Refresh Track 1 results with direct Linear substitution

Update all eight verified prediction files, measured WBM fallback rate, main shared-participant bootstrap reference and accompanying Track-1 documentation. Other tracks and Space code are unchanged.

downstream/README.md

@@ -22,16 +22,18 @@ server-side against the `linear` baseline.
 The substrate parquets are the canonical inputs for:
 
 - The OpenMHC HF Space (`MyHeartCounts/OpenMHC`) — the live leaderboard table (skill /
-  fair-skill / average-rank vs the `linear` baseline). The Space **reduces these substrates
-  live on every page load** — point estimate vs the `linear` baseline — exactly like the
+  worst-subgroup skill / average rank vs the `linear` baseline). The Space **reduces these substrates
+  on startup or first access and caches the results in-process** — point estimate vs the `linear` baseline — exactly like the
   imputation and forecasting tracks. So once your `<method>.parquet` (+ `.meta.json`) is on
   the dataset, **it goes straight onto the board** — reduced live vs `linear`, with no
   precomputed rows file and no offline maintainer reduce (the board picks it up on the
   Space's next restart). (The paper's 95% confidence intervals come from the separate
   `downstream/bootstrap/` reference; the live board shows the point value.)
 - Independent re-aggregation (the reducers in `scripts/paper_results/downstream/`).
-- The cluster-bootstrap reference at `downstream/bootstrap/` (per-draw CIs) is reduced
-  from these substrates, so any change here must be matched by a bootstrap refresh.
+- The current main bootstrap reference at `downstream/bootstrap/` uses the same
+  saved evaluation predictions, before float32 pair serialization. Refresh it
+  together with the prediction files. It contains all-participant skill/rank CIs;
+  it does not contain subgroup or worst-subgroup CIs.
 
 ## Loading
 
@@ -60,7 +62,7 @@ print(json.loads(open(meta_p).read()))
 ## Pooled substrate
 
 The pooled per-user frame across all methods is the concatenation of the per-method
-parquets (93,528 rows/method for the canonical 32-task config):
+parquets (92,997 rows/method (30,999 participant–task pairs) for the canonical 32-task config):
 
 ```python
 import glob, pandas as pd
@@ -74,7 +76,9 @@ pooled = pd.concat(
 
 Each sidecar carries `fallback_rate` — the fraction of the method's
 test predictions the harness left non-finite and substituted with the `linear` baseline
-before scoring. `wbm` is the only non-zero method (it embeds only participants with a
+before scoring, using the raw Linear score directly without percentile ranking
+or separate normalization. The pooled rate is the total number of substituted
+participant–task pairs divided by the total evaluated pairs. `wbm` is the only non-zero method (it embeds only participants with a
 full weekly window); the rest are `0.0`. A high rate means the headline scores partly
 reflect the baseline's performance on the substituted cells and should be read with
 caution.
@@ -98,26 +102,33 @@ METHOD=xgboost MHC_DATA_DIR=<data> PREDICTIONS_DIR=results/eval/final/prediction
 PYTHONPATH=src python scripts/paper_results/downstream/bootstrap_downstream_draws.py \
   --predictions_dir results/eval/final/predictions --csvs_dir results/eval/final \
   --methods linear multirocket xgboost lsm2 gru_d wbm toto chronos2 \
+  --fairness_attributes --split-file <canonical-split.json> \
   --output results/paper/bootstrap_draws.parquet
 
 # (3) Build the per-method substrate parquets (+ provenance sidecars).
-python scripts/paper_results/downstream/parity/produce_per_method_per_user_pairs.py \
+python scripts/paper_results/downstream/leaderboard/produce_per_method_per_user_pairs.py \
   --predictions-dir results/eval/final/predictions --out-dir results/leaderboard_downstream
 
-# (4) Parity gate — the substrate must equal the predictions, and a substrate-driven
-#     bootstrap must reproduce results/paper/bootstrap_draws.parquet.
-python scripts/paper_results/downstream/parity/parity_substrate.py
+# (4) Parity gate — check the serialized prediction pairs and point scores.
+#     Bootstrap reproducibility must account for float32 pair serialization.
+python scripts/paper_results/downstream/dev/parity_substrate.py
 
 # (5) Upload the per-method substrates (HF auth required).
-for m in multirocket xgboost lsm2 gru_d wbm toto chronos2; do
+for m in linear multirocket xgboost lsm2 gru_d wbm toto chronos2; do
   python tools/upload_leaderboard_substrate.py --dir results/leaderboard_downstream \
     --method "$m" --track downstream --name "<Display>" --type "<Type>" \
-    --submitter "OpenMHC team" --subtrack static
+    --submitter "OpenMHC team" --subtrack static --fallback-rate <measured-pooled-rate>
 done
 
 # (6) Upload the bootstrap-draws reference (-> downstream/bootstrap/).
 python tools/upload_leaderboard_bootstrap.py --dir results/paper --track downstream
 ```
 
