Restore research-first task descriptions across the v1 collection

README.md

@@ -17,32 +17,39 @@ configs:
   default: true
 ---
 
-# BrainPilotBench — task data
+# BrainPilotBench v1 — task data
 
-Public data for the BrainPilotBench v1 task collection: 112 runnable tasks across six data categories.
-This repository hosts public data for 111 tasks. H1 has a source-only entry because its source agreement prohibits redistribution.
+Public data for research tasks in brain science and human behavior.
+Agents use these data to develop neural decoders, predict human responses, and analyze relationships between brain measurements.
+The [BrainPilotBench source repository](https://github.com/NeuroAIHub/BrainPilotBench/tree/7306cc0b0e8958cc8f84918b68446310d8dbc662) provides task instructions, environments, submission formats, and scoring code.
 
-The collection combines 108 new tasks with four retained v0 tasks.
-Agent benchmark results for the 108 additions are not yet available.
-Data preparation and fixed-reference checks do not establish Agent research performance.
+**112 tasks · 6 data categories · public data for 111 tasks**
 
-## Find a task
+> **Agent benchmark results for the 108 new v1 tasks are not yet available.**
+> The four retained v0 tasks have separate historical results. Data preparation and fixed-reference checks are not Agent benchmark results.
 
-The Dataset Viewer shows one catalog row per task. It does not combine recordings or behavioral trials into one sample table.
+## Explore the tasks
 
-| Data category | Tasks | Directory |
+| Data category | Tasks | Research problems |
 |---|---:|---|
-| EEG | 25 | [Browse tasks](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg) |
-| Psychology | 76 | [Browse tasks](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology) |
-| MRI | 2 | [Browse tasks](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/mri) |
-| Connectomics | 1 | [Browse tasks](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/connectomics) |
-| Calcium imaging | 1 | [Browse tasks](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/calcium-imaging) |
-| Invasive recordings | 7 | [Browse tasks](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive) |
+| [EEG](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg) | 25 | Classify motor imagery, evoked responses, sleep stages, and other scalp-EEG conditions. |
+| [Psychology](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology) | 76 | Model participants' recorded choices and responses from causal trial history. |
+| [MRI](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/mri) | 2 | Predict age from structural MRI and tonic pain from fMRI connectivity. |
+| [Connectomics](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/connectomics) | 1 | Compare anatomical connectivity with measured signal propagation in C. elegans. |
+| [Calcium imaging](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/calcium-imaging) | 1 | Analyze RSC place-cell activity and decode behavior. |
+| [Invasive recordings](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive) | 7 | Decode movement, muscle activity, handwriting, film conditions, and birdsong from invasive neural recordings. |
 
-The [machine-readable catalog](catalog.jsonl) provides task descriptions, metrics, sources, and access status.
-The [collection manifest](collection.json) records the collection totals.
+Start with [human casino choices](tasks/psychology/psych101-lefebvre2017behavioural-exp1/README.md),
+[MRI age prediction](tasks/mri/brain-age-mri/README.md), or
+[ECoG finger-flexion decoding](tasks/invasive/ieeg-finger-flexion-bciciv4/README.md).
+Each task card explains the research question, available data, and evaluation, then links to its full guide.
 
-## Directory layout
+The Dataset Viewer displays one row per task. It is a task directory, not a combined table of neural recordings or behavioral trials.
+The [catalog](catalog.jsonl) and [collection manifest](collection.json) provide machine-readable entries and counts.
+
+## Get data for one task
+
+Each task has one stable location:
 
 ```text
 tasks/<data-category>/<task-id>/
@@ -51,32 +58,27 @@ tasks/<data-category>/<task-id>/
   public/<native-relative-path>
 ```
 
-Each task has one stable directory. Upload dates and storage batches do not define separate benchmark editions.
-Native NPY, NPZ, JSONL, MAT, ZIP, and other task formats remain unchanged.
-The collection contains 668 public data files, totaling 37,945,533,281 logical bytes.
-Private evaluation files, owner mappings, and secret keys are excluded.
-
-## Use the data
+Read the task card and source terms before download.
+Use its versioned GitHub guide to install the environment and prepare a workspace.
+The task's `data.lock` pins exact HF commits, file sizes, and SHA-256 digests.
+Native scientific formats remain unchanged.
 
-Read the task README and its source terms before download.
-Use the [BrainPilotBench source checkout](https://github.com/NeuroAIHub/BrainPilotBench/tree/a79bd29f4cb533f63013b9d4bc0b11fba5e66b53) for environments, protocols, and the CLI.
-Its data locks bind exact HF commits, byte counts, and SHA-256 digests.
-The v1 task collection is on GitHub main. This source snapshot adds the unified download and preparation workflow.
-The workflow update is on a separate feature branch, not yet merged into main. No new npm release is available.
+Public files are hosted for 111 tasks, including BCI IV-4 and six invasive-recording tasks.
+H1 provides [source download and local preparation instructions](tasks/invasive/falcon-h1-reach-grasp/README.md) because its Data Use Agreement prohibits redistribution.
+The film task's public package supports Agent preparation. Block-first scoring still requires a trusted evaluator mapping that is not hosted here.
+Private evaluation files, owner mappings, and keys are excluded from this data repository.
 
-The task manifests describe public files only. Private evaluation and official source-benchmark scores remain separate.
-BCI IV-4 includes the prepared Train/Valid split, not its original evaluation interval.
-H1 provides source links and a split/build recipe, not a mirror.
-The film task needs its trusted evaluator mapping for block-first scoring. That mapping is not public.
+## Versions, source terms, and results
 
-## Versions and source terms
+The v1 collection contains 108 new tasks and four retained v0 tasks.
+Individual task versions define scientific protocols. HF commit IDs identify data snapshots.
+Task cards link to versioned source guides. Directory navigation links use the current `main` tree.
 
-`v1` names the task collection. Individual task versions identify scientific protocols.
-HF commit IDs identify immutable storage snapshots. Use commit-pinned links for reproducible downloads.
-The directory navigation links use `main`, which can change.
-
-The [attribution guide](ATTRIBUTION.md) records source-specific notices and prior publication provenance.
-No repository-wide license supersedes the source data terms.
+The [attribution guide](ATTRIBUTION.md) records source-specific terms and citations.
 The root LICENSE does not grant an Apache license for all task data.
-The migration preserves existing data bytes and scientific splits.
-Earlier immutable commit URLs remain part of the publication history.
+BCI IV-4 retains its competition source terms. The film package excludes the copyrighted movie.
+
+The unified access update is under review in [PR #77](https://github.com/NeuroAIHub/BrainPilotBench/pull/77).
+Use the linked source snapshot for its documented workflow until that update reaches GitHub main.
+The [data preparation record](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/docs/data-preparation.md) separates delivery checks from Agent evaluation.
+Earlier immutable data URLs remain available in the publication history.

tasks/calcium-imaging/neuro-rsc-place-cell/README.md

@@ -1,32 +1,26 @@
 # neuro-rsc-place-cell
 
-[Task catalog](../../../README.md)
+The agent performs five analyses of RSC place-cell dynamics using the staged recording.
 
-RSC 位置细胞动态分析——依据 mouse RSC Ca²⁺ imaging + VR belt 行为数据(VRBeltReframe.mat),完成位置细胞筛选/可视化/贝叶斯解码/trial-bin 相关/firing-rate 动态,交 benchmark_summary.json + report.md + 图。
+Historical Agent results for this retained task are recorded in the v0 report.
 
