Measuring the Incremental Value of Text in Probabilistic Financial Forecasting: Scoring Incentives, Temporal Knowledge and Reader Variability

Does reading dated financial documents improve a probabilistic forecast beyond its numerical engine? We study this question on Agenthon 2026 Track 2 using seven text integrations, a text-blind control, and a leaderboard-matched gate ablation. On 94 locally reconstructed practice cards, the best stated single and paired configurations improve the mean baseline-normalized score by at most 0.006 (a post-hoc unanimity threshold reaches $-0.010$); their paired intervals include zero and the gains do not survive selection adjustment. A gate-enabled submission scores 0.8961 versus 0.8928 for its text-blind twin in one live reading, with lower scores better. Three audits qualify these results. First, outcome-dependent score normalization changes the target of evaluation: a Gaussian single-cell example rewards underdispersion and raises the accuracy required for a fixed directional adjustment. Second, a no-document date intervention is consistent with temporal recall, but does not causally separate reading from memory. Third, reader reruns and calendar-selected windows reveal instability and limited transfer. Two further experiments localize much of the engine's window loss to joint-score sensitivity and find no established benefit from model-recalled analogues. Rescoring the engine ladder, the text mechanisms and the rules on a companion metric with a fixed pre-as-of scale removes the rules' gain and turns the text-blind event count from the worst arm into the best, so both effects belong to the outcome-normalized metric rather than to the forecasts. We therefore report a scoped negative result, not evidence that the documents contain no predictive information. The contribution is a measurement framework linking matched text ablations, score-specific decision thresholds, temporal provenance and reader-level uncertainty. A locked, adequately powered evaluation beyond all documented training cutoffs remains necessary.