STARLI / YULI / BCI 200 / FAMILY 11 AI NEURAL DECODING F101–F105 / FIVE DISTINCT SCIENTIFIC CONTRACTS F101 Deep Learning For BCI Signal Decoding At The Limit Primary endpoint: balanced accuracy on unobserved participants [%]. Blind challenge: a blind reassignment of the entirely withheld participant cohort. Rival: reference-channel leakage and subject identity encoded in artifacts. F102 Self-Supervised Neural Signal Representations At The Limit Primary endpoint: pretraining-to-scratch held-out accuracy gain [percentage points]. Blind challenge: a concealed swap of the pretraining recording domain. Rival: implicit label exposure or recording-session identifiers. F103 Cross-Subject Transfer In BCI Decoding At The Limit Primary endpoint: transfer-minus-local error difference [percentage points]. Blind challenge: a concealed exchange of target test participants after tuning freeze. Rival: participant-ID recognition replacing neural task transfer. F104 Low-Data Neural Decoding At The Limit Primary endpoint: held-out expected loss versus labeled-sample budget [dimensionless loss]. Blind challenge: a blinded reduction of labeled training examples by block. Rival: augmentation leakage and preferentially easy active-learning samples. F105 Long-Term Drift Adaptation In Neural Decoding At The Limit Primary endpoint: calibration-error change per elapsed month [percentage points/month]. Blind challenge: a blinded change of electrode contact and reference montage. Rival: fatigue, attention and practice confounding an apparent model drift. FAMILY METHOD / NONCLINICAL BOUNDARY Ten distinct prospective 777-page research volumes. Each contains 14 research preface pages, 30 mathematical plates, 82 chapters with 410 investigations, 200 three-letter Research Gates and 33 Phoenix concepts. Subject-disjoint EEG testing, model calibration, strong classical comparison and lawful data handling are mandatory. No new experiments or therapeutic effects are reported.   STARLI / YULI / BCI 200 / FAMILY 11 AI NEURAL DECODING F106–F110 / FIVE DISTINCT SCIENTIFIC CONTRACTS F106 Explainable AI For BCI Decoding At The Limit Primary endpoint: masked-channel fidelity decrement [percentage points]. Blind challenge: a blinded mask over selected EEG time windows with a sham mask control. Rival: saliency driven by acquisition noise rather than task-bearing information. F107 Multimodal Neural Signal Fusion At The Limit Primary endpoint: conditional incremental label information from additional channels [bits/sample]. Blind challenge: a hidden offset between acquisition clocks in one modality. Rival: higher model capacity or common trigger confounding mistaken for useful fusion. F108 Uncertainty Calibration In Neural Decoding At The Limit Primary endpoint: expected confidence-calibration error [%]. Blind challenge: an undisclosed class-prior shift at test time. Rival: temperature scaling overfit to a reused calibration cohort. F109 Out-Of-Distribution Detection In AI BCI Decoding At The Limit Primary endpoint: OOD detection recall at a frozen false-alarm rate [%]. Blind challenge: a sealed injection of an unseen experimental task. Rival: high-confidence errors and session markers misread as novelty. F110 Trustworthy AI Neural Decoding Validation At The Limit Primary endpoint: second-site minus founder balanced-accuracy difference [percentage points]. Blind challenge: a blinded substitution of the preprocessing software stack. Rival: nonidentical participant splits and shared baseline data leakage. EVIDENCE LINEAGE / OPEN SCIENTIFIC REVIEW Author method lineage: STARLI F02, 10.5281/zenodo.23486020 (research design, not clinical or EEG proof). External precedents: Schirrmeister et al. 10.1002/hbm.23730; Lawhern et al. 10.1088/1741-2552/aace8c; Willett et al. 10.1038/s41586-023-06377-x (intracortical speech BCI, different modality); JCGM 10.59161/JCGM100-2008E. Proposed gate mechanisms, formula adaptations and clinical extrapolations remain unvalidated. Join the STARLI Research Community: https://zenodo.org/communities/starli-research-gate/ We do not need to know which door future generations will choose to open. We only hope that when they open it, they will find a little light to guide their way. 🌈🚪✨

BibTeX

@article{allodium:10.5281/zenodo.23989857,
