Upload 2 files

app.py

@@ -18,7 +18,6 @@ import numpy as np
 import gradio as gr
 
 HERE = os.path.dirname(os.path.abspath(__file__))
-DATA_PATH = os.path.join(HERE, "data", "app_players.csv")
 
 # Hudl IQ palette (sampled from hudl.com/products/football-iq)
 ORANGE = "#FF6300"   # Hudl brand orange
@@ -33,14 +32,30 @@ TIER_ORDER = ["Elite", "Star", "Above Avg", "Starter", "Rotation", "Below"]
 TIER_COLORS = {"Elite": ORANGE, "Star": "#FF8A3D", "Above Avg": BLUE,
                "Starter": "#9DAAB8", "Rotation": "#7A8A99", "Below": "#4E5D6C"}
 
-if os.path.exists(DATA_PATH):
-    df = pd.read_csv(DATA_PATH)
-else:
-    df = pd.DataFrame()
+# ── season data: one flat file per season in data/ ─────────────────────────
+SEASON_FILES = {"2022": "app_players.csv", "2023": "app_players_2023.csv"}
+DFS = {}
+for _season, _fname in SEASON_FILES.items():
+    _p = os.path.join(HERE, "data", _fname)
+    if os.path.exists(_p):
+        _d = pd.read_csv(_p)
+        # decode-file placeholder weights (e.g., 0 lbs) display as n/a instead
+        _d.loc[_d["weight_pounds"].notna() & (_d["weight_pounds"] < 100), "weight_pounds"] = np.nan
+        DFS[_season] = _d
+SEASONS = list(DFS.keys()) or ["2022"]
+DEFAULT_SEASON = SEASONS[-1]  # latest season present
 
 
-def _teams():
-    return ["All"] + sorted(df["team_short_name"].dropna().unique().tolist()) if len(df) else ["All"]
+def _df(season):
+    return DFS.get(season, pd.DataFrame())
+
+
+def _teams(d):
+    return ["All"] + sorted(d["team_short_name"].dropna().unique().tolist()) if len(d) else ["All"]
+
+
+def _players(d):
+    return sorted(d["player_name"].dropna().unique().tolist()) if len(d) else []
 
 
 def top_route(p):
@@ -54,6 +69,13 @@ def top_route(p):
     return name.replace("_", "/")
 
 
+def _route_choices(d):
+    if not len(d):
+        return ["Any"]
+    routes = sorted({top_route(p) for _, p in d.iterrows()} - {"n/a"})
+    return ["Any"] + routes
+
+
 # ── card + carousel rendering ──────────────────────────────────────────────
 
 def _subtitle(p):
@@ -121,11 +143,12 @@ def _spotlight_cards(q, pos_label):
     return cards
 
 
-def spotlight_row():
+def spotlight_row(season):
     """Leaderboard hero row: WR and TE callouts side by side, six across."""
-    if not len(df):
-        return "<p>data/app_players.csv not found</p>"
-    q = df[df["n"] >= 15]
+    d = _df(season)
+    if not len(d):
+        return f"<p>data/{SEASON_FILES.get(season, '?')} not found</p>"
+    q = d[d["n"] >= 15]
     cards = []
     for pos in ("WR", "TE"):
         cards.extend(_spotlight_cards(q[q["position"] == pos], pos))
@@ -140,10 +163,11 @@ LB_LABELS = ["Player", "Team", "Pos", "Targets", "Separation Rating", "Tier",
              "Percentile", "Sched Rank", "RSOE (shrunk)"]
 
 
-def leaderboard(search, team, pos, tier, min_targets):
-    if not len(df):
-        return pd.DataFrame({"error": ["data/app_players.csv not found — run the prep notebook and add it"]})
-    out = df.copy()
+def leaderboard(season, search, team, pos, tier, min_targets):
+    d = _df(season)
+    if not len(d):
+        return pd.DataFrame({"error": [f"data/{SEASON_FILES.get(season, '?')} not found — run the prep notebook and add it"]})
+    out = d.copy()
     if search:
         out = out[out["player_name"].str.contains(search, case=False, na=False)]
     if team != "All":
@@ -163,19 +187,13 @@ DRAFT_CLASSES = ["SR", "RS-SR", "Graduate"]
 ELIG_CHOICES = ["Any", "Returning (transfer candidates)", "Seniors & grads (draft)"]
 
