Sync local: train/test range split + metadata consistency fixes (measured/best_baseline, ranges) + field cleanup (drop problem_statement/auditor_notes/reference_baseline_id) + trim redundant range prose

baseball_pythagorean_winpct_lahman__win_pct/data/test.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6d974472e98b7386716cb942788233a1cb3789f9cae81719eb2531984db61646
-size 15962
+oid sha256:7ad82a9f599360efd361f23bd2be24853657c5426b86cabe28ab2a5cc96131aa
+size 19686

baseball_pythagorean_winpct_lahman__win_pct/data/train.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d5718100d60c7ff2e1a0ed725df485e7bb5ad1410f26f7fdaf3c90491916e008
-size 34267
+oid sha256:6087c6ad2640cda40a6e48c8a28804f79f05c5a5909cafcab51cdfb26cd1bc2d
+size 42219

baseball_pythagorean_winpct_lahman__win_pct/formulas/pythagenpat.py

@@ -0,0 +1,69 @@
+"""Pythagenpat won-loss formula (David Smyth, 2003) — run-environment-dependent
+Pythagorean exponent.
+
+The constant-exponent Pythagorean formula (James gamma=2; Miller gamma~1.79) is
+systematically off in extreme run environments. Pythagenpat makes the exponent
+a function of the run environment, measured by runs-per-game
+RPG = (R + RA) / G:
+
+    gamma = RPG ** pat_exp
+    win_pct = R^gamma / (R^gamma + RA^gamma)
+
+with the single empirical exponent pat_exp = 0.287. Davenport (Baseball
+Prospectus) endorsed this form as "simpler, more elegant, and [it] gets the
+better answer over a wider range of runs scored than Pythagenport".
+
+David Smyth / "Patriot" (2003), popularised via Baseball Prospectus and
+Tom Tango's blog; see also the Pythagorean expectation review on Wikipedia
+(reference/summary_formula_pythagorean.md).
+
+LAW_CONSTANTS
+-------------
+- pat_exp = 0.287: the Pythagenpat exponent applied to runs-per-game.
+  (Single law constant.)
+
+OTHER_CONSTANTS / LOCAL_FITTABLE
+--------------------------------
+None. The formula is dimensionless and uses a single league-universal
+exponent rule (no per-team/per-season refit). Type I.
+
+Column mapping (paper -> CSV):
+  R  (season total runs scored)   -> X[:, 0]
+  RA (season total runs allowed)  -> X[:, 1]
+  G  (games played in the season) -> X[:, 2]   (sets the run environment)
+
+Notes:
+- Unlike the scale-invariant constant-exponent forms, Pythagenpat needs the
+  games count G to form runs-per-game RPG = (R+RA)/G.
+- Reduces to a constant exponent only if RPG is constant; here it adapts to
+  each season's scoring level, improving fit in deadball / steroids-era
+  extremes.
+"""
+
+import numpy as np
+
+USED_INPUTS  = ["R", "RA", "G"]
+PAPER_REF    = "summary_formula_pythagorean.md"
+EQUATION_LOC = "Pythagenpat (Smyth 2003): gamma = ((R+RA)/G)^0.287; win = R^gamma/(R^gamma+RA^gamma)"
+
+LAW_CONSTANTS = {
+    "pat_exp": 0.287,   # Smyth (2003) Pythagenpat exponent on runs-per-game
+}
+OTHER_CONSTANTS = {}
+LOCAL_FITTABLE  = {}    # Type I
+
+
+def predict(X: np.ndarray, pat_exp: float) -> np.ndarray:
+    """Pythagenpat win percentage with a run-environment-dependent exponent.
+
+    X: (n, 3) — columns R (runs scored), RA (runs allowed), G (games).
+    Returns array of shape (n,) — win_pct in (0, 1).
+    """
+    R  = np.asarray(X[:, 0], dtype=float)
+    RA = np.asarray(X[:, 1], dtype=float)
+    G  = np.asarray(X[:, 2], dtype=float)
+    rpg = (R + RA) / G
+    gamma = np.power(rpg, pat_exp)
+    Rg  = np.power(R,  gamma)
+    RAg = np.power(RA, gamma)
+    return Rg / (Rg + RAg)

baseball_pythagorean_winpct_lahman__win_pct/formulas/pythagenport.py

@@ -0,0 +1,67 @@
+"""Pythagenport won-loss formula (Clay Davenport & Keith Woolner, Baseball
+Prospectus, 1999) — run-environment-dependent Pythagorean exponent.
+
+The first widely-used variable-exponent refinement of the Pythagorean formula.
+The exponent grows logarithmically with the run environment (runs-per-game
+RPG = (R + RA) / G):
+
+    gamma = slope * log10(RPG) + offset       (slope = 1.5, offset = 0.45)
+    win_pct = R^gamma / (R^gamma + RA^gamma)
+
+Davenport & Woolner reported it works within run environments of roughly
+4-40 runs per game and lowers RMSE versus the fixed exponent 2. (Later
+superseded for simplicity by Pythagenpat; see formulas/pythagenpat.py.)
+
+Davenport, C. & Woolner, K. (1999), Baseball Prospectus; reproduced in the
+Pythagorean expectation review on Wikipedia
+(reference/summary_formula_pythagorean.md).
+
+LAW_CONSTANTS
+-------------
+- slope  = 1.5  : coefficient on log10(runs-per-game)
+- offset = 0.45 : additive exponent offset
+  (Two law constants — the published Davenport-Woolner coefficients.)
+
+OTHER_CONSTANTS / LOCAL_FITTABLE
+--------------------------------
+None. League-universal exponent rule, no per-team/per-season refit. Type I.
+
+Column mapping (paper -> CSV):
+  R  (season total runs scored)   -> X[:, 0]
+  RA (season total runs allowed)  -> X[:, 1]
+  G  (games played in the season) -> X[:, 2]   (sets the run environment)
+
+Notes:
+- Implemented in the DIRECT Pythagorean form: the exponent gamma applies to R
+  and RA directly (gamma ~ 1.6-2.7 over MLB history), NOT to a log-odds /
+  sigma-of-log10 variant. Needs the games count G to form RPG = (R+RA)/G.
+"""
+
+import numpy as np
+
+USED_INPUTS  = ["R", "RA", "G"]
+PAPER_REF    = "summary_formula_pythagorean.md"
+EQUATION_LOC = "Pythagenport (Davenport & Woolner 1999): gamma = 1.5*log10((R+RA)/G)+0.45"
+
+LAW_CONSTANTS = {
+    "slope":  1.5,    # Davenport-Woolner coefficient on log10(RPG)
+    "offset": 0.45,   # Davenport-Woolner exponent offset
+}
+OTHER_CONSTANTS = {}
+LOCAL_FITTABLE  = {}    # Type I
+
+
+def predict(X: np.ndarray, slope: float, offset: float) -> np.ndarray:
+    """Pythagenport win percentage with a log-RPG-dependent exponent.
+
+    X: (n, 3) — columns R (runs scored), RA (runs allowed), G (games).
+    Returns array of shape (n,) — win_pct in (0, 1).
+    """
+    R  = np.asarray(X[:, 0], dtype=float)
+    RA = np.asarray(X[:, 1], dtype=float)
+    G  = np.asarray(X[:, 2], dtype=float)
+    rpg = (R + RA) / G
+    gamma = slope * np.log10(rpg) + offset
+    Rg  = np.power(R,  gamma)
+    RAg = np.power(RA, gamma)
+    return Rg / (Rg + RAg)

baseball_pythagorean_winpct_lahman__win_pct/formulas/reference_metrics.json

@@ -10,15 +10,15 @@
       },
       "local_fittable": [],
       "metrics": {
-        "log_mae": 0.018073578706717868,
+        "rmse": 0.02581841049545397,
         "mae": 0.02035253889196662,
-        "mape": 0.0416209828963854,
-        "mdae": 0.016923316811408906,
         "mse": 0.0006665903205117677,
-        "n_finite": 938,
+        "mdae": 0.016923316811408906,
+        "smape": 0.041594077969012565,
+        "mape": 0.0416209828963854,
+        "log_mae": 0.018073578706717868,
         "r2": 0.8723463499046444,
-        "rmse": 0.02581841049545397,
-        "smape": 0.041594077969012565
+        "n_finite": 938
       },
       "other_constants": {},
       "paper_ref": "summary_formula_miller_2007.md"
@@ -33,24 +33,71 @@
       },
       "local_fittable": [],
       "metrics": {
-        "log_mae": 0.0179324763805722,
+        "rmse": 0.025367518227170174,
         "mae": 0.02012816862599297,
-        "mape": 0.041496949291881115,
-        "mdae": 0.01727645471286568,
         "mse": 0.000643510981005811,
-        "n_finite": 938,
+        "mdae": 0.01727645471286568,
+        "smape": 0.04127015407815537,
+        "mape": 0.041496949291881115,
+        "log_mae": 0.0179324763805722,
         "r2": 0.8767660989454998,
-        "rmse": 0.025367518227170174,
-        "smape": 0.04127015407815537
+        "n_finite": 938
       },
       "other_constants": {},
       "paper_ref": "summary_formula_miller_2007.md"
+    },
+    "pythagenport": {
+      "equation_loc": "Pythagenport (Davenport & Woolner 1999): gamma = 1.5*log10((R+RA)/G)+0.45",
+      "error": null,
+      "failed": false,
+      "kind": "reference",
+      "law_constants": {
+        "slope": 1.5,
+        "offset": 0.45
+      },
+      "local_fittable": [],
+      "metrics": {
+        "rmse": 0.025279362728894618,
+        "mae": 0.020038159053313454,
+        "mse": 0.0006390461799790264,
+        "mdae": 0.017035062462379347,
+        "smape": 0.04101285775536915,
+        "mape": 0.04112307462482883,
+        "log_mae": 0.017820564889759735,
+        "r2": 0.8776211190837775,
+        "n_finite": 938
+      },
+      "other_constants": {},
+      "paper_ref": "summary_formula_pythagorean.md"
+    },
+    "pythagenpat": {
+      "equation_loc": "Pythagenpat (Smyth 2003): gamma = ((R+RA)/G)^0.287; win = R^g/(R^g+RA^g)",
+      "error": null,
+      "failed": false,
+      "kind": "reference",
+      "law_constants": {
+        "pat_exp": 0.287
+      },
+      "local_fittable": [],
+      "metrics": {
+        "rmse": 0.02526633301488769,
+        "mae": 0.02003358099347757,
+        "mse": 0.0006383875840192037,
+        "mdae": 0.01712980262652508,
+        "smape": 0.041005661779374376,
+        "mape": 0.04112240460441076,
+        "log_mae": 0.017817425550055974,
+        "r2": 0.8777472417945676,
+        "n_finite": 938
+      },
+      "other_constants": {},
+      "paper_ref": "summary_formula_pythagorean.md"
     }
   },
   "derived_caps": {
     "fit_timeout_seconds": null,
     "max_init_size_per_param": 1,
-    "max_law_constants": 1,
+    "max_law_constants": 2,
     "max_local_params": 0
   },
   "metric_declared": "rmse",
@@ -58,4 +105,4 @@
   "reference_baseline_id": null,
   "task": "baseball_pythagorean_winpct_lahman__win_pct",
   "type": "typeI"
-}
+}

baseball_pythagorean_winpct_lahman__win_pct/metadata.yaml

@@ -15,19 +15,32 @@ target:
   symbol: "W%"
   unit: ""
   description: Season win percentage — fraction of decided games won, W / (W + L).
-  range: [0.12, 0.90]
+  range:
+    train: [0.12987, 0.898734]
+    test:  [0.265432, 0.716667]
 
 inputs:
   - name: R
     symbol: RS
     unit: runs
     description: Season total runs scored by the team.
-    range: [219, 1220]
+    range:
+      train: [238, 1220]
+      test:  [219, 1009]
   - name: RA
     symbol: RA
     unit: runs
     description: Season total runs allowed by the team.
-    range: [209, 1252]
+    range:
+      train: [229, 1252]
+      test:  [209, 1103]
+  - name: G
+    symbol: G
+    unit: games
+    description: Games played in the season. Sets the run environment via runs-per-game RPG = (R + RA) / G, used by the run-environment-dependent exponent forms (Pythagenport, Pythagenpat). The constant-exponent forms (James, Miller) ignore it. Does not leak the target — G is the games total, not the win/loss split, so win_pct = W/(W+L) cannot be recovered from it.
+    range:
+      train: [52, 165]
+      test:  [58, 163]
 
 data_files:
   train: data/train.csv   # seasons 1872-1989 (historical era)
@@ -53,13 +66,19 @@ n_test:  938
 #   2026-05-28: removed davenport_2014 baseline (unsourced LAW a=1.50; σ-of-log10 form gives
 #     effective γ≈0.65 vs Davenport-Woolner 1999 published γ≈1.83 in direct-Pythagorean form;
 #     no PDF on disk; half-integrated wiring).
+#   2026-06-04: released input G (games); added run-environment-dependent baselines
+#     pythagenport (Davenport-Woolner 1999, γ=1.5·log10(RPG)+0.45, 2 law constants) and
+#     pythagenpat (Smyth 2003, γ=RPG^0.287, 1 constant), RPG=(R+RA)/G — the CORRECT
+#     direct-Pythagorean form (not the σ-of-log10 mistake of the removed davenport_2014).
+#     Raised caps.max_law_constants 1→2 for Pythagenport. pythagenpat is the new best
+#     baseline (rmse 0.02527). G does not leak win_pct (games total, not W/L split).
 # Final priors: 3 distractors, 0 candidates.
 priors:
   - name: games_per_season_modern
     value: 162.0
     unit: games
-    description: Number of regular-season games per team in the modern MLB schedule (since 1961 AL / 1962 NL). Not consumed by the Pythagorean formula, which is scale-invariant in R and RA.
-    source: Historical MLB schedule; not used directly in the Pythagorean formula (formula is scale-invariant in R, RA).
+    description: Number of regular-season games per team in the modern MLB schedule (since 1961 AL / 1962 NL). A fixed constant — distractor; the run-environment forms use the actual per-season games input G (which varies, e.g. 52–165 here), not this constant, and the constant-exponent forms ignore games entirely.
+    source: Historical MLB schedule. Distractor — not a parameter of any baseline (variable-exponent forms read the per-row G input instead).
     _role: distractor
   - name: innings_per_game
     value: 9.0
@@ -83,7 +102,7 @@ references:
     n_other_constants: 0
     n_local_params: 0
     measured:
-      rmse: 0.0258
+      rmse: 0.02582
       r2: 0.8723
   - id: miller_2007
     label: "Miller (2007) — Pythagorean formula, gamma = 1.79 (least-squares)"
@@ -93,11 +112,31 @@ references:
     n_other_constants: 0
     n_local_params: 0
     measured:
-      rmse: 0.0254
+      rmse: 0.02537
       r2: 0.8768
+  - id: pythagenport
+    label: "Davenport & Woolner (1999) — Pythagenport, gamma = 1.5*log10(RPG)+0.45, RPG=(R+RA)/G"
+    formula_file: formulas/pythagenport.py
+    reference_pdf: null
+    n_law_constants: 2
+    n_other_constants: 0
+    n_local_params: 0
+    measured:
+      rmse: 0.02528
+      r2: 0.8776
+  - id: pythagenpat
+    label: "Smyth (2003) — Pythagenpat, gamma = ((R+RA)/G)^0.287 (run-environment exponent)"
+    formula_file: formulas/pythagenpat.py
+    reference_pdf: null
+    n_law_constants: 1
+    n_other_constants: 0
+    n_local_params: 0
+    measured:
+      rmse: 0.02527
+      r2: 0.8777
 
 caps:
-  max_law_constants:        1
+  max_law_constants:        2   # raised 1→2 (2026-06-04) for Pythagenport (slope, offset)
   max_local_params:         0
   max_init_size_per_param:  1
   fit_timeout_seconds:      null
@@ -116,4 +155,4 @@ validity_rubrics:
   - "predicted win_pct is above 0.5 when R exceeds RA and below 0.5 when R is less than RA"
   - "swapping R and RA maps the prediction to its complement about 0.5"
 
-best_baseline: 0.0253675    # miller_2007 — rmse (lower is better); regenerated 2026-05-30
+best_baseline: 0.0252660    # pythagenpat — rmse (lower is better); regenerated 2026-06-04 after adding G + Pythagenport/Pythagenpat

baseball_pythagorean_winpct_lahman__win_pct/prep_data.py

@@ -11,6 +11,12 @@ License: CC-BY-SA-3.0 (Lahman Baseball Database).
 Column 0  win_pct  : float, dimensionless — win fraction W/(W+L) (SR target)
 Column 1  R        : int → float, runs — season total runs scored by the team
 Column 2  RA       : int → float, runs — season total runs allowed by the team
+Column 3  G        : int → float, games — games played in the season. Released so
+                     the run-environment-dependent exponent baselines work
+                     (Pythagenport, Davenport 1999; Pythagenpat, Smyth 2003),
+                     whose exponent is set by runs-per-game RPG = (R+RA)/G.
+                     Does NOT leak win_pct: G is the games total, not the W/L
+                     split, so win_pct = W/(W+L) cannot be recovered from G.
 
