mirror of
https://github.com/harvard-edge/cs249r_book.git
synced 2026-04-30 09:38:38 -05:00
6f5732558f
Add all Vol1 (labs 01-16) and Vol2 (labs 01-17) interactive Marimo labs as the first full first-pass implementation of the ML Systems curriculum labs. Each lab follows the PROTOCOL 2-Act structure (35-40 min): - Act I: Calibration with prediction lock → instruments → overlay - Act II: Design challenge with failure states and reflection Key pedagogical instruments introduced progressively: - Vol1: D·A·M Triad, Iron Law, Memory Ledger, Roofline, Amdahl's Law, Little's Law, P99 Histogram, Compression Frontier, Chouldechova theorem - Vol2: NVLink vs PCIe cliff, Bisection BW, Young-Daly T*, Parallelism Paradox, AllReduce ring vs tree, KV-cache model, Jevons Paradox, DP ε-δ tradeoff, SLO composition, Adversarial Pareto, two-volume synthesis capstone All 35 staged files pass AST syntax verification (36/36 including lab_00). Also includes: - labs/LABS_SPEC.md: authoritative sub-agent brief for all lab conventions - labs/core/style.py: expanded unified design system with semantic color tokens
1570 lines
74 KiB
Python
1570 lines
74 KiB
Python
import marimo
|
||
|
||
__generated_with = "0.19.6"
|
||
app = marimo.App(width="full")
|
||
|
||
# ─────────────────────────────────────────────────────────────────────────────
|
||
# LAB 15: THERE IS NO FREE FAIRNESS
|
||
#
|
||
# Chapter: Responsible Engineering (@sec-responsible-engineering)
|
||
# Core Invariant:
|
||
# Chouldechova's (2017) impossibility theorem — when base rates differ between
|
||
# groups, you CANNOT simultaneously achieve equal false positive rates, equal
|
||
# false negative rates, AND equal positive predictive value. Equal accuracy
|
||
# is not evidence of equal treatment.
|
||
#
|
||
# 2 Contexts: Cloud (H100) vs Mobile (Smartphone NPU)
|
||
#
|
||
# Act I — The Fairness Illusion (12–15 min)
|
||
# Stakeholder: Product Compliance Officer
|
||
# Scenario: Loan approval model, 85% accuracy on both groups. Legal says
|
||
# compliant. Advocates say unfair. Who is right?
|
||
# Prediction: What does equal accuracy actually guarantee?
|
||
# Instrument: Fairness metric explorer — sliders for base_rate_a, base_rate_b,
|
||
# model_threshold; per-group confusion matrices; FPR, FNR, PPV, equalized
|
||
# odds gap; prediction-vs-reality overlay.
|
||
#
|
||
# Act II — The Audit-Accuracy Tradeoff (20–25 min)
|
||
# Stakeholder: Engineering VP
|
||
# Scenario: Resume screening at scale. Choose fairness strategy + mitigation
|
||
# method + audit frequency. Design the deployment.
|
||
# Failure state: equalized_odds_gap > 10% AND context is high-stakes.
|
||
# Reflection: Why demographic parity can produce unfair outcomes.
|
||
#
|
||
# Design Ledger: chapter=15
|
||
# context, fairness_criterion, equalized_odds_gap, audit_cost_k,
|
||
# act1_prediction, act1_correct, act2_result, act2_decision,
|
||
# constraint_hit, regulatory_risk
|
||
# ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
|
||
# ── CELL 0: SETUP (hide_code=False — leave visible) ───────────────────────────
|
||
@app.cell
|
||
def _():
|
||
import marimo as mo
|
||
import sys
|
||
import math
|
||
from pathlib import Path
|
||
import plotly.graph_objects as go
|
||
import numpy as np
|
||
|
||
_root = Path(__file__).resolve().parents[2]
|
||
if str(_root) not in sys.path:
|
||
sys.path.insert(0, str(_root))
|
||
|
||
from labs.core.state import DesignLedger
|
||
from labs.core.style import COLORS, LAB_CSS, apply_plotly_theme
|
||
|
||
# ── Hardware constants — plain floats, sources annotated ──────────────────
|
||
H100_BW_GBS = 3350 # GB/s — H100 SXM5 HBM3e, NVIDIA spec
|
||
H100_TFLOPS_FP16 = 1979 # TFLOPS FP16 — NVIDIA H100 datasheet
|
||
H100_RAM_GB = 80 # GB HBM — NVIDIA H100 SXM5
|
||
MOBILE_BW_GBS = 68 # GB/s — Apple A17-class smartphone NPU
|
||
MOBILE_TOPS_INT8 = 35 # TOPS INT8 — Apple A17-class NPU
|
||
MOBILE_RAM_GB = 8 # GB — typical flagship smartphone
|
||
|
||
# ── Domain constants — responsible_engr.qmd scenario ─────────────────────
|
||
# Baseline model accuracy on both groups (the "85% parity illusion" scenario)
|
||
BASE_ACCURACY = 0.85 # overall accuracy on each group
|
||
|
||
# EEOC 4/5ths (80%) rule: selection rate ratio < 0.8 triggers review
|
||
# Equivalent to equalized odds gap threshold for high-stakes deployment
|
||
DISPARATE_IMPACT_THRESHOLD_PP = 10.0 # pp — EEOC / OFCCP guidance for hiring
|
||
|
||
# Audit cost model: cloud vs mobile context
|
||
# Cloud: H100 batch re-evaluation, near-realtime, ~$2/hr compute
|
||
CLOUD_AUDIT_COST_K_USD = 1.5 # $K per audit run (cloud, automated)
|
||
# Mobile: delayed centralized audit, human review + compute
|
||
MOBILE_AUDIT_COST_K_USD = 4.0 # $K per audit run (mobile, manual + compute)
|
||
|
||
# Annual audit frequencies (runs/year)
|
||
AUDIT_FREQ_OPTIONS = {
|
||
"Continuous (52×/yr)": 52,
|
||
"Monthly (12×/yr)": 12,
|
||
"Quarterly (4×/yr)": 4,
|
||
"Annual (1×/yr)": 1,
|
||
}
|
||
|
||
ledger = DesignLedger()
|
||
return (
|
||
mo, go, np, math,
|
||
ledger, COLORS, LAB_CSS, apply_plotly_theme,
|
||
H100_BW_GBS, H100_TFLOPS_FP16, H100_RAM_GB,
|
||
MOBILE_BW_GBS, MOBILE_TOPS_INT8, MOBILE_RAM_GB,
|
||
BASE_ACCURACY, DISPARATE_IMPACT_THRESHOLD_PP,
|
||
CLOUD_AUDIT_COST_K_USD, MOBILE_AUDIT_COST_K_USD,
|
||
AUDIT_FREQ_OPTIONS,
|
||
)
|
||
|
||
|
||
# ── CELL 1: HEADER ────────────────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo, LAB_CSS, COLORS):
|
||
mo.vstack([
|
||
LAB_CSS,
|
||
mo.Html(f"""
|
||
<div style="background: linear-gradient(135deg, #0f172a 0%, #1e293b 100%);
|
||
padding: 36px 44px; border-radius: 16px; color: white;
|
||
box-shadow: 0 8px 32px rgba(0,0,0,0.3);">
|
||
<div style="font-size: 0.72rem; font-weight: 700; letter-spacing: 0.18em;
|
||
color: #475569; text-transform: uppercase; margin-bottom: 10px;">
|
||
Machine Learning Systems · Volume I · Lab 15
|
||
</div>
|
||
<h1 style="margin: 0 0 10px 0; font-size: 2.4rem; font-weight: 900;
|
||
color: #f8fafc; line-height: 1.1; letter-spacing: -0.02em;">
|
||
There Is No Free Fairness
|
||
</h1>
|
||
<p style="margin: 0 0 20px 0; font-size: 1.05rem; color: #94a3b8;
|
||
max-width: 720px; line-height: 1.65;">
|
||
Your model has <strong style="color:#f8fafc;">85% accuracy on both groups</strong>.
|
||
Legal says compliant. Advocates say unfair. Someone is about to be proven wrong.
|
||
Chouldechova (2017) settled this debate mathematically.
|
||
</p>
|
||
<div style="display: flex; gap: 12px; flex-wrap: wrap;">
|
||
<span style="background: rgba(99,102,241,0.15); color: #a5b4fc;
|
||
padding: 5px 14px; border-radius: 20px; font-size: 0.8rem;
|
||
font-weight: 600; border: 1px solid rgba(99,102,241,0.25);">
|
||
Act I: The Fairness Illusion · Act II: Audit-Accuracy Tradeoff
|
||
</span>
|
||
<span style="background: rgba(16,185,129,0.15); color: #6ee7b7;
|
||
padding: 5px 14px; border-radius: 20px; font-size: 0.8rem;
|
||
font-weight: 600; border: 1px solid rgba(16,185,129,0.25);">
|
||
35–40 min
|
||
</span>
|
||
<span style="background: rgba(203,32,45,0.15); color: #fca5a5;
|
||
padding: 5px 14px; border-radius: 20px; font-size: 0.8rem;
|
||
font-weight: 600; border: 1px solid rgba(203,32,45,0.25);">
|
||
Disparate impact failure state active
|
||
</span>
|
||
<span style="background: rgba(204,85,0,0.15); color: #fdba74;
|
||
padding: 5px 14px; border-radius: 20px; font-size: 0.8rem;
|
||
font-weight: 600; border: 1px solid rgba(204,85,0,0.25);">
|
||
{COLORS['Mobile'] and 'Cloud vs Mobile context'}
|
||
</span>
|
||
</div>
|
||
</div>
|
||
"""),
|
||
])
|
||
return
|
||
|
||
|
||
# ── CELL 2: RECOMMENDED READING ───────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
mo.callout(mo.md("""
|
||
**Recommended Reading** — Complete the following before this lab:
|
||
|
||
- **@sec-responsible-engineering-fairness-metrics** — Demographic parity, equalized odds,
|
||
calibration, and individual fairness: formal definitions and what each criterion
|
||
requires from the underlying data distribution.
