Upload 8 files

README.md

@@ -1,15 +1,53 @@
 ---
-title: Reachfirst
-emoji: 💻
-colorFrom: indigo
-colorTo: green
+title: Crosswind Workforce Partnership ReachFirst
+emoji: 📬
+colorFrom: green
+colorTo: gray
 sdk: gradio
 sdk_version: 6.31.0
-python_version: '3.13'
+python_version: "3.12"
 app_file: app.py
 pinned: false
 license: cc-by-4.0
-short_description: income eligibility prediction
+models:
+  - drdavidprivacy/reachfirst-model
+datasets:
+  - drdavidprivacy/reachfirst-data
+  - drdavidprivacy/practicum-case-packs
+short_description: ReachFirst v1, the signed model and its dossier
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# ReachFirst v1: Crosswind Workforce Partnership
+
+**Governance analyst:** [YOUR NAME]
+
+**Program:** AI Governance Practicum, Track 1, cohort 1
+
+This page is the public home of a governed system. The model is ReachFirst v1, the version the Chief AI Officer signed in Week 6. The app has three pages: the model (its fixed weights and three example records), monitoring (upload a monthly file and read it against the training file and the tripwires), and the dossier (the six documents below).
+
+## The signed version
+
+- File: `reachfirst_v1.pkl`, 4,300 bytes, SHA-256 `120234d9f9a111028a2d605d2779a9d8e151ff8d6f85df4ec0497f651595b953`
+- Operating point: the top 20 percent of the ranked file
+- Libraries: scikit-learn 1.9.1, fairlearn 0.15.0
+- Data: the remediated working file in the program's dataset repository, loaded by link, never copied here: https://huggingface.co/datasets/drdavidprivacy/practicum-case-packs/resolve/main/crosswind_pilot_extract_remediated.csv (MD5 `ca9ed6d4b9dfd59f3f586b09da05b31f`)
+
+- Model repository (the model card and the weights): https://huggingface.co/drdavidprivacy/reachfirst-model. Dataset repository (the datasheet and the remediation log): https://huggingface.co/datasets/drdavidprivacy/reachfirst-data
+
+## The dossier
+
+The six deliverables, in course order. Each opens in the browser.
+- [Use Case Assessment (Week 1)](docs/01_Use-Case-Assessment.pdf)
+- [EDA Observation Log (Week 2)](docs/02_EDA-Observation-Log.pdf)
+- [Dataset Datasheet (Week 3)](docs/03_Dataset-Datasheet.pdf)
+- [Model Selection Rationale (Week 4)](docs/04_Model-Selection-Rationale.pdf)
+- [Model Card (Week 6)](docs/05_Model-Card.pdf)
+- [Go/No-Go Decision Record (Week 7)](docs/06_Go-No-Go-Decision-Record.pdf)
+
+## What this page is not
+
+A score says where a household's record ranks on one list. It is not a decision about the household. The decisions, and the conditions on them, are in the Go/No-Go Decision Record above.
+
+## Source
+
+Built for the AI Governance Practicum (AI Governance Collective). The organization, the people, and the program are fictional. The data is a public dataset, remediated and documented in the dossier's datasheet.

app.py

@@ -0,0 +1,348 @@
+"""The public page of a governed system: the signed model, its monitoring page, and the dossier.
+Built for the AI Governance Practicum (AI Governance Collective). One app.py serves all three organizations; space_config.json says which."""
+import json, os, pickle, io, warnings
+import numpy as np, pandas as pd
+import gradio as gr
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+warnings.filterwarnings("ignore")
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+
+# Hugging Face's free tier runs Gradio Spaces on ZeroGPU, which refuses to start an app that declares no GPU function.
+# This app needs no GPU: the declaration below satisfies the check and is never called. On a plain CPU Space it is skipped.
+try:
+    import spaces
+    @spaces.GPU
+    def _gpu_not_used():
+        return "This app runs on the CPU."
+except Exception:
+    pass
+CFG = json.load(open(os.path.join(HERE, "space_config.json")))
+REF = json.load(open(os.path.join(HERE, "reference.json")))
+MODEL = pickle.load(open(os.path.join(HERE, CFG["model_file"]), "rb"))
+RECAL = pickle.load(open(os.path.join(HERE, CFG["recal_file"]), "rb")) if CFG.get("recal_file") else None
+REASONS = pd.read_csv(os.path.join(HERE, CFG["reasons_file"])) if CFG.get("reasons_file") else None
+EXAMPLES = pd.read_csv(os.path.join(HERE, "examples.csv"), keep_default_na=False)
+INPUTS, KEPT, SHARE = CFG["inputs"], CFG["kept"], CFG["share"]
+CAL_FLOOR = 25   # a calibration band is read only when it expects at least this many outcomes; a smaller band says nothing either way
+SPACE_ID = os.environ.get("SPACE_ID", "")
+WEEK6 = CFG["week6"]
+
+# ----------------------------------------------------------------------------- scoring
+def score_frame(X):
+    """The production score. Where the Week 6 memo ratified fitted steps (the recalibration and the odds correction), they run here."""
+    X = X[INPUTS]
+    if RECAL is not None:
+        pf = RECAL.predict_proba(X)[:, 1]; f = CFG["odds_factor"]
+        return pf * f / (pf * f + 1 - pf)
+    return MODEL.predict_proba(X)[:, 1]
+
+def contributions(X):
+    """Each input's fixed weight times the record's value on the scaled, remover-adjusted inputs. The record's own reasons."""
+    Z = X[INPUTS]
+    for name, step in MODEL.steps[:-1]:
+        Z = step.transform(Z)
+    coef = MODEL.named_steps["model"].coef_[0]
+    return pd.DataFrame(np.asarray(Z) * coef, columns=KEPT, index=X.index)
+
+def allowed_reasons(con):
+    """Drop the columns a reason may never cite: a medication the record marks No (Alderwyn, Week 6 memo, ruling 5)."""
+    drugs = CFG.get("drug_fields", [])
+    if not drugs: return con
+    keep = [c for c in con.index if not any(c == d + "_No" for d in drugs)]
+    return con[keep]
+
+def allowed_reasons_frame(frame):
+    drugs = CFG.get("drug_fields", [])
+    return frame[[c for c in frame.columns if not any(c == d + "_No" for d in drugs)]] if drugs else frame
+
+def readable(col):
+    if REASONS is not None:
+        r = REASONS.loc[(REASONS.input == col) | ((REASONS.input + "_" + REASONS.value.astype(str)) == col), "reason"]
+        if len(r): return r.iloc[0]
+    for c in CFG["categorical"]:
+        if col.startswith(c + "_"): return f"{c} is {col[len(c) + 1:]}"
+    return col
+
+def coerce(df):
+    """Read a file the way the working file is read: numeric inputs as numbers, categorical inputs as text."""