 Dropped from working CSV (not written to released files):
   'W'          — tautology leak: win_pct = W/(W+L) by definition (data_spec §9.4)
@@ -18,7 +24,6 @@ Dropped from working CSV (not written to released files):
   'yearID'     — auxiliary bookkeeping; not a declared SR input
   'lgID'       — categorical; not a declared SR input
   'teamID'     — string id; not a declared SR input
-  'G'          — games played; auxiliary (not in any baseline's USED_INPUTS)
   All other Lahman columns (HR, BB, ERA, …) — not relevant to Pythagorean formula
 
 === TYPE I vs TYPE II DECISION ===
@@ -108,7 +113,7 @@ EXPECTED_SHA256 = "827bef5af12158eeb3d18a4a3f54ff846e395965b47c09b2564ad7bb29c5e
 MIN_W_PLUS_L    = 50        # minimum decided games; excludes ultra-short seasons
 TRAIN_YEAR_MAX  = 1989      # inclusive; train = yearID <= TRAIN_YEAR_MAX
 TEST_YEAR_MIN   = 1990      # inclusive; test  = yearID >= TEST_YEAR_MIN
-OUT_COLS        = ["win_pct", "R", "RA"]
+OUT_COLS        = ["win_pct", "R", "RA", "G"]
 
 
 # ---------------------------------------------------------------------------

baseball_pythagorean_winpct_lahman__win_pct/reference/README.md

@@ -5,8 +5,15 @@
 | `miller_2007.pdf` | Steven J. Miller (2007). "A Derivation of the Pythagorean Won-Loss Formula in Baseball." *Chance* 20(1):40–48. American Statistical Association. | **Primary baseline.** Provides the Weibull-distribution derivation of the Pythagorean formula; Eq. (1.2) PDF p. 1 and Theorem 2.2 / Eq. (2.6) PDF p. 4. Empirical gamma = 1.79 (least-squares). |
 | `miller_2007.txt` | Text extract of `miller_2007.pdf` via `pdf_to_text.py`. | Auto-generated; used for coefficient grepping. |
 | `miller_2007.bib` | BibTeX entry for Miller (2007). | `refkey = miller2007pythag`. |
-| `refs.bib` | Combined bibliography for all references cited in this task (Miller 2007, James 1980 citation, Davenport-Woolner 1999 citation, Lahman DB). | Retained from SRC for completeness; per-paper `.bib` files are the authoritative splits. |
+| `refs.bib` | Combined bibliography (Miller 2007, James 1980, Davenport-Woolner 1999, Smyth 2003, Lahman DB). | Retained from SRC for completeness; per-paper `.bib` files are the authoritative splits. |
 | `summary_formula_miller_2007.md` | Page-level summary of Miller (2007) for this benchmark. | Records Eq. (1.2) provenance, gamma values, and column mapping. |
+| `summary_formula_pythagorean.md` | Summary of the full constant- and variable-exponent ladder (James / Miller / Pythagenport / Pythagenpat). | Equations, the RPG=(R+RA)/G dependence, properties, and measured fit. Referenced by `pythagenport.py` / `pythagenpat.py`. |
+
+Baselines: `james_1980` (γ=2), `miller_2007` (γ=1.79), `pythagenport`
+(Davenport-Woolner 1999, γ=1.5·log10(RPG)+0.45), `pythagenpat` (Smyth 2003,
+γ=RPG^0.287). The two variable-exponent forms use the released games input `G`
+via RPG=(R+RA)/G; `pythagenpat` is the best baseline. W and L remain withheld
+(they would leak `win_pct=W/(W+L)`).
 
 ---
 
@@ -17,5 +24,7 @@
 - James (1980) Baseball Abstract: citation-only; no DOI or PDF mirrored.
 - Davenport-Woolner (1999) Baseball Prospectus:
   https://www.baseballprospectus.com/news/article/342/ — citation-only.
+- Smyth (2003) Pythagenpat: citation-only; see
+  https://en.wikipedia.org/wiki/Pythagorean_expectation
 - Lahman Database: https://sabr.org/lahman-database/ ;
   https://github.com/cbwinslow/baseballdatabank

baseball_pythagorean_winpct_lahman__win_pct/reference/refs.bib

@@ -2,10 +2,12 @@
 %
 % Single OA primary reference is Miller 2007 (Chance), whose author-mirror
 % PDF at Williams College is bundled in this directory as miller_2007.pdf.
-% The James (1980) "Baseball Abstract" and Davenport-Woolner (1999)
-% "Revisiting the Pythagorean Theorem" entries are citation-only (no PDF
-% is mirrored); they are the historical origins of the formula and the
-% gamma=1.83 refinement respectively.
+% The James (1980) "Baseball Abstract", Davenport-Woolner (1999)
+% "Revisiting the Pythagorean Theorem" (Pythagenport), and Smyth (2003)
+% "Pythagenpat" entries are citation-only (no PDF is mirrored); they are the
+% historical origin (gamma=2), the first run-environment-dependent exponent
+% (gamma = 1.5*log10(RPG)+0.45), and its simpler successor
+% (gamma = RPG^0.287), respectively, with RPG = (R+RA)/G.
 
 @article{miller2007pythag,
   author       = {Steven J. Miller},
@@ -37,6 +39,15 @@
   note         = {Refines the exponent to gamma about 1.83 by fitting MLB team-season records.  Citation-only; no PDF is mirrored.  This is the source of the modern saber-metrics "Pythagenport" exponent that improves on Bill James' gamma = 2 in modern high-scoring eras.}
 }
 
+@misc{smyth2003pythagenpat,
+  author       = {David Smyth},
+  title        = {{Pythagenpat}: A Run-Environment-Dependent {Pythagorean} Exponent},
+  year         = {2003},
+  howpublished = {Sabermetrics community (David Smyth / "Patriot"); popularised via Baseball Prospectus and Tom Tango's blog},
+  url          = {https://en.wikipedia.org/wiki/Pythagorean_expectation},
+  note         = {Sets the Pythagorean exponent from the run environment as gamma = ((R+RA)/G)^0.287, where (R+RA)/G is runs per game.  Endorsed by Davenport as "simpler, more elegant, and [it] gets the better answer over a wider range of runs scored than Pythagenport".  Citation-only; no PDF mirrored.  Requires the games count G (released as a task input).}
+}
+
 @misc{lahman2025db,
   author       = {Sean Lahman and {Chadwick Baseball Bureau}},
   title        = {Lahman Baseball Database ({baseballdatabank}), {Teams.csv}},

baseball_pythagorean_winpct_lahman__win_pct/reference/summary_formula_pythagorean.md

@@ -0,0 +1,62 @@
+# Pythagorean win-expectation: the constant- and variable-exponent family
+
+This task's target is season win fraction `win_pct = W/(W+L)`, predicted from
+runs scored `R` and runs allowed `RA` (and, for the variable-exponent forms,
+games played `G`). All four reference baselines share the **Pythagorean ratio
+form**
+
+    win_pct = R^gamma / (R^gamma + RA^gamma)
+
+and differ only in how the exponent `gamma` is set. (W and L are NOT released —
+`win_pct = W/(W+L)` is an arithmetic identity, so they would leak the target.)
+
+## The exponent ladder
+
+| baseline | exponent `gamma` | constants | source |
+|---|---|---|---|
+| `james_1980` | `2` (fixed) | 1 | Bill James, 1980 — gives the "Pythagorean" name (sum of squares) |
+| `miller_2007` | `1.79` (fixed) | 1 | Miller 2007, least-squares over 14 AL-2004 teams (Weibull derivation) |
+| `pythagenport` | `1.5·log10(RPG) + 0.45` | 2 | Davenport & Woolner, Baseball Prospectus 1999 |
+| `pythagenpat` | `RPG^0.287` | 1 | David Smyth, 2003 |
+
+where **`RPG = (R + RA) / G`** is runs per game (needs `G`).
+
+## Why the variable-exponent forms
+
+A fixed exponent is systematically off in extreme run environments (deadball
+era ~3.5 R/G; steroids era ~5.4 R/G). Making `gamma` grow with the run
+environment corrects this. Davenport reports Pythagenport works for ~4–40 R/G
+and lowers RMSE versus exponent 2; he later endorsed Pythagenpat as "simpler,
+more elegant, and [it] gets the better answer over a wider range of runs scored
+than Pythagenport". Pythagenpat is the modern sabermetric standard.
+
+## Structural properties (shared by all four; checked by the task rubrics)
+
+- bounded in (0, 1); equals 0.5 when R = RA;
+- monotone increasing in R, decreasing in RA;
+- swapping R↔RA maps the prediction to its complement about 0.5
+  (the exponent depends only on the symmetric `R+RA`, so it is unchanged);
+- the constant-exponent forms are scale-invariant in (R, RA); the
+  variable-exponent forms additionally use `G` to fix the run environment.
+
+## Measured fit on this task's test set (1990–2021, 938 team-seasons)
+
+| baseline | rmse | r2 |
+|---|---|---|
+| james_1980   | 0.02582 | 0.8723 |
+| miller_2007  | 0.02537 | 0.8768 |
+| pythagenport | 0.02528 | 0.8776 |
+| pythagenpat  | 0.02527 | 0.8777 |
+
+The variable-exponent forms fit best but only marginally: the RMSE floor
+(~0.025) is set by the irreducible season-to-season luck/variance of win% at a
+given run ratio, not by the exponent. `pythagenpat` is the best baseline
+(lowest rmse with a single constant).
+
+## Sources
+
+- Miller, S. J. (2007), *Chance* 20(1):40–48 (`miller_2007.pdf`).
+- James, B. (1980), *Baseball Abstract* (citation-only).
+- Davenport, C. & Woolner, K. (1999), Baseball Prospectus (citation-only).
+- Smyth, D. (2003), "Pythagenpat" (citation-only).
+- Pythagorean expectation review: https://en.wikipedia.org/wiki/Pythagorean_expectation

binary_pulsar_post_keplerian_atnf__PBDOT/metadata.yaml

@@ -9,20 +9,6 @@ license: "GPL-3.0-or-later (ATNF psrcat v2.8.0 distribution, CSIRO/ATNF); pulsar
 context: >-
   A binary pulsar is a compact system in which a pulsar orbits another neutron star, studied in relativistic astrophysics as a test of gravity.
 
-problem_statement: >-
-  Predict the GR-driven orbital-period derivative (PBDOT, dimensionless
-  s/s, always negative) of compact-DNS binary pulsars catalogued in
-  ATNF psrcat v2.8.0 from (Pb [days], e [dimensionless]) alone. The
-  split places near-circular systems (e < 0.05, predominantly recycled
-  millisecond pulsars in NS-WD orbits whose Pb-decay is dominated by
-  Galactic-acceleration kinematic terms rather than GR quadrupole
-  losses) in train (71 rows) and the eccentric GR-dominated compact
-  DNS systems (e >= 0.05, PBDOT < 0, Pb < 1 d) in test (10 rows, after FM-B8 dedup of the J0737-3039A/B physical-alias double-count).
-  The hard constraints to recover are the GR quadrupole prefactor
-  (192 pi / 5), the eccentricity-enhancement polynomial coefficients
-  (73/24, 37/96, exponent 7/2), and the chirp-mass scaling (5/3 power)
-  of Peters 1963 / Will 2014.
-
 target:
   name: PBDOT
   symbol: dot_P_b
@@ -36,8 +22,8 @@ target:
   # cluster contamination); the GR-valid regime is the test subset
   # [-5.294e-12, -1.366e-13]. See prep_data.py "Honest train-set limitation".
   range:
-    - -3.5e-9
-    - 6.0e-11
+    train: [-3.5e-09, 6e-11]
+    test:  [-5.294e-12, -1.366e-13]
 
 # === Wave-17 column cleanup, 2026-05-27 ===
 # Dropped omdot, gamma -- 74.6 % / 98.6 % NaN in train (most ATNF binaries
@@ -55,15 +41,15 @@ inputs:
     unit: d
     description: Keplerian orbital period of the binary system.
     range:
-      - 0.06
-      - 1236.73
+      train: [0.066231, 175.461]
+      test:  [0.078488, 0.420737]
   - name: e
     symbol: e
     unit: ""
     description: Orbital eccentricity, dimensionless and bounded in [0, 1).
     range:
-      - 0.0
-      - 0.91
+      train: [0, 0.027412]
+      test:  [0.063837, 0.681395]
 
 data_files:
   train: data/train.csv   # near-circular MSP-dominated split: e < 0.05, 71 rows
@@ -198,7 +184,6 @@ caps:
   max_init_size_per_param: 1
   fit_timeout_seconds: null
 