|
||
- **@sec-responsible-engineering-impossibility** — Chouldechova (2017) impossibility
|
||
theorem: when base rates differ, FPR equality, FNR equality, and PPV equality cannot
|
||
all hold simultaneously. This is a mathematical proof, not a policy preference.
|
||
- **@sec-responsible-engineering-audit-pipelines** — Continuous monitoring, audit
|
||
frequency, mitigation methods, and the compute and accuracy costs of fairness
|
||
enforcement at deployment scale.
|
||
"""), kind="info")
|
||
return
|
||
|
||
|
||
# ── CELL 3: CONTEXT TOGGLE ────────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
context_toggle = mo.ui.radio(
|
||
options={
|
||
"Cloud (H100 — automated continuous audit)": "cloud",
|
||
"Mobile (Smartphone NPU — delayed centralized audit)": "mobile",
|
||
},
|
||
value="Cloud (H100 — automated continuous audit)",
|
||
label="Deployment context:",
|
||
inline=True,
|
||
)
|
||
mo.vstack([
|
||
mo.md("---"),
|
||
mo.md(
|
||
"**Select your deployment context.** "
|
||
"Cloud enables near-realtime bias monitoring at low marginal cost. "
|
||
"Mobile deployments face delayed feedback loops and higher per-audit cost."
|
||
),
|
||
context_toggle,
|
||
])
|
||
return (context_toggle,)
|
||
|
||
|
||
# ─────────────────────────────────────────────────────────────────────────────
|
||
# ACT I — THE FAIRNESS ILLUSION
|
||
# ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
# ── CELL 4: ACT I SCENARIO ────────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo, COLORS):
|
||
_color = COLORS["OrangeLine"]
|
||
mo.vstack([
|
||
mo.md("---"),
|
||
mo.md("## Act I — The Fairness Illusion"),
|
||
mo.Html(f"""
|
||
<div style="border-left: 4px solid {_color}; background: #fff7ed;
|
||
border-radius: 0 10px 10px 0; padding: 16px 22px; margin: 12px 0;">
|
||
<div style="font-size: 0.72rem; font-weight: 700; color: {_color};
|
||
text-transform: uppercase; letter-spacing: 0.1em; margin-bottom: 6px;">
|
||
Incoming Message · Product Compliance Officer
|
||
</div>
|
||
<div style="font-style: italic; font-size: 1.0rem; color: #1e293b; line-height: 1.65;">
|
||
"Our loan approval model has 85% accuracy on both demographic groups.
|
||
Our legal team says we're compliant — equal accuracy means equal treatment.
|
||
But advocacy groups are threatening litigation, saying our false positive
|
||
and false negative rates differ by group. Who is right?
|
||
Do we have a problem or don't we?"
|
||
</div>
|
||
</div>
|
||
"""),
|
||
mo.md("""
|
||
The Compliance Officer is measuring the right thing — accuracy — but measuring it
|
||
at the wrong level of resolution. Equal overall accuracy is consistent with wildly
|
||
different error distributions between groups, especially when base rates differ.
|
||
|
||
Before touching the explorer, commit to a prediction.
|
||
"""),
|
||
])
|
||
return
|
||
|
||
|
||
# ── CELL 5: ACT I PREDICTION LOCK ─────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
act1_prediction = mo.ui.radio(
|
||
options={
|
||
"A) 85% accuracy on both groups proves fairness — equal accuracy means equal treatment": "option_a",
|
||
"B) Need to check precision (PPV) — equal accuracy can mask unequal precision rates": "option_b",
|
||
"C) Equal accuracy is necessary but not sufficient — must check FPR AND FNR separately": "option_c",
|
||
"D) Accuracy alone is sufficient for legal compliance under disparate impact doctrine": "option_d",
|
||
},
|
||
label=(
|
||
"The Compliance Officer claims equal accuracy proves fairness. "
|
||
"What is the most complete and correct assessment?"
|
||
),
|
||
)
|
||
mo.vstack([
|
||
mo.Html("""
|
||
<div style="background: #1e293b; border-radius: 12px; padding: 20px;
|
||
border-left: 4px solid #f59e0b; margin: 8px 0;">
|
||
<div style="font-size: 0.72rem; font-weight: 700; color: #fbbf24;
|
||
text-transform: uppercase; letter-spacing: 0.1em; margin-bottom: 10px;">
|
||
Prediction Lock — Act I
|
||
</div>
|
||
<div style="color: #e2e8f0; font-size: 0.88rem; margin-bottom: 12px;">
|
||
Commit before touching any sliders. The Fairness Metric Explorer
|
||
unlocks once you select an answer.
|
||
</div>
|
||
</div>
|
||
"""),
|
||
act1_prediction,
|
||
])
|
||
return (act1_prediction,)
|
||
|
||
|
||
# ── CELL 6: ACT I GATE ────────────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo, act1_prediction):
|
||
mo.stop(
|
||
act1_prediction.value is None,
|
||
mo.callout(
|
||
mo.md("Select your prediction above to unlock the Fairness Metric Explorer."),
|
||
kind="warn",
|
||
),
|
||
)
|
||
return
|
||
|
||
|
||
# ── CELL 7: ACT I CONTROLS ────────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
# Base rate sliders: the "positive rate" in each group's true population
|
||
# Source: responsible_engr.qmd — base rate is P(Y=1|Group=G)
|
||
# Range: 5–50% represents realistic lending/credit scenarios
|
||
base_rate_a = mo.ui.slider(
|
||
start=5, stop=50, value=10, step=1,
|
||
label="Base rate — Group A (true positive rate in population, %)",
|
||
)
|
||
base_rate_b = mo.ui.slider(
|
||
start=5, stop=50, value=40, step=1,
|
||
label="Base rate — Group B (true positive rate in population, %)",
|
||
)
|
||
# Model threshold: decision boundary applied uniformly to both groups
|
||
# Source: responsible_engr.qmd — single shared threshold is a policy choice
|
||
model_threshold = mo.ui.slider(
|
||
start=0.10, stop=0.90, value=0.50, step=0.05,
|
||
label="Model classification threshold (applied uniformly to both groups)",
|
||
)
|
||
mo.vstack([
|
||
mo.md("### Fairness Metric Explorer"),
|
||
mo.md(
|
||
"Set the **true base rate** for each group (the actual fraction of creditworthy "
|
||
"applicants in each population) and the **model threshold**. "
|
||
"The model is calibrated — its score reflects true probability — but the same "
|
||
"threshold is applied to both groups. Watch what happens to FPR and FNR."
|
||
),
|
||
mo.hstack([base_rate_a, base_rate_b], justify="start", gap="2rem"),
|
||
model_threshold,
|
||
])
|
||
return (base_rate_a, base_rate_b, model_threshold)
|
||
|
||
|
||
# ── CELL 8: ACT I PHYSICS ENGINE + CONFUSION MATRICES ─────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(
|
||
mo, go, np,
|
||
base_rate_a, base_rate_b, model_threshold,
|
||
apply_plotly_theme,
|
||
):
|
||
# ── Simulation physics ─────────────────────────────────────────────────────
|
||
# Model: calibrated binary classifier, score s ~ Beta(alpha, beta) for each group.
|
||
# Calibration means P(Y=1|s=p) = p, so the threshold determines the operating point.
|
||
#
|
||
# For a calibrated model with base rate r and threshold t:
|
||
# TPR = P(s > t | Y=1) — sensitivity
|
||
# FPR = P(s > t | Y=0) — false positive rate (false alarm)
|
||
# FNR = 1 - TPR — miss rate
|
||
# PPV = r*TPR / (r*TPR + (1-r)*FPR) — precision / positive predictive value
|
||
#
|
||
# Source: responsible_engr.qmd — confusion matrix decomposition by group,
|
||
# Chouldechova (2017) impossibility theorem derivation.
|
||
#
|
||
# We model TPR and FPR as functions of threshold for a logistic-like score distribution.
|
||
# Using a simple normal approximation: scores for positives ~ N(0.7, 0.15),
|
||
# scores for negatives ~ N(0.3, 0.15)
|
||
# This is a calibrated model because mean(positive scores) > mean(negative scores)
|
||
# in proportion to base rates, and both groups share the SAME score distributions
|
||
# (this is the "calibrated but disparate impact" scenario).
|
||
|
||
def _compute_group_metrics(base_rate_pct, threshold):
|
||
"""
|
||
Compute confusion matrix entries and derived metrics for one group.