+The CLI examples require the corrected shared-participant bootstrap implementation.
+The Space's pinned point reducer is validated separately and does not generate CIs.
+The bootstrap uploader expects `bootstrap_draws.parquet.meta.json`; explicitly
+stage the generated `bootstrap_draws.meta.json` under that name, or upload it
+directly as `downstream/bootstrap/draws.meta.json`.
+
 Steps (2) and (6) keep the bootstrap CIs on the same canonical predictions as the point
 numbers — run them together whenever the substrates change.

downstream/SCHEMA.md

@@ -13,7 +13,7 @@ The parquet is produced by the evaluation run — pass both `output_dir=` and
 `method_name="<method>"` to `openmhc.evaluate_prediction` and it writes
 `<method>.parquet` with the `method` column set to `<method>`. Upload that file
 verbatim; do not hand-author it. The maintainers concatenate every
-`downstream/*.parquet` and recompute the paired skill score, fair skill score,
+`downstream/*.parquet` and recompute the paired skill score, worst-subgroup skill,
 and average rank against the `linear` baseline during ingestion, so the columns
 and dtypes below must match exactly.
 
@@ -49,25 +49,26 @@ compression.
 | `subgroup_value` | string (dict) | `all` for the global cell; otherwise the subgroup level (age bucket, sex value, or `unknown`) |
 | `user_id` | string (dict) | participant id from the canonical split |
 | `y_true` | float32 | ground-truth label for this `(task, user)` |
-| `y_pred` | float32 | the model's discrete prediction (class for binary/ordinal/multiclass; point value for regression) |
-| `y_proba` | float32 | the model's continuous score — class-1 probability for binary, point prediction otherwise (the column the ranking/correlation metrics read) |
+| `y_pred` | float32 | stored model prediction; the continuous score used by Spearman for ordinal tasks and by Pearson for regression |
+| `y_proba` | float32 | continuous binary score used by average precision; ordinal/regression primary metrics read `y_pred` |
 
 ### Value semantics
 
 - **Missing-prediction fallback is already applied.** A participant the model
   could not produce a finite prediction for has been scored against the `linear`
-  baseline before this file is written (the fraction substituted is reported in
+  baseline by direct raw-score substitution before this file is written (the fraction substituted is reported in
   `fallback_rate` — see the sidecar). There are no NaN pairs.
 - **Structurally-absent tasks are omitted, not NaN-filled.** A task with no
   eligible test cohort simply has no rows.
 