-reference_baseline_id: peters_1963
 # === Validity rubric ===
 # Frozen list — the scoring-time judge (eval_consistency.py) scores a
 # submission by how many of THESE criteria it satisfies:
@@ -213,18 +198,3 @@ validity_rubrics:
   - "predicted PBDOT remains defined throughout the physical input domain Pb > 0 and 0 <= e < 1"
 
 best_baseline: 1.21727e-13    # peters_1963 — rmse (lower is better); regenerated 2026-05-30
-
-wave17_notes: >-
-  Wave-17 rescue (2026-05-27). User flagged Wave-16 will_2014 baseline as
-  algebraic re-arrangement of damour_1991 ("ladder padding", numerically
-  identical predictions to 6 digits). User also rejected the omdot/gamma
-  74.6%/98.6% NaN columns as "documented but not fixed". Wave-17 actions:
-  (1) deleted will_2014.py and damour_1991.py from formulas/; (2) dropped
-  omdot, gamma from CSV schema; (3) added peters_1963.py using only (Pb, e)
-  + canonical DNS chirp mass M_c=1.20 M_sun (Antoniadis 2013); (4) refreshed
-  ATNF data to psrcat v2.8.0 (+1 DNS: J1946+2052); (5) added Pb<1 d test
-  filter (drops 4 cluster/Be-star contaminated systems). Result: 0-NaN
-  dataset, 10 GR-clean test rows (11 from the v2.8.0 refresh minus 1 J0737-3039A/B
-  physical-alias dedup per FM-B8), R^2 0.994 (was 0.78), single
-  live baseline (SB tag accepted -- no genuinely-distinct 2nd form exists
-  with [Pb, e]-only inputs). See WAVE17_ATTEMPT.md for full diff.

bird_flight_speed_alerstam__Ue/metadata.yaml

@@ -16,19 +16,25 @@ target:
   symbol: Ue
   unit: m/s
   description: Mean equivalent airspeed corrected to sea-level air density, measured by tracking radar during migratory flapping flight.
-  range: [6.70, 23.60]
+  range:
+    train: [6.7, 20.1]
+    test:  [7.7, 23.6]
 
 inputs:
   - name: mass_kg
     symbol: m
     unit: kg
     description: Mean total body mass of the species.
-    range: [0.009, 10.597]
+    range:
+      train: [0.009, 0.49]
+      test:  [0.523, 10.597]
   - name: wing_area_m2
     symbol: S
     unit: m^2
     description: Mean projected wing area of the species (both wings plus body between wings).
-    range: [0.0073, 0.8824]
+    range:
+      train: [0.0073, 0.1568]
+      test:  [0.0474, 0.8824]
 
 # Dataset — Type I two-file layout
 data_files:

bns_merger_disk_ejecta__Mdisk/metadata.yaml

@@ -7,51 +7,51 @@ target:
   unit: M_sun
   description: Gravitational mass of the bound, post-merger accretion disk surrounding the merger remnant, measured in solar masses.
   range:
-  - 0.0
-  - 0.31
+    train: [0.0001, 0.2282]
+    test:  [3e-05, 0.307517]
 inputs:
 - name: M1
   symbol: M_1
   unit: M_sun
   description: Gravitational mass of the lighter pre-merger neutron star, in solar masses, under the convention M_1 <= M_2.
   range:
-  - 1.0
-  - 1.71
+    train: [1.2, 1.71]
+    test:  [1, 1.44]
 - name: M2
   symbol: M_2
   unit: M_sun
   description: Gravitational mass of the heavier pre-merger neutron star, in solar masses, under the convention M_1 <= M_2.
   range:
-  - 1.2
-  - 1.86
+    train: [1.2, 1.71]
+    test:  [1.32, 1.856]
 - name: C1
   symbol: C_1
   unit: ""
   description: Dimensionless compactness of the lighter pre-merger neutron star, defined as G M / (R c^2) for that star.
   range:
-  - 0.11
-  - 0.21
+    train: [0.135, 0.205]
+    test:  [0.119838, 0.195]
 - name: C2
   symbol: C_2
   unit: ""
   description: Dimensionless compactness of the heavier pre-merger neutron star, defined as G M / (R c^2) for that star.
   range:
-  - 0.13
-  - 0.24
+    train: [0.138, 0.219]
+    test:  [0.147, 0.230916]
 - name: q
   symbol: q
   unit: ""
   description: Mass ratio of the binary, defined as the lighter over the heavier neutron-star gravitational mass, so 0 < q <= 1.
   range:
-  - 0.54
-  - 1.0
+    train: [0.774194, 1]
+    test:  [0.549569, 1]
 - name: Lambda_tilde
   symbol: Λ̃
   unit: ""
   description: Binary-averaged dimensionless tidal deformability, the mass-weighted combination of the two stars' individual tidal deformabilities that enters the leading finite-size term of the gravitational waveform.
   range:
-  - 116.0
-  - 1726.28
+    train: [116, 1439]
+    test:  [208, 1726.28]
 
 priors: []
 

bns_merger_disk_ejecta__Mej/metadata.yaml

@@ -7,65 +7,65 @@ target:
   unit: M_sun
   description: Dynamical ejecta mass expelled at merger, in solar masses.
   range:
-  - "9e-06"
-  - 0.019
+    train: [9.88678e-06, 0.00415131]
+    test:  [0.00012, 0.01802]
 inputs:
 - name: M1
   symbol: M_1
   unit: M_sun
   description: Gravitational mass of the primary (heavier) neutron star.
   range:
-  - 1.2
-  - 1.86
+    train: [1.2, 1.71]
+    test:  [1.32, 1.856]
 - name: M2
   symbol: M_2
   unit: M_sun
   description: Gravitational mass of the secondary (lighter) neutron star.
   range:
-  - 1.02
-  - 1.71
+    train: [1.18, 1.71]
+    test:  [1.02, 1.44]
 - name: Mb1
   symbol: M^b_1
   unit: M_sun
   description: Baryonic (rest) mass of the primary neutron star.
   range:
-  - 0.97
-  - 2.13
+    train: [0.974405, 1.92768]
+    test:  [1.233, 2.124]
 - name: Mb2
   symbol: M^b_2
   unit: M_sun
   description: Baryonic (rest) mass of the secondary neutron star.
   range:
-  - 0.88
-  - 1.93
+    train: [0.887277, 1.92768]
+    test:  [1.086, 1.844]
 - name: C1
   symbol: C_1
   unit: ""
   description: Compactness of the primary neutron star, defined as G*M/(R*c^2); dimensionless.
   range:
-  - 0.134
-  - 0.231
+    train: [0.13477, 0.205]
+    test:  [0.137626, 0.230916]
 - name: C2
   symbol: C_2
   unit: ""
   description: Compactness of the secondary neutron star, defined as G*M/(R*c^2); dimensionless.
   range:
-  - 0.119
-  - 0.205
+    train: [0.125, 0.205]
+    test:  [0.119838, 0.175]
 - name: q
   symbol: q
   unit: ""
   description: Binary mass ratio (lighter over heavier component), dimensionless and in (0, 1].
   range:
-  - 0.549
-  - 1.0
+    train: [0.766234, 1]
+    test:  [0.549569, 1]
 - name: Lambda_tilde
   symbol: Λ̃
   unit: ""
   description: Effective binary tidal deformability combining the two component tidal deformabilities; dimensionless.
   range:
-  - 116.0
-  - 1689.0
+    train: [116.006, 1688.41]
+    test:  [340.802, 1437.16]
 
 # FM-H3 backfill (wave-11, 2026-05-26): empty priors block added for schema symmetry with GOLD.
 priors: []

bns_merger_disk_ejecta__vej/metadata.yaml

@@ -7,51 +7,51 @@ target:
   unit: c
   description: Mass-averaged terminal velocity of the dynamical ejecta in units of the speed of light.
   range:
-  - 0.11
-  - 0.45
+    train: [0.119726, 0.45]
+    test:  [0.113706, 0.34]
 inputs:
 - name: q
   symbol: q
   unit: ""
   description: Binary mass ratio of the heavier to the lighter neutron star (≥ 1).
   range:
-  - 1.0
-  - 1.82
+    train: [1, 1.3353]
+    test:  [1.4267, 2.05932]
 - name: Lambda_tilde
   symbol: Λ̃
   unit: ""
   description: Reduced (mass-weighted) tidal deformability of the binary.
   range:
-  - 116.0
-  - 1689.0
+    train: [49.3937, 3196.45]
+    test:  [251.356, 2450.56]
 - name: M1
   symbol: M₁
   unit: M_sun
   description: Gravitational mass of the heavier neutron star.
   range:
-  - 1.2
-  - 1.86
+    train: [1.2, 2]
+    test:  [1.5, 2]
 - name: M2
   symbol: M₂
   unit: M_sun
   description: Gravitational mass of the lighter neutron star.
   range:
-  - 1.02
-  - 1.71
+    train: [1.18, 2]
+    test:  [0.944, 1.35]
 - name: C1
   symbol: C₁
   unit: ""
   description: Compactness of the heavier neutron star, defined as G M / (c² R).
   range:
-  - 0.13
-  - 0.24
+    train: [0.120542, 0.230834]
+    test:  [0.156, 0.230916]
 - name: C2
   symbol: C₂
   unit: ""
   description: Compactness of the lighter neutron star, defined as G M / (c² R).
   range:
-  - 0.11
-  - 0.21
+    train: [0.120542, 0.230834]
+    test:  [0.103, 0.150939]
 
 # FM-H3 backfill (wave-11, 2026-05-26): empty priors block added for schema symmetry with GOLD.
 priors: []

bns_postmerger_waveform_core__f2/metadata.yaml

@@ -14,44 +14,60 @@ target:
   symbol: f₂
   unit: kHz
   description: Dominant post-merger gravitational-wave spectral peak frequency of the BNS remnant, extracted from high-resolution NR simulations.
-  range: [1.77, 3.74]
+  range:
+    train: [1.87612, 3.56651]
+    test:  [1.77242, 3.73145]
 
 inputs:
   - name: mass
     symbol: M
     unit: M_sun
     description: Total gravitational mass of the binary system, m1 + m2.
-    range: [2.40, 3.41]
+    range:
+      train: [2.4, 3.40006]
+      test:  [2.49989, 2.88803]
   - name: q
     symbol: q
     unit: ""
     description: Binary mass ratio m2/m1 with primary mass m1 >= m2, so q lies in (0, 1].
-    range: [0.48, 1.00]
+    range:
+      train: [0.816955, 1]
+      test:  [0.485607, 0.799999]
   - name: kap2t
     symbol: κ₂ᵀ
     unit: ""
     description: Combined tidal coupling constant of the binary, an EOS-sensitive aggregate of component tidal deformabilities weighted by masses.
-    range: [68.00, 1132.00]
+    range:
+      train: [70.388, 436.106]
+      test:  [68.6903, 1131.9]
   - name: lamT
     symbol: Λ̃
     unit: ""
     description: Combined (mass-weighted) tidal deformability of the binary system.
-    range: [364.00, 6256.00]
+    range:
+      train: [375.403, 2325.9]
+      test:  [364.653, 6255.34]
   - name: m1
     symbol: m₁
     unit: M_sun
     description: Gravitational mass of the primary (heavier) neutron star.
-    range: [1.20, 1.95]
+    range:
+      train: [1.2, 1.70003]
+      test:  [1.39959, 1.94401]
   - name: m2
     symbol: m₂
     unit: M_sun
     description: Gravitational mass of the secondary (lighter) neutron star.
-    range: [0.94, 1.71]
+    range:
+      train: [1.10646, 1.70003]
+      test:  [0.944024, 1.2226]
   - name: M_chirp
     symbol: ℳ
     unit: M_sun
     description: Chirp mass of the binary, (m1 * m2)^(3/5) / (m1 + m2)^(1/5).
-    range: [1.04, 1.48]
+    range:
+      train: [1.04466, 1.47996]
+      test:  [1.06186, 1.18859]
 
 data_files:
   train: data/train.csv  # q >= 0.80 (near-equal-mass binaries)

budyko_partitioning_mach__et_over_p/metadata.yaml

@@ -16,54 +16,74 @@ target:
   symbol: E/P
   unit: ""
   description: Long-term mean annual ratio of actual evapotranspiration to precipitation; a dimensionless evaporative fraction bounded between 0 and 1 for nearly all natural catchments.
-  range: [0.0, 1.13]
+  range:
+    train: [0.001681, 0.959384]
+    test:  [0.501995, 0.998182]
 
 inputs:
   - name: pet_over_p
     symbol: φ
     unit: ""
     description: Climate aridity index — ratio of mean annual potential evapotranspiration to mean annual precipitation. Values below 1 indicate energy-limited (humid) conditions; values above 1 indicate water-limited (arid) conditions.
-    range: [0.22, 4.51]
+    range:
+      train: [0.223204, 1.49618]
+      test:  [1.50951, 4.50998]
   - name: ppt_mm_yr
     symbol: P
     unit: mm yr⁻¹
     description: Mean annual precipitation over the 1981–2010 period.
-    range: [251.91, 3118.13]
+    range:
+      train: [511, 3118.13]
+      test:  [251.91, 1021.43]
   - name: pet_mm_yr
     symbol: Ep
     unit: mm yr⁻¹
     description: Mean annual potential evapotranspiration estimated by the Hargreaves method over the 1981–2010 period.
-    range: [491.49, 1927.81]
+    range:
+      train: [491.49, 1793.35]
+      test:  [726.73, 1892.73]
   - name: tav_C
     symbol: T
     unit: °C
     description: Mean annual air temperature over the 1981–2010 period.
-    range: [-0.48, 22.78]
+    range:
+      train: [-0.48, 22.78]
+      test:  [2.37, 21.57]
   - name: area_sqkm
     symbol: A
     unit: km²
     description: Total drainage area of the catchment.
-    range: [4.74, 25546.04]
+    range:
+      train: [4.74, 14269.9]
+      test:  [11.53, 25546]
   - name: basin_slope
     symbol: S
     unit: "%"
     description: Mean slope of the catchment terrain, expressed as a percentage.
-    range: [0.02, 66.62]
+    range:
+      train: [0.02, 66.62]
+      test:  [0.68, 56.49]
   - name: elev_mean_m
     symbol: Z
     unit: m
     description: Mean elevation of the catchment above sea level.
-    range: [8.07, 3602.21]
+    range:
+      train: [8.07, 3602.21]
+      test:  [69.37, 3238.03]
   - name: clay_pct
     symbol: c
     unit: "%"
     description: Soil clay fraction averaged over the catchment, expressed as a percentage of total soil composition.
-    range: [2.31, 53.08]
+    range:
+      train: [2.32, 53.08]
+      test:  [10.18, 51.37]
   - name: bfi
     symbol: BFI
     unit: ""
     description: Base-flow index — dimensionless ratio of long-term base flow to total streamflow, indicating the groundwater contribution to catchment discharge.
-    range: [0.06, 0.86]
+    range:
+      train: [0.0784, 0.8544]
+      test:  [0.0616, 0.8096]
 