-Historical results for this retained task remain in the v0 report.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `neuro-rsc-place-cell` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2` |
-| Data category | Calcium imaging |
-| Access | `hosted_public` |
-| Primary metric | `See task protocol` (See task protocol) |
-
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/calcium-imaging/neuro-rsc-place-cell/public) contains 1 files (203,632,523 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
-
-## Protocol and source
+| Input | Mouse retrosplenial cortex Ca²⁺ imaging and virtual-reality belt behavior. |
+| Output | `benchmark_summary.json`, `report.md`, and `figures/*.png`. |
+| Evaluation | A deterministic check compares the summary with a frozen reference. Human review scores the report and figures separately. |
+| Metric | Equal-weight mean of `place_cell_ratio_ok`, `decoding_significance_ok`, and `decoding_error_within_tolerance`. Human dimensions do not enter this score. |
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/calcium-imaging/neuro-rsc-place-cell/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Data access
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/calcium-imaging/neuro-rsc-place-cell/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/calcium-imaging/neuro-rsc-place-cell/public) contains 1 file (203,632,523 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
 
 Original source terms apply. This catalog does not relicense the data.
+
+## Protocol and source
+
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/calcium-imaging/neuro-rsc-place-cell/README.md) defines the scientific protocol, preparation, and scoring.
+
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/calcium-imaging/neuro-rsc-place-cell/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/connectomics/celegans-structure-function-divergence/README.md

@@ -1,32 +1,30 @@
 # Anatomical Connectivity and Signal Propagation in C. elegans
 
-[Task catalog](../../../README.md)
+How closely does anatomical wiring match measured signal propagation in C. elegans? The agent quantifies differences between a connectome and a signaling atlas using a fixed analysis protocol. It compares connections, communities, hubs, and pharyngeal organization.
 
-Candidate integration of an analysis task comparing a measured C. elegans signaling atlas with electron-microscopy connectomes. The agent implements a reusable pipeline for edge ranking, binary-topology community alignment, hub agreement, and pharyngeal co-assignment. A separately isolated evaluator checks numerical outputs on held-out reconstructions; only the public reconstruction and task instructions are available during development.
+The evaluator runs the submitted analyzer on additional reconstructions and compares its outputs with independent calculations. The task measures differences between the networks, not predictions of one network from the other.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+Agent benchmark results are not yet available.
 
-## Task and data
+## Task at a glance
 
-| Field | Value |
+| Field | Details |
 |---|---|
 | Task ID | `celegans-structure-function-divergence` |
-| Collection | BrainPilotBench v1 |
 | Task version | `0.6` |
-| Data category | Connectomics |
-| Access | `hosted_public` |
-| Primary metric | `score` (maximize) |
+| Public data | A signal propagation atlas, one anatomical connectome reconstruction, and neuron metadata |
+| Submission | A reusable Python analyzer, a protocol declaration, and a scientific report |
+| Primary metric | `score` in [0, 1]: numerical-statistic agreement weighted by judgment skill, higher is better |
+| Evaluation | Statistics and judgments checked on evaluator-controlled reconstructions |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/connectomics/celegans-structure-function-divergence/public) contains 3 files (1,063,606 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/connectomics/celegans-structure-function-divergence/public) contains 3 files (1,063,606 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/connectomics/celegans-structure-function-divergence/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+Original source terms apply. This catalog does not relicense the data.
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/connectomics/celegans-structure-function-divergence/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+## Protocol and source
 
-Original source terms apply. This catalog does not relicense the data.
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/connectomics/celegans-structure-function-divergence/README.md) defines the scientific protocol, preparation, and scoring.
+
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/connectomics/celegans-structure-function-divergence/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/bciciv-2a/README.md

@@ -1,32 +1,26 @@
 # bciciv-2a — 4-class Motor Imagery decoding (BCI Competition IV 2a)
 
-[Task catalog](../../../README.md)
+The agent designs a PyTorch `nn.Module` **MIAgentModel** that classifies a `(B, 22, 512)` tensor of cue-relative, 4–38 Hz motor-imagery EEG (128 Hz, microvolts) into 4 classes (left hand / right hand / feet / tongue).
 
-4-class motor-imagery decoding on BCI Competition IV Dataset 2a (9 subjects, within-subject cross-session: session T for training, session E for blind test). The agent submits a single PyTorch `MIAgentModel` class (no direct reuse of braindecode's prebuilt decoders); the scorer trains and infers per subject, reporting per-subject accuracy and Cohen's kappa with mean_kappa as the primary metric.
+Historical Agent results for this retained task are recorded in the v0 report.
 
-Historical results for this retained task remain in the v0 report.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `bciciv-2a` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `See task protocol` (See task protocol) |
-
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/bciciv-2a/public) contains 1 files (202,714,474 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
-
-## Protocol and source
+| Input | Cue-relative, 4–38 Hz motor-imagery EEG: `(B, 22, 512)` at 128 Hz. |
+| Output | Four-class `MIAgentModel` prediction: left hand, right hand, feet, or tongue. |
+| Split | Session T has a 230/58 stratified train/valid split per subject. Private session E has 288 trials per subject. |
+| Metric | Mean Cohen’s kappa across nine subjects; the leaderboard `score` clips `mean_kappa` to `[0,1]`. |
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/bciciv-2a/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Data access
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/bciciv-2a/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/bciciv-2a/public) contains 1 file (202,714,474 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
 
 Original source terms apply. This catalog does not relicense the data.
+
+## Protocol and source
+
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/bciciv-2a/README.md) defines the scientific protocol, preparation, and scoring.
+
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/bciciv-2a/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-cvep-thielen2015/README.md

@@ -1,32 +1,29 @@
 # Thielen2015 cVEP decoding
 
-[Task catalog](../../../README.md)
+This task asks an agent to classify one-second cVEP EEG windows into 36 native classes. The source protocol keeps train and validation subjects separate.
 
-Thielen2015 cVEP; frozen cvep/thielen2015 protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-cvep-thielen2015` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `cvep/thielen2015` (`thielen2015`) |
+| EEG input | 64 channels · 120 Hz · 1 s · `[64, 120]` samples |
+| Public windows | 2,592 train · 1,296 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 36 native classes · normalized probabilities `[B, 36]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-cvep-thielen2015/public) contains 4 files (119,737,938 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-cvep-thielen2015/public) contains 4 files (119,737,938 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-cvep-thielen2015/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-cvep-thielen2015/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-cvep-thielen2015/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-cvep-thielen2015/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-dementia-dice/README.md

@@ -1,32 +1,29 @@
 # DICE EEG classification
 
-[Task catalog](../../../README.md)
+The DICE research task classifies five-second EEG windows from source-labeled Alzheimer's, dementia, and healthy groups. Train and validation subjects do not overlap. This is a research classification task, not clinical diagnostic validation.
 
-DICE dementia diagnosis; frozen dementia_diagnosis/default protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-dementia-dice` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `dementia_diagnosis/default` (`Miltiadous2023Dice`) |
+| EEG input | 19 channels · 120 Hz · 5 s · `[19, 600]` samples |
+| Public windows | 8,323 train · 2,863 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 3 native classes · normalized probabilities `[B, 3]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-dementia-dice/public) contains 4 files (510,942,124 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-dementia-dice/public) contains 4 files (510,942,124 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-dementia-dice/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-dementia-dice/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-dementia-dice/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-dementia-dice/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-erp-ern/README.md

@@ -1,32 +1,29 @@
 # ERN EEG classification
 
-[Task catalog](../../../README.md)
+This ERN task distinguishes one-second EEG windows labeled `Target` and `NonTarget` in the source. The validation subjects do not appear in the training split.
 
-ERP CORE ERN; frozen ern/default protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-erp-ern` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `ern/default` (`ERP_CORE`) |
+| EEG input | 30 channels · 120 Hz · 1 s · `[30, 120]` samples |
+| Public windows | 9,591 train · 3,194 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-ern/public) contains 4 files (185,087,647 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-ern/public) contains 4 files (185,087,647 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-ern/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-ern/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-erp-ern/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-ern/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-erp-lrp/README.md

@@ -1,32 +1,29 @@
 # LRP EEG classification
 
-[Task catalog](../../../README.md)
+This LRP task classifies one-second EEG windows as source-labeled `Target` or `NonTarget`. The native split tests the method on subjects held out from training.
 
-ERP CORE LRP; frozen lrp/default protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-erp-lrp` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `lrp/default` (`ERP_CORE`) |
+| EEG input | 30 channels · 120 Hz · 1 s · `[30, 120]` samples |
+| Public windows | 9,591 train · 3,194 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-lrp/public) contains 4 files (185,087,647 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-lrp/public) contains 4 files (185,087,647 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-lrp/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-lrp/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-erp-lrp/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-lrp/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-erp-mmn/README.md

@@ -1,32 +1,29 @@
 # Mismatch-negativity EEG classification
 
-[Task catalog](../../../README.md)
+This mismatch-negativity task distinguishes source-labeled `Target` and `NonTarget` EEG windows. Each window lasts one second, and train and validation use disjoint subject groups.
 
