Polish linked public card docs

README.md

@@ -40,8 +40,8 @@ mission-held-out validation, and transparent release boundaries.
 | Hugging Face dataset | Public processed fold package | Download selected LOMO feature matrices and result artifacts | [HF dataset card](docs/hf_dataset_card.md) |
 | v9 public bulk | Metadata catalog | Task catalog, source inventory, and baseline summaries | [v9 HF-style card](docs/v9_hf_dataset_card.md) |
 
-For linked methods, evaluation, and release cards, start with the
-[SpaceBio-Bench card pack](docs/SPACEBIOBENCH_TRANSPARENCY_CARD_PACK.md).
+For linked methods, evaluation, and release-status notes, start with the
+[SpaceBio-Bench public documentation map](docs/SPACEBIOBENCH_TRANSPARENCY_CARD_PACK.md).
 For machine-readable release status, see
 [release/release_manifest.json](release/release_manifest.json).
 
@@ -159,10 +159,11 @@ scripts/               Data, evaluation, upload, validation, and figure scripts
 | Need | Document |
 |---|---|
 | Public result source | [docs/CANONICAL_RESULTS_V7_1.md](docs/CANONICAL_RESULTS_V7_1.md) |
-| Methods, evaluation, and release cards | [docs/SPACEBIOBENCH_TRANSPARENCY_CARD_PACK.md](docs/SPACEBIOBENCH_TRANSPARENCY_CARD_PACK.md) |
+| Public documentation map | [docs/SPACEBIOBENCH_TRANSPARENCY_CARD_PACK.md](docs/SPACEBIOBENCH_TRANSPARENCY_CARD_PACK.md) |
 | System scope | [docs/SPACEBIOBENCH_SYSTEM_CARD.md](docs/SPACEBIOBENCH_SYSTEM_CARD.md) |
 | Evaluation interpretation | [docs/SPACEBIOBENCH_EVALUATION_CARD.md](docs/SPACEBIOBENCH_EVALUATION_CARD.md) |
-| Release readiness | [docs/SPACEBIOBENCH_RELEASE_READINESS_CARD.md](docs/SPACEBIOBENCH_RELEASE_READINESS_CARD.md) |
+| Release status | [docs/SPACEBIOBENCH_RELEASE_READINESS_CARD.md](docs/SPACEBIOBENCH_RELEASE_READINESS_CARD.md) |
+| Public statement guide | [docs/SPACEBIOBENCH_CLAIM_REGISTER.md](docs/SPACEBIOBENCH_CLAIM_REGISTER.md) |
 | Hugging Face dataset card source | [docs/hf_dataset_card.md](docs/hf_dataset_card.md) |
 | v9 metadata catalog card source | [docs/v9_hf_dataset_card.md](docs/v9_hf_dataset_card.md) |
 | Contributing and submissions | [CONTRIBUTING.md](CONTRIBUTING.md) and [docs/submission_format.md](docs/submission_format.md) |

docs/SPACEBIOBENCH_EVALUATION_CARD.md

@@ -1,72 +1,53 @@
 ---
 title: SpaceBio-Bench Evaluation Card
 page_type: evaluation_card
-status: public_review_ready
-last_reviewed: 2026-06-05
-claim_boundary: benchmark_evaluation_card_draft_no_new_result_claim
+status: public_ready
+last_reviewed: 2026-06-16
 ---
 
 # SpaceBio-Bench Evaluation Card
 
-## Evaluation Purpose
+## Purpose
 
-This card documents how SpaceBio-Bench evaluations should be interpreted. It
-separates task validity, fold structure, metric reporting, baseline status, and
-claim boundaries so that benchmark scores are not overread as biological
-mechanism, translational readiness, or model superiority claims.
-
-This card does not introduce new results. It summarizes evaluation evidence
-already recorded in the v7.1 canonical result surface and the v9 public bulk
-metadata-alpha scaffold.
-
-Branch note: on the default `main` branch, v9-specific evidence paths such as
-`v9/...` and `docs/V9_*` refer to the curated public bulk metadata-alpha subset
-included in this repository. Payload matrices and draft extension lanes are
-excluded from the current public-review path.
+This card explains how to read SpaceBio-Bench evaluations. It separates task
+definition, fold structure, metric reporting, baseline rows, and pooled
+summaries so readers can understand what a score means.
 