|
||
Source: responsible_engr.qmd — Section on fairness metric decomposition.
|
||
|
||
Uses a calibrated model where positive class scores ~ N(0.70, 0.15)
|
||
and negative class scores ~ N(0.30, 0.15). The same model/threshold
|
||
is applied to both groups, but the GROUP's base rate changes the
|
||
fraction of positives vs negatives, and hence the confusion matrix.
|
||
"""
|
||
from scipy.special import erfc
|
||
r = base_rate_pct / 100.0 # base rate as fraction
|
||
|
||
# TPR = P(score > t | Y=1) = P(N(0.70, 0.15) > t)
|
||
# FPR = P(score > t | Y=0) = P(N(0.30, 0.15) > t)
|
||
mu_pos, mu_neg, sigma = 0.70, 0.30, 0.15
|
||
tpr = 0.5 * erfc((threshold - mu_pos) / (sigma * np.sqrt(2)))
|
||
fpr = 0.5 * erfc((threshold - mu_neg) / (sigma * np.sqrt(2)))
|
||
|
||
# Clamp to valid range
|
||
tpr = float(np.clip(tpr, 0.001, 0.999))
|
||
fpr = float(np.clip(fpr, 0.001, 0.999))
|
||
fnr = 1.0 - tpr
|
||
|
||
# Accuracy = r*TPR + (1-r)*(1-FPR)
|
||
accuracy = r * tpr + (1.0 - r) * (1.0 - fpr)
|
||
|
||
# PPV = precision = TP / (TP + FP) = r*TPR / (r*TPR + (1-r)*FPR)
|
||
denom = r * tpr + (1.0 - r) * fpr
|
||
ppv = (r * tpr / denom) if denom > 1e-9 else 0.0
|
||
|
||
# Approval rate = P(score > t) = r*TPR + (1-r)*FPR
|
||
approval_rate = denom
|
||
|
||
return {
|
||
"accuracy": accuracy,
|
||
"tpr": tpr,
|
||
"fpr": fpr,
|
||
"fnr": fnr,
|
||
"ppv": ppv,
|
||
"approval_rate": approval_rate,
|
||
}
|
||
|
||
# ── Attempt scipy import; fall back to pure numpy ─────────────────────────
|
||
try:
|
||
from scipy.special import erfc as _erfc_check
|
||
_scipy_ok = True
|
||
except ImportError:
|
||
_scipy_ok = False
|
||
|
||
def _compute_group_metrics_numpy(base_rate_pct, threshold):
|
||
"""Fallback without scipy — uses numpy erf approximation."""
|
||
r = base_rate_pct / 100.0
|
||
mu_pos, mu_neg, sigma = 0.70, 0.30, 0.15
|
||
tpr = 0.5 * (1.0 - np.sign(threshold - mu_pos) *
|
||
(1.0 - np.exp(-(threshold - mu_pos)**2 / (2*sigma**2))))
|
||
# More robust: use numpy's erf
|
||
tpr = float(np.clip(0.5 * (1.0 + np.sign(mu_pos - threshold) *
|
||
(1.0 - np.exp(-(abs(threshold - mu_pos))**2 / sigma**2 * 0.5))), 0.001, 0.999))
|
||
fpr = float(np.clip(0.5 * (1.0 + np.sign(mu_neg - threshold) *
|
||
(1.0 - np.exp(-(abs(threshold - mu_neg))**2 / sigma**2 * 0.5))), 0.001, 0.999))
|
||
fnr = 1.0 - tpr
|
||
accuracy = r * tpr + (1.0 - r) * (1.0 - fpr)
|
||
denom = r * tpr + (1.0 - r) * fpr
|
||
ppv = (r * tpr / denom) if denom > 1e-9 else 0.0
|
||
approval_rate = denom
|
||
return {"accuracy": accuracy, "tpr": tpr, "fpr": fpr,
|
||
"fnr": fnr, "ppv": ppv, "approval_rate": approval_rate}
|
||
|
||
_compute_fn = _compute_group_metrics if _scipy_ok else _compute_group_metrics_numpy
|
||
|
||
_br_a = base_rate_a.value # e.g., 10%
|
||
_br_b = base_rate_b.value # e.g., 40%
|
||
_t = model_threshold.value # e.g., 0.50
|
||
|
||
_ga = _compute_fn(_br_a, _t)
|
||
_gb = _compute_fn(_br_b, _t)
|
||
|
||
# ── Derived gap metrics ────────────────────────────────────────────────────
|
||
_fpr_gap = abs(_ga["fpr"] - _gb["fpr"]) * 100 # pp
|
||
_fnr_gap = abs(_ga["fnr"] - _gb["fnr"]) * 100 # pp
|
||
_ppv_gap = abs(_ga["ppv"] - _gb["ppv"]) * 100 # pp
|
||
_acc_gap = abs(_ga["accuracy"] - _gb["accuracy"]) * 100 # pp
|
||
_eo_gap = (_fpr_gap + _fnr_gap) / 2.0 # equalized odds gap (average)
|
||
_app_gap = abs(_ga["approval_rate"] - _gb["approval_rate"]) * 100 # pp
|
||
|
||
# ── Color coding: green if gap ≤5pp, orange if 5–10pp, red if >10pp ───────
|
||
def _gap_color(gap_pp):
|
||
if gap_pp <= 5.0:
|
||
return "#22c55e" # green
|
||
elif gap_pp <= 10.0:
|
||
return "#f59e0b" # orange
|
||
return "#ef4444" # red
|
||
|
||
def _gap_label(gap_pp):
|
||
if gap_pp <= 5.0:
|
||
return "Within 5pp"
|
||
elif gap_pp <= 10.0:
|
||
return "Caution 5–10pp"
|
||
return "Gap >10pp"
|
||
|
||
# ── Pareto sweep: vary threshold 0.1→0.9, compute EO gap + accuracy ───────
|
||
_thresholds = np.linspace(0.1, 0.9, 80)
|
||
_pareto_acc = []
|
||
_pareto_eo = []
|
||
for _th in _thresholds:
|
||
_g_a = _compute_fn(_br_a, _th)
|
||
_g_b = _compute_fn(_br_b, _th)
|
||
_avg_acc = 0.5 * (_g_a["accuracy"] + _g_b["accuracy"])
|
||
_eo = 0.5 * (abs(_g_a["fpr"] - _g_b["fpr"]) + abs(_g_a["fnr"] - _g_b["fnr"])) * 100
|
||
_pareto_acc.append(_avg_acc * 100)
|
||
_pareto_eo.append(_eo)
|
||
|
||
# ── Figure: confusion matrices as bar chart side-by-side ──────────────────
|
||
_categories = ["Accuracy", "TPR", "FPR", "FNR", "PPV", "Approval Rate"]
|
||
_vals_a = [
|
||
_ga["accuracy"]*100, _ga["tpr"]*100, _ga["fpr"]*100,
|
||
_ga["fnr"]*100, _ga["ppv"]*100, _ga["approval_rate"]*100,
|
||
]
|
||
_vals_b = [
|
||
_gb["accuracy"]*100, _gb["tpr"]*100, _gb["fpr"]*100,
|
||
_gb["fnr"]*100, _gb["ppv"]*100, _gb["approval_rate"]*100,
|
||
]
|
||
|
||
_fig = go.Figure()
|
||
_fig.add_trace(go.Bar(
|
||
name=f"Group A (base rate {_br_a}%)",
|
||
x=_categories, y=_vals_a,
|
||
marker_color="#6366f1",
|
||
text=[f"{v:.1f}%" for v in _vals_a],
|
||
textposition="outside",
|
||
))
|
||
_fig.add_trace(go.Bar(
|
||
name=f"Group B (base rate {_br_b}%)",
|
||
x=_categories, y=_vals_b,
|
||
marker_color="#f59e0b",
|
||
text=[f"{v:.1f}%" for v in _vals_b],
|
||
textposition="outside",
|
||
))
|
||
_fig = apply_plotly_theme(_fig)
|
||
_fig.update_layout(
|
||
title=f"Per-Group Fairness Metrics (threshold = {_t:.2f})",
|
||
yaxis_title="Rate (%)",
|
||
barmode="group",
|
||
height=380,
|
||
legend=dict(orientation="h", yanchor="bottom", y=1.02, xanchor="right", x=1),
|
||
yaxis_range=[0, 115],
|
||
)
|
||
|
||
# ── Figure: accuracy vs EO gap Pareto frontier ────────────────────────────
|
||
_current_eo_for_plot = _eo_gap
|
||
_current_acc_for_plot = 0.5 * (_ga["accuracy"] + _gb["accuracy"]) * 100
|
||
|
||
_fig2 = go.Figure()
|
||
_fig2.add_trace(go.Scatter(
|
||
x=_pareto_acc, y=_pareto_eo,
|
||
mode="lines",
|
||
name="Accuracy vs EO Gap frontier",
|
||
line=dict(color="#6366f1", width=2),
|
||
))
|
||
_fig2.add_trace(go.Scatter(
|
||
x=[_current_acc_for_plot], y=[_current_eo_for_plot],
|
||
mode="markers",
|
||
name="Current threshold",
|
||
marker=dict(color="#f59e0b", size=12, symbol="star"),
|
||
))
|
||
# Mark the 10pp gap threshold line
|
||
_fig2.add_hline(
|
||
y=10, line_dash="dash", line_color="#ef4444",