+    df = df.copy()
+    for c in CFG["numeric"]:
+        if c in df: df[c] = pd.to_numeric(df[c], errors="coerce")
+    for c in CFG["categorical"]:
+        if c in df: df[c] = df[c].astype(str)
+    return df
+
+def groups_of(frame):
+    g = {}
+    for gname, d in CFG["groups"].items():
+        col = d["col"]
+        if col not in frame: continue
+        if "bands" in d:
+            v = pd.to_numeric(frame[col], errors="coerce"); lab = pd.Series(["unrecorded"] * len(frame), index=frame.index, dtype=object)
+            for lo, hi, name in d["bands"]: lab[(v >= lo) & (v <= hi)] = name
+            g[gname] = lab
+        else:
+            s = frame[col].astype(str); g[gname] = s.map(d["map"]).fillna(s) if d.get("map") else s
+    return pd.DataFrame(g, index=frame.index)
+
+def pct(x, d=1): return f"{x * 100:.{d}f} percent"
+
+# ----------------------------------------------------------------------------- the signed version (the collapsed panel at the top)
+def model_header():
+    a = CFG["artifacts"][CFG["model_file"]]
+    L = [f"**Organization:** {CFG['org']}. **Signed by:** {CFG['caio']} ({CFG['memo_ref']}). **Kept by:** {CFG['lead']}.",
+         f"**File:** `{CFG['model_file']}`, {a['bytes']:,} bytes, SHA-256 `{a['sha256']}`."]
+    if RECAL is not None:
+        r = CFG["artifacts"][CFG["recal_file"]]
+        L.append(f"**Fitted steps the memo ratified:** the isotonic recalibration (`{CFG['recal_file']}`, SHA-256 `{r['sha256']}`), then the odds correction to the {int(CFG['calibration']['book_rate']*100)} percent book rate (factor {CFG['odds_factor']}). Every score on this page has both applied.")
+    L.append(f"**What it is:** a logistic regression with {len(KEPT)} fitted columns from {len(INPUTS)} input fields. The weights are fixed. The same record gets the same score every time.")
+    L.append(f"**Operating point:** the top {int(SHARE*100)} percent of the ranked file gets {CFG['action']}. On the Week 6 test set ({WEEK6['test_rows']:,} {CFG['units']}) that was {WEEK6['flagged']:,} {CFG['units']}, of which {WEEK6['caught']:,} were {CFG['positive_label']}. Test-set ROC-AUC {WEEK6['roc_auc']:.3f}.")
+    L.append(f"**Libraries:** scikit-learn {CFG['versions']['sklearn']}, fairlearn {CFG['versions']['fairlearn']}. The model does not load on another version of scikit-learn, by design.")
+    return "\n\n".join(L)
+
+def weights_table(n=25):
+    return pd.DataFrame([dict(column=c, weight=round(w, 4), meaning=readable(c)) for c, w in CFG["weights"][:n]])
+
+# ----------------------------------------------------------------------------- tab 1: one record
+CAT = [c for c in INPUTS if c in CFG["categorical"]]
+NUM = [c for c in INPUTS if c in CFG["numeric"]]
+FORM = CAT + NUM   # the order of the form's controls
+
+def _typical():
+    d = {c: REF["numeric"][c]["median"] for c in NUM}
+    for c in CAT:
+        r = REF["categorical"][c]; d[c] = max(r, key=r.get)
+    return d
+
+def _case_row(i):
+    row = EXAMPLES.iloc[i]
+    return {c: (float(row[c]) if c in NUM else str(row[c])) for c in FORM}
+
+PRESETS = {f"Typical {CFG['unit']} (training-file medians and most common values)": _typical()}
+for i in range(len(EXAMPLES)):
+    PRESETS[f"{EXAMPLES['case'].iloc[i].capitalize()} (Week 6 case, rank {EXAMPLES['week6_rank'].iloc[i]} of {WEEK6['test_rows']:,})"] = _case_row(i)
+FIRST = list(PRESETS)[0]
+
+def apply_preset(name):
+    d = PRESETS[name]
+    return [d[c] for c in FORM]
+
+def field_label(c):
+    """The input's name, with the meaning of its codes where the program's group definitions give one (SEX: 1 male, 2 female)."""
+    for d in CFG["groups"].values():
+        if d["col"] == c and d.get("map"):
+            return c + " (" + ", ".join(f"{k} {v}" for k, v in d["map"].items()) + ")"
+    return c
+
+def choices_for(c):
+    ch = sorted(REF["categorical"][c], key=str)
+    for v in EXAMPLES[c].astype(str).unique():
+        if v not in ch: ch.append(v)
+    return ch
+
+def score_one(*vals):
+    row = {c: v for c, v in zip(FORM, vals)}
+    X = coerce(pd.DataFrame([row]))
+    p = float(score_frame(X)[0]); cut = REF["cut_score"]; inside = p >= cut
+    verdict = (f"## Score: **{p:.1%}**\n\n"
+               f"### {'🔴 Above the cut' if inside else '🟢 Below the cut'}  (Week 6 cut = {cut:.1%}, the score of the last {CFG['unit']} inside the top {int(SHARE*100)} percent)\n\n"
+               f"At the operating point the {CFG['memo_ref']} set, this {CFG['unit']} {'would' if inside else 'would not'} be {CFG['flagged_word']}: {CFG['action']}. "
+               f"The score says where the record ranks. It is not a decision about the {CFG['person']}.")
+    con = allowed_reasons(contributions(X).iloc[0])
+    up = con[con > 0].sort_values(ascending=False).head(6); down = con[con < 0].sort_values().head(4)
+    items = list(up.items()) + list(down.items())
+    labs = [readable(c) for c, _ in items][::-1]; v = [float(x) for _, x in items][::-1]
+    fig, ax = plt.subplots(figsize=(11, 0.45 * len(v) + 1.8))
+    ax.barh(range(len(v)), v, color=["#d62728" if x > 0 else "#2ca02c" for x in v])
+    ax.set_yticks(range(len(v))); ax.set_yticklabels(labs, fontsize=9); ax.axvline(0, color="black", lw=0.8)
+    ax.set_xlabel(f"Toward {CFG['positive_label']} (red) or away from it (green): the fixed weight times this record's value", fontsize=9)
+    ax.set_title(f"Why {CFG['system']} scored this {CFG['unit']} {p:.1%}: the record's own reasons", fontsize=11, loc="left")
+    fig.tight_layout()
+    return verdict, fig
+
+# ----------------------------------------------------------------------------- tab 2: batch upload
+def score_batch(file, top_n):
+    if file is None: return None, f"Upload a CSV of {CFG['units']} to score (the model's input columns, one row per {CFG['unit']}).", None
+    try: df = coerce(pd.read_csv(file.name if hasattr(file, "name") else file, low_memory=False, keep_default_na=False, na_values=[]))
+    except Exception as e: return None, f"Could not read the file: {e}", None
+    missing = [c for c in INPUTS if c not in df.columns]
+    if missing: return None, "The file is missing input columns: " + ", ".join(missing) + ". Nothing was scored.", None
+    s = score_frame(df); order = np.argsort(-s, kind="stable"); k = int(round(SHARE * len(df)))
+    idc = CFG.get("id_col"); ids = df[idc].values if idc and idc in df.columns else np.arange(1, len(df) + 1)
+    con = allowed_reasons_frame(contributions(df)); top = [readable(con.columns[i]) for i in np.argmax(con.values, axis=1)]
+    out = pd.DataFrame({(idc if idc and idc in df.columns else "row"): ids, "score": np.round(s, 4), "ranked_in": np.where(np.arange(len(df)) >= 0, "", ""), "top_reason": top})
+    out = out.iloc[order].reset_index(drop=True); out.insert(0, "rank", np.arange(1, len(df) + 1)); out["ranked_in"] = np.where(out["rank"] <= k, "yes", "no")
+    path = os.path.join("/tmp", "ranked_list.csv"); out.to_csv(path, index=False)
+    msg = (f"Scored **{len(df):,}** {CFG['units']}. The top {int(SHARE*100)} percent, the operating point, is **{k:,}** {CFG['units']}. "
+           f"Showing the top **{int(top_n)}** by score (the team's capacity). The download carries every record with its score, rank, whether it is ranked in, and its own top reason.")
