction.

Honest limitations

Calendar/diagonal spreads are refused, never approximated, absent a pricing model for the far leg.
No separate guardrail layer exists by choice — the deterministic gates above are the guardrail.
Calibration is young (paper account, one week); every number reported carries its sample size.