Large-scale artificial-intelligence (AI) development pipelines fail constantly, and each failure produces thousands of log lines that an engineer, or increasingly a large language model (LLM), must read in order to assign a root cause. Feeding whole logs to an LLM is economically impossible at fleet scale, and the truncation heuristics used in practice-keep the first or the last k tokens, or grep for ERROR-discard the decisive evidence surprisingly often. We formalise the resulting bottleneck as budget-constrained evidence distillation: given a token budget B, select the subset of log lines that maximises the diagnostic information handed to a downstream reasoner. We present Ledger, which (i) scores every line with a contrastive template salience learned from run-level pass/fail labels only, augmented by rarity, severity, temporal-locality and (optionally) entity-sequence and parameter-anomaly channels; (ii) selects lines by maximising a monotone submodular objective that combines facility-location coverage with a concave with severity filtering and 14.3% with tail truncation; the LLM was correct on every excerpt that contained the evidence, so the entire gap is attributable to selection. The conformal router captures 100% of never-before-seen failure modes while issuing 35% fewer LLM calls than confidence-threshold routing. We also report two negative results: the parameter-anomaly channel pays off only on the 4.5% of evidence lines whose template also occurs in passing runs, and strict within-group deduplication-a natural design choice-costs 15 points of recall.