+    return out.head(int(top_n)), msg, path
+
+# ----------------------------------------------------------------------------- tab 2: monitoring
+def psi(ref_shares, new_shares):
+    keys = set(ref_shares) | set(new_shares); eps = 1e-4; v = 0.0
+    for k in keys:
+        a, b = max(ref_shares.get(k, 0.0), eps), max(new_shares.get(k, 0.0), eps); v += (b - a) * np.log(b / a)
+    return float(v)
+
+def drift_table(df):
+    rows = []
+    for c in INPUTS:
+        if c not in df.columns: continue
+        watch = c in CFG["watch_fields"]
+        if c in REF["numeric"]:
+            r = REF["numeric"][c]; v = pd.to_numeric(df[c], errors="coerce"); edges = r["edges"]
+            if len(edges) < 3: continue
+            def shares(x):
+                b = np.clip(np.searchsorted(edges, x, side="right") - 1, 0, len(edges) - 2); return {str(i): float((b == i).mean()) for i in range(len(edges) - 1)}
+            ref_sh = r["shares"]; new_sh = shares(v.dropna().values)
+            idx = psi(ref_sh, new_sh)
+            rows.append(dict(input=c, kind="number", training_file=f"mean {r['mean']:,.2f}, median {r['median']:,.2f}", this_file=f"mean {v.mean():,.2f}, median {v.median():,.2f}", missing=int(v.isna().sum()), stability_index=round(idx, 3), reading=reading(idx), watched="yes" if watch else ""))
+        elif c in REF["categorical"]:
+            r = REF["categorical"][c]; vc = df[c].astype(str).value_counts(normalize=True); new_sh = {str(k): float(x) for k, x in vc.items()}
+            idx = psi(r, new_sh); top_r = max(r, key=r.get); top_n = max(new_sh, key=new_sh.get) if new_sh else ""
+            new_levels = [k for k in new_sh if k not in r]
+            rows.append(dict(input=c, kind="category", training_file=f"most common {top_r} ({pct(r[top_r])})", this_file=f"most common {top_n} ({pct(new_sh.get(top_n, 0))})" + (f"; new levels: {', '.join(new_levels)}" if new_levels else ""),
+                             missing=int((df[c].astype(str) == "nan").sum()), stability_index=round(idx, 3), reading=reading(idx), watched="yes" if watch else ""))
+    t = pd.DataFrame(rows)
+    return t.sort_values(["watched", "stability_index"], ascending=[False, False]).reset_index(drop=True) if len(t) else t
+
+def reading(idx):
+    return "stable" if idx < 0.10 else ("watch" if idx < 0.25 else "SHIFTED")
+
+def group_table(df, s, flag, y):
+    """Each group's rate at the cut beside the Week 6 test-set reading and the tripwire. With outcomes: the ruled check (false alarm or miss). Without: the share ranked in."""
+    G = groups_of(df); rows = []; check = CFG["group_check"]; line = CFG["group_tripwire"]
+    if y is not None:
+        pos, neg = y == 1, y == 0
+        port = (((flag == 1) & neg).sum() / max(neg.sum(), 1)) if check == "false_alarm" else (((flag == 0) & pos).sum() / max(pos.sum(), 1))
+    for gname in G.columns:
+        d = CFG["groups"][gname]; order = [b[2] for b in d["bands"]] if "bands" in d else d.get("order", [])
+        for lvl in sorted(G[gname].unique(), key=lambda s: (order.index(s) if s in order else len(order), str(s))):
+            m = (G[gname] == lvl).values; n = int(m.sum())
+            wk6 = REF["group_rates"].get(gname, {}).get(str(lvl), {})
+            covered = n >= CFG["min_group"]   # the memo's group floor, applied to the month's file
+            if y is not None:
+                rate = (((flag == 1) & neg & m).sum() / max((neg & m).sum(), 1)) if check == "false_alarm" else (((flag == 0) & pos & m).sum() / max((pos & m).sum(), 1))
+                base = wk6.get("false_alarm" if check == "false_alarm" else "miss", np.nan)
+                over = rate > port * (1 + line)
+                result = (f"not judged (under the group floor of {CFG['min_group']:,})" if not covered else ("TRIP" if over else "inside the line"))
+                rows.append({"group": gname, "level": lvl, "records this file": n, ("false alarm rate" if check == "false_alarm" else "miss rate"): pct(rate), "whole file": pct(port), "line": pct(port * (1 + line)), "Week 6 test set": pct(base) if base == base else "", "watched": "yes" if str(lvl) in CFG["watch_groups"] else "", "result": result})
+            else:
+                rate = float(flag[m].mean()); base = wk6.get("flagged_share", np.nan)
+                rows.append({"group": gname, "level": lvl, "records this file": n, "share ranked in": pct(rate), "whole file": pct(SHARE), "Week 6 test set": pct(base) if base == base else "", "watched": "yes" if str(lvl) in CFG["watch_groups"] else "", "result": "no outcomes yet: share ranked in only" if covered else f"not judged (under the group floor of {CFG['min_group']:,})"})
+    return pd.DataFrame(rows)
+
+def calibration_table(s, y):
+    cal = CFG["calibration"]; rows = []
+    for lo, hi, name in cal["bands"]:
+        m = (s >= lo) & (s < hi)
+        if m.sum() == 0: rows.append(dict(band=name, records=0, predicted="", observed="", gap_share_of_observed="", result="empty band")); continue
+        mp, ob = float(s[m].mean()), float(y[m].mean()); gap = abs(mp - ob); expected = mp * m.sum()
+        if expected < CAL_FLOOR: res = f"not read (fewer than {CAL_FLOOR} expected outcomes in the band)"
+        else: res = "no outcomes in band" if ob == 0 else ("pass" if gap / ob <= cal["tripwire"] else "TRIP")
+        rows.append(dict(band=name, records=int(m.sum()), predicted=pct(mp, 2), observed=pct(ob, 2), gap_share_of_observed=(pct(gap / ob, 1) if ob else ""), result=res))
+    return pd.DataFrame(rows)
+
+def monitor(file):
+    empty = (None, None, None)
+    if file is None: return "Upload a monthly file: one row per " + CFG["unit"] + ", the model's input columns, the measurement fields, and the outcome column once it is known.", *empty
+    try: df = coerce(pd.read_csv(file.name if hasattr(file, "name") else file, low_memory=False, keep_default_na=False, na_values=[]))
+    except Exception as e: return f"Could not read the file: {e}", *empty
+    missing = [c for c in INPUTS if c not in df.columns]
+    if missing: return "The file is missing input columns: " + ", ".join(missing) + ". The job would hold this file out of the cycle (the input contract).", *empty
+    s = score_frame(df); order = np.argsort(-s, kind="stable"); k = int(round(SHARE * len(df))); flag = np.zeros(len(df), int); flag[order[:k]] = 1
+    y = (df[CFG["outcome"]].astype(str) == str(CFG["positive"])).astype(int).values if CFG["outcome"] in df.columns else None
+    above_cut = float((s >= REF["cut_score"]).mean())
+    L = [f"**File:** {len(df):,} {CFG['units']}, {df.shape[1]} columns. Outcome column `{CFG['outcome']}`: " + ("present, so the ruled checks run." if y is not None else "absent, so only the input and score readings run. Upload again when the outcomes are known."),
+         f"**Scores:** mean {s.mean():.4f} (training-file positive rate {pct(REF['positive_rate'])}). {pct(above_cut)} of this file scores above the Week 6 cut of {REF['cut_score']:.4f}, against {pct(SHARE, 0)} by construction on the test set. The top {int(SHARE*100)} percent of this file is {k:,} {CFG['units']}.",
+         f"**Inputs:** the stability index compares each input's distribution with the training file's ({REF['rows']:,} rows). Under 0.10 is stable, 0.10 to 0.25 is watch, over 0.25 is shifted. These lines are the usual working lines, provisional until the Week 8 monitoring plan sets the program's own. Watched fields: {CFG['watch_fields_label']}."]
