CLQT measures a 0.30 gap between what LLM trading agents say and what they allocate

Look-ahead bias is the default condition of a backtest. Pull a price or fundamentals series today and you get the revised version, restatements included, published after the date the agent is pretending to trade on. An agent that beats the index by reading a later revision has proved something about the harness. The abstract puts the consequence plainly: apparent alpha can dissolve once look-ahead bias and realistic trading costs are controlled ^[4].

Sealing each cycle into a recompute-verifiable audit chain is the part of this design worth copying ^[8]. A cumulative return curve has to be taken on trust. A score built from the chain can be rejected by a third party who reruns the trail.

In the campaign the capability leader was not the Sharpe leader ^[14]. The value of individual modules showed up on the capability and behavioral axes in cases where returns alone could not separate the configurations ^[16]. An ablation grid scored only on return will report that such a component does nothing.

The stating-versus-doing gap is the figure I would take into a design review. It ran +0.30 in the year-long backtest and +0.23 in the live track ^[15], which is 0.07 lower, about 23 percent smaller ^[22], over a window roughly a thirteenth as long ^[23]. The abstract does not say what scale the gap is measured on.

Qu and Chen are explicit about what one study cannot settle. They write that a credible model ranking would require many models, regimes and repeated runs, beyond the scope of this work ^[20]. Their standard for calling a model superior is dominance on the relevant capability axes, consistently across sub-periods. Until that holds, they write, the benchmark's product is "a reusable map of limitations" ^[21].

The coverage claim is a claim about everyone else's benchmarks. To the authors' knowledge, no existing benchmark satisfies all six conditions simultaneously: enforcing point-in-time data semantics, modelling realistic transaction and financing costs across asset classes, evaluating decisions at the portfolio level, measuring strategy consistency across rounds, accumulating structured multi-tier memory, and treating tool orchestration as a benchmarkable capability ^[18]. Six conditions at once is the bar CLQT will be measured against as well. For its cost result to transfer to your own agent, the modelled transaction and financing components ^[6] would have to match your asset mix, your holding periods and your broker's terms. Your data layer would also have to refuse post-decision revisions the way the TimeGate does ^[5].