-### Subgroup rows are required for fairness
+### Subgroup rows are required for cohort views and worst-subgroup skill
 
 Every `(task, user)` cell appears three times: once with `subgroup_attr = "all"`
 (the global cell used for skill and rank) and once for each demographic attribute
-(`age_group`, `sex`) the user belongs to. The fairness skill score is computed
-only from the `age_group` / `sex` rows, so a submission that ships only
-`subgroup_attr = "all"` will score `NaN` fairness.
+(`age_group`, `sex`) the user belongs to. The board uses these memberships to compute skill and rank separately for
+age <40, age 40+, female and male participants, with a within-cohort Linear
+baseline. Worst-subgroup skill is the minimum of those four overall skill scores.
+All four cohort scores must be available; otherwise it is shown as missing.
 `evaluate_prediction(output_dir=...)` emits all three automatically.
 
 ## `<method>.meta.json`

downstream/bootstrap/README.md

@@ -1,73 +1,35 @@
 # Downstream bootstrap reference
 
-This subdir holds the **Phase-1 bootstrap reference** for the Track-1 (outcome
-prediction) leaderboard recompute — the long-format per-draw error frame that the
-skill / rank / fairness CIs reduce from.
-
-## Layout
-
-```
-downstream/bootstrap/
-├── draws.parquet      # per-(method, task, subgroup, draw) error E = 1 − metric
-└── draws.meta.json    # provenance: seed, n_boot, methods, n_tasks, fairness attrs
-```
-
-## What's it for
-
-Each row of `draws.parquet` is one bootstrap draw of one task for one method,
-carrying the per-draw error `E = 1 − metric` (binary AUPRC, ordinal Spearman,
-regression Pearson). Phase-2 reduces the same frame three ways, all paired vs the
-`linear` baseline:
-
-- **Skill score** `S = 1 − geomean_task(E_method / E_linear)` — domain-balanced
-  macro; per-(method, scope) mean / SE / percentile-CI across draws.
-- **Average rank** — per-(method, scope) mean of the per-draw cross-method rank
-  (lower `E` → rank 1).
-- **Fairness skill score** — per-attribute disparity ratio over the `age_group` /
-  `sex` subgroup rows, with BCa intervals.
-
-The per-method `downstream/<method>.parquet` substrate (the raw user pairs, one level
-up) is the *input* the draws were bootstrapped from; this frame is the precomputed
-result so consumers need not re-run the 1000-draw paired bootstrap.
-
-## Provenance
-
-Generated by `scripts/paper_results/downstream/bootstrap_downstream_draws.py` in the
-code repo. The current snapshot was built with:
-
-- `seed = 42`, `n_boot = 1000`
-- `split = test`, canonical `sharable_users_seed42_2026`
-- 8 methods (see `draws.meta.json:methods`), 32 tasks, baseline `linear`
-- fairness attributes `age_group`, `sex`
-
-Parity of this frame against the uploaded per-method substrate is enforced by
-`scripts/paper_results/downstream/parity/parity_substrate.py` (a substrate-driven
-bootstrap must reproduce these draws).
-
-## Loading
-
-```python
-from huggingface_hub import hf_hub_download
-import pandas as pd, json
-
-draws_path = hf_hub_download(
-    "MyHeartCounts/OpenMHC-leaderboard-data",
-    "downstream/bootstrap/draws.parquet",
-    repo_type="dataset",
-)
-meta_path = hf_hub_download(
-    "MyHeartCounts/OpenMHC-leaderboard-data",
-    "downstream/bootstrap/draws.meta.json",
-    repo_type="dataset",
-)
-draws = pd.read_parquet(draws_path)
-meta = json.loads(open(meta_path).read())
-print(meta["seed"], meta["n_boot"], len(meta["methods"]))
-```
-
-See [`SCHEMA.md`](SCHEMA.md) for the full column spec.
-
-## Uploaded with
-
-`tools/upload_leaderboard_bootstrap.py --track downstream` in the
-[code repo](https://github.com/AshleyLab/myheartcounts-dataset).
+`draws.parquet` contains the current Track-1 all-participant per-draw errors;
+`draws.meta.json` records the sampling protocol and provenance. The live Space
+computes point scores from `downstream/<method>.parquet`; it does not read this
+bootstrap reference or display confidence intervals.
+
+The reference covers eight methods, 32 tasks and the Linear baseline. Each of
+1,000 replicates (seed 42) samples the entire declared test split of 3,569
+participants once, sharing participant multiplicities across every task and
+method. Each task retains its fixed eligibility rule; 1,762 distinct participants
+contribute to at least one task. Resampled task sizes therefore vary. Undefined
+single-class binary task draws are omitted jointly across methods and recorded
+in the metadata. Draw -1 is the original-cohort point estimate.
+
+Errors are `E = 1 - metric`: average precision for binary tasks, Spearman for
+ordinal tasks and Pearson for regression. Skill uses clipped method/Linear error
+ratios (0.01 to 100), a geometric mean within each domain, and an equal-weight
+mean of the five domain skills. Average rank is also domain-balanced.
+
+This reference supplies main skill/rank percentile confidence intervals. It
+contains only `subgroup_attr="all"`, `subgroup_value="all"`; it does not supply
+demographic subgroup, disparity-ratio or worst-subgroup confidence intervals.
+The live worst-subgroup point score is computed separately from four cohort
+views (age <40, age 40+, female, male), each with its own Linear baseline.
+
+The draws were generated from the verified saved evaluation predictions. The
+uploaded prediction-pair files serialize those predictions as float32, so a
+bootstrap regenerated from the uploaded pairs need not match byte for byte.
+Live point-score parity is checked at the leaderboard's display precision.
+WBM missing predictions use direct raw Linear substitution; metadata records
+`fallback_policy="direct_substitution"`.
+
+See [SCHEMA.md](SCHEMA.md) for the columns. Keep both reference files synchronized
+with the prediction files when publishing a new Track-1 result revision.