+    if y is not None:
+        pos = y == 1
+        L.append(f"**At the cut:** {int((flag & pos).sum()):,} of the {int(pos.sum()):,} {CFG['positive_label']} {CFG['units']} are ranked in ({pct((flag & pos).sum() / max(pos.sum(), 1))}); the whole-file {'false alarm' if CFG['group_check'] == 'false_alarm' else 'miss'} rate is the line every group is read against, plus {int(CFG['group_tripwire']*100)} percent ({CFG['memo_ref']}).")
+    cal = calibration_table(s, y) if (CFG.get("calibration") and y is not None) else None
+    if cal is not None:
+        worst = [r for r in cal.result if r == "TRIP"]; read = [r for r in cal.result if r in ("TRIP", "pass")]
+        L.append(f"**Calibration:** predicted against observed by score band, line {int(CFG['calibration']['tripwire']*100)} percent of observed. Bands that expect fewer than {CAL_FLOOR} outcomes are not read. " + (f"{len(read)} band(s) read, every one inside the line." if not worst else f"{len(worst)} of {len(read)} band(s) read are over the line: the calibration condition (ruling 1) is not met for this file."))
+    return "\n\n".join(L), drift_table(df), group_table(df, s, flag, y), cal
+
+# ----------------------------------------------------------------------------- tab 4: the dossier
+def dossier_markdown():
+    L = [f"## The dossier: {CFG['system']} at {CFG['org']}", "", "The six deliverables of the practicum, in course order. Each opens in the browser.", ""]
+    for f, label in CFG["docs"]:
+        p = os.path.join(HERE, "docs", f)
+        if SPACE_ID: url = f"https://huggingface.co/spaces/{SPACE_ID}/resolve/main/docs/{f}"
+        else: url = f"docs/{f}"
+        L.append(f"- [{label}]({url})" + ("" if os.path.exists(p) else "  (not uploaded yet)"))
+    L += ["", f"**Monitoring targets from the {CFG['memo_ref']}:**"] + [f"- {m}" for m in CFG["monitoring_lines"]]
+    L += ["", f"**Data:** [{CFG['data_file']}]({CFG['data_url']}) in the program's dataset repository (MD5 `{CFG['md5']}`). The data is not copied into this page."]
+    return "\n".join(L)
+
+# ----------------------------------------------------------------------------- the page
+INTRO = f"""
+# {CFG['system']} {CFG['version']}: {CFG['org']}
+**Governance analyst: [YOUR NAME]** · a teaching deployment for the *AI Governance Practicum*, Track 1
+
+---
+
+{CFG['system']} estimates, for each {CFG['unit']}, the likelihood of {CFG['positive_label']}, ranks the file, and sends the top {int(SHARE*100)} percent to a person for {CFG['action']}. This page runs the signed version ({CFG['memo_ref']}). It is for **education**, not real decisions. A score says where a {CFG['person']}'s record ranks on one list. It is not a decision about the {CFG['person']}.
+"""
+
+with gr.Blocks(title=f"{CFG['org']} {CFG['system']}") as demo:
+    gr.Markdown(INTRO)
+    with gr.Accordion("About the signed version: the file, its hash, the fitted steps, the weights", open=False):
+        gr.Markdown(model_header())
+        gr.Markdown("**The fixed weights.** The largest by size, with what each column means. A positive weight pushes the score up.")
+        gr.Dataframe(weights_table(), interactive=False, wrap=True)
+
+    with gr.Tab(f"1 · Single {CFG['unit']}"):
+        gr.Markdown(f"Score one {CFG['unit']}, then read the **reasons**. Load a Week 6 case to fill the form, or edit any field and score again.")
+        with gr.Row():
+            preset = gr.Dropdown(list(PRESETS), label="Load an example", value=FIRST)
+        d0 = PRESETS[FIRST]
+        widgets = {}
+        with gr.Accordion("Categorical inputs", open=True):
+            for i in range(0, len(CAT), 4):
+                with gr.Row():
+                    for c in CAT[i:i + 4]:
+                        widgets[c] = gr.Dropdown(choices_for(c), value=d0[c], label=field_label(c), allow_custom_value=True)
+        with gr.Accordion("Numeric inputs", open=(len(CAT) == 0)):
+            for i in range(0, len(NUM), 4):
+                with gr.Row():
+                    for c in NUM[i:i + 4]:
+                        widgets[c] = gr.Number(value=d0[c], label=c)
+        ordered = [widgets[c] for c in FORM]
+        score_btn = gr.Button(f"Score this {CFG['unit']}", variant="primary")
+        out_md = gr.Markdown(); out_plot = gr.Plot()
+        preset.change(apply_preset, preset, ordered)
+        score_btn.click(score_one, ordered, [out_md, out_plot])
+
+    with gr.Tab("2 · Batch upload"):
+        gr.Markdown(f"Upload a CSV of {CFG['units']} (the same columns as the working file). {CFG['system']} scores every row, ranks the file, and marks the top {int(SHARE*100)} percent, the operating point the {CFG['memo_ref']} set. Top-N is how many of the ranked list the team can work. `sample_month.csv` in this repository is a file to try.")
+        with gr.Row():
+            up = gr.File(label=f"{CFG['units'].capitalize()} CSV", file_types=[".csv"])
+            with gr.Column():
+                topn = gr.Slider(5, 100, value=20, step=5, label="Top-N by capacity (how many the team can work)")
+        batch_btn = gr.Button("Score & rank", variant="primary")
+        batch_msg = gr.Markdown(); batch_tbl = gr.Dataframe(interactive=False, wrap=True); batch_dl = gr.File(label="Download the full ranked list")
+        batch_btn.click(score_batch, [up, topn], [batch_tbl, batch_msg, batch_dl])
+
+    with gr.Tab("3 · Monitoring"):
+        gr.Markdown(f"**One month at a time.** Upload the month's file. The page reads it against the training file and the tripwires the {CFG['memo_ref']} set: the inputs' distributions, the scores, the groups at the cut, " + ("and the calibration table." if CFG.get("calibration") else "and the ruled group check.") + f" `sample_month.csv` in this repository is a stand-in drawn from the Week 6 test set, so the page can be tried before a monthly file exists.")
