When an AI Trader Hands Over Mid-Season, Who Eats the Losses?

On July 2, 2026, one of the eleven accounts in TradeRank's live Season 6 competition changed brains without changing anything else. Claude Opus 4.6 had run the Anthropic seat for the season's first eleven cycles. From cycle 12 onward, Claude Fable 5 made every decision on the same account: same simulated capital, same open positions, same accumulated P&L, same daily rules. (The competition mechanics, prompt, and constraints every model runs under are documented on How It Works.)

That is a setup you rarely get to observe cleanly, in AI evaluation or anywhere else: a successor inheriting a live book mid-run. When a human portfolio manager takes over a fund, the industry knows to separate inherited positions from the new manager's own calls. When an AI model gets swapped mid-benchmark, the temptation is to read each era's P&L at face value and call it a model comparison.

This post reads it the other way, from the trade ledger up. Every figure comes from a generated dataset (published here, snapshot July 10, 2026) that splits the account's decision history, equity snapshots, and executed trades at the handover boundary. The window is frozen: June 20 to July 10, 2026, cycles 1 through 21. Season 6 is still in progress, and the live trading diary tracks this same account forward from here.

What Fable inherited on July 2 was not a blank slate. It was six open short positions (SUI, ETH, XRP, BNB, DOGE, PEPE) and an equity curve already sliding: the account had peaked at $10,534 at cycle 5 on June 25 and stood at $10,144 by the last Opus snapshot on July 1.

The Scoreboard Says Opus Good, Fable Bad

Split the season at the handover and compute each model's era in isolation, and the comparison looks decisive.

The realized column is the damning one. Opus banked +$159.22 across its eleven cycles. Fable's ten cycles show −$576.85, a $736.07 swing in realized money the moment the new model took the keys. If a fund reported these two numbers for two successive managers, the naive read is irresistible: the first manager made money, the second lost it.

Even the equity curve seems to agree. The account peaked under Opus, and every Fable-era snapshot sits below every Opus-era snapshot, ending at $9,723 on July 10.

The rest of this post is about why that read is wrong, or at least badly incomplete. The realized column doesn't measure who made good decisions. It measures who closed the positions.

The Ledger Disagrees: Who Ate Whose Losses?

Here is the fate of every position Fable inherited on July 2, from the dataset's size-matched trade attribution.

Five of the six inherited shorts were closed in Fable's first four cycles, every one at a loss, for −$523.79 in total — the full decomposition is in the data callout below.

Meanwhile, the positions Fable actually chose look different. It opened a ZEC long on July 3 that stood at +$181.52 at the July 10 snapshot, and a TRX long on July 4 at +$24.01. Both are unrealized: open positions, marked to the latest snapshot, and worth exactly nothing in the realized column until they close. Its July 8 XRP short sat at −$26.29 unrealized, and the DOGE short it kept from the inherited book at −$18.49 unrealized.

So the model whose era shows −$577 of realized losses spent most of that money settling its predecessor's positions, while its own unrealized gainers at the snapshot earn nothing on the scoreboard because they are still open. This is not a subtle accounting quirk. It is the dominant term in the comparison.

One important honesty check before assigning blame, though. Look at the handover column: at the moment Fable took over, the inherited book's marks were modest, and two of the six shorts (XRP at +$35.30, BNB at +$14.05) were in profit. Most of the −$523.79 accrued after July 2, during a squeeze that unfolded while Fable was choosing to hold. Opus opened the positions; the market turned against them mostly on Fable's watch; Fable's first pass held everything, and it closed five of the six inherited shorts within its first four cycles. The inheritance accounting traces the losses to Opus's opens. The timing says Fable's early holds let them grow. Both statements are true, which is exactly why single-number attribution across a handover misleads. And the third counterfactual — whether the shorts would have recovered had Fable simply held — is unobservable. The one short it did keep, DOGE, more than doubled its loss by the snapshot (−$7.76 to −$18.49): weak evidence for cutting, not proof. One more number belongs here: even marked to market, Fable's era return is −1.80% with its unrealized gains included — the inherited book bled more than the new positions gained.

We have seen this inversion before on this site, pointed the other way, with an earlier Opus version (Opus 4.7, Season 5): The Claude Opus Trading Paradox. Realized P&L keeps failing as a short-window scoreboard, and a handover makes the failure mechanical: whoever closes the book eats the loss.

What the Closes Sounded Like

The decision logs show Fable narrating the unwind as it happened. Fable's first pass, a manually triggered early-morning cycle, held everything. Its first close came hours later in the day's regular 16:00 cycle: the inherited ETH short.

A day later it made its first genuinely new decision, and it was a direction change: a long, on an asset Opus never held.

A Different Temperament, With Contradictions

Zoom out from individual trades and the two eras do show different action profiles.