downstream/bootstrap/SCHEMA.md

@@ -9,7 +9,7 @@ downstream/bootstrap/draws.meta.json    # provenance sidecar
 
 This is the companion to the per-method `downstream/<method>.parquet` substrate
 (see `../SCHEMA.md`). The substrate is the raw per-user pairs; this is the
-**Phase-1 per-draw error frame** the skill / rank / fairness CIs reduce from, so a
+**Phase-1 per-draw error frame** the main skill / rank CIs reduce from, so a
 consumer can recompute the leaderboard intervals without re-running the (paired,
 1000-draw) bootstrap over the pairs.
 
@@ -26,8 +26,8 @@ precomputed here; Phase-2 derives all three from `E` (every metric is paired vs
 | `task` | string | benchmark task name (one of the 32 `BENCHMARK_TASKS`) |
 | `task_type` | string | `binary`, `ordinal`, or `regression` |
 | `domain` | string | task domain: `Demographics`, `Medical conditions`, `Body metrics and biomarkers`, `Mental well-being`, `Sleep and lifestyle` |
-| `subgroup_attr` | string | `all` (global cell), `age_group`, or `sex` |
-| `subgroup_value` | string | `all` for the global cell; otherwise the subgroup level (age bucket, sex value, or `unknown`) |
+| `subgroup_attr` | string | `all` in this reference |
+| `subgroup_value` | string | `all` in this reference |
 | `draw` | int | `-1` for the point estimate (full cohort, no resampling), else the bootstrap-draw index in `[0, n_boot)` |
 | `E` | float32 | per-draw error `E = 1 − metric` for this cell, evaluated on the resampled cohort |
 
@@ -39,11 +39,11 @@ precomputed here; Phase-2 derives all three from `E` (every metric is paired vs
   (domain-balanced, clipped) and the cross-method rank are well-defined per draw.
 - **`draw = -1`** is the point estimate (the metric on the full test cohort); draws
   `0 … n_boot−1` are the paired bootstrap resamples.
-- **Paired resamples** — for each task the same `seed=42` resample indices are reused
-  across all methods, so per-draw cross-method comparisons (skill ratios, ranks,
-  subgroup disparities) are valid.
-- No NaN-filling: a `(task, subgroup_value)` cell with no eligible cohort simply has
-  no rows.
+- **Paired resamples** — each draw samples the entire declared test population
+  once; participant multiplicities are shared across all tasks and methods.
+  Task eligibility is fixed, while resampled task sizes vary. Seed: 42.
+- No NaN-filling: undefined metric cells are omitted. Single-class binary
+  task draws are omitted jointly across all methods; see metadata omissions.
 