+        mf = gr.File(label="Monthly file (CSV)", file_types=[".csv"])
+        mon_btn = gr.Button("Read this month", variant="primary")
+        m_msg = gr.Markdown()
+        m_drift = gr.Dataframe(interactive=False, label="Inputs against the training file", wrap=True)
+        m_groups = gr.Dataframe(interactive=False, label="Groups at the cut, against the whole file and the line", wrap=True)
+        m_cal = gr.Dataframe(interactive=False, label="Calibration by score band", visible=bool(CFG.get("calibration")))
+        mon_btn.click(monitor, mf, [m_msg, m_drift, m_groups, m_cal])
+
+    with gr.Tab("4 · The dossier"):
+        gr.Markdown(dossier_markdown())
+
+    gr.Markdown(f"---\n*{CFG['system']} is a fictional model built on a public dataset for the AI Governance Practicum (AI Governance Collective). The organization, the people, and the program are fictional. Data: [{CFG['data_file']}]({CFG['data_url']}).*")
+
+if __name__ == "__main__":
+    demo.launch()

examples.csv

@@ -0,0 +1,4 @@
+case,case_note,age,workclass,education,marital-status,occupation,capital-gain,capital-loss,hours-per-week,relationship_model,capital_gain_topcoded,income,sex,race,native_country_report,week6_rank
+caught,"caught (ranked in, at or below the line)",20,Private,Some-college,Never-married,Priv-house-serv,0,0,45,Own-child,0,<=50K,Female,White,United-States,975
+missed,"missed (ranked out, at or below the line; the first record below the cut)",28,Private,Some-college,Never-married,Adm-clerical,0,0,40,Other-relative,0,<=50K,Male,Black,United-States,1953
+false alarm,"false alarm (ranked in, above the line; the highest-scored above-the-line record)",33,Private,7th-8th,Never-married,Farming-fishing,0,0,40,Not-in-family,0,>50K,Male,White,Mexico,417

reachfirst_v1.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:120234d9f9a111028a2d605d2779a9d8e151ff8d6f85df4ec0497f651595b953
+size 4300

reference.json

@@ -0,0 +1,442 @@
+{
+ "rows": 39032,
+ "test_rows": 9758,
+ "cut_score": 0.9864791972217691,
+ "share": 0.2,
+ "positive_rate": 0.7605810616929699,
+ "numeric": {
+  "age": {
+   "mean": 38.57839721254356,
+   "std": 13.709782584667774,
+   "median": 37.0,
+   "edges": [
+    17.0,
+    22.0,
+    26.0,
+    30.0,
+    33.0,
+    37.0,
+    41.0,
+    45.0,
+    50.0,
+    58.0,
+    90.0
+   ],
+   "shares": {
+    "0": 0.09804775568764092,
+    "1": 0.10040479606476738,
+    "2": 0.09971305595408896,
+    "3": 0.07949887271981963,
+    "4": 0.11008915761426522,
+    "5": 0.10045603607296577,
+    "6": 0.09397417503586801,
+    "7": 0.10086595613855298,
+    "8": 0.11239495798319328,
+    "9": 0.10455523672883787
+   }
+  },
+  "capital-gain": {
+   "mean": 1078.3918067226891,
+   "std": 7454.4260044832945,
+   "median": 0.0,
+   "edges": [
+    0.0,
+    99999.0
+   ],
+   "shares": {
+    "0": 1.0
+   }
+  },
+  "capital-loss": {
+   "mean": 86.47678827628613,
+   "std": 401.1241413791568,
+   "median": 0.0,
+   "edges": [
+    0.0,
+    4356.0
+   ],
+   "shares": {
+    "0": 1.0
+   }
+  },
+  "hours-per-week": {
+   "mean": 40.41898954703833,
+   "std": 12.400096371479366,
+   "median": 40.0,
+   "edges": [
+    1.0,
+    24.0,
+    35.0,
+    40.0,
+    48.0,
+    55.0,
+    99.0
+   ],
+   "shares": {
+    "0": 0.09469153515064563,
+    "1": 0.07765423242467719,
+    "2": 0.06730375076860012,
+    "3": 0.5458085673293708,
+    "4": 0.10965361754457881,
+    "5": 0.10488829678212748
+   }
+  }
+ },
+ "categorical": {
+  "workclass": {
+   "Private": 0.6928417708546833,
+   "Self-emp-not-inc": 0.07947325271572044,
+   "Local-gov": 0.06507481041196966,
+   "?": 0.05741442918630867,
+   "State-gov": 0.04078704652592744,
+   "Self-emp-inc": 0.03484320557491289,
+   "Federal-gov": 0.02897622463619594,
+   "Without-pay": 0.0004099200655872105,
+   "Never-worked": 0.0001793400286944046
+  },
+  "education": {
+   "HS-grad": 0.32232527157204344,
+   "Some-college": 0.2220998155359705,
+   "Bachelors": 0.16540274646443942,
+   "Masters": 0.054647468743594996,
+   "Assoc-voc": 0.041863086698093875,
+   "11th": 0.037046525927444146,
+   "Assoc-acdm": 0.033101045296167246,
+   "10th": 0.028258864521418323,
+   "7th-8th": 0.01916376306620209,
+   "Prof-school": 0.017114162738266037,
+   "9th": 0.015884402541504406,
+   "12th": 0.013783562205369953,
+   "Doctorate": 0.012425701988112318,
+   "5th-6th": 0.010401721664275465,
+   "1st-4th": 0.005021520803443328,
+   "Preschool": 0.0014603402336544375
+  },
+  "marital-status": {
+   "Married-civ-spouse": 0.4578038532486165,
+   "Never-married": 0.3324195531871285,
+   "Divorced": 0.13376204140192663,
+   "Widowed": 0.031461365033818404,
+   "Separated": 0.030769624923139987,
+   "Married-spouse-absent": 0.012963722074195532,
+   "Married-AF-spouse": 0.000819840131174421
+  },
+  "occupation": {
+   "Prof-specialty": 0.12689588030334084,
+   "Exec-managerial": 0.12435949989751999,
+   "Craft-repair": 0.12433387989342078,
+   "Adm-clerical": 0.11511067841770854,