-ERP CORE MMN; frozen mismatch_negativity/default protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-erp-mmn` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `mismatch_negativity/default` (`ERP_CORE`) |
+| EEG input | 30 channels · 120 Hz · 1 s · `[30, 120]` samples |
+| Public windows | 23,835 train · 8,000 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-mmn/public) contains 4 files (460,888,332 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-mmn/public) contains 4 files (460,888,332 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-mmn/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-mmn/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-erp-mmn/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-mmn/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-erp-n170/README.md

@@ -1,32 +1,29 @@
 # N170 EEG classification
 
-[Task catalog](../../../README.md)
+This N170 task distinguishes one-second EEG windows labeled `Target` and `NonTarget` in the source. Its validation split contains subjects absent from training.
 
-ERP CORE N170; frozen n170/default protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-erp-n170` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `n170/default` (`ERP_CORE`) |
+| EEG input | 30 channels · 120 Hz · 1 s · `[30, 120]` samples |
+| Public windows | 7,680 train · 2,560 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-n170/public) contains 4 files (148,243,682 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-n170/public) contains 4 files (148,243,682 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-n170/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-n170/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-erp-n170/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-n170/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-erp-n2pc/README.md

@@ -1,32 +1,29 @@
 # N2pc EEG classification
 
-[Task catalog](../../../README.md)
+This N2pc task classifies one-second EEG windows as source-labeled `Target` or `NonTarget`. The task uses a subject-disjoint public validation split.
 
-ERP CORE N2pc; frozen n2pc/default protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-erp-n2pc` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `n2pc/default` (`ERP_CORE`) |
+| EEG input | 30 channels · 120 Hz · 1 s · `[30, 120]` samples |
+| Public windows | 7,677 train · 2,570 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-n2pc/public) contains 4 files (148,345,021 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-n2pc/public) contains 4 files (148,345,021 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-n2pc/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-n2pc/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-erp-n2pc/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-n2pc/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-erp-n400/README.md

@@ -1,32 +1,29 @@
 # N400 EEG classification
 
-[Task catalog](../../../README.md)
+This N400 task distinguishes one-second EEG windows labeled `Target` and `NonTarget` in the source. The source protocol separates training and validation subjects.
 
-ERP CORE N400; frozen n400/default protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-erp-n400` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `n400/default` (`ERP_CORE`) |
+| EEG input | 30 channels · 120 Hz · 1 s · `[30, 120]` samples |
+| Public windows | 2,880 train · 960 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-n400/public) contains 4 files (55,590,922 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-n400/public) contains 4 files (55,590,922 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-n400/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-n400/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-erp-n400/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-n400/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-erp-p3/README.md

@@ -1,32 +1,29 @@
 # Kappenman2021 P3 EEG classification
 
-[Task catalog](../../../README.md)
+This P3 task uses one-second EEG windows from the Kappenman2021 source case and two native classes. Its public validation subjects are separate from the training subjects.
 
-ERP CORE P3; frozen p3/kappenman2021p3 protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-erp-p3` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `p3/kappenman2021p3` (`ERP_CORE`) |
+| EEG input | 30 channels · 120 Hz · 1 s · `[30, 120]` samples |
+| Public windows | 4,800 train · 1,600 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-p3/public) contains 4 files (92,652,002 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-erp-p3/public) contains 4 files (92,652,002 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-p3/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-p3/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-erp-p3/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-erp-p3/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-me-highgamma2017/README.md

@@ -1,32 +1,29 @@
 # High-Gamma motor-execution classification
 
-[Task catalog](../../../README.md)
+This task classifies three-second EEG windows from the High-Gamma motor-execution protocol into four native classes. The native split separates subjects between training and validation.
 
-High-Gamma motor execution; frozen motor_execution/schirrmeister2017 protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-me-highgamma2017` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `motor_execution/schirrmeister2017` (`highgamma-schirrmeister2017`) |
+| EEG input | 128 channels · 120 Hz · 3 s · `[128, 360]` samples |
+| Public windows | 8,111 train · 2,960 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 4 native classes · normalized probabilities `[B, 4]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-me-highgamma2017/public) contains 4 files (2,041,458,389 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-me-highgamma2017/public) contains 4 files (2,041,458,389 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-me-highgamma2017/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-me-highgamma2017/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-me-highgamma2017/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-me-highgamma2017/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-me-ofner2017/README.md

@@ -1,32 +1,29 @@
 # Ofner2017 motor-execution classification
 
-[Task catalog](../../../README.md)
+This Ofner2017 task asks an agent to classify three-second motor-execution EEG windows into seven native classes. Training and validation use different subjects.
 
-Ofner2017 motor execution; frozen motor_execution/ofner2017 protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-me-ofner2017` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `motor_execution/ofner2017` (`ofner2017-execution`) |
+| EEG input | 61 channels · 120 Hz · 3 s · `[61, 360]` samples |
+| Public windows | 3,780 train · 1,260 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 7 native classes · normalized probabilities `[B, 7]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-me-ofner2017/public) contains 4 files (443,100,882 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-me-ofner2017/public) contains 4 files (443,100,882 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-me-ofner2017/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-me-ofner2017/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-me-ofner2017/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-me-ofner2017/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-mental-arithmetic-zyma2019/README.md

@@ -1,32 +1,29 @@
 # Zyma2019 mental-arithmetic classification
 
-[Task catalog](../../../README.md)
+This Zyma2019 task classifies five-second mental-arithmetic EEG windows into two native classes. Its public validation split is subject-disjoint.
 
-Zyma mental arithmetic; frozen mental_arithmetic/default protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-mental-arithmetic-zyma2019` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `mental_arithmetic/default` (`Zyma2019Electroencephalograms`) |
+| EEG input | 20 channels · 120 Hz · 5 s · `[20, 600]` samples |
+| Public windows | 986 train · 337 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mental-arithmetic-zyma2019/public) contains 4 files (63,605,750 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mental-arithmetic-zyma2019/public) contains 4 files (63,605,750 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mental-arithmetic-zyma2019/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mental-arithmetic-zyma2019/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-mental-arithmetic-zyma2019/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mental-arithmetic-zyma2019/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-mi-barachant2012/README.md

@@ -1,32 +1,29 @@
 # AlexMI motor-imagery classification
 
-[Task catalog](../../../README.md)
+This AlexMI task asks an agent to classify four-second motor-imagery EEG windows into three native classes. The split separates trigger indices and excludes overlapping windows on the same timeline.
 
-AlexMI motor imagery; frozen motor_imagery/barachant2012 protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-mi-barachant2012` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `motor_imagery/barachant2012` (`barachant2012`) |
+| EEG input | 16 channels · 120 Hz · 4 s · `[16, 480]` samples |
+| Public windows | 288 train · 96 validation |
+| Native split | Trigger-index-disjoint native SklearnSplit. Same-timeline windows do not overlap. |
+| Output | 3 native classes · normalized probabilities `[B, 3]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-barachant2012/public) contains 4 files (11,825,808 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-barachant2012/public) contains 4 files (11,825,808 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-barachant2012/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-barachant2012/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-mi-barachant2012/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-barachant2012/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-mi-bnci2014/README.md

@@ -1,32 +1,29 @@
 # BCI-IV-2a native motor-imagery classification
 
-[Task catalog](../../../README.md)
+This task classifies four-second BCI-IV-2a EEG windows into four source-defined motor-imagery classes. It retains the native subject grouping from the frozen AgentEEGBench protocol. It is not the legacy `bciciv-2a` task, which fits session T and scores session E per subject.
 
-BCI-IV-2a native subject-grouped motor-imagery task, shared with AgentEEGBench.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-mi-bnci2014` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `motor_imagery/tangermann2012` · BCI-IV-2a |
+| EEG input | 22 channels · 120 Hz · 4 s · `[22, 480]` samples |
+| Public windows | 2,880 train · 1,152 validation |
+| Native split | Subject-grouped · distinct from the legacy session T/E protocol |
+| Output | 4 native classes · normalized probabilities `[B, 4]` |
+| Metric | Public balanced accuracy (maximize) · task-specific scorer |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-bnci2014/public) contains 2 files (170,344,916 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-bnci2014/public) contains 2 files (170,344,916 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-bnci2014/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-bnci2014/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-mi-bnci2014/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-bnci2014/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-mi-cho2017/README.md

@@ -1,32 +1,29 @@
 # Cho2017 motor-imagery classification
 
-[Task catalog](../../../README.md)
+This Cho2017 task classifies four-second motor-imagery EEG windows into two native classes. The source protocol reserves separate subjects for validation.
 
