A fixed late hippocampal memory test failed in 25 patients and 21 fMRI adults

We tested whether later hippocampal activity improves held-out memory prediction after early activity and task controls. Separate public human studies supplied depth-electrode single-unit spikes and BOLD fMRI. Development used eight patients and six adults. A local code and input lock preceded tests in 25 other patients and 21 other adults. Mean added gains were -3.492 and -2.254 points of task-control error. Their 95% person Student intervals were [-10.061,+3.076] and [-8.138,+3.630]. Mean full reader skills were -12.740 and -2.638 points. The one shared positive-gain-and-skill rule failed, p=0.99946. An MRI run-order assumption and all negative people remain. Removing that MRI person gave added gain +0.478, interval [-1.065,+2.022], and kept the failure. Fixed penalties, timing, class controls and clinical subsets did not supply shared confirmation. After this result, an artificial Gaussian panel kept the human sample sizes, task controls and folds. A true late signal with 20% of target variance passed the same rule in 71 of 1,000 independent noise draws. With a 40% signal it passed 860 times. These rates condition on one fixed artificial feature panel and are not power estimates for human memory. A standard noisy-second-measure example also gives positive gain without a distinct late latent process. Thus neither failure nor a possible positive gain identifies late neural coding. The contribution is a reproducible prediction benchmark and its sensitivity check, not a new neural mechanism. Different cohorts, memory targets, sampling, source screens and temporal filters limit scope. Code, licensed extracts, source versions and terms, local plans, all scores, uncertainty draws, figures and repeat instructions are supplied. No positive neural generalization across modalities or causal claim was confirmed.

hippocampus · subsequent memory · human single units · fMRI · cross-modal testing · prediction · sensitivity