 # Dataset — Type I two-file layout
 data_files:
@@ -104,7 +124,7 @@ references:
     n_law_constants: 1    # alpha
     n_other_constants: 0
     n_local_params: 0
-    measured: {rmse: 0.1110, r2: 0.0420}
+    measured: {rmse: 0.107988, r2: -0.212707}
   - id: wang_tang_2014
     label: Wang & Tang 2014 — one-parameter Budyko Eq. (10) with epsilon = 0.55 (MOPEX)
     formula_file: formulas/wang_tang_2014.py
@@ -112,7 +132,7 @@ references:
     n_law_constants: 1    # epsilon
     n_other_constants: 0
     n_local_params: 0
-    measured: {rmse: 0.1099, r2: 0.0604}
+    measured: {rmse: 0.105706, r2: -0.161986}
   - id: yang_2008
     label: Yang et al. 2008 — analytically derived Budyko Eq. (25) with n = 2.0 (Figure 1, PDF p. 4)
     formula_file: formulas/yang_2008.py
@@ -120,7 +140,7 @@ references:
     n_law_constants: 1    # n
     n_other_constants: 0
     n_local_params: 0
-    measured: {rmse: 0.1217, r2: -0.1518}
+    measured: {rmse: 0.122283, r2: -0.555023}
 
 # Candidate prior constants — the `priors` prompt slot. `_role` is audit-only
 # (stripped before the priors are shown). This task offers NO candidate: the

catchment_mean_runoff_camels__q_mean/metadata.yaml

@@ -13,75 +13,103 @@ target:
   name: q_mean
   symbol: Q
   unit: mm/day
-  description: Long-term mean daily runoff per unit catchment area (observed streamflow / area).  Bounded above by mean precipitation (water-balance closure) and below by zero.  Spans 0.0046 (most arid test catchment) to 9.69 mm/day; the declared metric is rmse (the field-standard runoff error; a log metric is unsuitable because the demand-limit baseline legitimately predicts exactly zero in arid catchments).
-  range: [0.0, 9.69]
+  description: Long-term mean daily runoff per unit catchment area (observed streamflow / area).  Bounded above by mean precipitation (water-balance closure) and below by zero.  The most arid test catchments approach zero runoff; the declared metric is rmse (the field-standard runoff error; a log metric is unsuitable because the demand-limit baseline legitimately predicts exactly zero in arid catchments).
+  range:
+    train: [0.336318, 9.68844]
+    test:  [0.00455316, 1.94922]
 
 inputs:
   - name: p_mean
     symbol: P
     unit: mm/day
     description: Long-term mean daily precipitation over the catchment.  PRIMARY input — one of the two Budyko variables; all three reference baselines use it.
-    range: [0.64, 8.94]
+    range:
+      train: [1.88818, 8.93689]
+      test:  [0.644565, 3.8965]
   - name: pet_mean
     symbol: E_p
     unit: mm/day
     description: Long-term mean daily potential evapotranspiration (Priestley-Taylor) over the catchment.  PRIMARY input — the second Budyko variable; all three reference baselines use it.
-    range: [1.89, 4.75]
+    range:
+      train: [1.8991, 4.34899]
+      test:  [1.95294, 4.74391]
   - name: aridity
     symbol: φ
     unit: ""
     description: Aridity index φ = PET/P.  DISTRACTOR — exactly collinear with the ratio of the two primary inputs (aridity = pet_mean/p_mean), so it carries no independent information.  The Budyko curve IS a function of φ, but a formula built from p_mean and pet_mean already has it.
-    range: [0.22, 5.21]
+    range:
+      train: [0.220284, 1.10049]
+      test:  [1.1017, 5.20791]
   - name: p_seasonality
     symbol: s_P
     unit: ""
     description: Precipitation seasonality / phasing index (positive = precip in warm season, negative = cold season).  Candidate secondary predictor; not used by any reference baseline.
-    range: [-1.44, 0.93]
+    range:
+      train: [-1.35835, 0.824779]
+      test:  [-1.43546, 0.922019]
   - name: frac_snow
     symbol: f_s
     unit: ""
     description: Fraction of annual precipitation falling as snow.  Candidate secondary predictor; not used by any reference baseline.
-    range: [0.0, 0.91]
+    range:
+      train: [0, 0.777311]
+      test:  [0, 0.906327]
   - name: area_gages2
     symbol: A
     unit: km^2
     description: Catchment drainage area (GAGES-II estimate).  Candidate secondary predictor; not used by any reference baseline.
-    range: [4.03, 25791.0]
+    range:
+      train: [5.38, 14268.9]
+      test:  [4.03, 25791]
   - name: elev_mean
     symbol: z
     unit: m
     description: Catchment-mean surface elevation above sea level.  Candidate secondary predictor; not used by any reference baseline.
-    range: [10.21, 3571.18]
+    range:
+      train: [10.21, 3370.38]
+      test:  [25.46, 3571.18]
   - name: slope_mean
     symbol: S
     unit: m/km
     description: Catchment-mean topographic slope.  Candidate secondary predictor; not used by any reference baseline.
-    range: [0.82, 255.69]
+    range:
+      train: [0.82221, 255.688]
+      test:  [1.80947, 176.367]
   - name: soil_porosity
     symbol: n
     unit: ""
     description: Catchment-mean soil porosity (volumetric void fraction).  Candidate secondary predictor; not used by any reference baseline.
-    range: [0.37, 0.69]
+    range:
+      train: [0.373282, 0.680023]
+      test:  [0.383145, 0.477398]
   - name: max_water_content
     symbol: S_max
     unit: m
     description: Catchment-mean maximum soil water storage capacity (equivalent water depth).  Candidate secondary predictor; not used by any reference baseline.
-    range: [0.08, 1.06]
+    range:
+      train: [0.161366, 1.05198]
+      test:  [0.0866022, 0.751752]
   - name: geol_permeability
     symbol: log10_k
     unit: log10(m^2)
     description: Base-ten log of catchment-mean subsurface permeability.  Candidate secondary predictor; not used by any reference baseline.
-    range: [-16.5, -10.9]
+    range:
+      train: [-16.5, -10.9]
+      test:  [-16.2648, -11.8]
   - name: frac_forest
     symbol: f_for
     unit: ""
     description: Fraction of catchment area covered by forest.  Candidate secondary predictor; not used by any reference baseline.
-    range: [0.0, 1.0]
+    range:
+      train: [0.0044, 1]
+      test:  [0, 1]
   - name: lai_max
     symbol: L_max
     unit: ""
     description: Catchment-mean annual maximum leaf-area index.  Candidate secondary predictor; not used by any reference baseline.
-    range: [0.36, 5.59]
+    range:
+      train: [1.19995, 5.58212]
+      test:  [0.367125, 5.51154]
 
 # NOTE: the target is NOT algebraically reconstructable from the
 # released inputs.  Long-term runoff is Q = P − E_actual, but the

cepheid_period_luminosity__M_W/metadata.yaml

@@ -13,19 +13,25 @@ target:
   symbol: "M_W"
   unit: mag
   description: Absolute Gaia Wesenheit magnitude — a reddening-free combination of broadband Gaia photometry referenced to the source distance via parallax.
-  range: [-8.47, -4.38]
+  range:
+    train: [-5.98979, -4.49653]
+    test:  [-8.46608, -6.1097]
 
 inputs:
   - name: log_P
     symbol: "log P"
     unit: ""
     description: Base-10 logarithm of the fundamental-mode pulsation period, with the period measured in days.
-    range: [0.41, 1.66]
+    range:
+      train: [0.489958, 0.989183]
+      test:  [1.01456, 1.65315]
   - name: feh
     symbol: "[Fe/H]"
     unit: dex
     description: Photospheric iron abundance relative to solar on a logarithmic scale.
-    range: [-0.18, 0.33]
+    range:
+      train: [-0.18, 0.22]
+      test:  [0.02, 0.33]
 
 data_files:
   train: data/train.csv   # log_P < 1.0 (short-to-medium period; P < 10 d)

cloud_cover_parameterization__cloud_cover/metadata.yaml

@@ -13,49 +13,67 @@ target:
   symbol: C
   unit: ""
   description: Sub-grid cloud area fraction in a coarse-grained atmospheric volume, dimensionless in [0, 1].
-  range: [0.0, 1.0]
+  range:
+    train: [0, 1]
+    test:  [0, 1]
 
 inputs:
   - name: q_c
     symbol: q_c
     unit: kg/kg
     description: Cloud liquid water mixing ratio — mass of cloud liquid water per unit mass of moist air.
-    range: [0.0, 0.00048]
+    range:
+      train: [0, 0.000473]
+      test:  [0, 0.000393]
   - name: q_i
     symbol: q_i
     unit: kg/kg
     description: Cloud ice mixing ratio — mass of cloud ice per unit mass of moist air.
-    range: [0.0, 0.00014]
+    range:
+      train: [0, 0.000129]
+      test:  [0, 0.000129]
   - name: RH
     symbol: RH
     unit: ""
     description: Relative humidity with respect to liquid water, computed from specific humidity, temperature, and pressure.
-    range: [0.0003, 1.02]
+    range:
+      train: [0.000389, 1.0156]
+      test:  [0.000467, 0.999862]
   - name: T
     symbol: T
     unit: K
     description: Air temperature at the model level.
-    range: [191.0, 313.0]
+    range:
+      train: [191.311, 312.824]
+      test:  [191.666, 311.6]
   - name: p
     symbol: p
     unit: Pa
     description: Atmospheric pressure at the model level.
-    range: [1525.0, 101611.0]
+    range:
+      train: [1525.9, 101598]
+      test:  [1526.61, 101610]
   - name: q_v
     symbol: q_v
     unit: kg/kg
     description: Specific humidity — mass of water vapour per unit mass of moist air.
-    range: [0.000002, 0.0198]
+    range:
+      train: [2e-06, 0.019779]
+      test:  [2e-06, 0.019173]
   - name: z_g
     symbol: z_g
     unit: m
     description: Geometric height of the model level above the surface.
-    range: [20.0, 28194.0]
+    range:
+      train: [20, 28193.8]
+      test:  [20, 28193.8]
   - name: dz_RH
     symbol: dRH/dz
     unit: 1/m
     description: Vertical derivative of relative humidity along the geometric height coordinate, computed by cubic spline interpolation.
-    range: [-0.00283, 0.00065]
+    range:
+      train: [-0.002556, 0.000643]
+      test:  [-0.002824, 0.000589]
 
 # FM-H3 backfill (wave-11, 2026-05-26): empty priors block added for schema symmetry with GOLD.
 priors: []

commuting_gravity_lodes__flow_Tij/metadata.yaml

@@ -14,8 +14,8 @@ target:
     county to the work county, after a +1 shift to avoid log(0) for single-worker
     pairs.
   range:
-  - 0.301
-  - 5.508  # rounded up by 0.001 to ensure inclusion of the actual max 5.507104
+    train: [0.30103, 5.5071]
+    test:  [0.30103, 5.30913]
 inputs:
 - name: m_i
   symbol: "m\u1D62"
@@ -23,24 +23,24 @@ inputs:
   description: "Total resident workers of the origin county \u2014 the sum of outbound\
     \ commuting flows from county i across all destinations."
   range:
-  - 107
-  - 4167638
+    train: [107, 4167638]
+    test:  [107, 4167638]
 - name: m_j
   symbol: "m\u2C7C"
   unit: workers
   description: "Total jobs located in the destination county \u2014 the sum of inbound\
     \ commuting flows to county j from all origins."
   range:
-  - 107
-  - 4167638
+    train: [107, 4167638]
+    test:  [107, 4167638]
 - name: d_ij_km
   symbol: "d\u1D62\u2C7C"
   unit: km
   description: Haversine great-circle distance between the representative centroids
     of the home and work counties.
   range:
-  - 21.79
-  - 1223.2
+    train: [21.7917, 1186.97]
+    test:  [23.1785, 1223.19]
 - name: s_ij
   symbol: "s\u1D62\u2C7C"
   unit: workers
@@ -48,8 +48,8 @@ inputs:
     \ summed over all counties whose centroids lie strictly inside the disk of radius\
     \ d_ij centred on the home county, excluding both home and work counties."
   range:
-  - 0
-  - 20447583
+    train: [0, 20447583]
+    test:  [0, 20433795]
 data_files:
   train: data/train.csv
   test: data/test.csv
@@ -178,14 +178,3 @@ best_baseline: 0.536064    # lenormand_2016_gravity_pow — rmse (lower is bette
 # was the RMSE of predicting the TRAIN mean on the OOD test, NOT the section-C9-c
 # naive (test-mean) predictor. Under the correct naive, NO baseline beat naive on
 # the old OOD split -- which is exactly what flagged the degenerate split. See VERDICT.md.
-auditor_notes:
-  mixed_bank_disclosure:
-    severity: ACCEPTABLE_E6_MIXED_BANK
-    split: representative (deterministic md5 hash of county-pair id, ~15% holdout)
-    summary: >-
-      Representative split replacing a degenerate d>=600 km OOD split. Mixed bank:
-      gravity_pow R2=+0.540 and simini R2=+0.143 beat naive (test-mean RMSE 0.790);
-      extrad R2=-0.532 and gravity_exp R2=-3.040 are weaker published Lenormand
-      models with frozen scaling-law coefficients. Section 9.6 satisfied (>=1 beats
-      naive). zipf_1946 removed (un-normalised, R2=-35 in-sample).
-    citation: "AUDIT_CHECKLIST E6 (mixed-bank disclosure) + split-priority rule (representative split when source papers validate on the whole dataset)"

dna_melting_temperature_khandelwal__Tm/metadata.yaml

@@ -16,14 +16,18 @@ target:
   symbol: "T_m"
   unit: "°C"
   description: Experimentally measured DNA duplex melting temperature — the temperature at which 50% of double-stranded DNA has dissociated into single strands, determined by UV-absorbance hyperchromic effect.
-  range: [35.30, 88.60]
+  range:
+    train: [39.1, 88.6]
+    test:  [35.3, 74.1]
 
 inputs:
   - name: E
     symbol: E
     unit: dimensionless
     description: DNA strength parameter per base — sum of overlapping dinucleotide step strength values divided by sequence length, encoding stacking and hydrogen-bonding contributions to duplex stability.
-    range: [5.409091, 9.760000]
+    range:
+      train: [5.40909, 9.76]
+      test:  [6.06667, 9.6]
   - name: N
     symbol: "N"
     unit: bp
@@ -33,12 +37,16 @@ inputs:
     symbol: "[Na+]"
     unit: "M"
     description: Sodium ion concentration of the solution in Molar units, controlling electrostatic screening of the DNA backbone.
-    range: [0.07, 1.02]
+    range:
+      train: [0.07, 1.02]
+      test:  [0.07, 1.02]
   - name: dna_conc_M
     symbol: "C_DNA"
     unit: "M"
     description: Total nucleotide strand concentration in Molar units.
-    range: [0.000001, 0.000100]
+    range:
+      train: [1e-06, 0.0001]
+      test:  [2e-06, 2e-06]
 
 data_files:
   train: data/train.csv   # Tables S1+S2: 348 oligomers, N=16-30 bp
@@ -124,7 +132,6 @@ caps:
   fit_timeout_seconds: null
 
 metric: rmse
-reference_baseline_id: khandelwal_2010   # sole baseline; the metric anchor
 # === Validity rubric ===
 # Frozen list — the scoring-time judge (eval_consistency.py) scores a
 # submission by how many of THESE criteria it satisfies:

drosophila_isolation_by_distance_dest__M_stat/metadata.yaml

@@ -14,19 +14,25 @@ target:
   symbol: M
   unit: ""
   description: Rousset's linearised differentiation statistic, M_stat = F_ST / (1 - F_ST), per pair of populations.
-  range: [0.0, 0.2934]
+  range:
+    train: [0.000335, 0.229405]
+    test:  [1e-06, 0.293438]
 
 inputs:
   - name: log_distance_km
     symbol: ln(d)
     unit: ln(km)
     description: Natural logarithm of the great-circle distance between the two populations of the pair (clipped to d >= 1 km before taking the log).
-    range: [0.0, 9.1509]
+    range:
+      train: [6.74304, 7.85398]
+      test:  [0, 9.15094]
   - name: distance_km
     symbol: d
     unit: km
     description: Great-circle (Haversine) distance between the two populations.
-    range: [1.0, 9423.28]
+    range:
+      train: [848.132, 2575.97]
+      test:  [1, 9423.28]
 