 ## `draws.meta.json`
 
@@ -51,20 +51,25 @@ precomputed here; Phase-2 derives all three from `E` (every metric is paired vs
 {
   "n_boot": 1000,
   "seed": 42,
-  "baseline": "linear",                                 // skill / fairness baseline
+  "baseline": "linear",                                 // skill baseline
   "methods": ["linear", "multirocket", "lsm2", "toto",
               "chronos2", "xgboost", "wbm", "gru_d"],   // 8 entries (incl. the baseline)
   "n_tasks": 32,
-  "fairness_attributes": ["age_group", "sex"]
+  "fairness_subgroups_included": false,
+  "sampling_unit": "participant",
+  "sampling_population_n": 3569,
+  "shared_across": "all tasks and methods",
+  "fallback_policy": "direct_substitution"
 }
 ```
 
 ## Conventions
 
 - Evaluated against the canonical split `sharable_users_seed42_2026` (`test`).
-- Track-1 baseline for skill / fairness: `linear`.
-- Skill is **domain-balanced macro** (mean over the 5 domains' geomean ratios);
-  fairness uses BCa intervals over the `age_group` / `sex` subgroup rows.
+- Track-1 baseline for skill: `linear`.
+- Skill is **domain-balanced macro**: compute one minus the geometric mean
+  of clipped error ratios within each domain, then average the five domain skills.
+  The current reference contains no subgroup or worst-subgroup CIs.
 - Format: single Parquet, dictionary-encoded string columns, `float32` `E`, `zstd`
   compression.
 

downstream/bootstrap/draws.meta.json

@@ -1,6 +1,7 @@
 {
   "n_boot": 1000,
   "seed": 42,
+  "ci_level": 0.95,
   "baseline": "linear",
   "methods": [
     "linear",
@@ -13,8 +14,66 @@
     "gru_d"
   ],
   "n_tasks": 32,
-  "fairness_attributes": [
-    "age_group",
-    "sex"
-  ]
-}
+  "sampling_unit": "participant",
+  "sampling_population": "entire declared test split",
+  "sampling_population_n": 3569,
+  "eligible_union_n": 1762,
+  "shared_across": "all tasks and methods",
+  "task_sample_sizes": "vary by replicate",
+  "split_sha256": "66dd5443ddb22c5278ad7fd25a0427b35c8d5ecc770bf9af4257d5a729248c35",
+  "counts_sha256": "5ac059e66f7ddbac2d5a808d258d338b4d63ae152c986bb0d4c79298ab69047e",
+  "counts_dtype": "int16",
+  "counts_shape": [
+    1000,
+    3569
+  ],
+  "point_draw": -1,
+  "error_storage_dtype": "float32",
+  "clip_lower": 0.01,
+  "clip_upper": 100.0,
+  "aggregation": "equal domains; within-domain mean ranks and geometric mean error ratios",
+  "undefined_policy": "single-class task omitted jointly across all methods; no imputation",
+  "omissions": [
+    {
+      "draw": 502,
+      "task": "PH",
+      "reason": "single_class_labels",
+      "positive_sample_count": 0,
+      "sample_size": 1089,
+      "methods_omitted": 8
+    },
+    {
+      "draw": 741,
+      "task": "PH",
+      "reason": "single_class_labels",
+      "positive_sample_count": 0,
+      "sample_size": 1104,
+      "methods_omitted": 8
+    },
+    {
+      "draw": 945,
+      "task": "PH",
+      "reason": "single_class_labels",
+      "positive_sample_count": 0,
+      "sample_size": 1079,
+      "methods_omitted": 8
+    },