+   "Sales": 0.11293297806927649,
+   "Other-service": 0.10089157614265218,
+   "Machine-op-inspct": 0.06212850994056159,
+   "?": 0.057593769215003074,
+   "Transport-moving": 0.04880610780897725,
+   "Handlers-cleaners": 0.04176060668169707,
+   "Farming-fishing": 0.030205984832957573,
+   "Tech-support": 0.029232424677187948,
+   "Protective-serv": 0.02034228325476532,
+   "Priv-house-serv": 0.00509838081574093,
+   "Armed-Forces": 0.00030744004919040785
+  },
+  "relationship_model": {
+   "Spouse": 0.45127075220332036,
+   "Not-in-family": 0.25586698093871696,
+   "Own-child": 0.15820352531256404,
+   "Unmarried": 0.10358167657306826,
+   "Other-relative": 0.031077064972330395
+  },
+  "capital_gain_topcoded": {
+   "0": 0.9950040992006559,
+   "1": 0.0049959007993441275
+  }
+ },
+ "groups": {
+  "sex": {
+   "Male": 26130,
+   "Female": 12902
+  },
+  "race": {
+   "White": 33372,
+   "Black": 3721,
+   "Asian-Pac-Islander": 1217,
+   "Amer-Indian-Eskimo": 387,
+   "Other": 335
+  },
+  "age band": {
+   "25 to 34": 10071,
+   "35 to 44": 9748,
+   "45 to 54": 6959,
+   "under 25": 6808,
+   "55 to 64": 3799,
+   "65 and over": 1647
+  },
+  "marital status": {
+   "Married-civ-spouse": 17869,
+   "Never-married": 12975,
+   "Divorced": 5221,
+   "Widowed": 1228,
+   "Separated": 1201,
+   "Married-spouse-absent": 506,
+   "Married-AF-spouse": 32
+  },
+  "country of origin": {
+   "United-States": 35034,
+   "all other": 953,
+   "Mexico": 745,
+   "unrecorded": 682,
+   "Philippines": 234,
+   "Germany": 166,
+   "Canada": 148,
+   "Puerto-Rico": 147,
+   "India": 125,
+   "El-Salvador": 122,
+   "Cuba": 120,
+   "England": 105,
+   "China": 101,
+   "South": 96,
+   "Italy": 95,
+   "Dominican-Republic": 80,
+   "Jamaica": 79
+  }
+ },
+ "portfolio": {
+  "false_alarm": 0.0029965753424657533,
+  "miss": 0.7379412557262194,
+  "flagged": 1952,
+  "caught": 1945
+ },
+ "group_rates": {
+  "sex": {
+   "Male": {
+    "n": 6484,
+    "false_alarm": 0.0025380710659898475,
+    "miss": 0.7780239255649092,
+    "flagged_share": 0.155305367057372
+   },
+   "Female": {
+    "n": 3274,
+    "false_alarm": 0.00546448087431694,
+    "miss": 0.6757221458046767,
+    "flagged_share": 0.28863775198533903
+   }
+  },
+  "race": {
+   "White": {
+    "n": 8342,
+    "false_alarm": 0.0028142589118198874,
+    "miss": 0.7516908212560387,
+    "flagged_share": 0.18556701030927836
+   },
+   "Black": {
+    "n": 962,
+    "false_alarm": 0.0,
+    "miss": 0.6485849056603774,
+    "flagged_share": 0.3097713097713098
+   },
+   "Amer-Indian-Eskimo": {
+    "n": 83,
+    "false_alarm": 0.0,
+    "miss": 0.6153846153846154,
+    "flagged_share": 0.3614457831325301
+   },
+   "Asian-Pac-Islander": {
+    "n": 300,
+    "false_alarm": 0.013157894736842105,
+    "miss": 0.7410714285714286,
+    "flagged_share": 0.19666666666666666
+   },
+   "Other": {
+    "n": 71,
+    "false_alarm": 0.0,
+    "miss": 0.7258064516129032,
+    "flagged_share": 0.23943661971830985
+   }
+  },
+  "age band": {
+   "under 25": {
+    "n": 1601,
+    "false_alarm": 0.07142857142857142,
+    "miss": 0.3314429741650914,
+    "flagged_share": 0.6633354153653966
+   },
+   "25 to 34": {
+    "n": 2493,
+    "false_alarm": 0.007537688442211055,
+    "miss": 0.7527446300715991,
+    "flagged_share": 0.2089851584436422
+   },
+   "35 to 44": {
+    "n": 2436,
+    "false_alarm": 0.0,
+    "miss": 0.8971393791844188,
+    "flagged_share": 0.069376026272578
+   },
+   "45 to 54": {
+    "n": 1806,
+    "false_alarm": 0.0,
+    "miss": 0.9044585987261147,
+    "flagged_share": 0.05813953488372093
+   },
+   "55 to 64": {
+    "n": 983,
+    "false_alarm": 0.006134969325153374,
+    "miss": 0.9162861491628614,
+    "flagged_share": 0.05798575788402848
+   },
+   "65 and over": {
+    "n": 439,
+    "false_alarm": 0.01020408163265306,
+    "miss": 0.8914956011730205,
+    "flagged_share": 0.08656036446469248
+   }
+  },
+  "marital status": {
+   "Never-married": {
+    "n": 3107,
+    "false_alarm": 0.030534351145038167,
+    "miss": 0.47043010752688175,
+    "flagged_share": 0.5085291277759897
+   },
+   "Separated": {
+    "n": 329,
+    "false_alarm": 0.0,
+    "miss": 0.7339743589743589,
+    "flagged_share": 0.25227963525835867
+   },
+   "Married-civ-spouse": {
+    "n": 4497,
+    "false_alarm": 0.0,
+    "miss": 0.9971887550200803,
+    "flagged_share": 0.00155659328441183
+   },
+   "Married-spouse-absent": {
+    "n": 121,
+    "false_alarm": 0.0,
+    "miss": 0.6181818181818182,
+    "flagged_share": 0.34710743801652894
+   },
+   "Divorced": {
+    "n": 1409,
+    "false_alarm": 0.013986013986013986,
+    "miss": 0.8570300157977883,
+    "flagged_share": 0.12987934705464868
+   },
+   "Widowed": {
+    "n": 290,
+    "false_alarm": 0.04,
+    "miss": 0.7886792452830189,
+    "flagged_share": 0.19655172413793104
+   },
+   "Married-AF-spouse": {
+    "n": 5,
+    "false_alarm": 0.0,
+    "miss": 1.0,