-Cho2017 motor imagery; frozen motor_imagery/cho2017 protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-mi-cho2017` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `motor_imagery/cho2017` (`cho2017`) |
+| EEG input | 64 channels · 120 Hz · 4 s · `[64, 480]` samples |
+| Public windows | 6,120 train · 2,200 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-cho2017/public) contains 4 files (1,023,001,442 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-cho2017/public) contains 4 files (1,023,001,442 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-cho2017/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-cho2017/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-mi-cho2017/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-cho2017/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-mi-eegmmidb/README.md

@@ -1,32 +1,29 @@
 # EEGMMIDB motor-imagery classification
 
-[Task catalog](../../../README.md)
+This EEGMMIDB task uses four-second motor-imagery EEG windows and four native classes. Training and validation subjects do not overlap.
 
-EEGMMIDB motor imagery; frozen motor_imagery/schalk2004bci2000 protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-mi-eegmmidb` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `motor_imagery/schalk2004bci2000` (`schalk2004-eegmmidb`) |
+| EEG input | 64 channels · 120 Hz · 4 s · `[64, 480]` samples |
+| Public windows | 5,670 train · 1,888 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 4 native classes · normalized probabilities `[B, 4]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-eegmmidb/public) contains 4 files (929,308,208 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-eegmmidb/public) contains 4 files (929,308,208 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-eegmmidb/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-eegmmidb/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-mi-eegmmidb/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-eegmmidb/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-mi-openbmi2019/README.md

@@ -1,32 +1,29 @@
 # OpenBMI motor-imagery classification
 
-[Task catalog](../../../README.md)
+This OpenBMI motor-imagery task asks an agent to classify four-second EEG windows into two native classes. Its native split keeps validation subjects out of training.
 
-OpenBMI motor imagery; frozen motor_imagery/lee2019mi protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-mi-openbmi2019` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `motor_imagery/lee2019mi` (`OpenBMI_Lee2019`) |
+| EEG input | 62 channels · 120 Hz · 4 s · `[62, 480]` samples |
+| Public windows | 6,400 train · 2,200 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-openbmi2019/public) contains 4 files (1,024,405,402 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-openbmi2019/public) contains 4 files (1,024,405,402 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-openbmi2019/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-openbmi2019/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-mi-openbmi2019/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-openbmi2019/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-mi-zhou2016/README.md

@@ -1,32 +1,29 @@
 # Zhou2016 motor-imagery classification
 
-[Task catalog](../../../README.md)
+This Zhou2016 task classifies four-second motor-imagery EEG windows into three native classes. The source protocol places training and validation in separate sessions.
 
-Zhou2016 motor imagery; frozen motor_imagery/zhou2016 protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-mi-zhou2016` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `motor_imagery/zhou2016` (`zhou2016`) |
+| EEG input | 14 channels · 120 Hz · 4 s · `[14, 480]` samples |
+| Public windows | 614 train · 586 validation |
+| Native split | Session-disjoint native PredefinedSplit. |
+| Output | 3 native classes · normalized probabilities `[B, 3]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-zhou2016/public) contains 4 files (32,355,620 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-mi-zhou2016/public) contains 4 files (32,355,620 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-zhou2016/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-zhou2016/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-mi-zhou2016/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-mi-zhou2016/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-p3-openbmi2019/README.md

@@ -1,32 +1,29 @@
 # OpenBMI P3 EEG classification
 
-[Task catalog](../../../README.md)
+This OpenBMI P3 task uses one-second EEG windows and two native classes. Training and validation use disjoint subjects under the source protocol.
 
-OpenBMI P3; frozen p3/lee2019erp protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-p3-openbmi2019` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `p3/lee2019erp` (`OpenBMI_Lee2019`) |
+| EEG input | 62 channels · 120 Hz · 1 s · `[62, 120]` samples |
+| Public windows | 264,960 train · 91,080 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 2 native classes · normalized probabilities `[B, 2]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-p3-openbmi2019/public) contains 4 files (10,623,665,898 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-p3-openbmi2019/public) contains 4 files (10,623,665,898 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-p3-openbmi2019/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-p3-openbmi2019/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-p3-openbmi2019/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-p3-openbmi2019/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-sleep-stage-sleepedf/README.md

@@ -1,32 +1,29 @@
 # Sleep-EDF cassette sleep-stage classification
 
-[Task catalog](../../../README.md)
+This task classifies 30-second Sleep-EDF cassette EEG windows into N1, N2, N3, R, and W. It uses a full native export from 78 subjects. The native pipeline crops wake time to 30 minutes and uses a subject-disjoint split.
 
-Sleep-EDF full native subject-disjoint sleep staging with public train/validation; two EEG channels, 30-second windows, and five-class balanced accuracy.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-sleep-stage-sleepedf` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `sleep_stage/default` from `Kemp2000Analysis` |
+| EEG input | 2 channels · 120 Hz · 30 s · `[2, 3600]` samples |
+| Public windows | 116,493 train · 38,071 validation |
+| Native split | Subject-disjoint `SklearnSplit` · 78 source subjects |
+| Output | 5 class probabilities · N1=0, N2=1, N3=2, R=3, W=4 |
+| Metric | Five-class balanced accuracy (maximize) · complete-class profile |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-sleep-stage-sleepedf/public) contains 4 files (4,453,124,284 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-sleep-stage-sleepedf/public) contains 4 files (4,453,124,284 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-sleep-stage-sleepedf/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-sleep-stage-sleepedf/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-sleep-stage-sleepedf/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-sleep-stage-sleepedf/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-ssvep-openbmi2019/README.md

@@ -1,32 +1,29 @@
 # OpenBMI SSVEP classification
 
-[Task catalog](../../../README.md)
+This OpenBMI SSVEP task asks an agent to classify four-second EEG windows into four native classes. The public validation subjects are separate from training subjects.
 
-OpenBMI SSVEP; frozen ssvep/lee2019ssvep protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-ssvep-openbmi2019` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `ssvep/lee2019ssvep` (`OpenBMI_Lee2019`) |
+| EEG input | 62 channels · 120 Hz · 4 s · `[62, 480]` samples |
+| Public windows | 6,400 train · 2,200 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 4 native classes · normalized probabilities `[B, 4]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-ssvep-openbmi2019/public) contains 4 files (1,024,405,402 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-ssvep-openbmi2019/public) contains 4 files (1,024,405,402 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-ssvep-openbmi2019/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-ssvep-openbmi2019/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-ssvep-openbmi2019/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-ssvep-openbmi2019/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-ssvep-wang2017/README.md

@@ -1,32 +1,29 @@
 # Wang2017 SSVEP classification
 
-[Task catalog](../../../README.md)
+This Wang2017 SSVEP task classifies four-second EEG windows into 40 native classes. The native split evaluates the chosen method on different subjects.
 
-Wang2017 SSVEP; frozen ssvep/default protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-ssvep-wang2017` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `ssvep/default` (`wang2017`) |
+| EEG input | 64 channels · 120 Hz · 4 s · `[64, 480]` samples |
+| Public windows | 4,800 train · 1,680 validation |
+| Native split | Subject-disjoint native SklearnSplit. |
+| Output | 40 native classes · normalized probabilities `[B, 40]` |
+| Metric | Balanced accuracy (maximize). Complete-class per-class recall. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-ssvep-wang2017/public) contains 4 files (796,760,562 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-ssvep-wang2017/public) contains 4 files (796,760,562 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-ssvep-wang2017/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-ssvep-wang2017/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-ssvep-wang2017/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-ssvep-wang2017/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/eeg-typing-spanishbcbl/README.md

@@ -1,32 +1,29 @@
 # SpanishBCBL EEG typing classification
 
-[Task catalog](../../../README.md)
+This task classifies one-second typing-related EEG windows into 29 source-defined output classes. The mapping includes characters, space, a number group, and a special-symbol group. The native SimilaritySplit separates train and validation by sentence-similarity clusters.
 