 
-def _route_choices():
-    if not len(df):
-        return ["Any"]
-    routes = sorted({top_route(p) for _, p in df.iterrows()} - {"n/a"})
-    return ["Any"] + routes
-
-
-def recruit_search(pos, elig, tier, min_rating, min_targets, route, min_contested,
+def recruit_search(season, pos, elig, tier, min_rating, min_targets, route, min_contested,
                    min_rz, min_height, min_weight):
     """The recruiting finder: filter the board by what a staff is shopping for."""
-    if not len(df):
-        return "data/app_players.csv not found", "", pd.DataFrame()
-    s = df.copy()
+    d = _df(season)
+    if not len(d):
+        return f"data/{SEASON_FILES.get(season, '?')} not found", "", pd.DataFrame()
+    s = d.copy()
     s["top_route"] = s.apply(top_route, axis=1)
     if pos != "All":
         s = s[s["position"] == pos]
@@ -210,18 +228,19 @@ def recruit_search(pos, elig, tier, min_rating, min_targets, route, min_conteste
     view = s[["player_name", "team_short_name", "position", "class_type", "sep_rating",
               "tier", "shrunk", "n", "catches", "top_route", "contested_win_rate",
               "rz_rating"]].copy()
-    view.columns = ["Player", "2022 Team", "Pos", "Class", "Separation Rating", "Tier",
+    view.columns = ["Player", f"{season} Team", "Pos", "Class", "Separation Rating", "Tier",
                     "RSOE", "Targets", "Catches", "Most frequent route",
                     "Contested win %", "RZ rating"]
     return summary, _carousel(cards), view.reset_index(drop=True)
 
 
-def player_card(name):
-    if not len(df) or not name:
+def player_card(season, name):
+    d = _df(season)
+    if not len(d) or not name:
         return "Select a player."
-    m = df[df["player_name"] == name]
+    m = d[d["player_name"] == name]
     if not len(m):
-        return f"No player named {name}."
+        return f"No player named {name} in {season}."
     p = m.iloc[0]
     rz = f"{int(p['rz_rating'])} ({int(p['rz_n'])} RZ targets)" if pd.notna(p.get("rz_rating")) else "not enough red zone targets"
     cw = f"{p['contested_win_rate']:.0f}% on {int(p['contested_att'])} contested" if pd.notna(p.get("contested_win_rate")) else "no contested targets"
@@ -230,7 +249,7 @@ def player_card(name):
         hw = f"{int(p['height_inches'] // 12)}'{int(p['height_inches'] % 12)}\" · {int(p['weight_pounds'])} lbs" if pd.notna(p.get("weight_pounds")) else ""
     cls = p.get("class_type") if pd.notna(p.get("class_type")) else "n/a"
     card = f"""
-## {p['player_name']} — {p['team_short_name']} ({p['position']})
+## {p['player_name']} — {p['team_short_name']} ({p['position']}) · {season}
 # Separation Rating: {int(p['sep_rating'])}  ·  {p['tier']}
 | | |
 |---|---|
@@ -251,12 +270,16 @@ def player_card(name):
     return card
 
 
-# ── Start Here explainers ──────────────────────────────────────────────────
+# ── Start Here explainers (pinned to 2022, the season the numbers cite) ────
+
+EXPLAIN_SEASON = "2022" if "2022" in DFS else DEFAULT_SEASON
+
 
 def _named(name):
-    if not len(df):
+    d = _df(EXPLAIN_SEASON)
+    if not len(d):
         return None
-    m = df[df["player_name"] == name]
+    m = d[d["player_name"] == name]
     return m.iloc[0] if len(m) else None
 
 
@@ -266,7 +289,8 @@ def _fmt(p):
 
 def build_explainers():
     """(question, answer) pairs for the Start Here accordions, using live numbers."""
-    top = df[df["n"] >= 15].sort_values(["sep_rating", "shrunk"], ascending=False).iloc[0] if len(df) else None
+    d = _df(EXPLAIN_SEASON)
+    top = d[d["n"] >= 15].sort_values(["sep_rating", "shrunk"], ascending=False).iloc[0] if len(d) else None
     hyatt, mayer, tinsley = _named("Jalin Hyatt"), _named("Michael Mayer"), _named("Mitchell Tinsley")
 
     q1 = ("What is RSOE?",
@@ -290,10 +314,10 @@ def build_explainers():
           "target count next to the number.\n\n"
           + (f"Face validity check: {_fmt(hyatt)}, the 2022 Biletnikoff winner, rates "
              f"**{int(hyatt['sep_rating'])}**, and his coverage schedule ranked "
-             f"{int(hyatt['sched_rank'])} of {len(df)} in difficulty (1 = toughest), so the number is "
+             f"{int(hyatt['sched_rank'])} of {len(d)} in difficulty (1 = toughest), so the number is "
              f"not a soft-schedule artifact." if hyatt is not None else ""))
 
