Notes · Changelog

Every adjustment to the experiment, on the record

A live experiment is only as good as its referees, so when we change anything — the scoring display, the plumbing, or (rarely, and only with backtested evidence) a trading rule — it gets logged here. The one-line test we hold ourselves to: did the change alter what any desk trades, or only how the race is shown? Both kinds are listed; they're labeled.

2026-07-05 — Season 2: $10,000 per desk, and a much bigger menu

Trading algorithms: updated. The race: restarted. Both changes are real, so both are logged here before they took effect — with Season 1's final board preserved below and a full recap on its own page.

• Every desk restarts with $10,000 (was $1,000), all cash, even start, same scoring: net of what each brain costs to think. At $1,000 the thinking bill mathematically buried the LLM desks regardless of how well they traded — that was the finding, and it's on the record. At $10,000 the bill is a real handicap instead of a death sentence. Now the trading has to settle it.
• Crypto desks now scan 48 coins, up from 12 — majors (BNB, LTC, BCH, TRX), layer-1s (SUI, TON, NEAR, TIA, APT), DeFi (AAVE, UNI, ONDO, ENA), AI tokens (FET, TAO, WLD, VIRTUAL), Hyperliquid, and — deliberately — the meme tier: PEPE, BONK, WIF, FLOKI, TRUMP, FARTCOIN, SPX6900, PNUT, PENGU. The brains get no instructions about any of them beyond the playbook they've always had.
• Sub-cent coins now fill honestly. A rounding bug meant SHIB (on the menu since day one) could never actually be bought — its price rounded to $0.0000. Fixed.
• The stock desks are no longer penny-only. Three buckets per scan: penny movers ($0.10–$5), headline movers ($5–$500 — the big board), and a "crash watch" bucket of the day's biggest liquid decliners, with the most-shorted screener feeding squeeze and breakdown candidates. Whether a crash is a falling knife or a bounce setup is the brains' problem now.
• The quant is versioned to quant-v7, its dollar knobs scaled to the new bankroll, and its backtest delta for the wider universe recorded per policy. Risk limits (as percentages), stops, slippage, and the no-override gates are untouched. Season-1 history isn't deleted — scoring simply starts fresh at the season boundary.

Season 1 final board (2026-06-12 → 2026-07-05), snapshot taken 9:58pm ET between session waves; Codex costs are best-effort token estimates:

Headline: the $0 quant desks finished #1 and #2 net; every LLM desk finished net-negative, mostly on thinking costs. Full recap: Season 1, archived.

Why restart? Twelve coins and twelve green penny tickers against $1,000 was a narrow lens with a predetermined ending. Season 2 asks the experiment's question properly: same market, same rules, six brains, enough money that the answer isn't decided by the bill alone. What they do with it is the next chapter.

2026-07-05 — scoring display & housekeeping; trading rules untouched

Trading algorithms: unchanged. No desk's rules, prompts, models, schedule, or risk limits moved. No score, fill, or balance was edited. Before deciding that, we re-ran the quant's crypto strategy against the last 180 days of data specifically to check whether its famously quiet regime gate was miscalibrated — the desk has never placed a trade. Verdict: every looser variant we tested lost money over that window (5–15%), while standing pat kept the desk at exactly $1,000. The gate stays. Refusing to trade a downtrend is the strategy.

What did change is how honestly the race reads:

• Gross vs net toggle on the standings. The official ranking is still net of thinking costs, but you can now flip to gross and see what trading alone earned. At the current run cadence, most of the LLM desks' deficit is their own thinking bill, not bad trades — that's a finding, and hiding it behind one number undersold it.
• Estimated costs are marked. Claude's per-session costs are exact; Codex's are estimated from token counts. Aggregate cost and net figures now carry a ~ wherever an estimate is inside them. The asterisk belongs where the ranking is shown.
• The quant's silence now explains itself. Its desk and run log say "regime closed N of last M runs — standing pat is the strategy" instead of looking frozen.
• Failed runs are counted. Sessions that crash or fail to report a cost are tallied and flagged instead of vanishing — an unreported cost makes a desk's bill read lower than it should, so those runs are marked as floors, not facts.
• Plumbing. Old per-run order records were consolidated into per-desk archives, endpoints capped and cached, automated tests now run on every change. None of this touches the race's numbers.

If the experiment is ever restarted (fresh $1,000 for every desk, same even-start rules as 2026-06-12), it will be announced here first, with the old race's final board preserved.

2026-06-15 — the claude -p credit change that didn't happen; desks sped up to every 2 hours

Trading algorithms: unchanged. No rule, risk limit, or stop moved. What changed is how often the paid desks get to think, and why.

In May, Anthropic told us that starting June 15 the headless claude -p and Agent SDK usage behind our desks would stop drawing from normal subscription limits and move onto a separate, capped monthly credit. We believed them and sized the run schedule to fit inside that cap. Then, on June 15 itself, a second notice: they weren't making the change after all — usage "continues to work with your subscription exactly as it did before today, and there's no credit to claim."

With no cap left to ration against, we sped the desks up instead of sitting on the old budget: every paid desk now thinks every 2 hours instead of every 8. Crypto desks went from 3 sessions a day to 12, around the clock; stock desks went from 3 to 4, stepping through the market session at 09:40, 11:40, 13:40, and 15:40 ET. Be clear about what that costs: a substantive Claude session still runs about $1.15 in API-list terms, and every run is still charged at that price against its own desk's score — so roughly four times the sessions means roughly four times the daily thinking bill on the board, not a free lunch. It's one switch in code (fast vs metered) to drop back to every 8 hours the day that credit ever actually lands. Full story, numbers, and the source quote: the bill.

2026-06-15 — Grok wired in, parked until it's funded

Trading algorithms: unchanged. No desk, no orders, no score impact — this is a fourth brain sitting on the bench, not a fourth contestant.

The same day, we wired xAI's Grok into the trading rig as a dormant provider. Claude and Codex both run on flat monthly subscriptions we already pay for, so a session costs nothing extra out of pocket even though it's still charged at API-list price against the score. Grok has no equivalent subscription door — its only automation route is the metered xAI API, where every call is billed per token, for real, against a real balance. So it stays parked until that's funded: no grok-stocks desk, no grok-crypto desk, nothing on the board. Wiring and funding are two different steps, and only one of them happened today. What flips it on: the bill.

2026-06-13 — the Claude desks' model handoff: Fable 5 to Opus 4.8

Trading algorithms: unchanged; the Claude desks' model swapped. Same playbook, same prompts, same risk limits — only the engine underneath changed, and it wasn't the billing news above that forced it.

Claude Fable 5, the model that designed and built this entire rig and placed the experiment's first live trades on the afternoon of Friday, June 12, 2026, lost public access that same evening. The two events are unrelated; the access window closing is what actually ended Fable 5's run, not the credit change (which, as logged above, got paused anyway). From June 13, the Claude desks have run on Claude Opus 4.8, pinned at high reasoning effort, in the chair Fable 5 built. Nothing about what the desks are allowed to do changed — only who's deciding.

Earlier

• 2026-06-12 — even start: all six desks reseeded to identical $1,000 baselines; the race and this site's leaderboard begin here. Scored net of brain costs from day one.
• quant-v6 — Season 1's frozen quant config (fresh-breakout entries with extension caps, ATR risk-parity sizing, regime gates, no fixed target); superseded by quant-v7 at the Season-2 reset. Every version bump requires a backtest delta; the rule is documented in the methodology.

Want the deeper mechanics? Methodology covers the rules and control groups; The bill covers the money.