-SpanishBCBL EEG typing; frozen typing/default protocol adapted for public train/valid model validation.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `eeg-typing-spanishbcbl` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Scientific data | `typing/default` (`Levy2026NoninvasiveEeg`) |
+| EEG input | 61 channels · 120 Hz · 1 s · `[61, 120]` samples |
+| Public windows | 87,656 train · 32,909 validation |
+| Native split | Sentence-similarity cluster-disjoint native SimilaritySplit. |
+| Output | 29 native classes · normalized probabilities `[B, 29]` |
+| Metric | Balanced accuracy (maximize). Sparse macro over observed true or predicted classes. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-typing-spanishbcbl/public) contains 4 files (3,539,526,604 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/eeg-typing-spanishbcbl/public) contains 4 files (3,539,526,604 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-typing-spanishbcbl/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-typing-spanishbcbl/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/eeg-typing-spanishbcbl/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/eeg-typing-spanishbcbl/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/eeg/sleep-edf/README.md

@@ -1,32 +1,26 @@
 # sleep-edf — 5-class Sleep Staging (Sleep-EDF Cassette)
 
-[Task catalog](../../../README.md)
+The agent designs a PyTorch `nn.Module` **SleepAgentModel** that classifies a `(B, 1, 3000)` tensor of preprocessed Fpz-Cz sleep EEG (30 s epoch at 100 Hz, microvolts) into 5 sleep stages (Wake / N1 / N2 / N3 / REM).
 
-5-class sleep-staging on Sleep-EDF Cassette (20 subjects, single-channel Fpz-Cz EEG, 30 s epochs → Wake / N1 / N2 / N3 / REM). Subjects 0-13 are the training pool, 14-15 are validation, and 16-19 are the private held-out test set. The agent submits a PyTorch `SleepAgentModel` class (no direct reuse of prebuilt sleep-staging models); the scorer trains one shared model with class-weighted CE and early stopping, then blind-tests on subjects 16-19, reporting Cohen's kappa (primary) alongside accuracy / balanced accuracy / macro F1 / per-class recall.
+Historical Agent results for this retained task are recorded in the v0 report.
 
-Historical results for this retained task remain in the v0 report.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `sleep-edf` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2` |
-| Data category | EEG |
-| Access | `hosted_public` |
-| Primary metric | `See task protocol` (See task protocol) |
-
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/sleep-edf/public) contains 2 files (1,225,526,664 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
-
-## Protocol and source
+| Input | Fpz-Cz sleep EEG: a 30-second, 100 Hz epoch shaped `(B, 1, 3000)`. |
+| Output | One of five sleep stages: Wake, N1, N2, N3, or REM. |
+| Split | Subjects 0–13 train the model; 14–15 support early stopping; 16–19 are private Test. |
+| Metric | `test_kappa` across subjects 16–19; the leaderboard `score` clips it to `[0,1]`. |
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/sleep-edf/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Data access
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/sleep-edf/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/eeg/sleep-edf/public) contains 2 files (1,225,526,664 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
 
 Original source terms apply. This catalog does not relicense the data.
+
+## Protocol and source
+
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/eeg/sleep-edf/README.md) defines the scientific protocol, preparation, and scoring.
+
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/eeg/sleep-edf/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/invasive/falcon-b1-birdsong/README.md

@@ -1,34 +1,32 @@
 # B1: birdsong spectrogram
 
-[Task catalog](../../../README.md)
+A Neuropixels probe recorded 85 threshold-crossing channels from a singing zebra finch. The task predicts one vocalization spectrogram for each complete song motif. It is not a human speech task.
 
-Predict a zebra finch vocalization spectrogram from one complete motif of intracortical threshold crossings at a time.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `falcon-b1-birdsong` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation-motif` |
-| Data category | Invasive recordings |
-| Access | `hosted_public` |
-| Primary metric | `normalized_spectrogram_MSE` (minimize) |
+| Input | One `float32[27000,85]` neural motif: 900 ms sampled at 30 kHz. |
+| Output | One finite `float32[880,158]` vocalization spectrogram per motif. |
+| Split | Two earlier recording dates for Train, the last held-in date for Valid, and three calibration dates for evaluator-only Test. |
+| Metric | Native-mask normalized spectrogram MSE per motif, then the equal-weight date mean. Lower is better. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/falcon-b1-birdsong/public) contains 11 files (170,230,855 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+| Role | Dates | Complete motifs | Neural samples |
+|---|---:|---:|---:|
+| Train | 2 | 40 | 1,080,000 |
+| Valid | 1 | 8 | 216,000 |
+| Evaluator-only Test | 3 | 9 | 243,000 |
 
-This task uses a BrainPilotBench date split. Its scores are not official FALCON leaderboard scores.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/falcon-b1-birdsong/public) contains 11 files (170,230,855 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+The DANDI record lists CC BY 4.0. See the source attribution and notice of transformations.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/intracortical/falcon-b1-birdsong/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://dandiarchive.org/dandiset/001046/draft) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/intracortical/falcon-b1-birdsong/README.md) defines the scientific protocol, preparation, and scoring.
 
-The DANDI record lists CC BY 4.0. See the source attribution and notice of transformations.
+[Source record](https://dandiarchive.org/dandiset/001046/draft) · [Source attribution](../../../ATTRIBUTION.md)

tasks/invasive/falcon-h1-reach-grasp/README.md

@@ -1,33 +1,32 @@
 # H1: attempted reach and grasp
 
-[Task catalog](../../../README.md)
+A human participant watched a virtual arm and attempted arm and grasp movements. Utah arrays recorded threshold crossings in motor cortex. The task predicts seven commanded velocities, not measured arm movement. Read the [FALCON dataset description](https://snel-repo.github.io/falcon/datasets.html) and the [native loader](https://github.com/snel-repo/falcon-challenge/blob/770c9b9597b8f58bf67b5a9ec6401ef705a2cb0f/falcon_challenge/dataloaders.py#L93-L106).
 
-Develop a causal intracortical decoder for seven attempted arm and grasp velocities, with chronological training and public validation on a later recording date.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `falcon-h1-reach-grasp` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation-stream` |
-| Data category | Invasive recordings |
-| Access | `source_only` |
-| Primary metric | `R2_variance_weighted` (maximize) |
+| Input | One current <code>float32[176]</code> threshold-crossing bin every 20 ms, with a trial-start marker. |
+| Output | Finite <code>float32[7]</code> commanded velocity components for the current bin. |
+| Split | Five earlier source dates for Train, one held-in date for Valid, and seven later calibration dates for evaluator-only Test. |
+| Metric | Native-mask <code>R2_variance_weighted</code> per date, then the equal-weight mean across dates. Higher is better. |
 
-This task uses source download and local preparation. No source or derived H1 arrays are hosted here.
-Read and comply with the [source agreement](https://dandiarchive.org/dandiset/000954/draft) before download.
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/intracortical/falcon-h1-reach-grasp/README.md) defines the exact date split and local build commands.
+| Role | Source date IDs | Files |
+|---|---|---:|
+| Train | <code>19250101</code>, <code>19250108</code>, <code>19250113</code>, <code>19250115</code>, <code>19250119</code> | 11 |
+| Valid | <code>19250120</code> | 2 |
+| Evaluator-only Test | <code>19250126</code>, <code>19250127</code>, <code>19250129</code>, <code>19250202</code>, <code>19250203</code>, <code>19250206</code>, <code>19250209</code> | 14 |
 
-This task uses a BrainPilotBench date split. Its scores are not official FALCON leaderboard scores.
+## Data access
 
-## Protocol and source
+This task uses source download and local preparation. No H1 source or derived arrays are hosted here. Read the [source agreement](https://dandiarchive.org/dandiset/000954/draft) before download.
+
+The source description prohibits redistribution under a Data Use Agreement. No H1 data are mirrored.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/intracortical/falcon-h1-reach-grasp/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://dandiarchive.org/dandiset/000954/draft) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/intracortical/falcon-h1-reach-grasp/README.md) defines the scientific protocol, preparation, and scoring.
 