|
||
annotation_text="10pp gap threshold (EEOC guidance)",
|
||
annotation_position="top right",
|
||
)
|
||
_fig2 = apply_plotly_theme(_fig2)
|
||
_fig2.update_layout(
|
||
title="Accuracy vs Equalized Odds Gap — Pareto Frontier",
|
||
xaxis_title="Average Accuracy (%)",
|
||
yaxis_title="Equalized Odds Gap (pp)",
|
||
height=340,
|
||
)
|
||
|
||
# ── Physics formula display ────────────────────────────────────────────────
|
||
_physics_md = f"""
|
||
### Physics
|
||
|
||
```
|
||
Group A (base rate = {_br_a}%):
|
||
TPR = P(score > {_t:.2f} | Y=1) = {_ga['tpr']*100:.1f}%
|
||
FPR = P(score > {_t:.2f} | Y=0) = {_ga['fpr']*100:.1f}%
|
||
FNR = 1 - TPR = {_ga['fnr']*100:.1f}%
|
||
PPV = r·TPR / (r·TPR + (1-r)·FPR)
|
||
= {_br_a/100:.2f}×{_ga['tpr']:.3f} / ({_br_a/100:.2f}×{_ga['tpr']:.3f} + {1-_br_a/100:.2f}×{_ga['fpr']:.3f})
|
||
= {_ga['ppv']*100:.1f}%
|
||
Accuracy = r·TPR + (1-r)·(1-FPR) = {_ga['accuracy']*100:.1f}%
|
||
|
||
Group B (base rate = {_br_b}%):
|
||
TPR = {_gb['tpr']*100:.1f}% | FPR = {_gb['fpr']*100:.1f}%
|
||
FNR = {_gb['fnr']*100:.1f}% | PPV = {_gb['ppv']*100:.1f}%
|
||
Accuracy = {_gb['accuracy']*100:.1f}%
|
||
```
|
||
|
||
### Gap Summary
|
||
|
||
| Metric | Group A | Group B | Gap | Status |
|
||
|--------|---------|---------|-----|--------|
|
||
| Accuracy | {_ga['accuracy']*100:.1f}% | {_gb['accuracy']*100:.1f}% | {_acc_gap:.1f}pp | {_gap_label(_acc_gap)} |
|
||
| FPR | {_ga['fpr']*100:.1f}% | {_gb['fpr']*100:.1f}% | {_fpr_gap:.1f}pp | {_gap_label(_fpr_gap)} |
|
||
| FNR | {_ga['fnr']*100:.1f}% | {_gb['fnr']*100:.1f}% | {_fnr_gap:.1f}pp | {_gap_label(_fnr_gap)} |
|
||
| PPV | {_ga['ppv']*100:.1f}% | {_gb['ppv']*100:.1f}% | {_ppv_gap:.1f}pp | {_gap_label(_ppv_gap)} |
|
||
| Equalized Odds Gap | — | — | {_eo_gap:.1f}pp | {_gap_label(_eo_gap)} |
|
||
"""
|
||
|
||
# ── Metric cards ──────────────────────────────────────────────────────────
|
||
_cards_html = f"""
|
||
<div style="display: flex; gap: 16px; flex-wrap: wrap; justify-content: center; margin: 16px 0;">
|
||
<div style="padding: 18px 22px; border: 1px solid #e2e8f0; border-radius: 10px;
|
||
width: 160px; text-align: center; background: white;">
|
||
<div style="color: #64748b; font-size: 0.8rem; margin-bottom: 4px;">Accuracy Gap</div>
|
||
<div style="font-size: 1.9rem; font-weight: 800; color: {_gap_color(_acc_gap)};">
|
||
{_acc_gap:.1f}pp
|
||
</div>
|
||
<div style="font-size: 0.72rem; color: #94a3b8;">{_gap_label(_acc_gap)}</div>
|
||
</div>
|
||
<div style="padding: 18px 22px; border: 1px solid #e2e8f0; border-radius: 10px;
|
||
width: 160px; text-align: center; background: white;">
|
||
<div style="color: #64748b; font-size: 0.8rem; margin-bottom: 4px;">FPR Gap</div>
|
||
<div style="font-size: 1.9rem; font-weight: 800; color: {_gap_color(_fpr_gap)};">
|
||
{_fpr_gap:.1f}pp
|
||
</div>
|
||
<div style="font-size: 0.72rem; color: #94a3b8;">{_gap_label(_fpr_gap)}</div>
|
||
</div>
|
||
<div style="padding: 18px 22px; border: 1px solid #e2e8f0; border-radius: 10px;
|
||
width: 160px; text-align: center; background: white;">
|
||
<div style="color: #64748b; font-size: 0.8rem; margin-bottom: 4px;">FNR Gap</div>
|
||
<div style="font-size: 1.9rem; font-weight: 800; color: {_gap_color(_fnr_gap)};">
|
||
{_fnr_gap:.1f}pp
|
||
</div>
|
||
<div style="font-size: 0.72rem; color: #94a3b8;">{_gap_label(_fnr_gap)}</div>
|
||
</div>
|
||
<div style="padding: 18px 22px; border: 1px solid #e2e8f0; border-radius: 10px;
|
||
width: 160px; text-align: center; background: white;">
|
||
<div style="color: #64748b; font-size: 0.8rem; margin-bottom: 4px;">PPV Gap</div>
|
||
<div style="font-size: 1.9rem; font-weight: 800; color: {_gap_color(_ppv_gap)};">
|
||
{_ppv_gap:.1f}pp
|
||
</div>
|
||
<div style="font-size: 0.72rem; color: #94a3b8;">{_gap_label(_ppv_gap)}</div>
|
||
</div>
|
||
<div style="padding: 18px 22px; border: 1px solid #e2e8f0; border-radius: 10px;
|
||
width: 160px; text-align: center; background: white;">
|
||
<div style="color: #64748b; font-size: 0.8rem; margin-bottom: 4px;">EO Gap</div>
|
||
<div style="font-size: 1.9rem; font-weight: 800; color: {_gap_color(_eo_gap)};">
|
||
{_eo_gap:.1f}pp
|
||
</div>
|
||
<div style="font-size: 0.72rem; color: #94a3b8;">avg FPR+FNR gap</div>
|
||
</div>
|
||
</div>
|
||
"""
|
||
|
||
mo.vstack([
|
||
mo.md(_physics_md),
|
||
mo.Html(_cards_html),
|
||
mo.ui.plotly(_fig),
|
||
mo.ui.plotly(_fig2),
|
||
])
|
||
return (
|
||
_ga, _gb,
|
||
_fpr_gap, _fnr_gap, _ppv_gap, _acc_gap, _eo_gap,
|
||
_br_a, _br_b, _t,
|
||
)
|
||
|
||
|
||
# ── CELL 9: ACT I MATH PEEK ───────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
mo.accordion({
|
||
"The governing equations (Chouldechova 2017 impossibility)": mo.md("""
|
||
**Fairness metric definitions:**
|
||
|
||
Let `r` = base rate for a group, `t` = model threshold.
|
||
For a calibrated classifier with score distribution `s`:
|
||
|
||
```
|
||
TPR (sensitivity) = P(s > t | Y = 1)
|
||
FPR (false alarm) = P(s > t | Y = 0)
|
||
FNR (miss rate) = 1 - TPR
|
||
PPV (precision) = P(Y = 1 | s > t)
|
||
= r · TPR / (r · TPR + (1-r) · FPR)
|
||
Accuracy = r · TPR + (1-r) · (1 - FPR)
|
||
```
|
||
|
||
**Chouldechova (2017) Impossibility Theorem:**
|
||
|
||
When `r_A ≠ r_B` (base rates differ between groups A and B),
|
||
a single shared model threshold `t` applied to a calibrated classifier
|
||
**cannot simultaneously satisfy all three of:**
|
||
|
||
```
|
||
1. Calibration: PPV_A = PPV_B
|
||
2. Equal FPR: FPR_A = FPR_B
|
||
3. Equal FNR: FNR_A = FNR_B
|
||
```
|
||
|
||
**Proof sketch (from the paper):**
|
||
|
||
From the PPV formula: `FPR = r · TPR / (PPV · (1-r)) - r · TPR / (1-r)`
|
||
|
||
If PPV is equal across groups and TPR (= 1 - FNR) is equal across groups,
|
||
then FPR can only be equal if `r_A = r_B`. Since base rates differ by
|
||
assumption, at least one of the three equalities must be violated.
|
||
|
||
**Implication for the lab:**
|
||
|
||
Equal accuracy is consistent with large FPR or FNR gaps — because accuracy
|
||
weighs positives and negatives by their base-rate frequencies, which differ
|
||
by group. Two groups can have identical accuracy (85%) while one group's
|
||
false alarm rate is many times higher than the other's.