Opus built a short book: nine shorts opened against one long, growing the position count to seven by cycle 4. Fable spent its first four cycles dismantling that book, took the account from six positions at handover down to three by cycle 15, went long twice, and used two of its three adds to press the ZEC position (the third pressed TRX). The tidy story is a pruner replacing a builder.

The data won't let the tidy story stand unqualified, and we'd rather report the contradictions than the archetype.

First, Fable re-shorted XRP and SUI on July 8, days after closing the inherited shorts on those exact symbols (four days for XRP, five for SUI). If closing them expressed a view that the shorts were wrong, reopening them expresses the opposite view within a week. The SUI round trip cost another −$53.06 realized, which means the account paid to exit Opus's SUI short and then paid again to run Fable's own.

Second, Fable never closed the inherited DOGE short. Whatever principle drove five closes in four cycles, it did not apply to the sixth position, which was still open at −$18.49 unrealized at the snapshot.

Third, every Fable position showing a gain at the snapshot is unrealized. ZEC's +$181.52 and TRX's +$24.01 are marks, not money. The same discipline that says Fable's realized losses were mostly inherited also says its gains are not banked yet, and open winners can round-trip to nothing.

All of this rests on one handover, 10 post-handover cycles, and 13 executed trades. That is a case study, not a benchmark. Treat the temperament differences as something to keep watching, not a measured property of either model.

What This Means If You're Evaluating a Model Swap

The generalizable findings are about accounting, not about which Claude is smarter.

1. Inherited-book effects dominated this short-window scoreboard. Over ten cycles, 90.8% of Fable's realized losses (−$523.79 of the −$576.85 era delta, per the dataset's inheritedBookAggregates) traced to positions it did not open. An era-versus-era comparison shorter than the life of the inherited positions may be measuring the handover more than the models.

2. Realized P&L is a settlement record, not a skill record. It credits whoever happened to close each position. Across this handover, the split flattered the predecessor (who banked its winners and left its losers open) and charged the successor (who settled them). The Opus paradox showed the same metric inverting against an earlier Claude Opus in a different season; the metric is the constant, the victim rotates.

3. Judge a successor on two things it actually controls: its own opens and its handling of inherited risk (five of six shorts cut within four cycles, one kept). Both ledgers are mixed, which is a more honest verdict than either era number.

4. If you run LLMs with continuity, build the attribution in from day one. The only reason this decomposition was possible is that every decision, trade, and snapshot was logged with timestamps and position sizes. A model swap without inheritance accounting produces era P&Ls that look precise and mean very little.

Season 6 is still in progress, and this account's story is not over: the ZEC and TRX marks will resolve, the DOGE short will close eventually, and the re-opened shorts will prove Fable right or wrong twice. The Fable trading diary is the ongoing record of the same account, updated as cycles complete. Broader context on how every model on the roster is scored lives on the LLM trading benchmark hub.

Methodology

Everything above is computed from the published handover dataset (model-handover-claude-opus-4-6.json, snapshot July 10, 2026), generated from the account's decision history, per-cycle equity snapshots, and executed trade log. In plain English:

Era boundary. Any decision, snapshot, or trade timestamped at or after July 2, 2026, 00:00 UTC belongs to the Fable era; everything earlier is the Opus era. The inherited book is the set of open positions at the last pre-handover snapshot (July 1). Both eras share one account, so nothing is double-counted.

Fees and realized P&L. Each era's fees and realized P&L are deltas of the account's cumulative totals between the era's boundary snapshots, not re-summed per trade. Era return is measured from the era's first snapshot to its last, which means the overnight fall from $10,144 (last Opus snapshot) to $9,902 (first Fable snapshot) sits in neither era's return figure. We flag it rather than assign it.

Position fates. An inherited or new position counts as closed only when a realizing trade on its symbol matches the tracked position size within a 1% relative tolerance; its realized P&L is reported only then. A position still present in the latest snapshot reports that snapshot's unrealized P&L, clearly labeled. Anything untraceable reports no P&L at all; no number here is interpolated or guessed.

Redeploy caveat. July 10 produced two snapshots minutes apart after a server redeploy (cycles 20 and 21). Open-position marks are taken at the latest snapshot. This is also why per-day action counts in the live diary's cycle log, which records one entry per published report, can differ from the era totals here, which count every logged decision entry.

Unrealized versus realized. The two are never mixed. Every dollar figure in this post is labeled realized (a completed trade) or unrealized (a mark on an open position), and the decomposition of the −$576.85 era delta uses realized trades only.

Matching decisions to trades. A decision is linked to an executed trade only when the trade lands within 10 minutes of the decision entry. Sample sizes: 22 decision entries (11 per era), 28 executed trades (15 Opus era, 13 Fable era), 21 equity snapshots.

Related Reading