  title = {STARLI / YULI / BCI 200 / FAMILY 11 AI NEURAL DECODING F101–F105 / FIVE DISTINCT SCIENTIFIC CONTRACTS F101  Deep Learning For BCI Signal Decoding At The Limit Primary endpoint: balanced accuracy on unobserved participants [\%]. Blind challenge: a blind reassignment of the entirely withheld participant cohort. Rival: reference-channel leakage and subject identity encoded in artifacts. F102  Self-Supervised Neural Signal Representations At The Limit Primary endpoint: pretraining-to-scratch held-out accuracy gain [percentage points]. Blind challenge: a concealed swap of the pretraining recording domain. Rival: implicit label exposure or recording-session identifiers. F103  Cross-Subject Transfer In BCI Decoding At The Limit Primary endpoint: transfer-minus-local error difference [percentage points]. Blind challenge: a concealed exchange of target test participants after tuning freeze. Rival: participant-ID recognition replacing neural task transfer. F104  Low-Data Neural Decoding At The Limit Primary endpoint: held-out expected loss versus labeled-sample budget [dimensionless loss]. Blind challenge: a blinded reduction of labeled training examples by block. Rival: augmentation leakage and preferentially easy active-learning samples. F105  Long-Term Drift Adaptation In Neural Decoding At The Limit Primary endpoint: calibration-error change per elapsed month [percentage points/month]. Blind challenge: a blinded change of electrode contact and reference montage. Rival: fatigue, attention and practice confounding an apparent model drift. FAMILY METHOD / NONCLINICAL BOUNDARY Ten distinct prospective 777-page research volumes. Each contains 14 research preface pages, 30 mathematical plates, 82 chapters with 410 investigations, 200 three-letter Research Gates and 33 Phoenix concepts. Subject-disjoint EEG testing, model calibration, strong classical comparison and lawful data handling are mandatory. No new experiments or therapeutic effects are reported.   STARLI / YULI / BCI 200 / FAMILY 11 AI NEURAL DECODING F106–F110 / FIVE DISTINCT SCIENTIFIC CONTRACTS F106  Explainable AI For BCI Decoding At The Limit Primary endpoint: masked-channel fidelity decrement [percentage points]. Blind challenge: a blinded mask over selected EEG time windows with a sham mask control. Rival: saliency driven by acquisition noise rather than task-bearing information. F107  Multimodal Neural Signal Fusion At The Limit Primary endpoint: conditional incremental label information from additional channels [bits/sample]. Blind challenge: a hidden offset between acquisition clocks in one modality. Rival: higher model capacity or common trigger confounding mistaken for useful fusion. F108  Uncertainty Calibration In Neural Decoding At The Limit Primary endpoint: expected confidence-calibration error [\%]. Blind challenge: an undisclosed class-prior shift at test time. Rival: temperature scaling overfit to a reused calibration cohort. F109  Out-Of-Distribution Detection In AI BCI Decoding At The Limit Primary endpoint: OOD detection recall at a frozen false-alarm rate [\%]. Blind challenge: a sealed injection of an unseen experimental task. Rival: high-confidence errors and session markers misread as novelty. F110  Trustworthy AI Neural Decoding Validation At The Limit Primary endpoint: second-site minus founder balanced-accuracy difference [percentage points]. Blind challenge: a blinded substitution of the preprocessing software stack. Rival: nonidentical participant splits and shared baseline data leakage. EVIDENCE LINEAGE / OPEN SCIENTIFIC REVIEW Author method lineage: STARLI F02, 10.5281/zenodo.23486020 (research design, not clinical or EEG proof). External precedents: Schirrmeister et al. 10.1002/hbm.23730; Lawhern et al. 10.1088/1741-2552/aace8c; Willett et al. 10.1038/s41586-023-06377-x (intracortical speech BCI, different modality); JCGM 10.59161/JCGM100-2008E. Proposed gate mechanisms, formula adaptations and clinical extrapolations remain unvalidated. Join the STARLI Research Community: https://zenodo.org/communities/starli-research-gate/ We do not need to know which door future generations will choose to open. We only hope that when they open it, they will find a little light to guide their way. 🌈🚪✨},
  author = {33},
  year = {2026},
  doi = {10.5281/zenodo.23989857},
  url = {https://doi.org/10.5281/zenodo.23989857}