|
||
"""),
|
||
})
|
||
return
|
||
|
||
|
||
# ── CELL 10: ACT I PREDICTION-VS-REALITY OVERLAY ──────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo, act1_prediction, _acc_gap, _fpr_gap, _fnr_gap):
|
||
# The key insight: equal accuracy does NOT imply equal FPR or FNR
|
||
# The correct answer is C: must check FPR AND FNR separately.
|
||
_a1_correct = act1_prediction.value == "option_c"
|
||
|
||
_explanations = {
|
||
"option_a": (
|
||
"**Incorrect.** Equal accuracy is not sufficient for fairness. "
|
||
"When base rates differ, the same accuracy score can conceal large "
|
||
"asymmetries in who bears the false positive vs. false negative burden. "
|
||
"Accuracy weights each type of error by base-rate frequency, so groups "
|
||
"with different base rates can score identically in accuracy while "
|
||
"experiencing very different error rates. "
|
||
f"In your current configuration: FPR gap = **{_fpr_gap:.1f}pp**, "
|
||
f"FNR gap = **{_fnr_gap:.1f}pp**, Accuracy gap = **{_acc_gap:.1f}pp**.",
|
||
"warn"
|
||
),
|
||
"option_b": (
|
||
"**Partially correct, but incomplete.** Checking precision (PPV) is "
|
||
"important, but precision alone does not capture the full picture. "
|
||
"Chouldechova (2017) shows the impossibility involves FPR AND FNR "
|
||
"jointly — checking only PPV leaves out one of the key error asymmetries. "
|
||
"The complete answer requires checking FPR and FNR separately.",
|
||
"warn"
|
||
),
|
||
"option_c": (
|
||
"**Correct.** This is exactly the lesson Chouldechova (2017) formalizes. "
|
||
"Equal accuracy is necessary but not sufficient. When base rates differ, "
|
||
"the same model threshold produces structurally different FPR and FNR across "
|
||
"groups — regardless of overall accuracy. The compliance officer needs to "
|
||
"audit error rates *by group*, not just accuracy *by group*. "
|
||
f"In the default configuration (base rate A=10%, B=40%): "
|
||
f"FPR gap = {_fpr_gap:.1f}pp, FNR gap = {_fnr_gap:.1f}pp, "
|
||
f"Accuracy gap = {_acc_gap:.1f}pp.",
|
||
"success"
|
||
),
|
||
"option_d": (
|
||
"**Incorrect.** Accuracy is not the standard that disparate impact doctrine "
|
||
"applies. Courts and regulators examine selection rates, false positive rates, "
|
||
"and the four-fifths rule — none of which are equivalent to overall accuracy. "
|
||
"Equal accuracy while maintaining a 20pp FPR gap would almost certainly "
|
||
"not satisfy a disparate impact claim.",
|
||
"warn"
|
||
),
|
||
}
|
||
|
||
_explanation, _kind = _explanations.get(
|
||
act1_prediction.value, ("No prediction selected.", "info")
|
||
)
|
||
|
||
_overlay_md = (
|
||
f"**Prediction-vs-Reality:** You predicted *{act1_prediction.value.upper().replace('_', ' ')}*. "
|
||
f"Current configuration: accuracy gap = {_acc_gap:.1f}pp, FPR gap = {_fpr_gap:.1f}pp, "
|
||
f"FNR gap = {_fnr_gap:.1f}pp. "
|
||
"Notice that accuracy gap can be near zero while FPR and FNR gaps remain substantial."
|
||
)
|
||
|
||
mo.vstack([
|
||
mo.callout(mo.md(_explanation), kind=_kind),
|
||
mo.callout(mo.md(_overlay_md), kind="info"),
|
||
])
|
||
return (_a1_correct,)
|
||
|
||
|
||
# ── CELL 11: ACT I REFLECTION ─────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
act1_reflection = mo.ui.radio(
|
||
options={
|
||
"A) Perfect fairness is always achievable with enough data and a better model": "ref_a",
|
||
"B) When base rates differ, you cannot simultaneously achieve calibration, equal FPR, AND equal FNR": "ref_b",
|
||
"C) Accuracy and fairness are always in tension (too general — not the theorem's claim)": "ref_c",
|
||
"D) Fairness metrics are mathematically equivalent and differ only in emphasis": "ref_d",
|
||
},
|
||
label="Reflection: What does Chouldechova's (2017) impossibility theorem state?",
|
||
)
|
||
mo.vstack([
|
||
mo.md("---"),
|
||
mo.md("### Act I Reflection"),
|
||
act1_reflection,
|
||
])
|
||
return (act1_reflection,)
|
||
|
||
|
||
# ── CELL 12: ACT I REFLECTION REVEAL ─────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo, act1_reflection):
|
||
mo.stop(
|
||
act1_reflection.value is None,
|
||
mo.callout(mo.md("Select your reflection answer to continue to Act II."), kind="warn"),
|
||
)
|
||
|
||
_ref_explanations = {
|
||
"ref_a": (
|
||
"**Incorrect.** Chouldechova (2017) proves this is impossible regardless of "
|
||
"data quantity or model quality, as long as base rates differ. More data does "
|
||
"not fix the structural incompatibility — it only makes the model more "
|
||
"precisely wrong in the same direction.",
|
||
"warn"
|
||
),
|
||
"ref_b": (
|
||
"**Correct.** This is the exact claim of the theorem. For any calibrated "
|
||
"classifier applied with a shared threshold: if `r_A ≠ r_B`, then at most "
|
||
"two of {calibration, equal FPR, equal FNR} can hold simultaneously. "
|
||
"There is no engineering solution — only a choice of which criterion to "
|
||
"prioritize and which to sacrifice.",
|
||
"success"
|
||
),
|
||
"ref_c": (
|
||
"**Too general.** The statement that accuracy and fairness are always in "
|
||
"tension is a common heuristic, but it is not the theorem. Chouldechova's "
|
||
"contribution is more precise: it identifies the specific three-way "
|
||
"incompatibility and the exact condition (differing base rates) that triggers it.",
|
||
"warn"
|
||
),
|
||
"ref_d": (
|
||
"**Incorrect.** The entire point of the theorem is that fairness metrics "
|
||
"are NOT equivalent. Demographic parity, equalized odds, and calibration "
|
||
"each capture different aspects of fair treatment, and they actively "
|
||
"conflict when base rates differ. Choosing a metric is a policy decision, "
|
||
"not a technical one.",
|
||
"warn"
|
||
),
|
||
}
|
||
|
||
_expl, _kind = _ref_explanations.get(
|
||
act1_reflection.value, ("No answer selected.", "info")
|
||
)
|
||
mo.callout(mo.md(_expl), kind=_kind)
|
||
return
|
||
|
||
|
||
# ─────────────────────────────────────────────────────────────────────────────
|
||
# ACT II — THE AUDIT-ACCURACY TRADEOFF
|
||
# ─────────────────────────────────────────────────────────────────────────────
|
||
|
||
# ── CELL 13: ACT II SCENARIO ──────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo, COLORS):
|
||
_color = COLORS["BlueLine"]
|
||
mo.vstack([
|
||
mo.md("---"),
|
||
mo.md("## Act II — The Audit-Accuracy Tradeoff"),
|
||
mo.Html(f"""
|
||
<div style="border-left: 4px solid {_color}; background: #eff6ff;
|
||
border-radius: 0 10px 10px 0; padding: 16px 22px; margin: 12px 0;">
|
||
<div style="font-size: 0.72rem; font-weight: 700; color: {_color};
|
||
text-transform: uppercase; letter-spacing: 0.1em; margin-bottom: 6px;">
|
||
Incoming Message · Engineering VP
|
||
</div>
|
||
<div style="font-style: italic; font-size: 1.0rem; color: #1e293b; line-height: 1.65;">
|
||
"We're deploying a resume screening model at scale — 50,000 applications
|
||
per quarter. We need 80% recall on qualified candidates. We have $50K
|
||
in annual audit budget. Three options are on the table: (1) accuracy-optimized
|
||
model — no fairness constraint, maximize recall; (2) demographic parity
|
||
constraint — equal selection rates across demographic groups; (3) individual
|
||
fairness with equalized odds — similar candidates treated similarly, equal
|
||
FPR and FNR across groups. Design the deployment. Which strategy is most
|
||
defensible, and does it fit the budget?"
|
||
</div>
|
||
</div>
|
||
"""),
|
||
mo.md("""
|
||
Each strategy has a different accuracy cost, audit cost, and legal exposure.
|
||
The VP needs a defensible choice — not just the highest accuracy.
|
||
Commit to a strategy prediction before using the design instruments.
|
||
"""),
|
||
])
|
||
return
|
||
|
||
|
||
# ── CELL 14: ACT II PREDICTION LOCK ───────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
act2_prediction = mo.ui.radio(
|
||
options={
|
||
"A) Option 1 — maximize accuracy, handle fairness concerns in post-hoc legal review": "option_a",
|
||
"B) Option 2 — demographic parity is the legally safest strategy for a hiring context": "option_b",
|
||
"C) Option 3 — individual fairness with equalized odds and regular audit is most defensible": "option_c",
|
||
"D) All three strategies are equivalent in practice — the choice is arbitrary": "option_d",
|
||
},
|
||
label=(
|
||
"Which deployment strategy is most defensible for a high-stakes hiring context "
|
||
"under EEOC and disparate impact doctrine?"