 ## Evaluation Surfaces
 
-| Surface | Current status | Evaluation unit | Result boundary |
-|---|---|---|---|
-| v1-v7 / v7.1 canonical surface | Canonical historical result surface | Tissue, method, feature type, held-out validation, and FM snapshot rows | Cross-mission transcriptomics benchmark summary |
-| v9 public bulk alpha | Metadata-only alpha scaffold | Task manifest, LOMO fold, baseline run, prediction row, run manifest | Workflow and provenance evidence, not leaderboard ranking |
-| v9 draft extension lanes | Diagnostic or feasibility scaffolds | Asset inventory, metric spec, payload audit, draft task manifest, or diagnostic run | Draft-only; no public benchmark score claim |
+| Surface | Evaluation unit | Public interpretation |
+|---|---|---|
+| v1-v7 / v7.1 canonical surface | Tissue, method, feature type, held-out validation, and foundation-model snapshot rows | Historical cross-mission transcriptomics benchmark summary |
+| Hugging Face public fold package | Public LOMO fold with matrices, labels, metadata, and selected genes | Processed data package for reproducible task-level evaluation |
+| v9 public bulk catalog | Task manifest, LOMO fold, baseline run, prediction row, metric file, and run record | Metadata catalog and reference baseline surface |
 
 ## Evaluation Flow
 
-| Stage | Evidence to inspect | Interpretation control |
+| Stage | Evidence to inspect | Why it matters |
 |---|---|---|
-| 1. Source inventory | OSDR accessions, tissue labels, mission labels, access status, and checksum-manifest evidence | Confirms the public data source before interpreting any score |
-| 2. Task manifest | Task id, tissue, feature namespace, source ids, label map, and metric ids | Defines what the evaluation is actually testing |
-| 3. Held-out mission fold | Train/test mission split, row counts, and selected-gene counts | Keeps mission-held-out validation separate from random-split performance |
-| 4. Prediction and metric files | Baseline or submitted predictions, task/fold ids, AUROC, macro-F1, balanced accuracy, calibration | Ties every metric to a concrete task and fold surface |
-| 5. Per-task interpretation | Tissue-specific and fold-specific behavior | Prevents pooled means from hiding failures or confounding |
-| 6. Pooled summary | Aggregate result only after task/fold checks | Allows navigation-level summaries with mission, tissue, baseline, and payload caveats |
-| 7. Claim register language | Allowed, blocked, and future-only wording | Converts evaluation evidence into release-safe public claims |
+| 1. Source rows | OSDR accessions, tissue labels, mission labels, and access status | Identifies the public data behind a task |
+| 2. Task manifest | Task ID, tissue, feature namespace, label map, and metric IDs | Defines what the evaluation is testing |
+| 3. Held-out mission fold | Train/test mission split, row counts, and selected-gene counts | Separates mission-held-out validation from random-split performance |
+| 4. Prediction and metric files | Predictions, task/fold IDs, AUROC, macro-F1, balanced accuracy, and calibration | Ties every score to a concrete task and fold |
+| 5. Per-task interpretation | Tissue-specific and fold-specific behavior | Preserves mission and tissue variability |
+| 6. Pooled summary | Aggregate metrics after task and fold checks | Gives a navigation-level summary |
 
-The evaluation flow is intentionally claim-aware. A score is first interpreted
-at the task and fold level, then summarized only with caveats about mission,
-tissue, payload, baseline, and release-surface boundaries.
+## v9 Public Bulk Evaluation Unit
 
-## Current v9 Public Bulk Evaluation Unit
+The v9 public bulk catalog organizes mission-held-out classification tasks for
+public mouse bulk RNA-seq sources.
 
-The current v9 public bulk lane evaluates mission-held-out classification tasks
-for public mouse bulk RNA-seq sources. The unit hierarchy is:
-
-| Unit | Meaning | Evidence |
+| Unit | Meaning | Public files |
 |---|---|---|
-| Task | Tissue-specific bulk LOMO task with source ids, missions, feature namespace, and metric ids | `v9/task_manifest_index.csv`; `v9/task_manifests/*.json` |
+| Task | Tissue-specific bulk LOMO task with missions, labels, feature namespace, and metric IDs | `v9/task_manifest_index.csv`; `v9/task_manifests/*.json` |
 | Fold | One held-out mission with train/test row counts and selected-gene counts | `v9/task_data_index.csv` |
 | Baseline run | One baseline family evaluated on one task | `v9/reports/bulk_lomo_baseline_summary.csv` |
-| Prediction file | Per-sample prediction output for a task/baseline run | per-baseline `predictions.csv` |
+| Prediction file | Per-sample prediction output for a task and baseline run | per-baseline `predictions.csv` |
 | Metrics file | Task-level metric output for a baseline run | per-baseline `metrics.json` |
-| Run manifest | Provenance for a baseline execution | per-baseline `run_manifest.json` |
+| Run record | Execution metadata for a baseline run | per-baseline `run_manifest.json` |
 