 # Dataset — Type I, two-file flat layout.  Source:
 # data_raw/pairwise_F_ST.csv (1378 unordered pairs of 53 populations,

eclipsing_binary_mass_luminosity_debcat__log_L_Lsun/metadata.yaml

@@ -14,19 +14,25 @@ target:
   symbol: log₁₀(L/L☉)
   unit: dex
   description: Base-10 logarithm of bolometric stellar luminosity expressed in solar luminosities.
-  range: [-2.32, 5.26]
+  range:
+    train: [-2.3, 5.187]
+    test:  [-2.313, 5.253]
 
 inputs:
   - name: log_M_Msun
     symbol: log₁₀(M/M☉)
     unit: dex
-    description: Base-10 logarithm of stellar mass expressed in solar masses.  Train [-0.68, 1.44] (pre-2018 stars, M = 0.21-27.3 M_sun); test [-0.98, 1.37] (post-2018 stars, M = 0.11-23.5 M_sun); both splits cover the full mass range.
-    range: [-0.98, 1.44]
+    description: Base-10 logarithm of stellar mass expressed in solar masses.  Train = pre-2018 stars (M = 0.21-27.3 M_sun); test = post-2018 stars (M = 0.11-23.5 M_sun); both splits cover the full mass range.
+    range:
+      train: [-0.678195, 1.43569]
+      test:  [-0.976336, 1.37088]
   - name: metallicity
     symbol: "[M/H]"
     unit: dex
     description: Logarithmic stellar metallicity relative to the Sun.  Missing (NaN) for stars without a published spectroscopic determination (~50% of rows).  Distractor input — no reference baseline uses it.
-    range: [-1.86, 0.55]
+    range:
+      train: [-1.86, 0.55]
+      test:  [-1.47, 0.51]
 
 # NOTE: log_R_Rsun and T_eff_K are NOT released as inputs.  DEBCat's
 # log_L is derived from R and T_eff via the Stefan-Boltzmann radiation

exfor_neutron_capture_resonance_gold__sigma_E/metadata.yaml

@@ -13,15 +13,19 @@ target:
   name: log10_sigma_b
   symbol: log10(sigma)
   unit: log10(barn)
-  description: Base-10 logarithm of the Au-197(n,gamma) inclusive capture cross section in barns.  The log transform is needed because sigma spans >5 decades from the thermal background (~10-100 b) to the 4.9 eV resonance peak (~33,000 b).  Range -1.18 to 4.45 in the E < 50 eV window.
-  range: [-1.184, 4.446]
+  description: Base-10 logarithm of the Au-197(n,gamma) inclusive capture cross section in barns.  The log transform is needed because sigma spans >5 decades from the thermal background (~10-100 b) to the 4.9 eV resonance peak (~33,000 b), measured in the E < 50 eV window.
+  range:
+    train: [-0.866238, 4.44623]
+    test:  [-1.18417, 4.3353]
 
 inputs:
   - name: E_eV
     symbol: E
     unit: eV
-    description: Incident neutron kinetic energy in electronvolts.  The single SR feature — both the 1/v law and the Breit-Wigner resonance are univariate functions of E.  Spans 0.04 meV (sub-thermal) to ~50 eV in this window.
-    range: [3.99e-05, 49.975]
+    description: Incident neutron kinetic energy in electronvolts.  The single SR feature — both the 1/v law and the Breit-Wigner resonance are univariate functions of E.  Spans the sub-thermal to ~50 eV window.
+    range:
+      train: [3.99e-05, 49.975]
+      test:  [6.88e-05, 49.308]
 
 # NOTE: the target is NOT circular.  sigma is the experimentally MEASURED
 # cross section (EXFOR), E is the independently-set incident neutron

exoplanet_mass_radius__R/metadata.yaml

@@ -17,7 +17,9 @@ target:
   unit: R_earth
   description: Radius of the object in Earth radii, measured from transit photometry
     (planets) or eclipsing/light-curve analysis (brown dwarfs).
-  range: [0.30, 87.3]
+  range:
+    train: [0.3098, 87.2059]
+    test:  [0.338, 30.2643]
 
 inputs:
   - name: M
@@ -25,7 +27,9 @@ inputs:
     unit: M_earth
     description: Mass of the object in Earth masses; directly measured from radial
       velocity, transit timing variations, or binary orbital dynamics.
-    range: [0.02, 24982.0]
+    range:
+      train: [0.0553, 24981.3]
+      test:  [0.02, 23805.3]
 
 # Dataset — Type I two-file layout
 data_files:

galaxy_mass_metallicity_sdss__oh/metadata.yaml

@@ -17,19 +17,25 @@ target:
   symbol: 12 + log(O/H)
   unit: dex
   description: Gas-phase oxygen abundance of the galaxy's interstellar medium, expressed as 12 plus the base-10 logarithm of the number-density ratio of oxygen to hydrogen atoms.
-  range: [7.67, 9.4666]
+  range:
+    train: [7.67, 9.4666]
+    test:  [7.67, 9.432]
 
 inputs:
   - name: log_Mstar
     symbol: log₁₀ M★
     unit: log(M_sun)
     description: Base-10 logarithm of the galaxy's total stellar mass in solar masses.
-    range: [8.50, 11.50]
+    range:
+      train: [8.5001, 11.4996]
+      test:  [8.5005, 11.4985]
   - name: log_SFR
     symbol: log₁₀ SFR
     unit: log(M_sun/yr)
     description: Base-10 logarithm of the galaxy's total star-formation rate in solar masses per year.
-    range: [-4.01, 3.88]
+    range:
+      train: [-4.003, 3.8753]
+      test:  [-3.3512, 3.8751]
 
 data_files:
   train: data/train.csv   # deterministic mass-stratified representative split (~80%)
@@ -190,7 +196,6 @@ caps:
   fit_timeout_seconds:     null
 
 metric: rmse
-reference_baseline_id: mannucci_2010_mzr   # lowest RMSE / highest R² on the representative split
 # === Validity rubric ===
 # Frozen list — the scoring-time judge (eval_consistency.py) scores a
 # submission by how many of THESE criteria it satisfies:

galaxy_stellar_mass_sdss__log_M_star/metadata.yaml

@@ -19,7 +19,9 @@ target:
     Log base-10 of total galaxy stellar mass in solar masses, estimated by the
     MPA-JHU pipeline via Bayesian spectral fitting of the 4000 Angstrom break
     strength and the Hδ_A absorption index under the Kroupa initial mass function.
-  range: [7.02, 12.48]   # train [7.029926, 12.46617]; test [8.118873, 12.46157]; round outward
+  range:
+    train: [7.02993, 12.4662]
+    test:  [8.11887, 12.4616]
 
 inputs:
   - name: u
@@ -27,38 +29,50 @@ inputs:
     unit: mag
     description: >-
       Galactic-extinction-corrected SDSS u-band apparent model magnitude on the AB system.
-    range: [15.36, 22.96]   # full: [15.36988, 22.94913]; round outward
+    range:
+      train: [15.5337, 22.8882]
+      test:  [15.3699, 22.9491]
   - name: g
     symbol: "g"
     unit: mag
     description: >-
       Galactic-extinction-corrected SDSS g-band apparent model magnitude on the AB system.
-    range: [14.86, 22.52]   # full: [14.86303, 22.51956]
+    range:
+      train: [14.863, 22.5196]
+      test:  [15.2487, 22.3044]
   - name: r
     symbol: "r"
     unit: mag
     description: >-
       Galactic-extinction-corrected SDSS r-band apparent model magnitude on the AB system.
-    range: [14.41, 21.92]   # full: [14.41012, 21.91952]
+    range:
+      train: [14.4101, 21.5607]
+      test:  [15.1033, 21.9195]
   - name: i
     symbol: "i"
     unit: mag
     description: >-
       Galactic-extinction-corrected SDSS i-band apparent model magnitude on the AB system.
-    range: [14.08, 21.97]   # full: [14.08313, 21.96579]
+    range:
+      train: [14.0831, 21.9029]
+      test:  [14.6102, 21.9658]
   - name: z_mag
     symbol: "z"
     unit: mag
     description: >-
       Galactic-extinction-corrected SDSS z-band apparent model magnitude on the AB system.
       Column named z_mag to avoid ambiguity with the redshift symbol z.
-    range: [13.76, 21.96]   # full: [13.76824, 21.95789]
+    range:
+      train: [13.7682, 21.9579]
+      test:  [14.3054, 21.8377]
   - name: redshift
     symbol: "z"
     unit: ""
     description: >-
       Spectroscopic redshift of the galaxy measured by the SDSS spectrograph.
-    range: [0.02, 0.31]   # full: [0.020002, 0.299998]; round outward
+    range:
+      train: [0.020002, 0.157737]
+      test:  [0.157737, 0.299998]
   - name: DM
     symbol: "μ"
     unit: mag
@@ -66,7 +80,9 @@ inputs:
       Distance modulus computed from spectroscopic redshift under a flat LCDM
       cosmology (H0=70 km/s/Mpc, Omega_m=0.3, Omega_Lambda=0.7), enabling
       conversion from apparent to absolute magnitudes.
-    range: [34.69, 40.96]   # full: [34.6972, 40.9555]; round outward
+    range:
+      train: [34.6972, 39.3857]
+      test:  [39.3857, 40.9555]
 
 # Dataset — Type I two-file layout
 # FM G1-borderline upgrade (2026-05-26, wave-3 subagent K finding + wave-4 subagent O fix):

geomagnetic_storm_dst__dDst_dt/metadata.yaml

@@ -17,7 +17,9 @@ target:
     finite difference of the hourly Dst series: (Dst[i+1] - Dst[i-1]) / 2.
     Negative during storm main phase (ring-current intensification);
     positive during recovery.
-  range: [-133.0, 62.5]
+  range:
+    train: [-133, 62.5]
+    test:  [-35, 13]
 
 inputs:
   - name: Dst
@@ -27,26 +29,34 @@ inputs:
       Disturbance storm-time index measuring the globally averaged
       horizontal magnetic perturbation at low latitudes; negative values
       indicate ring-current intensification.
-    range: [-422.0, 77.0]
+    range:
+      train: [-422, 77]
+      test:  [-82, 40]
   - name: Ey
     symbol: "E_y"
     unit: mV/m
     description: >-
       Dawn-to-dusk component of the solar wind convective electric field
       at L1, the primary driver of ring-current injection.
-    range: [-35.94, 47.11]
+    range:
+      train: [-35.94, 47.11]
+      test:  [-7.12, 8.13]
   - name: P_dyn
     symbol: "P_dyn"
     unit: nPa
     description: >-
       Solar wind dynamic pressure at L1, contributing to the magnetopause
       current correction of the Dst index.
-    range: [0.03, 71.04]
+    range:
+      train: [0.03, 71.04]
+      test:  [0.16, 10.68]
   - name: V_sw
     symbol: "V_sw"
     unit: km/s
     description: Solar wind bulk flow speed at L1.
-    range: [228.0, 1189.0]
+    range:
+      train: [228, 1189]
+      test:  [273, 614]
   - name: Bz_GSM
     symbol: "B_z"
     unit: nT
@@ -54,22 +64,30 @@ inputs:
       North-south component of the interplanetary magnetic field in
       geocentric solar magnetospheric coordinates; negative (southward)
       values drive dayside reconnection.
-    range: [-57.8, 41.6]
+    range:
+      train: [-57.8, 41.6]
+      test:  [-18.3, 15.9]
   - name: n_sw
     symbol: "n_sw"
     unit: cm^-3
     description: Solar wind proton number density at L1.
-    range: [0.1, 137.2]
+    range:
+      train: [0.1, 137.2]
+      test:  [0.6, 44.7]
   - name: B_mag
     symbol: "|B|"
     unit: nT
     description: Magnitude of the interplanetary magnetic field at L1.
-    range: [0.4, 62.0]
+    range:
+      train: [0.4, 62]
+      test:  [0.7, 20.6]
   - name: T_sw
     symbol: "T_sw"
     unit: K
     description: Solar wind proton temperature at L1.
-    range: [3299.0, 897673.0]
+    range:
+      train: [3299, 897673]
+      test:  [4002, 531324]
 
 data_files:
   train: data/train.csv   # years 1995-2021 Mar; temporal OOD split

ghg_emission_drivers__CDE/metadata.yaml

@@ -13,39 +13,53 @@ target:
   symbol: "I_CO2"
   unit: kg CO2
   description: National annual CO2 emissions from fossil fuel combustion and industrial processes, excluding land-use change and forestry
-  range: [14000000.0, 10600000000000.0]
+  range:
+    train: [1.4319e+07, 1.05304e+13]
+    test:  [4.09376e+08, 6.14233e+12]
 
 inputs:
   - name: GDP
     symbol: "A"
     unit: USD per capita (constant 2015 USD)
     description: Gross domestic product per capita in constant 2015 US dollars
-    range: [166.71, 112418.0]
+    range:
+      train: [166.71, 19979.6]
+      test:  [20014.5, 112418]
   - name: TP
     symbol: "P"
     unit: persons
     description: Total national population
-    range: [52597.0, 1379860000.0]
+    range:
+      train: [63066, 1379860000]
+      test:  [52597, 321815121]
   - name: AP
     symbol: "AP"
     unit: "%"
     description: Share of population aged 15 to 64 years (working-age or active population)
-    range: [45.77, 85.25]
+    range:
+      train: [45.7769, 75.4748]
+      test:  [55.0479, 85.2495]
   - name: DP
     symbol: "DP"
     unit: persons per km2
     description: Population density as number of inhabitants per square kilometre of land area
-    range: [1.35, 6944.1]
+    range:
+      train: [1.35625, 1224.42]
+      test:  [2.22136, 6944.1]
   - name: UR
     symbol: "UR"
     unit: "%"
     description: Percentage of the total population living in urban areas
-    range: [5.27, 100.0]
+    range:
+      train: [5.27494, 97.9803]
+      test:  [27.0586, 100]
   - name: AT
     symbol: "AT"
     unit: "degrees C"
     description: Annual mean surface air temperature averaged over the national territory
-    range: [-4.97, 29.78]
+    range:
+      train: [-4.90051, 29.7776]
+      test:  [-4.96912, 28.9267]
 