}

RIS

TY  - JOUR
TI  - STARLI / YULI / BCI 200 / FAMILY 11 AI NEURAL DECODING F101–F105 / FIVE DISTINCT SCIENTIFIC CONTRACTS F101  Deep Learning For BCI Signal Decoding At The Limit Primary endpoint: balanced accuracy on unobserved participants [%]. Blind challenge: a blind reassignment of the entirely withheld participant cohort. Rival: reference-channel leakage and subject identity encoded in artifacts. F102  Self-Supervised Neural Signal Representations At The Limit Primary endpoint: pretraining-to-scratch held-out accuracy gain [percentage points]. Blind challenge: a concealed swap of the pretraining recording domain. Rival: implicit label exposure or recording-session identifiers. F103  Cross-Subject Transfer In BCI Decoding At The Limit Primary endpoint: transfer-minus-local error difference [percentage points]. Blind challenge: a concealed exchange of target test participants after tuning freeze. Rival: participant-ID recognition replacing neural task transfer. F104  Low-Data Neural Decoding At The Limit Primary endpoint: held-out expected loss versus labeled-sample budget [dimensionless loss]. Blind challenge: a blinded reduction of labeled training examples by block. Rival: augmentation leakage and preferentially easy active-learning samples. F105  Long-Term Drift Adaptation In Neural Decoding At The Limit Primary endpoint: calibration-error change per elapsed month [percentage points/month]. Blind challenge: a blinded change of electrode contact and reference montage. Rival: fatigue, attention and practice confounding an apparent model drift. FAMILY METHOD / NONCLINICAL BOUNDARY Ten distinct prospective 777-page research volumes. Each contains 14 research preface pages, 30 mathematical plates, 82 chapters with 410 investigations, 200 three-letter Research Gates and 33 Phoenix concepts. Subject-disjoint EEG testing, model calibration, strong classical comparison and lawful data handling are mandatory. No new experiments or therapeutic effects are reported.   STARLI / YULI / BCI 200 / FAMILY 11 AI NEURAL DECODING F106–F110 / FIVE DISTINCT SCIENTIFIC CONTRACTS F106  Explainable AI For BCI Decoding At The Limit Primary endpoint: masked-channel fidelity decrement [percentage points]. Blind challenge: a blinded mask over selected EEG time windows with a sham mask control. Rival: saliency driven by acquisition noise rather than task-bearing information. F107  Multimodal Neural Signal Fusion At The Limit Primary endpoint: conditional incremental label information from additional channels [bits/sample]. Blind challenge: a hidden offset between acquisition clocks in one modality. Rival: higher model capacity or common trigger confounding mistaken for useful fusion. F108  Uncertainty Calibration In Neural Decoding At The Limit Primary endpoint: expected confidence-calibration error [%]. Blind challenge: an undisclosed class-prior shift at test time. Rival: temperature scaling overfit to a reused calibration cohort. F109  Out-Of-Distribution Detection In AI BCI Decoding At The Limit Primary endpoint: OOD detection recall at a frozen false-alarm rate [%]. Blind challenge: a sealed injection of an unseen experimental task. Rival: high-confidence errors and session markers misread as novelty. F110  Trustworthy AI Neural Decoding Validation At The Limit Primary endpoint: second-site minus founder balanced-accuracy difference [percentage points]. Blind challenge: a blinded substitution of the preprocessing software stack. Rival: nonidentical participant splits and shared baseline data leakage. EVIDENCE LINEAGE / OPEN SCIENTIFIC REVIEW Author method lineage: STARLI F02, 10.5281/zenodo.23486020 (research design, not clinical or EEG proof). External precedents: Schirrmeister et al. 10.1002/hbm.23730; Lawhern et al. 10.1088/1741-2552/aace8c; Willett et al. 10.1038/s41586-023-06377-x (intracortical speech BCI, different modality); JCGM 10.59161/JCGM100-2008E. Proposed gate mechanisms, formula adaptations and clinical extrapolations remain unvalidated. Join the STARLI Research Community: https://zenodo.org/communities/starli-research-gate/ We do not need to know which door future generations will choose to open. We only hope that when they open it, they will find a little light to guide their way. 🌈🚪✨
AU  - 33
PY  - 2026
DO  - 10.5281/zenodo.23989857
