The Deflated Sharpe Ratio (DSR) is the standard defence against backtest overfitting. It raises the significance bar with the number of strategies tried, and recent work proposes logging every trial so that DSR can be applied to the true search intensity. We show that counting the backtests actually run is the wrong currency whenever the researcher is adaptive, that is, whenever each new strategy is chosen after seeing earlier validation results. This describes everyday quantitative practice, and it describes LLM research agents and RL alpha miners by construction. First, we prove that a one-line "combine what worked" procedure over k pure-noise strategies drives the validation t-statistic to 2k/π (1+o p (1)), independently of cross-strategy correlation. The DSR bar for the same complete ledger grows only as √ 2 ln k, so DSR certifies noise with probability tending to one. Second, we prove a matching positive result. If the holdout can reveal only a transcript from a set T , then certifying at t > Ḡ-1 (α/|T |) is valid for every adaptive analyst. DSR is the non-adaptive special case |T | = N , and overfitting scales as Θ(√ bits revealed) rather than ln(queries). Under full-precision Sharpe feedback, the only generic valid correction is the classical span (Scheffé/GRS) bound, which grows as √ dim and is unusable at realistic scale. This motivates a sealed holdout: an opaque type whose only operation is a budgeted PASS/FAIL verdict. On synthetic markets, and on 423 real industry strategies (1970-2026) under a stationarybootstrap null, ordinary adaptive workflows such as "rank on validation, keep the top j" are certified by query-counted DSR on 84-100% of pure-noise replicates, and sealed-verdict variants hold false certification at 0.8-3.0%. With frontier code-executing agents (Claude Code), the outcome depends on the instruction. Asked for a robust strategy, the agents usually decline to tune on the holdout. Instructed to maximise validation Sharpe, Claude Opus autonomously executes our attack in about a minute: it sweeps the single signals and then sign-weights the strongest ones, reaching validation t of up to 15.7 on pure noise, against a test t of about 0. Ledger DSR certifies these searches at the attack step under optional stopping (4 of 12 episodes, against 0 of 12 for a non-adaptive baseline) and is rescued only when continued search inflates its variance estimate. A GRPO alpha miner with an AlphaGen-style pool reward learns the same ensemble from reward alone in one of three seeds (177 signals, validation t = 12.0, test t =-0.56), and ledger DSR certifies it. In research campaigns that reuse one holdout across rounds, a researcher who builds on earlier full-precision results makes a false discovery in 100% of 5-round null campaigns, whether certifying by a Bonferroni bar or by cumulative-ledger DSR; splitting the holdout instead cuts power to 0.03 at 10 rounds. The sealed verdict stays at 1-3% while adapting, and keeps power of 0.65-0.79.