-    med_n = int(df["n"].median()) if len(df) else 0
+    med_n = int(d["n"].median()) if len(d) else 0
     q2b = ("How do we know the rating is skill and not luck?",
            "This was the central test in developing the metric: prove that average RSOE measures the "
            "player, not his bounces.\n\n"
@@ -314,6 +338,9 @@ def build_explainers():
            "gets pulled hard; a 110-target player barely moves. So the rating you see already carries the "
            "appropriate skepticism, and a 90 on 25 targets means the raw number was spectacular enough to "
            "survive aggressive shrinkage.\n\n"
+           "**Season two confirmed it.** With 2023 in hand, a player's 2022 rating predicts his 2023 "
+           "result at roughly 85% of the theoretical maximum his sample size allows. The signal is real "
+           "and it persists.\n\n"
            + (f"**The honest limit**: the median player on this board has {med_n} targets, well short of "
               f"100, so most single-season ratings lean heavily on shrinkage. That is the reality of "
               f"one-season data, not a flaw of the metric, and it is the case for adding more seasons. "
@@ -364,9 +391,10 @@ def build_explainers():
 
 HOW_TO = """
 ### How to use the app
+- **Season**: the selector at the top switches every tab between the 2022 and 2023 boards.
 - **Leaderboard**: the spotlight row shows WR and TE standouts. Cut the board with search, team, position, tier, and min targets; the table sorts by Separation Rating.
 - **Player card**: below the leaderboard, pick a name for the full profile: rating, red zone rating, contested win rate, schedule rank, season line, and RSOE split by route and by coverage.
-- **Recruiting Finder**: set the profile your room needs (position, eligibility, tier, primary route, rating floor, contested, red zone, size) and hit **Find players**. The team column shows 2022 teams.
+- **Recruiting Finder**: set the profile your room needs (position, eligibility, tier, primary route, rating floor, contested, red zone, size) and hit **Find players**.
 - **Mind the sample**: ratings are half signal at about 100 targets. Shrinkage protects against small-sample flukes, but always glance at the target count on the card.
 - **Compare across conferences freely**: every rating is opponent-adjusted, so a Sun Belt 90 and an SEC 90 mean the same thing. That is the recruiting edge.
 """
@@ -386,10 +414,13 @@ METHODOLOGY = """
 - Predicted separation tracks reality across every zone of the field (red zone compression: actual 3.79 to 2.60,
   expected 3.81 to 2.66, bias under 0.06 yards everywhere).
 - Rankings survive model swaps (0.85-0.98 correlation across four model generations) and screen exclusion.
+- The same frozen model spec generalizes across seasons (2022 R-squared 0.283, 2023 R-squared 0.295), and a
+  player's 2022 rating predicts his 2023 result at roughly 85% of the ceiling his sample size allows.
 - Face validity: the 2022 Biletnikoff winner ranks top-10; famous contested-catch specialists correctly rank
   low on separation and are flagged for the contested-catch lane instead.
 - The same framework produces defender ratings (separation allowed on-field): the 2022 leaders include
-  Kool-Aid McKinstry, Brian Branch, and the Alabama secondary.
+  Kool-Aid McKinstry, Brian Branch, and the Alabama secondary; the 2023 leaders are headlined by the
+  Florida State secondary.
 
 ### Honest limits
 - One-play separation is noisy: a player's rating is half signal at roughly 100 targets, which is why shrinkage
@@ -428,9 +459,11 @@ THEME = gr.themes.Base(
 )
 
 with gr.Blocks(title="Separation Rating", theme=THEME, css=CSS, js=FORCE_DARK) as demo:
-    gr.Markdown("# **Separation** Rating\n**Who really gets open?** Scheme- and opponent-adjusted receiver "
-                "evaluation for D1 football, built on Hudl IQ data. UVA MSBA x Hudl.",
-                elem_id="hero-title")
+    with gr.Row():
+        gr.Markdown("# **Separation** Rating\n**Who really gets open?** Scheme- and opponent-adjusted receiver "
+                    "evaluation for D1 football, built on Hudl IQ data. UVA MSBA x Hudl.",
+                    elem_id="hero-title", scale=4)
+        season = gr.Radio(SEASONS, value=DEFAULT_SEASON, label="Season", scale=1)
 
     with gr.Tab("Start Here"):
         gr.Markdown("## The metric, in plain football terms\n"
@@ -441,21 +474,21 @@ with gr.Blocks(title="Separation Rating", theme=THEME, css=CSS, js=FORCE_DARK) a
         gr.Markdown(HOW_TO)
 