 # FM-H3 backfill (wave-11, 2026-05-26): empty priors block added for schema symmetry with GOLD.
 priors: []

ghg_emission_drivers__EC/metadata.yaml

@@ -13,39 +13,53 @@ target:
   symbol: "I_E"
   unit: TJ
   description: Annual national primary energy consumption from all sources.
-  range: [271.74, 137592706.0]
+  range:
+    train: [271.743, 1.37593e+08]
+    test:  [6275.1, 1.04414e+08]
 
 inputs:
   - name: GDP
     symbol: "A"
     unit: USD per capita (constant 2015 USD (World Bank WDI base year))
     description: Per-capita gross domestic product, constant 2015 USD (World Bank WDI base year).
-    range: [166.71, 112417.88]
+    range:
+      train: [166.71, 13145.8]
+      test:  [13165.8, 112418]
   - name: TP
     symbol: "P"
     unit: persons
     description: Total resident population.
-    range: [48232.0, 1379860000.0]
+    range:
+      train: [52878, 1379860000]
+      test:  [48232, 321815121]
   - name: AP
     symbol: "f_{active}"
     unit: "%"
     description: Share of population aged 15 to 64 (working-age fraction).
-    range: [45.77, 85.25]
+    range:
+      train: [45.7769, 72.8645]
+      test:  [54.1385, 85.2495]
   - name: DP
     symbol: "\\rho"
     unit: persons per km2
     description: Population density in persons per square kilometre of land area.
-    range: [0.13, 6944.10]
+    range:
+      train: [1.35625, 1224.42]
+      test:  [0.136436, 6944.1]
   - name: UR
     symbol: "f_{urban}"
     unit: "%"
     description: Share of population living in urban areas.
-    range: [5.27, 100.00]
+    range:
+      train: [5.27494, 92.9604]
+      test:  [25, 100]
   - name: AT
     symbol: "T"
     unit: "degrees C"
     description: Mean annual surface air temperature at the country level.
-    range: [-20.45, 29.78]
+    range:
+      train: [-4.90051, 29.7776]
+      test:  [-20.4459, 28.9267]
 
 priors: []
 
@@ -70,8 +84,8 @@ references:
     n_other_constants: 1
     n_local_params: 0
     measured:
-      rmse: 6823023.853269
-      r2: 0.883284
+      smape: 0.472866
+      r2: 0.886862
   - id: vazquez_2022_bms
     label: BMS symbolic regression expression for log(EC), Vazquez 2022 Table 3
     formula_file: formulas/vazquez_2022_bms.py
@@ -80,8 +94,8 @@ references:
     n_other_constants: 1   # AT_K_OFFSET = 273.15 (Celsius → Kelvin for AT)
     n_local_params: 0
     measured:
-      rmse: 3699955.079356
-      r2: 0.965678
+      smape: 0.485155
+      r2: 0.959087
     # 2026-05-25 structural fix (relay-A gen 9 wave 81): predict() and docstring
     # repaired after opus PDF audit found that the outer-exponent denominator
     # was a3 (wrong; should be AT in Kelvin) and the inner-base leading

ghg_emission_drivers__ME/metadata.yaml

@@ -13,39 +13,53 @@ target:
   symbol: "I_CH4"
   unit: kg CH4
   description: National annual CH4 emissions from all sectors excluding land-use change and forestry
-  range: [478974.0, 56670782000.0]
+  range:
+    train: [478974, 5.66708e+10]
+    test:  [500007, 3.14636e+10]
 
 inputs:
   - name: GDP
     symbol: "A"
     unit: USD per capita (constant 2015 USD (World Bank WDI base year))
     description: Gross domestic product per capita in constant 2015 US dollars
-    range: [166.71, 112418.0]
+    range:
+      train: [166.71, 19979.6]
+      test:  [20014.5, 112418]
   - name: TP
     symbol: "P"
     unit: persons
     description: Total national population
-    range: [52597.0, 1379860000.0]
+    range:
+      train: [63066, 1379860000]
+      test:  [52597, 321815121]
   - name: AP
     symbol: "AP"
     unit: "%"
     description: Share of population aged 15 to 64 years (working-age or active population)
-    range: [45.77, 85.25]
+    range:
+      train: [45.7769, 75.4748]
+      test:  [55.0479, 85.2495]
   - name: DP
     symbol: "DP"
     unit: persons per km2
     description: Population density as number of inhabitants per square kilometre of land area
-    range: [1.35, 6944.1]
+    range:
+      train: [1.35625, 1224.42]
+      test:  [2.22136, 6944.1]
   - name: UR
     symbol: "UR"
     unit: "%"
     description: Percentage of the total population living in urban areas
-    range: [5.27, 100.0]
+    range:
+      train: [5.27494, 97.9803]
+      test:  [27.0586, 100]
   - name: AT
     symbol: "AT"
     unit: "degrees C"
     description: Annual mean surface air temperature averaged over the national territory
-    range: [-4.97, 29.78]
+    range:
+      train: [-4.90051, 29.7776]
+      test:  [-4.96912, 28.9267]
 
 # FM-H3 backfill (wave-11, 2026-05-26): empty priors block added for schema symmetry with GOLD.
 priors: []
@@ -97,7 +111,6 @@ caps:
   fit_timeout_seconds: null
 
 metric: smape
-reference_baseline_id: vazquez_2022_stirpat   # strongest baseline (lowest SMAPE / highest R²); the metric anchor
 # === Validity rubric ===
 # Frozen list — the scoring-time judge (eval_consistency.py) scores a
 # submission by how many of THESE criteria it satisfies:

ghg_emission_drivers__NOE/metadata.yaml

@@ -13,39 +13,53 @@ target:
   symbol: "I_N2O"
   unit: kg N2O
   description: National annual N2O emissions from all sectors excluding land-use change and forestry, measured in kilograms of N2O per country per year
-  range: [11392.0, 2173932900.0]
+  range:
+    train: [11392.1, 2.17393e+09]
+    test:  [45141.4, 1.6905e+09]
 
 inputs:
   - name: GDP
     symbol: "A"
     unit: USD per capita (constant 2015 USD (World Bank WDI base year))
     description: Gross domestic product per capita in constant 2015 US dollars
-    range: [166.71, 112418.0]
+    range:
+      train: [166.71, 19979.6]
+      test:  [20014.5, 112418]
   - name: TP
     symbol: "P"
     unit: persons
     description: Total national population
-    range: [52597.0, 1379860000.0]
+    range:
+      train: [63066, 1379860000]
+      test:  [52597, 321815121]
   - name: AP
     symbol: "AP"
     unit: "%"
     description: Share of population aged 15 to 64 years (working-age or active population)
-    range: [45.77, 85.25]
+    range:
+      train: [45.7769, 75.4748]
+      test:  [55.0479, 85.2495]
   - name: DP
     symbol: "DP"
     unit: persons per km2
     description: Population density as number of inhabitants per square kilometre of land area
-    range: [1.35, 6944.1]
+    range:
+      train: [1.35625, 1224.42]
+      test:  [2.22136, 6944.1]
   - name: UR
     symbol: "UR"
     unit: "%"
     description: Percentage of the total population living in urban areas
-    range: [5.27, 100.0]
+    range:
+      train: [5.27494, 97.9803]
+      test:  [27.0586, 100]
   - name: AT
     symbol: "AT"
     unit: "degrees C"
     description: Annual mean surface air temperature averaged over the national territory
-    range: [-4.97, 29.78]
+    range:
+      train: [-4.90051, 29.7776]
+      test:  [-4.96912, 28.9267]
 
 # FM-H3 backfill (wave-11, 2026-05-26): empty priors block added for schema symmetry with GOLD.
 priors: []
@@ -85,35 +99,6 @@ references:
     measured:
       smape: 0.587733
       r2: 0.793058
-auditor_notes:
-  neg_r2_disclosure:
-    severity: ACCEPTABLE_PER_E6_MIXED_BANK
-    baselines_affected:
-      - id: vazquez_2022_bms
-        pooled_r2: -0.006502
-        pooled_smape: 1.456827
-    best_baseline_id: vazquez_2022_stirpat
-    best_baseline_smape: 0.587733
-    best_baseline_r2: 0.793058
-    summary: >-
-      One baseline (vazquez_2022_bms) yields a near-naive negative pooled R²=-0.007 on the
-      high-GDP OOD test split. It is a verified-published (Vazquez 2022 Table 3, NOE column)
-      but weak Bayesian-symbolic-regression expression. The bank satisfies §9.6 via
-      vazquez_2022_stirpat (R²=0.793, sMAPE=0.588). Metric is sMAPE (NOE spans ~5 orders of
-      magnitude). §9.14 (feature-OOD legitimately negative) applies.
-    citation: "AUDIT_CHECKLIST.md §E6 (mixed-bank partial-negative-R²) + data_spec §9.14"
-  york_2003_removed:
-    date: 2026-05-29
-    by: "逐条审核 / 逐条11 (claude-opus, effort=max)"
-    reason: >-
-      york_2003 dropped to .dropped_baselines/. It was a CROSS-TARGET GRAFT — York 2003's
-      1996 CO2 Model-3 coefficients (Table 1) applied to the N2O target (FM C9) — plus a C11
-      centered-intercept-on-raw-inputs error. Its in-distribution adequacy (train R²=0.41) was
-      an accident of the two errors partially cancelling; with the mathematically-correct
-      uncentered intercept (+0.899) it collapses to sMAPE=2.0 / R²≈-1e9. Its docstring also
-      made a factually-incorrect C11 claim ("both produce sMAPE≈2.0"). In the score it floored
-      at 0.0 (no ladder value). Bank is now STIRPAT (anchor) + BMS, both verified-published
-      from Vazquez 2022 on the actual N2O target.
 
 caps:
   max_law_constants: 7

gravity_wgs84_somigliana__g0/metadata.yaml

@@ -15,14 +15,18 @@ target:
   symbol: "γ"
   unit: m/s^2
   description: Normal gravity at the WGS84 reference ellipsoid surface — observed gravity reduced to sea level via free-air correction.
-  range: [9.785, 9.840]
+  range:
+    train: [9.78542, 9.83323]
+    test:  [9.81836, 9.83737]
 
 inputs:
   - name: latitude
     symbol: "φ"
     unit: degrees
     description: Geodetic latitude of the gravity station, positive north.
-    range: [20.0, 80.0]
+    range:
+      train: [20.0083, 59.9973]
+      test:  [60, 79.9917]
 
 # Dataset — Type I two-file layout
 data_files:

ground_motion_ita18__log10_pga/metadata.yaml

@@ -15,14 +15,18 @@ target:
   description: >-
     Log base 10 of the RotD50 peak ground acceleration of the horizontal components,
     in cm/s^2. RotD50 is the 50th percentile over all horizontal azimuths.
-  range: [-2.03, 3.12]
+  range:
+    train: [-2.02326, 2.75797]
+    test:  [-0.443219, 3.10983]
 
 inputs:
   - name: Mw
     symbol: "M_w"
     unit: ""
     description: Moment magnitude of the earthquake.
-    range: [4.0, 6.9]
+    range:
+      train: [4, 5.7]
+      test:  [5.8, 6.9]
   - name: R_km
     symbol: R
     unit: km
@@ -30,12 +34,16 @@ inputs:
       Joyner-Boore distance (horizontal distance to the surface projection of the
       fault rupture) in km; epicentral distance used as fallback when fault geometry
       is unavailable.
-    range: [0.0, 199.61]
+    range:
+      train: [0, 199.6]
+      test:  [0, 199.61]
   - name: depth_km
     symbol: h
     unit: km
     description: Hypocentral depth of the earthquake in km.
-    range: [0.21, 29.0]
+    range:
+      train: [0.218, 29]
+      test:  [5, 22]
   - name: ec8_A
     symbol: "C_A"
     unit: ""

gutenberg_richter_b_value_usgs__gr_freq_mag/metadata.yaml

@@ -15,14 +15,18 @@ target:
   symbol: "log₁₀N"
   unit: ""
   description: Log base-10 of the cumulative annual earthquake rate for events with moment magnitude at or above the threshold.
-  range: [-1.35, 2.43]
+  range:
+    train: [0.60392, 2.42092]
+    test:  [-1.3479, 0.518384]
 
 inputs:
   - name: M_threshold
     symbol: "M"
     unit: ""
     description: Moment-magnitude threshold — minimum Mw for which log10_N is the cumulative annual event rate.
-    range: [5.6, 9.1]
+    range:
+      train: [5.6, 7.4]
+      test:  [7.5, 9.1]
 
 # Dataset — Type I two-file layout
 data_files:

hacks_law_river_length_hydrosheds__main_length_km/metadata.yaml

@@ -1,14 +1,16 @@
 task_id: hacks_law_river_length_hydrosheds__main_length_km
 domain: earth_science / hydrology
 license: HydroSHEDS-v1-Derivative
-license_notes: "Distributed under HydroSHEDS V1 custom license (TechDoc v1.4 Appendix
-  A). §7.1 grants free non-commercial and commercial use; §2.1.2 permits distribution
-  of Licensed Materials incorporated into Derivative Works. The benchmark CSV is a
-  materially transformed Derivative Work (filtered: NEXT_DOWN==0, UPLAND_SKM>=100,
-  DIST_UP_KM>=5; column-renamed; aggregated from 9 shapefiles; embedded in a benchmark
-  harness with formulas, metadata, and evaluation scripts). License determination
-  via §2.1.2 Derivative Work + §7.1 general grant + Feist 1991 numeric-facts doctrine.
-  DEC filed 2026-05-25; GM-approved 2026-05-26. See audit/proposed_GM/hacks_law_license_decision_2026-05-25.md."
+# license_notes: Distributed under HydroSHEDS V1 custom license (TechDoc v1.4
+#   Appendix A). §7.1 grants free non-commercial and commercial use; §2.1.2
+#   permits distribution of Licensed Materials incorporated into Derivative Works.
+#   The benchmark CSV is a materially transformed Derivative Work (filtered:
+#   NEXT_DOWN==0, UPLAND_SKM>=100, DIST_UP_KM>=5; column-renamed; aggregated from
+#   9 shapefiles; embedded in a benchmark harness with formulas, metadata, and
+#   evaluation scripts). License determination via §2.1.2 Derivative Work + §7.1
+#   general grant + Feist 1991 numeric-facts doctrine. DEC filed 2026-05-25;
+#   GM-approved 2026-05-26. See
+#   audit/proposed_GM/hacks_law_license_decision_2026-05-25.md.
 