|
||
),
|
||
)
|
||
mo.vstack([
|
||
mo.Html("""
|
||
<div style="background: #1e293b; border-radius: 12px; padding: 20px;
|
||
border-left: 4px solid #6366f1; margin: 8px 0;">
|
||
<div style="font-size: 0.72rem; font-weight: 700; color: #a5b4fc;
|
||
text-transform: uppercase; letter-spacing: 0.1em; margin-bottom: 10px;">
|
||
Prediction Lock — Act II
|
||
</div>
|
||
<div style="color: #e2e8f0; font-size: 0.88rem; margin-bottom: 12px;">
|
||
Commit before adjusting the design instruments.
|
||
</div>
|
||
</div>
|
||
"""),
|
||
act2_prediction,
|
||
])
|
||
return (act2_prediction,)
|
||
|
||
|
||
# ── CELL 15: ACT II GATE ──────────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo, act2_prediction):
|
||
mo.stop(
|
||
act2_prediction.value is None,
|
||
mo.callout(
|
||
mo.md("Select your strategy prediction above to unlock the Audit Design Cockpit."),
|
||
kind="warn",
|
||
),
|
||
)
|
||
return
|
||
|
||
|
||
# ── CELL 16: ACT II CONTROLS ──────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
# Fairness criterion selection
|
||
# Source: responsible_engr.qmd — the three main criteria for hiring contexts
|
||
fairness_criterion = mo.ui.dropdown(
|
||
options={
|
||
"No constraint (accuracy only)": "accuracy",
|
||
"Demographic parity (equal selection rates)": "demographic_parity",
|
||
"Equalized odds (equal FPR + FNR)": "equalized_odds",
|
||
"Individual fairness (similar candidates treated similarly)": "individual",
|
||
},
|
||
value="No constraint (accuracy only)",
|
||
label="Fairness criterion",
|
||
)
|
||
# Bias mitigation method
|
||
# Source: responsible_engr.qmd — three standard mitigation approaches
|
||
bias_mitigation = mo.ui.dropdown(
|
||
options={
|
||
"None": "none",
|
||
"Reweighting (pre-processing)": "reweighting",
|
||
"Adversarial debiasing (in-processing)": "adversarial",
|
||
"Post-processing threshold adjustment": "postprocessing",
|
||
},
|
||
value="None",
|
||
label="Bias mitigation method",
|
||
)
|
||
# Audit frequency: how often the deployed model is re-evaluated for bias
|
||
# Source: responsible_engr.qmd — audit cost model
|
||
audit_frequency = mo.ui.dropdown(
|
||
options={
|
||
"Continuous (52×/yr)": "continuous",
|
||
"Monthly (12×/yr)": "monthly",
|
||
"Quarterly (4×/yr)": "quarterly",
|
||
"Annual (1×/yr)": "annual",
|
||
},
|
||
value="Quarterly (4×/yr)",
|
||
label="Audit frequency",
|
||
)
|
||
mo.vstack([
|
||
mo.md("### Audit Design Cockpit"),
|
||
mo.md(
|
||
"Select a **fairness criterion**, a **bias mitigation method**, and an "
|
||
"**audit frequency**. The instruments will show the resulting accuracy cost, "
|
||
"equalized odds gap, annual audit cost, and regulatory risk level."
|
||
),
|
||
mo.hstack([fairness_criterion, bias_mitigation, audit_frequency],
|
||
justify="start", gap="2rem"),
|
||
])
|
||
return (fairness_criterion, bias_mitigation, audit_frequency)
|
||
|
||
|
||
# ── CELL 17: ACT II PHYSICS ENGINE ────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(
|
||
mo, go, np, math,
|
||
context_toggle,
|
||
fairness_criterion, bias_mitigation, audit_frequency,
|
||
_br_a, _br_b,
|
||
DISPARATE_IMPACT_THRESHOLD_PP,
|
||
CLOUD_AUDIT_COST_K_USD, MOBILE_AUDIT_COST_K_USD,
|
||
apply_plotly_theme,
|
||
):
|
||
# ── Context-dependent parameters ──────────────────────────────────────────
|
||
_ctx = context_toggle.value # "cloud" or "mobile"
|
||
_cost_per_audit_k = CLOUD_AUDIT_COST_K_USD if _ctx == "cloud" else MOBILE_AUDIT_COST_K_USD
|
||
|
||
# ── Audit frequency → runs per year ───────────────────────────────────────
|
||
_freq_map = {"continuous": 52, "monthly": 12, "quarterly": 4, "annual": 1}
|
||
_audits_per_year = _freq_map[audit_frequency.value]
|
||
|
||
# ── Fairness criterion: base accuracy and baseline EO gap ─────────────────
|
||
# Source: responsible_engr.qmd — Table of fairness-accuracy tradeoffs
|
||
# These are physics-based estimates calibrated to a base rate gap of 30pp
|
||
# (Group A=10%, Group B=40%) from Act I defaults.
|
||
#
|
||
# Accuracy-only model: maximizes recall, no fairness constraint.
|
||
# EO gap is large because threshold optimized for accuracy, not parity.
|
||
#
|
||
# Demographic parity: forces equal selection rates.
|
||
# This requires different thresholds per group → can harm accuracy more.
|
||
# EO gap may actually WORSEN because FPR/FNR are not directly controlled.
|
||
#
|
||
# Equalized odds: forces equal FPR and FNR.
|
||
# Requires per-group threshold calibration → moderate accuracy cost.
|
||
# EO gap is directly minimized.
|
||
#
|
||
# Individual fairness: similar candidates get similar scores.
|
||
# Requires metric learning → highest implementation cost, best defensibility.
|
||
# EO gap depends on implementation quality.
|
||
_base_rate_gap_pp = abs(_br_b - _br_a) # gap between group base rates
|
||
|
||
# Base accuracy and EO gap by criterion
|
||
# EO gap is modeled as a function of base rate gap and constraint choice
|
||
_criterion = fairness_criterion.value
|
||
if _criterion == "accuracy":
|
||
_base_accuracy_pct = 87.5 # unconstrained, high accuracy
|
||
_base_eo_gap = 0.38 * _base_rate_gap_pp # large EO gap when base rates differ
|
||
elif _criterion == "demographic_parity":
|
||
_base_accuracy_pct = 83.0 # accuracy cost from forcing equal selection rates
|
||
_base_eo_gap = 0.50 * _base_rate_gap_pp # DP can WORSEN EO gap (see theorem)
|
||
elif _criterion == "equalized_odds":
|
||
_base_accuracy_pct = 84.5 # moderate accuracy cost
|
||
_base_eo_gap = 0.12 * _base_rate_gap_pp # EO gap directly constrained
|
||
else: # individual
|
||
_base_accuracy_pct = 84.0 # similar to EO
|
||
_base_eo_gap = 0.10 * _base_rate_gap_pp # best EO gap, defensible in court
|
||
|
||
# ── Bias mitigation: reduces EO gap further, with additional accuracy cost ─
|
||
# Source: responsible_engr.qmd — mitigation method effectiveness table
|
||
_mitigation = bias_mitigation.value
|
||
_mitigation_gap_reduction = {
|
||
"none": 0.0,
|
||
"reweighting": 0.25, # reduces gap by 25%, accuracy cost ~0.5pp
|
||
"adversarial": 0.45, # reduces gap by 45%, accuracy cost ~1.5pp
|
||
"postprocessing": 0.35, # reduces gap by 35%, accuracy cost ~0.8pp
|
||
}[_mitigation]
|
||
_mitigation_acc_cost = {
|
||
"none": 0.0,
|
||
"reweighting": 0.5,
|
||
"adversarial": 1.5,
|
||
"postprocessing": 0.8,
|
||
}[_mitigation]
|
||
|
||
_final_eo_gap = _base_eo_gap * (1.0 - _mitigation_gap_reduction)
|
||
_final_accuracy_pct = _base_accuracy_pct - _mitigation_acc_cost
|
||
|
||
# ── Audit cost model ───────────────────────────────────────────────────────
|
||
# Source: responsible_engr.qmd — audit cost = compute + human review
|
||
# Mobile has 2.7× higher cost due to data collection and manual review overhead
|
||
_audit_cost_k = _cost_per_audit_k * _audits_per_year # total annual audit cost ($K)
|
||
|
||
# ── Regulatory risk model ─────────────────────────────────────────────────
|
||
# Source: responsible_engr.qmd — EEOC disparate impact guidelines
|
||
# Risk is a function of EO gap AND audit frequency
|
||
# An unaudited high-gap model is maximum risk; well-audited low-gap model is low risk
|
||
_above_threshold = _final_eo_gap > DISPARATE_IMPACT_THRESHOLD_PP
|
||
_audits_adequate = _audits_per_year >= 4 # quarterly minimum per OFCCP guidance
|
||
_disparate_impact_triggered = _above_threshold # the failure state
|
||
|
||
if not _above_threshold:
|
||
_reg_risk = "low"
|
||
elif _above_threshold and _audits_adequate:
|
||
_reg_risk = "medium"
|
||
else:
|
||
_reg_risk = "high"
|
||
|
||
# ── Budget check ──────────────────────────────────────────────────────────
|
||
_budget_k = 50.0 # $50K annual audit budget (from VP scenario)
|
||
_over_budget = _audit_cost_k > _budget_k