-The source description prohibits redistribution under a Data Use Agreement. No H1 data are mirrored.
+[Source record](https://dandiarchive.org/dandiset/000954/draft) · [Source attribution](../../../ATTRIBUTION.md)

tasks/invasive/falcon-h2-handwriting/README.md

@@ -1,34 +1,32 @@
 # H2: attempted handwriting
 
-[Task catalog](../../../README.md)
+A BrainGate2 participant attempted to write prompted English sentences while intracortical arrays recorded spike counts. The task decodes one sentence after each complete trial.
 
-Develop a streaming decoder for attempted handwriting from intracortical activity and return one English sentence at each complete trial boundary.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `falcon-h2-handwriting` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation-stream` |
-| Data category | Invasive recordings |
-| Access | `hosted_public` |
-| Primary metric | `WER` (minimize) |
+| Input | One current mask-selected `float32[192]` neural bin per call. Bins are 20 ms apart and mark trial starts. |
+| Output | One UTF-8 sentence string of at most 2,048 bytes after each complete trial. |
+| Split | 20 earlier recording dates for Train, one later held-in date for Valid, and five calibration dates for evaluator-only Test. |
+| Metric | Word error rate (WER) per date, then the equal-weight date mean. Lower is better. Character error rate (CER) is secondary. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/falcon-h2-handwriting/public) contains 86 files (1,110,017,522 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+| Role | Dates | Complete trials | Neural bins |
+|---|---:|---:|---:|
+| Train | 20 | 647 | 1,397,554 |
+| Valid | 1 | 24 | 43,929 |
+| Evaluator-only Test | 5 | 15 | 22,712 |
 
-This task uses a BrainPilotBench date split. Its scores are not official FALCON leaderboard scores.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/falcon-h2-handwriting/public) contains 86 files (1,110,017,522 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+The DANDI record lists CC BY 4.0. See the source attribution and notice of transformations.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/intracortical/falcon-h2-handwriting/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://dandiarchive.org/dandiset/000950/0.241029.1403) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/intracortical/falcon-h2-handwriting/README.md) defines the scientific protocol, preparation, and scoring.
 
-The DANDI record lists CC BY 4.0. See the source attribution and notice of transformations.
+[Source record](https://dandiarchive.org/dandiset/000950/0.241029.1403) · [Source attribution](../../../ATTRIBUTION.md)

tasks/invasive/falcon-m1-emg/README.md

@@ -1,34 +1,26 @@
 # M1-A: upper-limb EMG
 
-[Task catalog](../../../README.md)
+This task predicts 16 upper-limb muscle signals from 64 intracortical channels during macaque reach and grasp.
 
-Develop a causal 64-channel decoder for 16 upper-limb EMG signals. Train on earlier M1-A dates and validate on the final held-in date.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `falcon-m1-emg` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation-stream` |
-| Data category | Invasive recordings |
-| Access | `hosted_public` |
-| Primary metric | `R2_variance_weighted` (maximize) |
+| Input | One current `float32[64]` neural bin per call, with a trial-start marker. Bins are 20 ms apart. |
+| Output | Finite `float32[16]` preprocessed EMG values for the current bin. |
+| Split | Three earlier recording dates for Train, one date for Valid, and three later dates for evaluator-only Test. |
+| Metric | Native-mask variance-weighted R² for each date, then the equal-weight mean across dates. Higher is better. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/falcon-m1-emg/public) contains 18 files (134,424,422 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-This task uses a BrainPilotBench date split. Its scores are not official FALCON leaderboard scores.
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/falcon-m1-emg/public) contains 18 files (134,424,422 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
 
-## Protocol and source
+The DANDI record lists CC BY 4.0. See the source attribution and notice of transformations.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/intracortical/falcon-m1-emg/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://dandiarchive.org/dandiset/000941/0.241029.1405) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/intracortical/falcon-m1-emg/README.md) defines the scientific protocol, preparation, and scoring.
 
-The DANDI record lists CC BY 4.0. See the source attribution and notice of transformations.
+[Source record](https://dandiarchive.org/dandiset/000941/0.241029.1405) · [Source attribution](../../../ATTRIBUTION.md)

tasks/invasive/falcon-m2-finger-velocity/README.md

@@ -1,34 +1,26 @@
 # M2: finger-group velocity
 
-[Task catalog](../../../README.md)
+This task predicts two finger-group velocities from 96 macaque intracortical channels during finger control.
 
-Develop a causal 96-channel decoder for two finger-group velocities. Train on earlier M2 dates and validate on the final held-in date.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `falcon-m2-finger-velocity` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation-stream` |
-| Data category | Invasive recordings |
-| Access | `hosted_public` |
-| Primary metric | `R2_variance_weighted` (maximize) |
+| Input | One current `float32[96]` neural bin per call, with a trial-start marker. Bins are 20 ms apart. |
+| Output | Finite `float32[2]` finger-group velocities for the current bin. |
+| Split | Three earlier recording dates for Train, one date for Valid, and four later dates for evaluator-only Test. |
+| Metric | Native-mask variance-weighted R² for each date, then the equal-weight mean across dates. Higher is better. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/falcon-m2-finger-velocity/public) contains 30 files (13,234,392 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-This task uses a BrainPilotBench date split. Its scores are not official FALCON leaderboard scores.
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/falcon-m2-finger-velocity/public) contains 30 files (13,234,392 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
 
-## Protocol and source
+The DANDI record lists CC BY 4.0. See the source attribution and notice of transformations.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/intracortical/falcon-m2-finger-velocity/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://dandiarchive.org/dandiset/000953/draft) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/intracortical/falcon-m2-finger-velocity/README.md) defines the scientific protocol, preparation, and scoring.
 
-The DANDI record lists CC BY 4.0. See the source attribution and notice of transformations.
+[Source record](https://dandiarchive.org/dandiset/000953/draft) · [Source attribution](../../../ATTRIBUTION.md)

tasks/invasive/ieeg-film-speech-music/README.md

@@ -1,35 +1,28 @@
 # Film speech versus music from clinical iEEG
 
-[Task catalog](../../../README.md)
+This task classifies speech and music intervals in a film from clinical ECoG and sEEG recordings. It tests held-out film blocks within each participant.
 
-Predict speech or music from independent clinical iEEG windows in unseen segments of the same film for each participant.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `ieeg-film-speech-music` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-complete-guarded-windows` |
-| Data category | Invasive recordings |
-| Access | `hosted_public` |
-| Primary metric | `balanced_accuracy` (maximize) |
+| Input | One `float32[channels,512]` two-second neural window and an opaque participant ID. |
+| Output | One finite `float32` speech probability in `[0,1]` for each window. |
+| Split | Global film blocks: Train 1, 3, 6, 7, 8, 10, 13. Valid 4, 11. Evaluator-only Test 2, 5, 9, 12. |
+| Metric | Average window probabilities within each block. Compute balanced accuracy per participant, then average participants equally. Higher is better. |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/ieeg-film-speech-music/public) contains 257 files (922,341,200 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-The public package supports Agent preparation. Block-first scoring still requires the trusted evaluator mapping.
-The package excludes the movie, owner mapping, shuffle key, and Test data.
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/ieeg-film-speech-music/public) contains 257 files (922,341,200 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
 
-## Protocol and source
+The public package supports Agent preparation. Block-first scoring requires the trusted evaluator’s owner mapping. That mapping is not public.
+
+The pinned OpenNeuro metadata lists CC0. The copyrighted film is excluded.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/ieeg/ieeg-film-speech-music/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://doi.org/10.18112/openneuro.ds003688.v1.0.7) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/ieeg/ieeg-film-speech-music/README.md) defines the scientific protocol, preparation, and scoring.
 
-The pinned OpenNeuro metadata lists CC0. The copyrighted film is excluded.
+[Source record](https://doi.org/10.18112/openneuro.ds003688.v1.0.7) · [Source attribution](../../../ATTRIBUTION.md)

tasks/invasive/ieeg-finger-flexion-bciciv4/README.md

@@ -1,36 +1,26 @@
 # BCI Competition IV Dataset 4: finger flexion
 
-[Task catalog](../../../README.md)
+This task predicts continuous finger flexion from ECoG recordings in three participants. Each model returns five trajectories, while the score uses four fingers.
 
