The Signal / Transmission Systems · Build in public

Field Notes · Paper Trader · №021

A trading desk built to disprove its own edge.

Paper Trader is a paper-only trading OS that reads the market across a dozen sources, argues with itself before it acts, and keeps a ledger honest enough to tell me when I'm wrong. No real money. No fabricated backtests. Just the one question that actually matters: does it beat SPY, net of slippage?

ObserveReasonActMeasureRememberAdapt
ModePAPER · no real money
Equity$100,000
BenchmarkSPY
ScoringNet of slippage

I

The premise

Most trading bots are one script with a confident opinion and a returns screenshot you can't reproduce. I wanted the opposite: a system that behaves like a disciplined desk, not a lucky gambler. That means it's allowed to wait. It's allowed to sit out. And it's built to catch its own bad trades before they happen — not just find more of them.

The rule is painted across the entire stack, in the platform's own words: validation = beating SPY over weeks, net of slippage. Trade count is not edge.

II

What makes it different

Six choices separate this from the average bot. None of them are features you bolt on — they're the difference between a gambler and a desk.

Multi-model. A council deliberates before a trade — it isn't one model's hunch.not single-model
Stateful. It carries regime, posture, and open risk between decisions.not stateless
Memory-backed. It recalls what past setups actually did.not prompt-only
Governed. A deterministic policy has veto power over every action.not free-running
Self-healing. The runtime checkpoints, restarts, and resumes on its own.not manual-only
Measured. Every call is scored against what happened next.not vibes-based

III

How it thinks

The whole system runs one loop, over and over — and the models never pull the trigger. They inform the decision; a deterministic policy makes the actual call.

Data Awareness Reasoning Policy Execution Outcomes Learning Memory
Layer 01

Market Brain

Classifies the regime — risk-on, risk-off, high-vol — and sets posture and size. Equities and crypto are read separately.

Layer 02

Reasoning Council

Several models deliberate, then hand back the consensus, the dissent, and a chair's synthesis — not a single black-box answer.

Layer 03

Prediction Engine

Turns news and community chatter into ranked, testable hypotheses — then grades each one against what actually happened.

Layer 04

ML Layer

Learns a score from outcomes, with champion/challenger promotion: a new model only takes over if it genuinely beats the old one.

Layer 05

Policy & Risk

The deterministic chokepoint. Trailing stops, a daily circuit breaker, 1:2 reward:risk, liquidity gates. Built to fail safely.

Layer 06

Memory & Repair

Long-term recall plus a self-healing runtime that checkpoints and resumes — so the desk keeps running and remembering instead of forgetting.

IV

What it's made of

Six named bots, each with its own ledger, so strategies compete against each other instead of one pretending to be magic. A dozen live feeds — equities, crypto, filings, and the noise floor of the internet. And underneath it, a universe built to fight the biggest lie in backtesting:

21,692
securities in the point-in-time universe
9,695
delisting events — the dead names most tests ignore
899
S&P 500 membership intervals
398
removed names recovered
6
independent bots, independent ledgers
~12
live data feeds, equities to crypto to SEC

Most backtests only test on the companies that survived — which is like judging a strategy by asking the winners. Paper Trader tests against the graveyard too.

V

The honest ledger

Here's the part no trading ad will ever show you: what didn't work. Every verdict below is one the system reached by trading against itself — and, often, losing.

+Trend-timing to cash improved risk-adjusted returnsHeld up
+Tighter controls and sharper selectivity beat adding more signalsHeld up
Naive shorting did not magically winDisproven
Crypto momentum rotation failed hardDisproven
Overtrading quietly bled the accountDisproven
Ignoring survivorship bias faked great-looking resultsDisproven

Four red marks against two green ones. That's not a hole in the pitch — that is the product. A machine that tells you the truth about your ideas is worth more than one that flatters them.

VI

Why I'm building it in public

The edge is still being proven, not assumed — and I'd rather show you a machine that's honest about that than sell you a fantasy.

It runs the full loop now: observe, reason, act, measure, remember, adapt. The next milestones aren't more surface area — they're more resolved outcomes and deeper history. More truth, not more hype. That's the only thing that turns a well-built desk into a proven one.

Paper money Real discipline Zero placeholders