+    "flagged_share": 0.0
+   }
+  },
+  "country of origin": {
+   "United-States": {
+    "n": 8758,
+    "false_alarm": 0.0014012143858010276,
+    "miss": 0.7417258576394137,
+    "flagged_share": 0.1954784197305321
+   },
+   "all other": {
+    "n": 258,
+    "false_alarm": 0.0,
+    "miss": 0.7635467980295566,
+    "flagged_share": 0.18604651162790697
+   },
+   "Jamaica": {
+    "n": 27,
+    "false_alarm": 0.0,
+    "miss": 0.5833333333333334,
+    "flagged_share": 0.37037037037037035
+   },
+   "Canada": {
+    "n": 34,
+    "false_alarm": 0.0,
+    "miss": 1.0,
+    "flagged_share": 0.0
+   },
+   "unrecorded": {
+    "n": 174,
+    "false_alarm": 0.038461538461538464,
+    "miss": 0.7868852459016393,
+    "flagged_share": 0.16091954022988506
+   },
+   "Mexico": {
+    "n": 198,
+    "false_alarm": 0.07692307692307693,
+    "miss": 0.5621621621621622,
+    "flagged_share": 0.41414141414141414
+   },
+   "Germany": {
+    "n": 40,
+    "false_alarm": 0.0,
+    "miss": 0.8484848484848485,
+    "flagged_share": 0.125
+   },
+   "Puerto-Rico": {
+    "n": 37,
+    "false_alarm": 0.2,
+    "miss": 0.75,
+    "flagged_share": 0.24324324324324326
+   },
+   "Italy": {
+    "n": 10,
+    "false_alarm": 0.0,
+    "miss": 0.8333333333333334,
+    "flagged_share": 0.1
+   },
+   "Philippines": {
+    "n": 60,
+    "false_alarm": 0.0,
+    "miss": 0.673469387755102,
+    "flagged_share": 0.26666666666666666
+   },
+   "Dominican-Republic": {
+    "n": 23,
+    "false_alarm": 0.0,
+    "miss": 0.45454545454545453,
+    "flagged_share": 0.5217391304347826
+   },
+   "Cuba": {
+    "n": 18,
+    "false_alarm": 0.0,
+    "miss": 0.7857142857142857,
+    "flagged_share": 0.16666666666666666
+   },
+   "China": {
+    "n": 21,
+    "false_alarm": 0.0,
+    "miss": 0.9333333333333333,
+    "flagged_share": 0.047619047619047616
+   },
+   "India": {
+    "n": 26,
+    "false_alarm": 0.0,
+    "miss": 0.6875,
+    "flagged_share": 0.19230769230769232
+   },
+   "England": {
+    "n": 22,
+    "false_alarm": 0.0,
+    "miss": 0.8333333333333334,
+    "flagged_share": 0.09090909090909091
+   },
+   "South": {
+    "n": 19,
+    "false_alarm": 0.0,
+    "miss": 0.8666666666666667,
+    "flagged_share": 0.10526315789473684
+   },
+   "El-Salvador": {
+    "n": 33,
+    "false_alarm": 0.0,
+    "miss": 0.5,
+    "flagged_share": 0.48484848484848486
+   }
+  }
+ }
+}

requirements.txt

@@ -0,0 +1,7 @@
+gradio==6.31.0
+scikit-learn==1.9.1
+fairlearn==0.15.0
+pandas>=2.2
+numpy>=1.26
+scipy>=1.11
+matplotlib>=3.8

space_config.json

@@ -0,0 +1,472 @@
+{
+ "org": "Crosswind Workforce Partnership",
+ "system": "ReachFirst",
+ "version": "v1",
+ "short": "reachfirst",
+ "emoji": "\ud83d\udcec",
+ "model_file": "reachfirst_v1.pkl",
+ "recal_file": null,
+ "odds_factor": null,
+ "reasons_file": null,
+ "data_file": "crosswind_pilot_extract_remediated.csv",
+ "md5": "ca9ed6d4b9dfd59f3f586b09da05b31f",
+ "id_col": null,
+ "outcome": "income",
+ "positive": "<=50K",
+ "positive_label": "income at or below the line",
+ "negative_label": "income above the line",
+ "unit": "record",
+ "units": "records",
+ "action": "a place on the first mailer list for the measured first cycle",
+ "flagged_word": "ranked in",
+ "person": "household",
+ "share": 0.2,
+ "book_size": 15000,
+ "capacity": 3000,
+ "sample_rows": 4000,
+ "groups": {
+  "sex": {
+   "col": "sex"
+  },
+  "race": {
+   "col": "race"
+  },
+  "age band": {
+   "col": "age",
+   "bands": [
+    [
+     0,
+     24,
+     "under 25"
+    ],
+    [
+     25,
+     34,
+     "25 to 34"
+    ],
+    [
+     35,
+     44,
+     "35 to 44"
+    ],
+    [
+     45,
+     54,
+     "45 to 54"
+    ],
+    [
+     55,
+     64,
+     "55 to 64"
+    ],
+    [
+     65,
+     200,
+     "65 and over"
+    ]
+   ]
+  },
+  "marital status": {
+   "col": "marital-status"
+  },
+  "country of origin": {
+   "col": "native_country_report",
+   "map": {
+    "?": "unrecorded"
+   }
+  }
+ },
+ "min_group": 500,
+ "group_check": "miss",
+ "group_tripwire": 0.2,
+ "watch_groups": [
+  "Married-civ-spouse",
+  "35 to 44",
+  "45 to 54",
+  "55 to 64"
+ ],
+ "calibration": null,
+ "watch_fields": [
+  "marital-status",
+  "age",
+  "relationship_model",
+  "occupation"
+ ],
+ "watch_fields_label": "marital status, age, household role, and occupation (Week 6 memo, rulings 1 and 4)",
+ "week6": {
+  "roc_auc": 0.905,
+  "flagged": 1952,
+  "caught": 1945,
+  "test_rows": 9758,
+  "test_positives": 7422
+ },
+ "lead": "Lauren Chen, AI Governance Lead",
+ "caio": "Denise Whitfield, Chief AI Officer",
+ "memo_ref": "Week 6 memo",
+ "monitoring_lines": [
+  "The share of at-or-below records left off the list, by age band and marital status, against the file's (ruling 1).",
+  "The measured first cycle: the list half against the current-practice half, married at-or-below households as the comparison group (ruling 3).",
+  "The sex-and-race recovery check, rerun at every retrain (ruling 4)."