+    {
+      "draw": 114,
+      "task": "Peripheral/Systemic Vascular Disease",
+      "reason": "single_class_labels",
+      "positive_sample_count": 0,
+      "sample_size": 1059,
+      "methods_omitted": 8
+    }
+  ],
+  "fairness_subgroups_included": false,
+  "corrected_methods": [
+    "wbm"
+  ],
+  "previous_aggregation_manifest_sha256": "94434aabbfc07655c0a2edcec00b7e2e59aac7b62aab5328112884d961d82033",
+  "evaluation_manifest_sha256": "5591c64949a4fe119e9ed79ee029b6e093d33c8793590f516ad81072f3d0d846",
+  "fallback_policy": "direct_substitution",
+  "parent_aggregation_sha256": "8cee66e651e7b8eb92d323f3aea7d7a8202836632b7c9359743d6fc2525293aa",
+  "model_refitting": false
+}

downstream/bootstrap/draws.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9f0b5d9a80583ae2c655d9ac035c5dae22bee355f535ec2704a7d812e723b4fa
-size 8899958
+oid sha256:3cbae19f68a7936faf93d636c22248e990b4a453e7fc179d138ced0a98327bc7
+size 1423232

downstream/chronos2.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1cc1fe7b603ba0a7591838bdf44404def4542e97f349f289a9f1205f940bbf21
-size 521035
+oid sha256:d23386b91d4c0f5a5cbc9bca2dce4c14d255ff1aa7982766c1e9e691e51e52fb
+size 512869

downstream/gru_d.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2efb2de58a1a0325e228c7bc0213d87440b1317199f4151fbf9165de71d1cc8a
-size 674246
+oid sha256:5f5cb814e460f3fce8e74cdea142d965accf61545d34f87892336918ae7acb31
+size 503204

downstream/linear.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:34f3c77c3ea1995017ddc0d0e722b22fb709cf242c658cc2d059cb53cd123277
-size 522461
+oid sha256:eee631a44f3316fbb124814d9e58f11bb4cd7486eed5bce71d6807e800e66e33
+size 513716

downstream/lsm2.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6486342eef73d29d9ce3c43a73483c057dbbc15e5474d5f3ceea1cd59e72ab54
-size 521795
+oid sha256:a8ba0f324cc3e40e71f28c3e268bf01b4a9ad609b9a22be51b4d7b423c6f07f8
+size 513751

downstream/multirocket.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fae69d6b12510dcfcc2ddb17abd57b6e0120517bf6bcf8b0ef29e8925691909b
-size 522780
+oid sha256:077e263025105eba57e72ff42b71f700382bdc37221518b08f370e82904748e6
+size 515397

downstream/toto.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:af056d1704b0492f2537ca4d5ebf9f8a17a5ee35414fbc3df074743bbc7dd72b
-size 520793
+oid sha256:70a745a21d4efd26e7a6d299a5eeb3e7a251c0242347b0c13dc24b043e138b46
+size 511983

downstream/wbm.meta.json

@@ -3,5 +3,5 @@
   "type": "Self-Supervised",
   "submitter": "OpenMHC team",
   "subtrack": "static",
-  "fallback_rate": 0.6276
-}
+  "fallback_rate": 0.6270524855640505
+}

downstream/wbm.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e4d732f999f55ab9614165da10e929f8ff50cfb7a88804a2814e0db52accbf69
-size 517877
+oid sha256:e7096d60c695c084acc28ba9210c8f49d7884b4bf658abd5697b8f69408eceb9
+size 515145

downstream/xgboost.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:32931df5f3cf125b0c78467acf0976da9cb79b83782202ced41e87e1769435a2
-size 519369
+oid sha256:254c3b7730df888293f229b7a5461fa970c64aae7e669ef4038ff45d0c903fbc
+size 510825