|
||
|
||
# ── Recall check ─────────────────────────────────────────────────────────
|
||
# VP requires 80% recall on qualified candidates
|
||
_recall_below_80 = _final_accuracy_pct < 80.0
|
||
|
||
# ── Color coding ─────────────────────────────────────────────────────────
|
||
def _color_for_gap(g):
|
||
if g <= 5.0:
|
||
return "#22c55e"
|
||
elif g <= 10.0:
|
||
return "#f59e0b"
|
||
return "#ef4444"
|
||
|
||
def _color_for_risk(risk):
|
||
return {"low": "#22c55e", "medium": "#f59e0b", "high": "#ef4444"}[risk]
|
||
|
||
_gap_color = _color_for_gap(_final_eo_gap)
|
||
_risk_color = _color_for_risk(_reg_risk)
|
||
_budget_color = "#22c55e" if not _over_budget else "#ef4444"
|
||
_acc_color = "#22c55e" if _final_accuracy_pct >= 84.0 else "#f59e0b" if _final_accuracy_pct >= 80.0 else "#ef4444"
|
||
|
||
# ── Pareto frontier: sweep all criterion × mitigation combinations ────────
|
||
_frontier_configs = [
|
||
("accuracy + none", 87.5, 0.38 * _base_rate_gap_pp * 1.00),
|
||
("accuracy + reweight", 87.0, 0.38 * _base_rate_gap_pp * 0.75),
|
||
("accuracy + adversarial",86.0, 0.38 * _base_rate_gap_pp * 0.55),
|
||
("accuracy + postproc", 86.7, 0.38 * _base_rate_gap_pp * 0.65),
|
||
("dem_parity + none", 83.0, 0.50 * _base_rate_gap_pp * 1.00),
|
||
("dem_parity + reweight", 82.5, 0.50 * _base_rate_gap_pp * 0.75),
|
||
("dem_parity + adversarial", 81.5, 0.50 * _base_rate_gap_pp * 0.55),
|
||
("dem_parity + postproc", 82.2, 0.50 * _base_rate_gap_pp * 0.65),
|
||
("eq_odds + none", 84.5, 0.12 * _base_rate_gap_pp * 1.00),
|
||
("eq_odds + reweight", 84.0, 0.12 * _base_rate_gap_pp * 0.75),
|
||
("eq_odds + adversarial", 83.0, 0.12 * _base_rate_gap_pp * 0.55),
|
||
("eq_odds + postproc", 83.7, 0.12 * _base_rate_gap_pp * 0.65),
|
||
("individual + none", 84.0, 0.10 * _base_rate_gap_pp * 1.00),
|
||
("individual + reweight", 83.5, 0.10 * _base_rate_gap_pp * 0.75),
|
||
("individual + adversarial", 82.5, 0.10 * _base_rate_gap_pp * 0.55),
|
||
("individual + postproc", 83.2, 0.10 * _base_rate_gap_pp * 0.65),
|
||
]
|
||
|
||
# Group colors for Pareto plot
|
||
_group_colors = {
|
||
"accuracy": "#94a3b8", # gray
|
||
"dem_parity": "#f59e0b", # amber
|
||
"eq_odds": "#6366f1", # indigo
|
||
"individual": "#22c55e", # green
|
||
}
|
||
|
||
_fig_pareto = go.Figure()
|
||
for _name, _acc, _gap in _frontier_configs:
|
||
_grp = _name.split(" + ")[0].replace("dem_parity", "dem_parity")
|
||
_grp_short = _grp.split("_")[0] if "_" not in _grp[:5] else (
|
||
"dem_parity" if "dem" in _grp else _grp
|
||
)
|
||
_grp_color = _group_colors.get(
|
||
"dem_parity" if "dem" in _name else (
|
||
"eq_odds" if "eq" in _name else (
|
||
"individual" if "individual" in _name else "accuracy"
|
||
)
|
||
), "#6366f1"
|
||
)
|
||
_fig_pareto.add_trace(go.Scatter(
|
||
x=[_acc], y=[_gap],
|
||
mode="markers",
|
||
name=_name,
|
||
marker=dict(color=_grp_color, size=10),
|
||
showlegend=False,
|
||
))
|
||
|
||
# Highlight current selection
|
||
_fig_pareto.add_trace(go.Scatter(
|
||
x=[_final_accuracy_pct], y=[_final_eo_gap],
|
||
mode="markers",
|
||
name="Your design",
|
||
marker=dict(color="#ffffff", size=16, symbol="star",
|
||
line=dict(color="#f59e0b", width=2)),
|
||
))
|
||
|
||
# EEOC threshold line
|
||
_fig_pareto.add_hline(
|
||
y=DISPARATE_IMPACT_THRESHOLD_PP,
|
||
line_dash="dash", line_color="#ef4444",
|
||
annotation_text=f"EEOC guidance threshold ({DISPARATE_IMPACT_THRESHOLD_PP:.0f}pp)",
|
||
annotation_position="top right",
|
||
)
|
||
_fig_pareto = apply_plotly_theme(_fig_pareto)
|
||
_fig_pareto.update_layout(
|
||
title="Fairness-Accuracy Pareto Frontier",
|
||
xaxis_title="Model Accuracy (%)",
|
||
yaxis_title="Equalized Odds Gap (pp) — lower is better",
|
||
height=380,
|
||
)
|
||
|
||
# ── Metric cards HTML ────────────────────────────────────────────────────
|
||
_ctx_display = "H100 Cloud" if _ctx == "cloud" else "Mobile NPU"
|
||
_cards = f"""
|
||
<div style="display: flex; gap: 16px; flex-wrap: wrap; justify-content: center; margin: 16px 0;">
|
||
<div style="padding: 18px 22px; border: 1px solid #e2e8f0; border-radius: 10px;
|
||
width: 165px; text-align: center; background: white;">
|
||
<div style="color: #64748b; font-size: 0.8rem; margin-bottom: 4px;">Accuracy</div>
|
||
<div style="font-size: 1.9rem; font-weight: 800; color: {_acc_color};">
|
||
{_final_accuracy_pct:.1f}%
|
||
</div>
|
||
<div style="font-size: 0.72rem; color: #94a3b8;">recall target 80%</div>
|
||
</div>
|
||
<div style="padding: 18px 22px; border: 1px solid #e2e8f0; border-radius: 10px;
|
||
width: 165px; text-align: center; background: white;">
|
||
<div style="color: #64748b; font-size: 0.8rem; margin-bottom: 4px;">EO Gap</div>
|
||
<div style="font-size: 1.9rem; font-weight: 800; color: {_gap_color};">
|
||
{_final_eo_gap:.1f}pp
|
||
</div>
|
||
<div style="font-size: 0.72rem; color: #94a3b8;">threshold: 10pp</div>
|
||
</div>
|
||
<div style="padding: 18px 22px; border: 1px solid #e2e8f0; border-radius: 10px;
|
||
width: 165px; text-align: center; background: white;">
|
||
<div style="color: #64748b; font-size: 0.8rem; margin-bottom: 4px;">Audit Cost / yr</div>
|
||
<div style="font-size: 1.9rem; font-weight: 800; color: {_budget_color};">
|
||
${_audit_cost_k:.0f}K
|
||
</div>
|
||
<div style="font-size: 0.72rem; color: #94a3b8;">budget: $50K ({_ctx_display})</div>
|
||
</div>
|
||
<div style="padding: 18px 22px; border: 1px solid #e2e8f0; border-radius: 10px;
|
||
width: 165px; text-align: center; background: white;">
|
||
<div style="color: #64748b; font-size: 0.8rem; margin-bottom: 4px;">Regulatory Risk</div>
|
||
<div style="font-size: 1.9rem; font-weight: 800; color: {_risk_color};">
|
||
{_reg_risk.upper()}
|
||
</div>
|
||
<div style="font-size: 0.72rem; color: #94a3b8;">EEOC / OFCCP</div>
|
||
</div>
|
||
</div>
|
||
"""
|
||
|
||
# ── Physics summary ───────────────────────────────────────────────────────
|
||
_phys_md = f"""
|
||
### Design Summary
|
||
|
||
```
|
||
Criterion: {fairness_criterion.value}
|
||
Mitigation: {bias_mitigation.value}
|
||
Audit freq: {audit_frequency.value} ({_audits_per_year}×/yr)
|
||
Context: {_ctx_display}
|
||
|
||
Accuracy = {_final_accuracy_pct:.1f}%
|
||
(base {_base_accuracy_pct:.1f}% − mitigation cost {_mitigation_acc_cost:.1f}pp)
|
||
|
||
EO Gap = {_final_eo_gap:.1f}pp
|
||
(base {_base_eo_gap:.1f}pp × (1 − {_mitigation_gap_reduction:.0%} mitigation))
|
||
|
||
Audit cost / yr = ${_cost_per_audit_k:.1f}K × {_audits_per_year} = ${_audit_cost_k:.1f}K
|
||
Budget remaining = ${ _budget_k - _audit_cost_k:+.1f}K
|
||
|
||
Regulatory risk = {_reg_risk.upper()}
|
||
(EO gap > {DISPARATE_IMPACT_THRESHOLD_PP:.0f}pp: {_above_threshold} | audits >= quarterly: {_audits_adequate})
|
||
```
|
||
"""
|
||
|
||
# ── Output ────────────────────────────────────────────────────────────────
|
||
_output_items = [
|
||
mo.md(_phys_md),
|
||
mo.Html(_cards),
|
||
mo.ui.plotly(_fig_pareto),
|
||
]
|
||
|
||
# ── FAILURE STATE: Disparate Impact triggered ─────────────────────────────
|
||
# Source: EEOC 4/5ths rule and OFCCP audit guidance
|
||
# Triggered when equalized odds gap > 10pp in a high-stakes (hiring) context
|
||
if _disparate_impact_triggered:
|
||
_output_items.append(
|
||
mo.callout(
|
||
mo.md(
|
||
f"**Disparate impact threshold exceeded.** "
|
||
f"Equalized odds gap: **{_final_eo_gap:.1f}pp**. "
|
||
f"High-stakes deployment (hiring) requires gap < {DISPARATE_IMPACT_THRESHOLD_PP:.0f}pp "
|
||
f"per EEOC guidelines. This deployment would face substantial regulatory risk. "
|
||
f"Reduce the gap by switching to equalized odds or individual fairness criterion, "
|
||
f"or apply adversarial debiasing to bring the gap below threshold."