 ## Current Task Inventory
 
-The v9 public bulk metadata-alpha scaffold currently indexes eight generated
-bulk LOMO task manifests and 33 fold definitions:
-
 | Task | Tissue | Variant | Missions | Folds | Sources |
 |---|---|---:|---:|---:|---:|
 | `A1_liver_bulk_lomo` | liver | canonical | 6 | 6 | 6 |
@@ -78,117 +59,74 @@ bulk LOMO task manifests and 33 fold definitions:
 | `A5_skin_bulk_lomo` | skin | canonical | 3 | 3 | 4 |
 | `A6_eye_bulk_lomo` | eye | canonical | 3 | 3 | 3 |
 
-These tasks are part of a metadata-only alpha snapshot. They do not imply a
-frozen payload release.
-
 ## Metrics
 
-Current v9 public bulk task manifests list these metric ids:
-
 | Metric | Interpretation | Reporting note |
 |---|---|---|
 | `macro_f1` | Class-balanced F1 summary across labels | Report per task; small folds can be unstable |
 | `balanced_accuracy` | Mean sensitivity across classes | Useful under class imbalance |
-| `auroc` | Rank-based discrimination between flight/control labels | Can look strong even when calibration or fold reliability is weak |
-| `calibration_error` | Probability calibration diagnostic | Lower is better; compare only across matched output formats |
+| `auroc` | Rank-based discrimination between flight/control labels | Compare with calibration and per-fold behavior |
+| `calibration_error` | Probability calibration diagnostic | Lower is better; compare matched output formats |
 | `mission_discrimination` | Diagnostic for mission-correlated structure where embeddings are available | Not emitted for every baseline family |
 
-Metric reporting should include the task id, fold family, baseline id, variant,
-and whether the result comes from a canonical result surface, metadata alpha, or
-draft diagnostic lane.
-
-## Current v9 Baseline Status
+Metric reports should include task ID, fold ID, baseline or method ID, variant,
+and release surface.
 
-The current v9 public bulk scaffold includes 24 evaluated baseline rows across
-three simple baseline families:
+## Current v9 Baseline Rows
 
-| Baseline family | Rows | Current interpretation |
+| Baseline family | Rows | Interpretation |
 |---|---:|---|
 | `logistic_regression_l2` | 8 | Simple linear workflow baseline |
-| `nearest_centroid` | 8 | Simple centroid-based diagnostic baseline |
+| `nearest_centroid` | 8 | Simple centroid-based baseline |
 | `pca_logistic_regression` | 8 | PCA feature-compression plus logistic baseline |
 
-Mean metrics across the eight task rows are recorded in the v9 dataset card:
+Mean metrics across the eight task rows:
 
 | Baseline | Macro-F1 | Balanced accuracy | AUROC | Calibration error | Mission discrimination |
 |---|---:|---:|---:|---:|---:|
 | Logistic regression L2 | 0.5532 | 0.5808 | 0.6870 | 0.3439 | NA |
 | Nearest centroid | 0.5383 | 0.5685 | 0.6321 | 0.1132 | 0.8733 |
 | PCA logistic regression | 0.5353 | 0.5619 | 0.6447 | 0.3747 | 0.9190 |
 
-These are scaffold baselines. They validate the task/evaluation plumbing and
-provide anchors for future methods, but they are not tuned leaderboard
-endpoints.
-
-## What This Enables
-
-The evaluation card gives readers a compact way to check whether a score is
-being read at the right level: task, fold, baseline run, or pooled summary. It
-also makes clear when a number is a workflow sanity check rather than a model
-ranking.
+These rows provide reproducible reference anchors. Use per-task rows together
+with pooled means when comparing methods.
 