 # Type I — each row is one independent terminal river outlet; no within-basin
 # repeated measures, no nested cluster structure. Both reference formulas are
@@ -25,8 +27,8 @@ target:
   unit: km
   description: Main-stem river length, measured along the longest channel from the drainage divide to the basin outlet.
   range:
-  - 11.6
-  - 6846.5
+    train: [11.6, 471.9]
+    test:  [41.4, 6846.5]
 
 inputs:
 - name: upland_area_skm
@@ -35,40 +37,40 @@ inputs:
   dtype: float
   description: Total upstream drainage area at the outlet (sum of all contributing catchment areas above this point).
   range:
-  - 100.0
-  - 5912761.0
+    train: [100, 1087.7]
+    test:  [1088.3, 5.91276e+06]
 - name: catch_area_skm
   symbol: A_c
   unit: km^2
   dtype: float
   description: Local catchment area of the outlet segment alone (the terminal reach's own contributing area, excluding upstream tributaries).
   range:
-  - 0.03
-  - 309.27
+    train: [0.06, 251.93]
+    test:  [0.03, 309.27]
 - name: segment_length_km
   symbol: L_s
   unit: km
   dtype: float
   description: Length of the terminal outlet segment alone (the final reach), which is typically a small fraction of the cumulative main-stem length.
   range:
-  - 0.1
-  - 340.08
+    train: [0.1, 340.08]
+    test:  [0.1, 89.01]
 - name: endorheic
   symbol: e
   unit: ""
   dtype: float
   description: Indicator (0 or 1) for whether the basin terminates at an inland sink (1) or reaches the ocean (0).
   range:
-  - 0.0
-  - 1.0
+    train: [0, 1]
+    test:  [0, 1]
 - name: strahler_order
   symbol: omega
   unit: ""
   dtype: float
   description: Strahler stream order of the outlet segment, an integer measure of branching depth in the channel network (1 = unbranched headwater; higher values denote progressively higher-order trunk channels).
   range:
-  - 1.0
-  - 10.0
+    train: [1, 5]
+    test:  [2, 10]
 
 # Type I two-file layout
 data_files:
@@ -109,8 +111,8 @@ references:
   n_other_constants: 0
   n_local_params: 0
   measured:
-    rmse: 237.8967
-    r2: 0.5048
+    log_mae: 0.098846
+    r2: 0.504788
 - id: dodds_rothman_2000
   label: Dodds & Rothman 2000 — maximal-basin Hack's law L = 3.0 A^0.5 (Annu. Rev. Earth Planet. Sci. 28, Eq. 36 / Fig. 8 caption, p. 23), fit to 37 of the world's largest basins (c_tilde=3.0, h_tilde=0.5).
   formula_file: formulas/dodds_rothman_2000.py
@@ -119,8 +121,8 @@ references:
   n_other_constants: 0
   n_local_params: 0
   measured:
-    rmse: 93.1996
-    r2: 0.9240
+    log_mae: 0.105328
+    r2: 0.923995
 
 caps:
   max_law_constants:       2

hea_hardness_lattice_distortion_couzinie__HV/metadata.yaml

@@ -17,7 +17,9 @@ target:
     high-entropy alloy, compiled from peer-reviewed experimental measurements
     in a curated mechanical-property database.
     Values are means over repeated measurements of nominally identical compositions.
-  range: [166.0, 857.0]
+  range:
+    train: [166, 850]
+    test:  [365, 857]
 
 inputs:
   - name: delta
@@ -27,7 +29,9 @@ inputs:
       Yang-Zhang atomic-size mismatch parameter, defined as
       100 × sqrt(Σ c_i (1 − r_i/r̄)²) with Goldschmidt CN-12 radii.
       Quantifies the degree of lattice distortion in the solid solution.
-    range: [2.31, 10.78]
+    range:
+      train: [2.48584, 10.7786]
+      test:  [2.31046, 10.6363]
   - name: VEC
     symbol: "VEC"
     unit: ""
@@ -35,7 +39,9 @@ inputs:
       Valence electron concentration (s + p + d electrons per atom),
       calculated as Σ c_i VEC_i. Controls the preferred crystal structure
       (FCC vs BCC) and electronic bonding strength.
-    range: [4.20, 6.50]
+    range:
+      train: [4.2, 6.5]
+      test:  [4.27273, 6.5]
   - name: dHmix
     symbol: "ΔH_mix"
     unit: "kJ/mol"
@@ -43,7 +49,9 @@ inputs:
       Miedema regular-solution enthalpy of mixing, Σ_{i<j} 4 ΔH_{AB} c_i c_j,
       computed from a Miedema-type pairwise mixing-enthalpy table. Negative values
       indicate attractive cross-element interactions that promote solid-solution stability.
-    range: [-39.46, 3.50]
+    range:
+      train: [-33.3907, 2.72]
+      test:  [-39.4502, 3.5]
 
 data_files:
   train: data/train.csv   # ETM-RHEA alloys (VEC<=6.5); representative within-domain subsample

her_hor_pt_acid_kucernak__current_density/formulas/reference_metrics.json

@@ -11,20 +11,23 @@
         "i0_ref": 62.16814883075403
       },
       "local_fittable": [],
-      "other_constants": {
-        "F": 96485.0,
-        "R": 8.314,
-        "T_ref": 298.15
-      },
-      "paper_ref": "summary_formula_dataset_kucernak_2016.md",
-      "pooled": {
+      "metrics": {
+        "log_mae": 0.4424651600346051,
         "mae": 52.243007750325226,
+        "mape": 1.2268313311893262,
+        "mdae": 40.793342265879815,
         "mse": 4475.39931157304,
         "n_finite": 334,
         "r2": 0.8286063504751774,
         "rmse": 66.89842532954748,
         "smape": 0.7630759654572248
-      }
+      },
+      "other_constants": {
+        "F": 96485.0,
+        "R": 8.314,
+        "T_ref": 298.15
+      },
+      "paper_ref": "summary_formula_dataset_kucernak_2016.md"
     },
     "kucernak_2016_hv": {
       "equation_loc": "Kucernak & Zalitis (2016) Eq 15 (theta, PDF p. 7), Eq 19 (j_HV dimensionless-scaled form, PDF p. 9), Eq 71 (temperature scaling, PDF p. 27), Eq 69 (Henry's law a_H2, PDF p. 27). Five parameters fit simultaneously to all 13 datasets (PDF p. 27, Sec. 4.1).",
@@ -39,20 +42,23 @@
         "k2_eq_ref": 0.009599691056554147
       },
       "local_fittable": [],
-      "other_constants": {
-        "F": 96485.0,
-        "R": 8.314,
-        "T_ref": 298.15
-      },
-      "paper_ref": "summary_formula_dataset_kucernak_2016.md",
-      "pooled": {
+      "metrics": {
+        "log_mae": 0.029248247153562446,
         "mae": 7.209053007960816,
+        "mape": 0.07489488403841417,
+        "mdae": 2.9716883836356836,
         "mse": 206.28811375951847,
         "n_finite": 334,
         "r2": 0.9920998171985668,
         "rmse": 14.362733505830931,
         "smape": 0.07452131927301364
-      }
+      },
+      "other_constants": {
+        "F": 96485.0,
+        "R": 8.314,
+        "T_ref": 298.15
+      },
+      "paper_ref": "summary_formula_dataset_kucernak_2016.md"
     }
   },
   "derived_caps": {

her_hor_pt_acid_kucernak__current_density/metadata.yaml

@@ -23,7 +23,9 @@ target:
   symbol: j
   unit: "A cm-2"
   description: Specific current density per real platinum surface area (from upd-H charge), measured by potentiostat chronoamperometry on a floating electrode. Positive = anodic (HOR direction). Negative = cathodic (HER direction).
-  range: [-486.0, 388.0]
+  range:
+    train: [-474.936, 387.63]
+    test:  [-485.646, 379.28]
 
 inputs:
   - name: eta
@@ -34,17 +36,23 @@ inputs:
       measurement is performed vs RHE and the equilibrium potential is
       E_eq = 0 V vs RHE, the overpotential equals the measured potential
       directly: eta = E_vs_RHE. Positive = anodic (HOR favoured).
-    range: [-0.05, 0.05]
+    range:
+      train: [-0.04992, 0.04977]
+      test:  [-0.04999, 0.04995]
   - name: T
     symbol: T
     unit: K
-    description: Temperature of the water-jacketed electrochemical cell. The T series spans 278–333 K in 4.0 M HClO4.
-    range: [278.0, 333.0]
+    description: Temperature of the water-jacketed electrochemical cell; the temperature series is measured in 4.0 M HClO4.
+    range:
+      train: [278, 333]
+      test:  [278, 333]
   - name: P_H2
     symbol: "P(H2)"
     unit: bar
-    description: H2 partial pressure, corrected for water vapour pressure at elevated temperatures. The p(H2) series spans 0.029–0.126 bar at T=298 K in 0.5 M HClO4; the T series is at p(H2)~0.8–0.99 bar.
-    range: [0.029, 0.99]
+    description: H2 partial pressure, corrected for water vapour pressure at elevated temperatures. The p(H2) series is measured at T=298 K in 0.5 M HClO4; the temperature series is at near-ambient p(H2).
+    range:
+      train: [0.029, 0.99]
+      test:  [0.029, 0.99]
   - name: c_acid
     symbol: "c(HClO4)"
     unit: "mol dm-3"
@@ -149,22 +157,3 @@ caps:
 
 # === Scoring ===
 metric: rmse
-
-# === Auditor notes ===
-auditor_notes: >-
-  Rebuilt Type II -> Type I (2026-06-02 audit). The paper fits all 13 polarisation
-  curves simultaneously with one shared 5-parameter set (PDF p. 27, Sec. 4.1; Table 4;
-  Fig. 8 "common set of parameters"), so the intrinsic kinetic constants are
-  cross-condition invariants and belong in LAW_CONSTANTS, not as per-cluster LOCAL.
-  The prior Type II buried them as LOCAL (Direction-A) and additionally placed the
-  CODATA F/R in LAW (Direction-B); both fixed. A global least-squares fit on train
-  recovers shape constants (beta=0.35, Ea=17.8 kJ/mol, K_std=3.30) matching the
-  paper's Table 4 (1/3, 18 kJ/mol, 3.09); the released floating-electrode currents
-  sit ~10^3x above the paper's example parameter scale (the paper's k2_eq~5.1e-6 does
-  not reproduce this data), so k2_eq_ref/i0_ref are fit to the released data. HV
-  (best) test R2=0.992 (rmse 14.4); BV the deliberately-inadequate limiting case
-  (single i0, no explicit P_H2 dependence) test R2=0.829 (rmse 66.9) — both beat the
-  naive constant predictor (test rmse 161.8). G_std~0.29 puts the data in the small-G
-  regime where k2_eq_ref and G_std are partially degenerate (only their product sets
-  the scale); HV's advantage over BV comes from the K_std denominator term (K_std=3.30
-  != 1), exactly the paper's point that the Butler-Volmer reduction does not hold on Pt.

island_species_richness_terceira__S_terceira/metadata.yaml

@@ -16,34 +16,46 @@ target:
   symbol: S
   unit: ""
   description: Arthropod species richness per 1-ha pitfall-trap site on Terceira Island, Azores.
-  range: [0.0, 11.0]
+  range:
+    train: [0, 10]
+    test:  [0, 11]
 
 inputs:
   - name: d
     symbol: D
     unit: ""
     description: Anthropogenic habitat-modification index — degree of disturbance along the gradient from undisturbed native forest to intensively managed pasture.
-    range: [14.70, 76.92]
+    range:
+      train: [15.0637, 75.5817]
+      test:  [14.7005, 76.9155]
   - name: h
     symbol: H
     unit: m
     description: Altitude above sea level.
-    range: [13.95, 934.86]
+    range:
+      train: [13.951, 930]
+      test:  [56.3745, 934.858]
   - name: p
     symbol: P
     unit: mm/yr
     description: Mean annual precipitation.
-    range: [1015.40, 3474.00]
+    range:
+      train: [1015.4, 3474]
+      test:  [1015.9, 3461.5]
   - name: t
     symbol: T
     unit: degC
     description: Mean annual temperature.
-    range: [12.00, 18.35]
+    range:
+      train: [12.05, 18.35]
+      test:  [12, 18.15]
   - name: sl
     symbol: Sl
     unit: degrees
     description: Terrain slope.
-    range: [0.40, 21.91]
+    range:
+      train: [1.08584, 21.9015]
+      test:  [0.400647, 17.0603]
 
 # Dataset -- Type I two-file layout
 data_files:
@@ -53,6 +65,9 @@ data_files:
 n_train: 34
 n_test:  18
 
+# Candidate prior constants — the `priors` prompt slot.
+priors: []   # no candidate constants offered for this task
+
 # Reference-baseline bank
 references:
   - id: cardoso_2020

lake_thermocline_depth_pilla__z_t/metadata.yaml

@@ -18,7 +18,9 @@ target:
     interpolated summer measurements per lake and year, aggregated by median.
     Values are on a 0.5 m grid due to the uniform interpolation spacing of the
     source dataset.
-  range: [1.5, 42.0]
+  range:
+    train: [1.5, 42]
+    test:  [1.5, 35.5]
 
 inputs:
   - name: A_s_km2
@@ -27,14 +29,18 @@ inputs:
     description: >-
       Lake surface area in square kilometres — the primary morphometric driver of
       wind-fetch-induced thermocline deepening.
-    range: [0.1, 895.5]
+    range:
+      train: [0.1, 895.47]
+      test:  [0.1, 895.47]
   - name: Z_max_m
     symbol: "Z_max"
     unit: m
     description: >-
       Lake maximum depth in metres — an auxiliary covariate providing a physical
       upper bound on thermocline depth.
-    range: [5.5, 594.0]
+    range:
+      train: [5.5, 594]
+      test:  [5.5, 594]
 
 data_files:
   train: data/train.csv    # rank-based stratified sample (4 of every 5 rows)

life_expectancy_preston__e0_vs_gdppc/metadata.yaml

@@ -9,20 +9,6 @@ license: "CC0-1.0 (per-country life expectancy and income figures are uncopyrigh
 context: >-
   This task draws on national-level demographic and development economics data from a mid-twentieth-century cross-section of countries.
 
