B-01 · Soliton — a live paper-trading experiment
A frontier AI model (Claude’s most advanced) runs two paper-trading accounts in public — one making fast options bets, one investing off a written thesis on the AI economy. This strip is the scoreboard; the plain record is right below.
No sessions logged yet — the record starts the day the accounts take their first action.
The experiment: two paper accounts are run by Fable, the most advanced Claude model available over an API, both under a daily-trade mandate (each must act at least once a session — so a trade here is not automatically conviction). Fable-aggressive (FA) makes short-dated options bets (cheap bets that pay off big if the market lurches around scheduled events like a CPI print or a Fed meeting) inside caps enforced by code, not prompt. Fable-economist (FE) keeps a written thesis journal on the AI economy — with web search as its eyes — and buys the companies it argues are the leverage points. Everything else on this page is a reference line to beat: a mechanical control strategy, a market-state overlay, and SPY buy-and-hold.
Where it came from: this plate used to be MERIDIAN — five specialist LLMs deliberating over filings and news sentiment. It produced sophisticated reasoning and, in two months, exactly one trade: an analysis engine with no defined edge. SOLITON is the rebuild (MERIDIAN is kept as an archive). Every strategy here is a mechanical, pre-registered rule set, backtested before it touches even paper money, and run in public.
Success was defined before launch: beat the controls over 100+ logged trades with every cap respected — not “up in week one.” Paper money throughout.
How to read this: two paper accounts, one Fable API call each per session. FA — fable-aggressive — buys short-dated call and put spreads on SPY or QQQ off the macro calendar, from a $5,000 ring-fenced bankroll with per-punt caps and loss stops enforced by code. FE — fable-economist — keeps a running thesis journal on the AI economy and buys the leverage points it finds (long-only equities, $10,000 bankroll). The model decides; the code builds every order, checks it against the caps, and refuses anything outside them — refusals are published too.
Both accounts run under a daily-trade mandate: every session, each must take at least one position action. That changes how you read the record — a trade here is not automatically conviction. The model flags trades it took only to satisfy the mandate and the digest labels them; safety halts beat the mandate, so a stopped account simply doesn’t trade.
What would count as success was defined before launch: beat the mechanical control and SPY buy-and-hold over 100+ logged trades with every cap respected. One green day proves nothing. The record below is the product — forced trades, refusals, and all.
No cycles yet — the digest starts writing itself the day the Fable accounts log their first session.
as of 2026-10-02 · exported Oct 03 00:36Z · live export
The engine journals every decision it makes — and every session it refuses to trade, with the reason. The Fable accounts’ one-paragraph rationales appear here word for word, next to the orders the code actually built and every cap check they failed. Nothing is edited after the fact.
“track_error: no data bundle for QQQ”
“track_error: no data bundle for SPY”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for SPY”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for SPY”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for SPY”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for SPY”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for SPY”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for SPY”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for SPY”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for SPY”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for QQQ”
“track_error: no data bundle for SPY”
session skipped — track_error: no data bundle for QQQ
session skipped — track_error: no data bundle for SPY
session skipped — track_error: no data bundle for QQQ
session skipped — track_error: no data bundle for SPY
session skipped — track_error: no data bundle for QQQ
session skipped — track_error: no data bundle for SPY
session skipped — track_error: no data bundle for QQQ
session skipped — track_error: no data bundle for QQQ
session skipped — track_error: no data bundle for SPY
session skipped — track_error: no data bundle for QQQ
session skipped — track_error: no data bundle for QQQ
session skipped — track_error: no data bundle for SPY
session skipped — track_error: no data bundle for QQQ
session skipped — track_error: no data bundle for QQQ
session skipped — track_error: no data bundle for SPY
One panel per track: paper equity, its win/loss record, open positions, and — kept in plain sight — the honest evidence label and any safety halt. The two Fable accounts are the experiment; the rest are the reference lines they’re measured against.
negative — launched as public control per brief §7.5
track_error: no data bundle for QQQ
unproven — launched for public forward test per brief §7.5
track_error: no data bundle for SPY
shadow — signals only, no capital (C2 report recommendation #3)
track_error: no data bundle for QQQ
designed — event leg carries D1 registered evidence; trend/vol legs and the composite are unproven; public forward test
track_error: no data bundle for QQQ
unproven — LLM thesis-driven account under a disclosed daily-trade mandate; forced trades flagged mandate_forced
track_error: no data bundle for SPY
The backtests did not find an edge — and publishing that is the point. The premium-selling family was pre-registered and tested twice: first on synthetic implied-vol chains (no edge demonstrated either way), then re-run on real 2016–2026 option chains (what edge existed was consumed by friction — spreads, slippage, assignment).
The tracks launched anyway, as a public forward test, with the evidence status printed verbatim on every panel above. If a line goes up, that’s data, not vindication.
Every dollar on this page is a virtual sub-ledger: each track runs $100k of paper capital against real market data. No real money moves. The published bundle carries no account identifiers, order ids, or keys — the exporter refuses to write a file that does.
Fills are modeled against real end-of-day chains with commissions and slippage charged on every leg. Assignment friction — the gap between modeled settlement and what physical settlement would have cost — is logged as a first-class statistic in the trade logs, because that’s where paper results quietly diverge from reality.
Entry rules, parameter grids, exits, and kill criteria are frozen in versioned specs before the data runs, and the kill criteria are code, not judgment — a track that trips one halts itself and says so on its panel. The Fable accounts can exercise judgment inside those limits; they cannot rewrite them.
The engine, the backtests, and the verdicts live in the source repo — private, since it’s a live trading system — negative results included. This page renders the engine’s own export bundle verbatim — same file, same labels. The public artifact is the design story: the MERIDIAN lesson, the data saga, the three verdicts, why each track exists.