-Predict five finger-flexion trajectories from human ECoG in three participants, with chronological public validation and a guarded within-participant split.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `ieeg-finger-flexion-bciciv4` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1-public-validation-offline` |
-| Data category | Invasive recordings |
-| Access | `hosted_public` |
-| Primary metric | `pearson_r` (maximize) |
-
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/ieeg-finger-flexion-bciciv4/public) contains 14 files (324,773,839 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
-
-Train uses samples `[0,319000)` and Valid uses `[321000,400000)` for all three participants.
-The 2,000-sample guard and original 200,000-sample evaluation interval remain outside these public files.
-The mean Pearson correlation scores thumb, index, middle, and little finger. Predictions retain five columns.
-
-## Protocol and source
+| Input | Offline `float32[T,C]` ECoG interval. `C` is 62, 48, or 64 channels for `sub1`, `sub2`, or `sub3`. |
+| Output | Finite `float32[T,5]` flexion in thumb, index, middle, ring, little-finger order at 1000 Hz. |
+| Split | For each participant, Train `[0,319000)` and Valid `[321000,400000)`. A 2,000-sample guard separates them. Original evaluation input stays outside this task's public export. |
+| Metric | Mean Pearson `r` across thumb, index, middle, and little finger for all three participants: 12 correlations, higher is better. |
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/ieeg/ieeg-finger-flexion-bciciv4/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Data access
 
-[Source reference](https://www.bbci.de/competition/iv/) · [Source attribution](../../../ATTRIBUTION.md)
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/invasive/ieeg-finger-flexion-bciciv4/public) contains 14 files (324,773,839 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
 
 Original competition attribution, citation, and publication-notification terms apply. No CC or Apache data license is asserted.
+
+## Protocol and source
+
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/ieeg/ieeg-finger-flexion-bciciv4/README.md) defines the scientific protocol, preparation, and scoring.
+
+[Source record](https://www.bbci.de/competition/iv/) · [Source attribution](../../../ATTRIBUTION.md)

tasks/mri/brain-age-mri/README.md

@@ -1,32 +1,28 @@
 # Age Prediction from Structural MRI
 
-[Task catalog](../../../README.md)
+The agent develops a model that predicts chronological age from T1-weighted brain MRI. Model development uses public images and age labels. Evaluation uses a separate set of held-out subjects.
 
-Predict chronological age from anonymized T1-weighted structural MRI. Build and freeze a model using 449 public subjects; the evaluator runs the submitted inference entry on 112 held-out images and computes subject-level MAE.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Details |
 |---|---|
 | Task ID | `brain-age-mri` |
-| Collection | BrainPilotBench v1 |
 | Task version | `0.1-candidate` |
-| Data category | MRI |
-| Access | `hosted_public` |
-| Primary metric | `test_mae` (minimize) |
+| Public data | MRI scans and chronological ages for 449 subjects |
+| Submission | A saved model, inference code, training code/configuration, public validation predictions, and a report |
+| Primary metric | Mean absolute error in years (`test_mae`), lower is better |
+| Evaluation | 112 held-out subjects with evaluator-controlled age labels |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/mri/brain-age-mri/public) contains 1 files (3,729,262,581 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/mri/brain-age-mri/public) contains 1 file (3,729,262,581 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/mri/brain-age-mri/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/mri/brain-age-mri/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/mri/brain-age-mri/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/mri/brain-age-mri/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/mri/tops-fmri/README.md

@@ -1,32 +1,26 @@
 # tops-fmri — Tonic Pain Signature (linear FC, external validation)
 
-[Task catalog](../../../README.md)
+The agent trains a **linear** 279 ROI / 38781 edge functional-connectivity signature on the public **Study3** fMRI dynamic-FC data (`dfc_5bin_dat` × `CAPS/REST` + pain ratings).
 
-用公开 Study3 fMRI dynamic-FC 训练一个 279 ROI / 38781 edge 的**线性** tonic-pain connectivity signature,提交权重 w_raw/b_raw 和 numpy-only apply_signature.py; benchmark 平台**私下**用 Study4/Study5 held-out features + labels 计算 Pearson r / AUC,聚合 0-1 分。agent 无法接触 held-out labels(工作区隔离 + scorer only)。
+Historical Agent results for this retained task are recorded in the v0 report.
 
-Historical results for this retained task remain in the v0 report.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
+| Field | Task contract |
 |---|---|
-| Task ID | `tops-fmri` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.1` |
-| Data category | MRI |
-| Access | `hosted_public` |
-| Primary metric | `See task protocol` (See task protocol) |
-
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/mri/tops-fmri/public) contains 2 files (965,667,246 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
-
-## Protocol and source
+| Input | Public Study3 fMRI dynamic functional connectivity: 279 ROIs, 38,781 edges, and pain ratings. |
+| Output | A linear connectivity signature with saved weights and a NumPy inference script. |
+| Evaluation | Study3 supports public development. Private Study4 and Study5 supply external evaluation labels. |
+| Metric | `score = 0.5·study4_score + 0.5·study5_score`; Study4 uses Pearson `r`, and Study5 uses AUC. |
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/mri/tops-fmri/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Data access
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/mri/tops-fmri/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/mri/tops-fmri/public) contains 2 files (965,667,246 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
 
 Original source terms apply. This catalog does not relicense the data.
+
+## Protocol and source
+
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/mri/tops-fmri/README.md) defines the scientific protocol, preparation, and scoring.
+
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/mri/tops-fmri/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-badham2017deficits-exp1/README.md

@@ -1,32 +1,28 @@
 # Modeling Human Category Learning from Feedback
 
-[Task catalog](../../../README.md)
+Participants classify geometric objects as E or K across four rule problems and receive feedback. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for badham2017deficits/exp1.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-badham2017deficits-exp1` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `badham2017deficits/exp1.csv` |
+| Public data | 68 train and 17 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-badham2017deficits-exp1/public) contains 2 files (2,304,677 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-badham2017deficits-exp1/public) contains 2 files (2,304,677 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-badham2017deficits-exp1/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-badham2017deficits-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-badham2017deficits-exp1/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-badham2017deficits-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-bahrami2020four-exp/README.md

@@ -1,32 +1,28 @@
 # Predicting Human Choices Among Four Rewards
 
-[Task catalog](../../../README.md)
+Participants repeatedly choose among four labeled options and observe the points earned. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for bahrami2020four/exp.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-bahrami2020four-exp` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `bahrami2020four/exp.csv` |
+| Public data | 695 train and 174 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-bahrami2020four-exp/public) contains 2 files (5,106,777 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-bahrami2020four-exp/public) contains 2 files (5,106,777 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-bahrami2020four-exp/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-bahrami2020four-exp/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-bahrami2020four-exp/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-bahrami2020four-exp/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-collsioo2023mcpl-exp1-775cd1984f/README.md

@@ -1,32 +1,28 @@
 # Human Hormone-Level Estimates from Verbal Cues
 
-[Task catalog](../../../README.md)
+Participants estimate a target hormone level from two verbally described hormone levels. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for collsiöö2023MCPL/exp1.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-collsioo2023mcpl-exp1-775cd1984f` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `collsiöö2023MCPL/exp1.csv` |
+| Public data | 56 train and 14 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-collsioo2023mcpl-exp1-775cd1984f/public) contains 2 files (3,027,152 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-collsioo2023mcpl-exp1-775cd1984f/public) contains 2 files (3,027,152 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-collsioo2023mcpl-exp1-775cd1984f/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-collsioo2023mcpl-exp1-775cd1984f/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-collsioo2023mcpl-exp1-775cd1984f/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-collsioo2023mcpl-exp1-775cd1984f/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-collsioo2023mcpl-exp2-6bd18ff5bf/README.md

@@ -1,32 +1,28 @@
 # Human Hormone-Level Estimates from Numeric Cues
 
-[Task catalog](../../../README.md)
+Participants estimate a target hormone level from two numbered hormone levels. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for collsiöö2023MCPL/exp2.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-collsioo2023mcpl-exp2-6bd18ff5bf` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `collsiöö2023MCPL/exp2.csv` |
+| Public data | 59 train and 15 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-collsioo2023mcpl-exp2-6bd18ff5bf/public) contains 2 files (3,201,056 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-collsioo2023mcpl-exp2-6bd18ff5bf/public) contains 2 files (3,201,056 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-collsioo2023mcpl-exp2-6bd18ff5bf/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-collsioo2023mcpl-exp2-6bd18ff5bf/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-collsioo2023mcpl-exp2-6bd18ff5bf/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-collsioo2023mcpl-exp2-6bd18ff5bf/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-collsioo2023mcpl-exp3-c372064286/README.md