     with gr.Tab("Leaderboard"):
-        gr.HTML(spotlight_row())
+        spot = gr.HTML(spotlight_row(DEFAULT_SEASON))
         with gr.Row():
             search = gr.Textbox(label="Search player", placeholder="e.g., Hyatt")
-            team = gr.Dropdown(_teams(), value="All", label="Team")
+            team = gr.Dropdown(_teams(_df(DEFAULT_SEASON)), value="All", label="Team")
             pos = gr.Dropdown(["All", "WR", "TE"], value="All", label="Position")
             tierdd = gr.Dropdown(["All"] + TIER_ORDER, value="All", label="Tier")
             min_t = gr.Slider(10, 80, value=15, step=5, label="Min targets")
-        table = gr.Dataframe(value=leaderboard("", "All", "All", "All", 15), interactive=False, wrap=True)
+        table = gr.Dataframe(value=leaderboard(DEFAULT_SEASON, "", "All", "All", "All", 15),
+                             interactive=False, wrap=True)
         for c in (search, team, pos, tierdd, min_t):
-            c.change(leaderboard, [search, team, pos, tierdd, min_t], table)
+            c.change(leaderboard, [season, search, team, pos, tierdd, min_t], table)
         gr.Markdown("---\n## Player card")
-        psel = gr.Dropdown(sorted(df["player_name"].dropna().unique().tolist()) if len(df) else [],
-                           label="Select player")
+        psel = gr.Dropdown(_players(_df(DEFAULT_SEASON)), label="Select player")
         card = gr.Markdown()
-        psel.change(player_card, psel, card)
+        psel.change(player_card, [season, psel], card)
 
     with gr.Tab("Recruiting Finder"):
         gr.Markdown("## Build your target board\n"
@@ -466,7 +499,7 @@ with gr.Blocks(title="Separation Rating", theme=THEME, css=CSS, js=FORCE_DARK) a
             r_pos = gr.Dropdown(["All", "WR", "TE"], value="All", label="Position")
             r_elig = gr.Dropdown(ELIG_CHOICES, value="Any", label="Eligibility")
             r_tier = gr.Dropdown(["All"] + TIER_ORDER, value="All", label="Tier")
-            r_route = gr.Dropdown(_route_choices(), value="Any", label="Primary route")
+            r_route = gr.Dropdown(_route_choices(_df(DEFAULT_SEASON)), value="Any", label="Primary route")
             r_rating = gr.Slider(40, 99, value=80, step=1, label="Min Separation Rating")
             r_targets = gr.Slider(10, 80, value=15, step=5, label="Min targets")
         with gr.Row():
@@ -476,20 +509,34 @@ with gr.Blocks(title="Separation Rating", theme=THEME, css=CSS, js=FORCE_DARK) a
             r_height = gr.Slider(66, 80, value=66, step=1, label="Min height, inches (66 = off)")
             r_weight = gr.Slider(150, 260, value=150, step=5, label="Min weight, lbs (150 = off)")
         r_btn = gr.Button("Find players", variant="primary")
-        _s0, _c0, _t0 = recruit_search("All", "Any", "All", 80, 15, "Any", 0, 40, 66, 150)
+        _s0, _c0, _t0 = recruit_search(DEFAULT_SEASON, "All", "Any", "All", 80, 15, "Any", 0, 40, 66, 150)
         r_summary = gr.Markdown(_s0)
         r_cards = gr.HTML(_c0)
         r_table = gr.Dataframe(value=_t0, interactive=False, wrap=True)
-        r_inputs = [r_pos, r_elig, r_tier, r_rating, r_targets, r_route,
+        r_inputs = [season, r_pos, r_elig, r_tier, r_rating, r_targets, r_route,
                     r_contested, r_rz, r_height, r_weight]
         r_btn.click(recruit_search, r_inputs, [r_summary, r_cards, r_table])
         gr.Markdown("*Low separation + high contested win rate is its own lane (the big-body "
                     "archetype), not a red flag: set the contested slider and drop the rating "
-                    "floor to shop that profile. Eligibility is class-based from 2022 rosters; "
+                    "floor to shop that profile. Eligibility is class-based from the season's rosters; "
                     "early draft declares and actual transfer decisions are not tracked.*")
 
     with gr.Accordion("Methodology & validation", open=False):
         gr.Markdown(METHODOLOGY)
 
+    def _on_season(s):
+        d = _df(s)
+        s0, c0, t0 = recruit_search(s, "All", "Any", "All", 80, 15, "Any", 0, 40, 66, 150)
+        return (spotlight_row(s),
+                gr.update(choices=_teams(d), value="All"),
+                leaderboard(s, "", "All", "All", "All", 15),
+                gr.update(choices=_players(d), value=None),
+                "Select a player.",
+                gr.update(choices=_route_choices(d), value="Any"),
+                s0, c0, t0)
+
+    season.change(_on_season, season,
+                  [spot, team, table, psel, card, r_route, r_summary, r_cards, r_table])
+
 if __name__ == "__main__":
     demo.launch()