From 75a64888cb3e117ee6d232adc7b48974647fbc1e Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Thu, 23 Jul 2026 08:17:56 -0500 Subject: figures: add audited oral-B-v2 confirmation panel --- results/figs/figure5_bci_v2_caption.md | 20 ++++++++++++++++++++ 1 file changed, 20 insertions(+) create mode 100644 results/figs/figure5_bci_v2_caption.md (limited to 'results/figs/figure5_bci_v2_caption.md') diff --git a/results/figs/figure5_bci_v2_caption.md b/results/figs/figure5_bci_v2_caption.md new file mode 100644 index 0000000..41e0dbd --- /dev/null +++ b/results/figs/figure5_bci_v2_caption.md @@ -0,0 +1,20 @@ +# Oral-B-v2 confirmation figure caption + +**Figure 5 | Role-vectorized somato-dendritic innovation carries +temporal-difference outcome surprise.** All summaries use the complete +untouched six-task by five-model confirmation; uncertainty is computed after +averaging models within each task seed. **a,** Mean episode success over +training. The learned causal-role vector is necessary, the forward-plasticity +lesion (omitted for visual overlap with zero) does not learn, and critic or +terminal-outcome training lesions are reported rather than hidden. Bands are +95% normal intervals over six task clusters. **b,** Ordinary apical traffic is +almost perfectly soma-correlated before subtraction; the innovation is below +the frozen correlation ceiling in every task cluster. **c,** Five targets are +chosen from fixed quantiles of 512 outcome-free cursor-max calibration trials. +The displayed rewarded fractions come from independent assay trajectories; +no evaluation label selects or reweights a target. **d,** Terminal innovation +decodes rewarded versus timeout outcomes, while the acute terminal-outcome +lesion removes role-aligned separation. The critic contribution tracks the +stored value prediction and attenuates expected outcomes. Exact role maps, +outcome labels, and gradients are diagnostic-only and never used by the +learner's updates. -- cgit v1.2.3