@@ -1,32 +1,28 @@
 # Human Hormone-Level Estimates from Two Cues
 
-[Task catalog](../../../README.md)
+Participants estimate a target hormone level from two verbally described hormone levels. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for collsiöö2023MCPL/exp3.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-collsioo2023mcpl-exp3-c372064286` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `collsiöö2023MCPL/exp3.csv` |
+| Public data | 70 train and 18 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-collsioo2023mcpl-exp3-c372064286/public) contains 2 files (2,025,867 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-collsioo2023mcpl-exp3-c372064286/public) contains 2 files (2,025,867 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-collsioo2023mcpl-exp3-c372064286/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-collsioo2023mcpl-exp3-c372064286/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-collsioo2023mcpl-exp3-c372064286/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-collsioo2023mcpl-exp3-c372064286/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-cox2017information-exp1/README.md

@@ -1,32 +1,28 @@
 # Modeling Human Recognition of Studied Word Pairs
 
-[Task catalog](../../../README.md)
+Participants study word pairs and judge whether a later pair appeared together in the study list. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for cox2017information/exp1.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-cox2017information-exp1` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `cox2017information/exp1.csv` |
+| Public data | 339 train and 85 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-cox2017information-exp1/public) contains 2 files (8,684,274 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-cox2017information-exp1/public) contains 2 files (8,684,274 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-cox2017information-exp1/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-cox2017information-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-cox2017information-exp1/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-cox2017information-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-enkavi2019adaptivenback-exp1/README.md

@@ -1,32 +1,28 @@
 # Predicting Human N-Back Letter Responses
 
-[Task catalog](../../../README.md)
+Participants compare each letter with the one shown N trials earlier and press a match or nonmatch key. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for enkavi2019adaptivenback/exp1.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-enkavi2019adaptivenback-exp1` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `enkavi2019adaptivenback/exp1.csv` |
+| Public data | 376 train and 94 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-enkavi2019adaptivenback-exp1/public) contains 2 files (8,466,948 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-enkavi2019adaptivenback-exp1/public) contains 2 files (8,466,948 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019adaptivenback-exp1/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019adaptivenback-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-enkavi2019adaptivenback-exp1/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019adaptivenback-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-enkavi2019digitspan-exp1/README.md

@@ -1,32 +1,28 @@
 # Modeling Human Digit-Sequence Recall
 
-[Task catalog](../../../README.md)
+Participants view a digit sequence and type the digits in their remembered order. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for enkavi2019digitspan/exp1.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-enkavi2019digitspan-exp1` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `enkavi2019digitspan/exp1.csv` |
+| Public data | 377 train and 95 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-enkavi2019digitspan-exp1/public) contains 2 files (5,692,403 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-enkavi2019digitspan-exp1/public) contains 2 files (5,692,403 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019digitspan-exp1/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019digitspan-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-enkavi2019digitspan-exp1/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019digitspan-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-enkavi2019gonogo-exp1/README.md

@@ -1,32 +1,28 @@
 # Predicting Human Go/No-Go Keypresses
 
-[Task catalog](../../../README.md)
+Participants press a key for one color and withhold a response for another color. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for enkavi2019gonogo/exp1.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-enkavi2019gonogo-exp1` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `enkavi2019gonogo/exp1.csv` |
+| Public data | 370 train and 93 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-enkavi2019gonogo-exp1/public) contains 2 files (7,564,952 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-enkavi2019gonogo-exp1/public) contains 2 files (7,564,952 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019gonogo-exp1/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019gonogo-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-enkavi2019gonogo-exp1/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019gonogo-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-enkavi2019recentprobes-exp1/README.md

@@ -1,32 +1,28 @@
 # Modeling Human Recognition of Recent Letters
 
-[Task catalog](../../../README.md)
+Participants remember six letters and judge whether a later probe letter belonged to that list. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for enkavi2019recentprobes/exp1.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-enkavi2019recentprobes-exp1` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `enkavi2019recentprobes/exp1.csv` |
+| Public data | 376 train and 95 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-enkavi2019recentprobes-exp1/public) contains 2 files (3,678,382 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-enkavi2019recentprobes-exp1/public) contains 2 files (3,678,382 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019recentprobes-exp1/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019recentprobes-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-enkavi2019recentprobes-exp1/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-enkavi2019recentprobes-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-feng2021dynamics-exp1/README.md

@@ -1,32 +1,28 @@
 # Predicting Human Choices Between Changing Slot Machines
 
-[Task catalog](../../../README.md)
+Participants choose between two slot machines over multiple games and observe point outcomes. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for feng2021dynamics/exp1.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-feng2021dynamics-exp1` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `feng2021dynamics/exp1.csv` |
+| Public data | 20 train and 6 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-feng2021dynamics-exp1/public) contains 2 files (3,114,689 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-feng2021dynamics-exp1/public) contains 2 files (3,114,689 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-feng2021dynamics-exp1/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-feng2021dynamics-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-feng2021dynamics-exp1/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-feng2021dynamics-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)

tasks/psychology/psych101-flesch2018comparing-exp1/README.md

@@ -1,32 +1,28 @@
 # Modeling Human Tree-Category Learning Across Gardens
 
-[Task catalog](../../../README.md)
+Participants accept or reject trees that vary in leaves and branches across two gardens, with feedback during training. The agent predicts participants' recorded responses from the information available before each response.
 
-Public Psych-101 marked-response UTF-8 byte+EOS likelihood for flesch2018comparing/exp1.csv. New protocol version; not numerically comparable with historical token or categorical scores.
+Agent benchmark results are not yet available.
 
-Agent benchmark results are not yet available. Preparation and fixed-reference checks are not Agent results.
+## Task at a glance
 
-## Task and data
-
-| Field | Value |
-|---|---|
+| Field | Details |
+| --- | --- |
 | Task ID | `psych101-flesch2018comparing-exp1` |
-| Collection | BrainPilotBench v1 |
-| Task version | `0.2-byte-next-v1` |
-| Data category | Psychology |
-| Access | `hosted_public` |
-| Primary metric | `mean_response_nll` (minimize) |
+| Source experiment | `flesch2018comparing/exp1.csv` |
+| Public data | 256 train and 64 valid participants |
+| Agent output | `report.md` and three `submission/seed-N/` model bundles |
+| Prediction protocol | `byte_next_v1`: next UTF-8 byte or EOS for marked responses |
+| Public metric | `mean_response_nll` ↓, nats per response · public validation only |
 
-The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-flesch2018comparing-exp1/public) contains 2 files (31,867,420 bytes).
-The [manifest](manifest.json) records each relative path, byte count, and SHA-256 digest.
-Use the task data lock in the BrainPilotBench source checkout for commit-pinned downloads.
-The `main` links on this page are navigation links, not immutable data identifiers.
+## Data access
 
-## Protocol and source
+The [public folder](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Public/tree/main/tasks/psychology/psych101-flesch2018comparing-exp1/public) contains 2 files (31,867,420 bytes). The [manifest](manifest.json) lists each path, byte count, and SHA-256 digest.
+
+Original source terms apply. This catalog does not relicense the data.
 
-The [task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-flesch2018comparing-exp1/README.md) describes the inputs, outputs, split, environment, and scoring interface.
-Scientific task versions remain separate from the collection name and storage layout.
+## Protocol and source
 
-[Source reference](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-flesch2018comparing-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)
+The [versioned task guide](https://github.com/NeuroAIHub/BrainPilotBench/blob/7306cc0b0e8958cc8f84918b68446310d8dbc662/tasks/psychology/psych101-flesch2018comparing-exp1/README.md) defines the scientific protocol, preparation, and scoring.
 
-Original source terms apply. This catalog does not relicense the data.
+[Frozen source record](https://github.com/NeuroAIHub/BrainPilotBench/blob/9d92b6738189d23607478ee1f71bdb6ff567d07b/tasks/psychology/psych101-flesch2018comparing-exp1/README.md) · [Source attribution](../../../ATTRIBUTION.md)