UR  - https://doi.org/10.5281/zenodo.23989857
ER  - 

APA

33 (2026). STARLI / YULI / BCI 200 / FAMILY 11 AI NEURAL DECODING F101–F105 / FIVE DISTINCT SCIENTIFIC CONTRACTS F101 Deep Learning For BCI Signal Decoding At The Limit Primary endpoint: balanced accuracy on unobserved participants [%]. Blind challenge: a blind reassignment of the entirely withheld participant cohort. Rival: reference-channel leakage and subject identity encoded in artifacts. F102 Self-Supervised Neural Signal Representations At The Limit Primary endpoint: pretraining-to-scratch held-out accuracy gain [percentage points]. Blind challenge: a concealed swap of the pretraining recording domain. Rival: implicit label exposure or recording-session identifiers. F103 Cross-Subject Transfer In BCI Decoding At The Limit Primary endpoint: transfer-minus-local error difference [percentage points]. Blind challenge: a concealed exchange of target test participants after tuning freeze. Rival: participant-ID recognition replacing neural task transfer. F104 Low-Data Neural Decoding At The Limit Primary endpoint: held-out expected loss versus labeled-sample budget [dimensionless loss]. Blind challenge: a blinded reduction of labeled training examples by block. Rival: augmentation leakage and preferentially easy active-learning samples. F105 Long-Term Drift Adaptation In Neural Decoding At The Limit Primary endpoint: calibration-error change per elapsed month [percentage points/month]. Blind challenge: a blinded change of electrode contact and reference montage. Rival: fatigue, attention and practice confounding an apparent model drift. FAMILY METHOD / NONCLINICAL BOUNDARY Ten distinct prospective 777-page research volumes. Each contains 14 research preface pages, 30 mathematical plates, 82 chapters with 410 investigations, 200 three-letter Research Gates and 33 Phoenix concepts. Subject-disjoint EEG testing, model calibration, strong classical comparison and lawful data handling are mandatory. No new experiments or therapeutic effects are reported.   STARLI / YULI / BCI 200 / FAMILY 11 AI NEURAL DECODING F106–F110 / FIVE DISTINCT SCIENTIFIC CONTRACTS F106 Explainable AI For BCI Decoding At The Limit Primary endpoint: masked-channel fidelity decrement [percentage points]. Blind challenge: a blinded mask over selected EEG time windows with a sham mask control. Rival: saliency driven by acquisition noise rather than task-bearing information. F107 Multimodal Neural Signal Fusion At The Limit Primary endpoint: conditional incremental label information from additional channels [bits/sample]. Blind challenge: a hidden offset between acquisition clocks in one modality. Rival: higher model capacity or common trigger confounding mistaken for useful fusion. F108 Uncertainty Calibration In Neural Decoding At The Limit Primary endpoint: expected confidence-calibration error [%]. Blind challenge: an undisclosed class-prior shift at test time. Rival: temperature scaling overfit to a reused calibration cohort. F109 Out-Of-Distribution Detection In AI BCI Decoding At The Limit Primary endpoint: OOD detection recall at a frozen false-alarm rate [%]. Blind challenge: a sealed injection of an unseen experimental task. Rival: high-confidence errors and session markers misread as novelty. F110 Trustworthy AI Neural Decoding Validation At The Limit Primary endpoint: second-site minus founder balanced-accuracy difference [percentage points]. Blind challenge: a blinded substitution of the preprocessing software stack. Rival: nonidentical participant splits and shared baseline data leakage. EVIDENCE LINEAGE / OPEN SCIENTIFIC REVIEW Author method lineage: STARLI F02, 10.5281/zenodo.23486020 (research design, not clinical or EEG proof). External precedents: Schirrmeister et al. 10.1002/hbm.23730; Lawhern et al. 10.1088/1741-2552/aace8c; Willett et al. 10.1038/s41586-023-06377-x (intracortical speech BCI, different modality); JCGM 10.59161/JCGM100-2008E. Proposed gate mechanisms, formula adaptations and clinical extrapolations remain unvalidated. Join the STARLI Research Community: https://zenodo.org/communities/starli-research-gate/ We do not need to know which door future generations will choose to open. We only hope that when they open it, they will find a little light to guide their way. 🌈🚪✨. https://doi.org/10.5281/zenodo.23989857