+ ],
+ "inputs": [
+  "age",
+  "workclass",
+  "education",
+  "marital-status",
+  "occupation",
+  "capital-gain",
+  "capital-loss",
+  "hours-per-week",
+  "relationship_model",
+  "capital_gain_topcoded"
+ ],
+ "categorical": [
+  "workclass",
+  "education",
+  "marital-status",
+  "occupation",
+  "relationship_model",
+  "capital_gain_topcoded"
+ ],
+ "numeric": [
+  "age",
+  "capital-gain",
+  "capital-loss",
+  "hours-per-week"
+ ],
+ "kept": [
+  "age",
+  "capital-gain",
+  "capital-loss",
+  "hours-per-week",
+  "workclass_?",
+  "workclass_Federal-gov",
+  "workclass_Local-gov",
+  "workclass_Never-worked",
+  "workclass_Private",
+  "workclass_Self-emp-inc",
+  "workclass_Self-emp-not-inc",
+  "workclass_State-gov",
+  "workclass_Without-pay",
+  "education_10th",
+  "education_11th",
+  "education_12th",
+  "education_1st-4th",
+  "education_5th-6th",
+  "education_7th-8th",
+  "education_9th",
+  "education_Assoc-acdm",
+  "education_Assoc-voc",
+  "education_Bachelors",
+  "education_Doctorate",
+  "education_HS-grad",
+  "education_Masters",
+  "education_Preschool",
+  "education_Prof-school",
+  "education_Some-college",
+  "marital-status_Divorced",
+  "marital-status_Married-AF-spouse",
+  "marital-status_Married-civ-spouse",
+  "marital-status_Married-spouse-absent",
+  "marital-status_Never-married",
+  "marital-status_Separated",
+  "marital-status_Widowed",
+  "occupation_?",
+  "occupation_Adm-clerical",
+  "occupation_Armed-Forces",
+  "occupation_Craft-repair",
+  "occupation_Exec-managerial",
+  "occupation_Farming-fishing",
+  "occupation_Handlers-cleaners",
+  "occupation_Machine-op-inspct",
+  "occupation_Other-service",
+  "occupation_Priv-house-serv",
+  "occupation_Prof-specialty",
+  "occupation_Protective-serv",
+  "occupation_Sales",
+  "occupation_Tech-support",
+  "occupation_Transport-moving",
+  "relationship_model_Not-in-family",
+  "relationship_model_Other-relative",
+  "relationship_model_Own-child",
+  "relationship_model_Spouse",
+  "relationship_model_Unmarried",
+  "capital_gain_topcoded_0",
+  "capital_gain_topcoded_1"
+ ],
+ "artifacts": {
+  "reachfirst_v1.pkl": {
+   "bytes": 4300,
+   "sha256": "120234d9f9a111028a2d605d2779a9d8e151ff8d6f85df4ec0497f651595b953"
+  }
+ },
+ "weights": [
+  [
+   "capital-gain",
+   -2.2417872489881026
+  ],
+  [
+   "education_Doctorate",
+   -1.6366401222699463
+  ],
+  [
+   "education_Prof-school",
+   -1.5886947281443564
+  ],
+  [
+   "marital-status_Married-civ-spouse",
+   -1.3318312733718085
+  ],
+  [
+   "education_Masters",
+   -1.226998809000688
+  ],
+  [
+   "education_7th-8th",
+   1.1948745792414959
+  ],
+  [
+   "education_9th",
+   1.0117126998478085
+  ],
+  [
+   "education_Bachelors",
+   -0.9637259660175791
+  ],
+  [
+   "education_5th-6th",
+   0.9516051530601645
+  ],
+  [
+   "occupation_Farming-fishing",
+   0.8772839801447048
+  ],
+  [
+   "occupation_Exec-managerial",
+   -0.8541365785148912
+  ],
+  [
+   "occupation_Other-service",
+   0.781840452365622
+  ],
+  [
+   "education_10th",
+   0.7805877804503274
+  ],
+  [
+   "education_1st-4th",
+   0.7792674440291991
+  ],
+  [
+   "education_11th",
+   0.7773197750619535
+  ],
+  [
+   "marital-status_Never-married",
+   0.7576080167618042
+  ],
+  [
+   "relationship_model_Spouse",
+   -0.7179494340103041
+  ],
+  [
+   "occupation_Prof-specialty",
+   -0.6729447708086544
+  ],
+  [
+   "occupation_Handlers-cleaners",
+   0.6242368981515822
+  ],
+  [
+   "occupation_Tech-support",
+   -0.6189663047020056
+  ],
+  [
+   "relationship_model_Own-child",
+   0.6110107663731764
+  ],
+  [
+   "marital-status_Married-AF-spouse",
+   -0.6036891114878192
+  ],
+  [
+   "workclass_Federal-gov",
+   -0.5782322983438535
+  ],
+  [
+   "occupation_Priv-house-serv",
+   0.49771933984626343
+  ],
+  [
+   "education_Preschool",
+   0.475000514022028
+  ],
+  [
+   "occupation_Protective-serv",
+   -0.4486092827356464
+  ],
+  [
+   "education_12th",
+   0.43801881119853986
+  ],
+  [
+   "relationship_model_Other-relative",
+   0.4366084364105382
+  ],
+  [
+   "marital-status_Widowed",
+   0.411442162972904
+  ],
+  [
+   "workclass_Self-emp-not-inc",
+   0.40584831710662533
+  ],
+  [
+   "education_Assoc-acdm",
+   -0.40047630749854773
+  ],
+  [
+   "occupation_Sales",
+   -0.3691730492412525
+  ],
+  [
+   "marital-status_Divorced",
+   0.35970764075843786
+  ],
+  [
+   "education_Assoc-voc",
+   -0.35460703817389366
+  ],
+  [
+   "relationship_model_Not-in-family",
+   -0.3468235390014372
+  ],
+  [
+   "hours-per-week",
+   -0.3376055865841321
+  ],
+  [
+   "workclass_Self-emp-inc",
+   -0.3133450413850334
+  ],
+  [
+   "age",
+   -0.3079133863771645
+  ],
+  [
+   "marital-status_Separated",
+   0.2891974899880865
+  ],
+  [
+   "occupation_?",
+   0.27540443434911954
+  ],
+  [
+   "workclass_?",
+   0.27236425493903366
+  ],
+  [
+   "occupation_Machine-op-inspct",
+   0.26824061777393843
+  ],
+  [
+   "education_Some-college",
+   -0.25756126820826947
+  ],
+  [
+   "capital-loss",
+   -0.24456817976716777
+  ],
+  [
+   "workclass_State-gov",
+   0.22032596360113468
+  ],
+  [
+   "marital-status_Married-spouse-absent",
+   0.19106420631129978
+  ],
+  [
+   "occupation_Craft-repair",
+   -0.12003826217984946
+  ],
+  [
+   "workclass_Without-pay",
+   0.10573384913493386
+  ],
+  [
+   "occupation_Armed-Forces",
+   -0.10136306490673276
+  ],
+  [
+   "education_HS-grad",
+   0.09381661433467584
+  ],
+  [
+   "relationship_model_Unmarried",
+   0.09065290216094356
+  ],
+  [
+   "workclass_Private",
+   -0.08277029477532862
+  ],
+  [
+   "occupation_Adm-clerical",
+   -0.0779884657765726
+  ],
+  [
+   "capital_gain_topcoded_0",
+   0.06912153062371537
+  ],
+  [
+   "workclass_Local-gov",
+   0.04053420224531096
+  ],
+  [
+   "occupation_Transport-moving",
+   0.011993188167275508
+  ],
+  [
+   "capital_gain_topcoded_1",
+   0.004377601309201683
+  ],
+  [
+   "workclass_Never-worked",
+   0.003040179410085815
+  ]
+ ],
+ "intercept": 2.818318687777577,
+ "data_url": "https://huggingface.co/datasets/drdavidprivacy/practicum-case-packs/resolve/main/crosswind_pilot_extract_remediated.csv",
+ "dataset_repo": "https://huggingface.co/datasets/drdavidprivacy/practicum-case-packs",
+ "versions": {
+  "sklearn": "1.9.1",
+  "fairlearn": "0.15.0",
+  "pandas": "3.0.5",
+  "numpy": "2.5.3"
+ },
+ "docs": [
+  [
+   "01_Use-Case-Assessment.pdf",
+   "Use Case Assessment (Week 1)"
+  ],
+  [
+   "02_EDA-Observation-Log.pdf",
+   "EDA Observation Log (Week 2)"
+  ],
+  [
+   "03_Dataset-Datasheet.pdf",
+   "Dataset Datasheet (Week 3)"
+  ],
+  [
+   "04_Model-Selection-Rationale.pdf",
+   "Model Selection Rationale (Week 4)"
+  ],
+  [
+   "05_Model-Card.pdf",
+   "Model Card (Week 6)"
+  ],
+  [
+   "06_Go-No-Go-Decision-Record.pdf",
+   "Go/No-Go Decision Record (Week 7)"
+  ]
+ ]
+}