 ## Reading Order For Results
 
-Read results in this order:
-
-1. Release surface and claim boundary.
+1. Release surface.
 2. Task and fold definition.
-3. Source provenance and payload verification status.
+3. Source rows and access status.
 4. Per-fold and per-task metrics.
-5. Baseline run manifest and prediction count.
-6. Pooled summary, only after the above checks.
+5. Baseline run record and prediction count.
+6. Pooled summary.
 
-Do not start with a pooled mean alone. Pooled summaries can hide fold-level
-failure, tissue-specific instability, mission-label confounding, or variant
-differences.
+Starting from the task and fold keeps tissue-specific behavior, mission-label
+structure, and variant differences visible.
 
 ## Leakage And Confounding Controls
 
-Current controls:
-
 - Mission-held-out fold definitions are explicit in `v9/task_data_index.csv`.
 - Task-source relationships are explicit in `v9/task_manifest_index.csv`.
-- v7.1 public wording requires subset notes when comparing mixed FM surfaces.
-- v9 public bulk alpha separates public bulk tasks from draft organoid,
-  single-cell, and multispecies lanes.
-- v9 alpha documents that payload-level hash verification is pending.
-
-Controls that remain future or lane-specific:
-
-- Payload-level hash verification for every distributed fold matrix.
-- Release-time revalidation of feature-selection and preprocessing leakage
-  controls for any frozen v9 payload bundle.
-- Adapter-specific validation for foundation models and virtual-cell models.
-- Human or biological validation for mechanistic interpretation claims.
+- v7.1 public wording includes subset notes for mixed foundation-model
+  comparisons.
+- v9 public bulk tasks are separated from extension workspaces for other
+  modalities.
+- Feature selection and preprocessing should be checked within each training
+  split for any newly packaged fold.
 
 ## Failure Modes To Watch
 
 | Failure mode | Why it matters | Mitigation |
 |---|---|---|
-| Pooled-score overclaim | A high mean can hide poor tissue or mission performance | Always report per-task metrics |
-| Mixed-surface FM comparison | Rows may come from different tissues, adapters, or versions | Label the reported surface for every row |
-| Mission confounding | Mission labels can encode hardware, vehicle, age, protocol, or processing | Treat mission shift as benchmark pressure, not pure biology |
-| Payload-boundary drift | Metadata alpha can be mistaken for frozen payload release | Keep payload hash status in every release-facing card |
-| Extension-lane leakage | Draft organoid, single-cell, or multispecies lanes can be pulled into public bulk claims | Keep release target and claim boundary explicit |
-| Biological interpretation overreach | Classification performance is not mechanistic proof | Separate benchmark evidence from mechanistic validation |
-
-## Boundary Summary
+| Pooled-score overread | A high mean can hide weak tissue or mission rows | Report per-task metrics |
+| Mixed-surface comparison | Rows can come from different tissues, adapters, or versions | Label the release surface for every row |
+| Mission confounding | Mission labels can encode hardware, vehicle, age, protocol, or processing | Interpret mission shift as the benchmark pressure |
+| Biological overinterpretation | Classification performance is not mechanistic validation | Separate benchmark scores from mechanistic follow-up |
 
-The current evaluation surfaces should not be read as evidence for:
-
-- A frozen v9 payload release.
-- A state-of-the-art model leaderboard.
-- A uniform 8-tissue foundation-model comparison.
-- Clinical, crew-health, countermeasure, intervention, or Mars-regime validity.
-- Biological mechanism proof from classifier performance.
-- Payload-level integrity beyond the currently documented checksum-manifest
-  evidence.
-
-## Reproducibility Evidence
-
-Local evidence:
+## Reproducibility Files
 
 - `docs/CANONICAL_RESULTS_V7_1.md`
+- `docs/hf_dataset_card.md`
 - `docs/v9_hf_dataset_card.md`
 - `v9/task_manifest_index.csv`
 - `v9/task_data_index.csv`
@@ -198,9 +136,3 @@ Local evidence:
 - `v9/reports/nearest_centroid/bulk_lomo_summary.csv`
 - `v9/reports/sklearn_baselines/bulk_lomo_summary.csv`
 - `v9/datapackage.draft.json`
-
-Companion cards:
-
-- `docs/SPACEBIOBENCH_SYSTEM_CARD.md`
-- `docs/SPACEBIOBENCH_CLAIM_REGISTER.md`
-- `docs/SPACEBIOBENCH_RELEASE_READINESS_CARD.md`