-problem_statement: >-
-  Predict the sex-averaged life expectancy at birth (e0) of a country
-  in the 1960s cross-section as a function of national income per head
-  expressed in constant 1963 US dollars. Each row is one country
-  (population >= 2 million). The task asks for a single closed-form
-  concave curve over the income range, fitted on a representative
-  income-stratified training sample and evaluated on held-out countries
-  spanning the same range — the within-cross-section curve recovery
-  Preston (1975) performed independently for each decade. (Earlier
-  versions used a cross-decade extrapolation split; that conflated the
-  curve with Preston's documented ~13-year cross-decade level shift and
-  was replaced by this single-cross-section design in the 2026-05-29
-  audit re-split — see prep_data.py.)
-
 target:
   name: e0_vs_gdppc
   symbol: e0
@@ -33,8 +19,8 @@ target:
     (1975) Appendix A, Tables A-1/A-2/A-3, hand-transcribed from the
     original UN Demographic Yearbook figures.
   range:
-    - 31.6
-    - 74.23
+    train: [31.6, 74.225]
+    test:  [36.75, 72.805]
 
 inputs:
   - name: gdppc
@@ -46,8 +32,8 @@ inputs:
       Source: Preston (1975) Appendix A (national income figures from
       UN Statistical Yearbook and Kuznets 1956).
     range:
-      - 40.0
-      - 2995.0
+      train: [44, 2995]
+      test:  [40, 1987]
 
 data_files:
   train: data/train.csv   # 1960s cross-section, representative income-stratified split: 38 rows
@@ -162,7 +148,6 @@ caps:
   max_init_size_per_param: 1
   fit_timeout_seconds: null
 
-reference_baseline_id: preston_1975_logistic
 # === Validity rubric ===
 # Frozen list — the scoring-time judge (eval_consistency.py) scores a
 # submission by how many of THESE criteria it satisfies:

liion_capacity_fade_nasa_pcoe__capacity_Ah/metadata.yaml

@@ -14,14 +14,18 @@ target:
   symbol: Q
   unit: Ah
   description: Discharge capacity per cycle (time-integrated absolute current over a full discharge event).
-  range: [1.1538, 2.0353]
+  range:
+    train: [1.33701, 1.88223]
+    test:  [1.15382, 2.03534]
 
 inputs:
   - name: cycle_index
     symbol: k
     unit: ""
     description: Per-cell discharge cycle counter, starting from 1; an integer index of how many full charge–discharge cycles the cell has completed.
-    range: [1, 168]
+    range:
+      train: [32, 128]
+      test:  [1, 168]
 
 # Dataset — Type I, two-file flat layout.  Source: NASA PCoE Battery
 # Dataset (Saha & Goebel 2007), 4 18650 NMC Li-ion cells (B0005,
@@ -76,7 +80,7 @@ references:
     n_law_constants: 4         # A, B, C, D
     n_other_constants: 0
     n_local_params: 0
-    measured: {rmse: 0.1207, r2: 0.7912}
+    measured: {rmse: 0.121531, r2: 0.788240}
 
 # === v2 anti-dump caps ===
 caps:

liquid_dispersion_glycols__refractive_index/formulas/reference_metrics.json

@@ -11,16 +11,19 @@
         "C_C": -0.000306
       },
       "local_fittable": [],
-      "other_constants": {},
-      "paper_ref": "summary_formula_cauchy_1836.md",
-      "pooled": {
+      "metrics": {
+        "log_mae": 0.001049107740481774,
         "mae": 0.003444374882347796,
+        "mape": 0.0024168419637406228,
+        "mdae": 0.0034000000000000696,
         "mse": 1.591393434745938e-05,
         "n_finite": 170,
         "r2": 0.1943706392108332,
         "rmse": 0.003989227287014289,
         "smape": 0.002415657424008889
-      }
+      },
+      "other_constants": {},
+      "paper_ref": "summary_formula_cauchy_1836.md"
     },
     "jakubczyk_2023": {
       "equation_loc": "Jakubczyk (2023) Eq. 3 (Sellmeier, PDF p. 3/8), Eq. 2 (thermal coefficient, PDF p. 3), Eq. 1 (linear T shift, PDF p. 3). LAW_CONSTANTS from Table 2 (Sellmeier) and Table 3 (thermal), ethylene glycol (EG/MEG) row, PDF p. 8.",
@@ -38,18 +41,21 @@
         "C_UV": 0.01332
       },
       "local_fittable": [],
-      "other_constants": {
-        "T_ref": 20.0
-      },
-      "paper_ref": "summary_formula_dataset_jakubczyk_2023.md",
-      "pooled": {
+      "metrics": {
+        "log_mae": 4.080840156334191e-05,
         "mae": 0.0001337898164209085,
+        "mape": 9.396491614168303e-05,
+        "mdae": 9.398540797711519e-05,
         "mse": 3.033315021543341e-08,
         "n_finite": 170,
         "r2": 0.9984644101272994,
         "rmse": 0.00017416414733071044,
         "smape": 9.39648168545186e-05
-      }
+      },
+      "other_constants": {
+        "T_ref": 20.0
+      },
+      "paper_ref": "summary_formula_dataset_jakubczyk_2023.md"
     }
   },
   "derived_caps": {

liquid_dispersion_glycols__refractive_index/metadata.yaml

@@ -16,19 +16,25 @@ target:
   symbol: "n"
   unit: ""
   description: Refractive index of ethylene glycol at the given wavelength and temperature.
-  range: [1.4160, 1.4509]
+  range:
+    train: [1.4233, 1.4509]
+    test:  [1.416, 1.4353]
 
 inputs:
   - name: lambda_um
     symbol: "lambda"
     unit: um
     description: Vacuum wavelength of light in micrometres.
-    range: [0.394, 1.071]
+    range:
+      train: [0.394, 0.635]
+      test:  [0.651, 1.071]
   - name: temperature_C
     symbol: "T"
     unit: degC
     description: Liquid temperature.
-    range: [1.0, 45.0]
+    range:
+      train: [1, 45]
+      test:  [1, 45]
 
 # Dataset - Type I two-file layout
 data_files:
@@ -49,7 +55,7 @@ references:
     n_law_constants: 8
     n_other_constants: 1
     n_local_params: 0
-    measured: {rmse: 0.0002, r2: 0.9985}
+    measured: {rmse: 0.000174164, r2: 0.998464}
   - id: cauchy_1836
     label: "Cauchy 1836 - three-term isothermal dispersion power series (Eq. 55)"
     formula_file: formulas/cauchy_1836.py
@@ -57,7 +63,7 @@ references:
     n_law_constants: 3
     n_other_constants: 0
     n_local_params: 0
-    measured: {rmse: 0.0040, r2: 0.1944}
+    measured: {rmse: 0.00398923, r2: 0.194371}
 
 # Candidate prior constants - the `priors` prompt slot.
 # Audit 2026-05-29 (逐条审核, effort=max): declared the Eq.1 reference temperature (20 degC) in
@@ -100,4 +106,4 @@ validity_rubrics:
   - "predicted refractive index remains greater than 1 across the transparent liquid (lambda, T) domain"
   - "predicted n is finite and smooth with no resonance poles or discontinuities inside the transparent wavelength-temperature domain [0.394, 1.071] um x [1, 45] C"
 
-best_baseline: 0.0002
+best_baseline: 0.000174164

lpbf_meltpool_rosenthal_inconel__depth/metadata.yaml

@@ -15,24 +15,32 @@ target:
   symbol: d
   unit: um
   description: Maximum depth of the laser melt-pool cross-section, measured by optical microscopy perpendicular to the scan direction.
-  range: [28.0, 232.0]
+  range:
+    train: [101.292, 231.633]
+    test:  [29, 170.41]
 
 inputs:
   - name: laser_power_W
     symbol: P
     unit: W
     description: Incident laser power delivered to the workpiece surface.
-    range: [137.0, 326.0]
+    range:
+      train: [245, 325]
+      test:  [137.9, 195]
   - name: scan_velocity_mm_per_s
     symbol: V
     unit: mm/s
     description: Laser beam scan velocity along the track direction.
-    range: [399.0, 1201.0]
+    range:
+      train: [800, 1200]
+      test:  [400, 1200]
   - name: beam_diameter_um
     symbol: D4sigma
     unit: um
     description: Laser beam D4-sigma diameter at the workpiece surface (four times the 1/e^2 radius).
-    range: [48.0, 171.0]
+    range:
+      train: [49, 82]
+      test:  [100, 170]
 
 # Dataset — Type I two-file layout
 # OOD split axis: beam_diameter_um / material system.
@@ -68,30 +76,6 @@ references:
       rmse: 35.525694
       r2: 0.268688
 
-auditor_notes:
-  neg_r2_disclosure:
-    severity: ACCEPTABLE_PER_E6_MIXED_BANK
-    baseline: king_2014_normalized_enthalpy
-    pooled_r2: -0.852250
-    pooled_rmse: 56.538059
-    best_positive_r2_baseline: promoppatum_2017_rosenthal
-    best_positive_r2: 0.268688
-    summary: >-
-      King 2014 normalized-enthalpy baseline yields pooled R² = -0.852 / RMSE = 56.54 µm
-      on the OOD test split. The mixed bank satisfies §9.6 via promoppatum_2017_rosenthal
-      (R²=0.269). Anchor-scale check: RMSE (56.54 µm) < target span (141.4 µm) — §9.18
-      anchor-scale violation NOT triggered. §9.14 (feature-OOD legitimate) applies.
-    citation: "AUDIT_CHECKLIST.md v0.5 §E6 (mixed-bank partial-negative-R² annotation) + data_spec §9.14 (cross-feature OOD legitimately produces negative pooled R²)"
-    root_cause: >-
-      King 2014 coefficients (316L stainless steel, Table 3, PDF p. 36) and the graphically
-      derived d/sigma linear fit (PDF §5.3, Fig. 8) were calibrated on 316L specimens.
-      The benchmark test set contains Inconel 718 specimens with substantially different
-      thermal properties (higher Tm, different absorptivity). The normalized-enthalpy
-      scaling does not account for alloy-specific melt-pool geometry; systematic under- or
-      over-prediction on IN718 is expected and constitutes a cross-material OOD probe.
-    # FM E6 annotation added by subagent L 2026-05-26 (wave-3 re-audit).
-    # Policy: §9.14 only (RMSE 56.54 < span 141.4, bank beats naive via promoppatum).
-
 # FM C12 fix (2026-05-26): removed 4 leaking priors that exactly matched
 # LAW_CONSTANTS in the shipped formulas:
 #   absorptivity_rosenthal_in718=0.4 → promoppatum_2017_rosenthal.py LAW_CONSTANTS["absorptivity"]

magnet_3c90_core_loss__log_P_vol/metadata.yaml

@@ -14,19 +14,25 @@ target:
   symbol: log10(P_vol)
   unit: log10(kW/m^3)
   description: Base-10 logarithm of the volumetric core loss in kW per cubic metre, computed as the area of the measured B-H hysteresis loop times the excitation frequency, normalised by the core volume.
-  range: [-0.2559, 3.3165]
+  range:
+    train: [-0.255893, 3.31654]
+    test:  [-0.096799, 3.31479]
 
 inputs:
   - name: f_Hz
     symbol: f
     unit: Hz
     description: Excitation frequency of the sinusoidal/triangular/trapezoidal flux waveform.
-    range: [49950.0, 446690.0]
+    range:
+      train: [49950, 446690]
+      test:  [49950, 446430]
   - name: B_peak_T
     symbol: B
     unit: T
     description: Peak flux density max|B(t)| of the excitation waveform.
-    range: [0.0107, 0.2815]
+    range:
+      train: [0.01073, 0.281542]
+      test:  [0.011096, 0.278339]
   - name: T_C
     symbol: T
     unit: degC

mars_crater_size_frequency__N_cum/metadata.yaml

@@ -14,24 +14,32 @@ target:
   symbol: N(>=D)
   unit: km^-2
   description: Cumulative areal density of Mars craters with diameter at or above the bin's lower edge (craters per km^2 of Mars surface).  Spans ~5 orders of magnitude across the diameter range; the declared metric is log_mae (log10-space mean absolute error), the natural error scale for a multi-decade positive target.
-  range: [1.38e-08, 2.66e-03]
+  range:
+    train: [1.26383e-06, 0.00265434]
+    test:  [1.38123e-08, 9.94485e-07]
 
 inputs:
   - name: D_lower_km
     symbol: D_lo
     unit: km
     description: Lower diameter edge of the logarithmic bin (km).  PRIMARY input — the cumulative SFD closed form N(>=D) is a function of this threshold diameter alone; all three reference baselines use this input only.
-    range: [1.0, 664.0]
+    range:
+      train: [1, 117.377]
+      test:  [128, 663.982]
   - name: D_upper_km
     symbol: D_up
     unit: km
     description: Upper diameter edge of the bin (= D_lower * 2^0.125).  DISTRACTOR — perfectly collinear with D_lower (their ratio is the fixed bin factor 2^0.125 ≈ 1.0905), so it carries no information beyond D_lower.  No reference baseline uses it.
-    range: [1.09, 724.08]
+    range:
+      train: [1.09051, 128]
+      test:  [139.585, 724.077]
   - name: D_geomean_km
     symbol: D_gm
     unit: km
     description: Geometric-mean diameter of the bin, sqrt(D_lower * D_upper).  DISTRACTOR — also collinear with D_lower (= D_lower * 2^0.0625).  The natural representative point of a log-spaced bin, but it adds no information for the cumulative target.  No reference baseline uses it.
-    range: [1.04, 689.97]
+    range:
+      train: [1.04427, 122.573]
+      test:  [133.667, 693.379]
 
 # NOTE: the aggregated source CSV carries three host-only columns that
 # are NOT released as inputs because each would shortcut the target:

materials_project_elastic_birch_murnaghan__K_VRH/formulas/reference_metrics.json

@@ -9,6 +9,17 @@
         "anderson_nafe_exponent": -1.3333333333333333
       },
       "local_fittable": [],
+      "metrics": {
+        "log_mae": 0.15478535717563074,
+        "mae": 42.370561483376456,
+        "mape": 0.41397223836735497,
+        "mdae": 29.17506039866212,
+        "mse": 3552.625510755104,
+        "n_finite": 177,
+        "r2": 0.33474272314988407,
+        "rmse": 59.603905163630884,
+        "smape": 0.34189276891894305
+      },
       "other_constants": {
         "a_by_csid_1": 3.677239,
         "a_by_csid_2": 3.671468,
@@ -19,15 +30,7 @@
         "a_by_csid_7": 3.675828,
         "a_global": 3.677239
       },
-      "paper_ref": "summary_birch_murnaghan_kvrh.md",
-      "pooled": {
-        "mae": 42.370561483376456,
-        "mse": 3552.625510755104,
-        "n_finite": 177,
-        "r2": 0.33474272314988407,
-        "rmse": 59.603905163630884,
-        "smape": 0.34189276891894305
-      }
+      "paper_ref": "summary_birch_murnaghan_kvrh.md"
     },
     "cohen_1985": {
       "equation_loc": "Cohen 1985 Phys. Rev. B 32:7988 \u2014 K = (1972 - 220*lambda)/d^3.5; this baseline uses lambda=0 (group-IV) per-paper prefactor 1972 with d = V_atomic^(1/3); documented in summary_birch_murnaghan_kvrh.md \u00a73",
@@ -39,16 +42,19 @@
         "cohen_prefactor": 1972.0
       },
       "local_fittable": [],
-      "other_constants": {},
-      "paper_ref": "summary_birch_murnaghan_kvrh.md",
-      "pooled": {
+      "metrics": {
+        "log_mae": 0.24815698656250143,
         "mae": 65.34665235943552,
+        "mape": 0.4638685618593104,
+        "mdae": 54.18721988785896,
         "mse": 7173.539695403433,
         "n_finite": 177,
         "r2": -0.3433021489861354,
         "rmse": 84.6967513863633,
         "smape": 0.542397306618077
-      }
+      },
+      "other_constants": {},
+      "paper_ref": "summary_birch_murnaghan_kvrh.md"
     }
   },
   "derived_caps": {