|
||
),
|
||
kind="danger",
|
||
)
|
||
)
|
||
else:
|
||
_output_items.append(
|
||
mo.callout(
|
||
mo.md(
|
||
f"EO gap **{_final_eo_gap:.1f}pp** is below the {DISPARATE_IMPACT_THRESHOLD_PP:.0f}pp "
|
||
f"threshold. Regulatory risk is **{_reg_risk}**. "
|
||
+ ("Audit frequency meets OFCCP quarterly minimum." if _audits_adequate else
|
||
"Warning: audit frequency below OFCCP quarterly minimum recommendation.")
|
||
),
|
||
kind="success" if _reg_risk == "low" else "warn",
|
||
)
|
||
)
|
||
|
||
# ── Budget warning ────────────────────────────────────────────────────────
|
||
if _over_budget:
|
||
_output_items.append(
|
||
mo.callout(
|
||
mo.md(
|
||
f"**Budget exceeded.** Annual audit cost **${_audit_cost_k:.0f}K** > "
|
||
f"${_budget_k:.0f}K budget on **{_ctx_display}**. "
|
||
"Reduce audit frequency or switch to cloud context for lower per-run cost."
|
||
),
|
||
kind="warn",
|
||
)
|
||
)
|
||
|
||
mo.vstack(_output_items)
|
||
return (
|
||
_final_eo_gap,
|
||
_final_accuracy_pct,
|
||
_audit_cost_k,
|
||
_reg_risk,
|
||
_disparate_impact_triggered,
|
||
_above_threshold,
|
||
_criterion,
|
||
)
|
||
|
||
|
||
# ── CELL 18: ACT II MATH PEEK ─────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
mo.accordion({
|
||
"The governing equations (fairness metrics and audit cost model)": mo.md("""
|
||
**Equalized Odds (Hardt et al. 2016):**
|
||
|
||
A classifier satisfies equalized odds if, for all groups `g` and labels `y ∈ {0,1}`:
|
||
```
|
||
P(ŷ = 1 | Y = y, Group = g) is equal across all groups g
|
||
```
|
||
Equivalently: both TPR and FPR are equal across groups.
|
||
|
||
**Equalized Odds Gap (lab's operationalization):**
|
||
```
|
||
EO_gap = (|FPR_A − FPR_B| + |FNR_A − FNR_B|) / 2
|
||
```
|
||
|
||
**Demographic Parity:**
|
||
```
|
||
DP: P(ŷ = 1 | Group = A) = P(ŷ = 1 | Group = B)
|
||
```
|
||
Note: DP forces equal selection rates, NOT equal error rates.
|
||
When qualified candidate rates differ, DP can require REJECTING more
|
||
qualified candidates from one group to equalize selection rates —
|
||
which is its own form of unfairness.
|
||
|
||
**Individual Fairness (Dwork et al. 2012 — Lipschitz condition):**
|
||
```
|
||
d_Y(M(x), M(x')) ≤ L · d_X(x, x')
|
||
```
|
||
Individuals at similar distances in feature space receive similar decisions.
|
||
Requires a task-specific metric `d_X` — the hardest criterion to specify
|
||
but the most defensible under equal treatment theory.
|
||
|
||
**Audit Cost Model:**
|
||
```
|
||
Annual audit cost = cost_per_run × runs_per_year
|
||
Cloud: cost_per_run ≈ $1,500 (automated, H100 batch evaluation)
|
||
Mobile: cost_per_run ≈ $4,000 (manual data collection + compute)
|
||
```
|
||
|
||
**Regulatory Risk Model (EEOC / OFCCP guidance):**
|
||
```
|
||
If EO_gap > 10pp AND context is high-stakes: HIGH risk
|
||
If EO_gap > 10pp AND quarterly+ audits: MEDIUM risk
|
||
If EO_gap ≤ 10pp: LOW risk
|
||
```
|
||
The 10pp threshold derives from EEOC's 4/5ths (80%) rule applied
|
||
to selection rate ratios, converted to equalized odds gap units.
|
||
"""),
|
||
})
|
||
return
|
||
|
||
|
||
# ── CELL 19: ACT II PREDICTION REVEAL ─────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo, act2_prediction, _final_eo_gap, _reg_risk, _criterion):
|
||
_a2_correct = act2_prediction.value == "option_c"
|
||
|
||
_a2_explanations = {
|
||
"option_a": (
|
||
"**Incorrect.** Option 1 (accuracy-only) does not avoid fairness liability — "
|
||
"it embeds it. Under EEOC disparate impact doctrine, a facially neutral "
|
||
"employment practice that causes adverse impact must be justified by "
|
||
"business necessity AND must use the least discriminatory alternative. "
|
||
"Post-hoc legal review is not a substitute for a defensible deployment design. "
|
||
f"Your current EO gap is **{_final_eo_gap:.1f}pp** — "
|
||
f"regulatory risk is **{_reg_risk}**.",
|
||
"warn"
|
||
),
|
||
"option_b": (
|
||
"**Incorrect.** Demographic parity is legally problematic in hiring contexts. "
|
||
"It forces equal selection rates regardless of qualified candidate rates. "
|
||
"If 40% of Group B applicants are qualified vs 10% of Group A, demographic "
|
||
"parity requires either rejecting many qualified Group B applicants or "
|
||
"accepting unqualified Group A applicants — both of which can constitute "
|
||
"disparate treatment claims. Equalized odds is more defensible because it "
|
||
"conditions on actual qualification (the true label), not just group membership.",
|
||
"warn"
|
||
),
|
||
"option_c": (
|
||
"**Correct.** Option 3 — individual fairness or equalized odds with regular "
|
||
"audit — is the most legally defensible approach. It satisfies equal treatment "
|
||
"theory (similar candidates get similar decisions) and can demonstrate "
|
||
"compliance through the audit record. The Pareto frontier shows equalized odds "
|
||
"achieves lower EO gap than demographic parity at similar accuracy cost, "
|
||
"and individual fairness further improves defensibility. "
|
||
f"Your current EO gap is **{_final_eo_gap:.1f}pp** — "
|
||
f"regulatory risk is **{_reg_risk}**.",
|
||
"success"
|
||
),
|
||
"option_d": (
|
||
"**Incorrect.** The three strategies produce materially different outcomes "
|
||
"on accuracy, EO gap, and regulatory risk. The Pareto frontier makes this "
|
||
"visible: demographic parity can produce WORSE EO gaps than equalized odds "
|
||
"at the same accuracy cost, because it constrains the wrong quantity "
|
||
"(selection rate, not error rate). The strategies are not equivalent.",
|
||
"warn"
|
||
),
|
||
}
|
||
|
||
_expl2, _kind2 = _a2_explanations.get(
|
||
act2_prediction.value, ("No prediction selected.", "info")
|
||
)
|
||
mo.callout(mo.md(_expl2), kind=_kind2)
|
||
return (_a2_correct,)
|
||
|
||
|
||
# ── CELL 20: ACT II REFLECTION ────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
act2_reflection = mo.ui.radio(
|
||
options={
|
||
"A) Demographic parity is never a valid fairness metric": "ref2_a",
|
||
"B) It forces equal selection rates regardless of base rates — if qualified candidates "
|
||
"differ by group, it means rejecting more qualified people in one group to hit the quota": "ref2_b",
|
||
"C) It always hurts overall accuracy by more than 10 percentage points": "ref2_c",
|
||
"D) Demographic parity violates GDPR by requiring protected attribute access": "ref2_d",
|
||
},
|
||
label=(
|
||
"Reflection: Why can demographic parity actually produce unfair outcomes "
|
||
"in a hiring context where qualified candidate rates differ by group?"
|
||
),
|
||
)
|
||
mo.vstack([
|
||
mo.md("---"),
|
||
mo.md("### Act II Reflection"),
|
||
act2_reflection,
|
||
])
|
||
return (act2_reflection,)
|
||
|
||
|
||
# ── CELL 21: ACT II REFLECTION REVEAL ────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo, act2_reflection):
|
||
mo.stop(
|
||
act2_reflection.value is None,
|
||
mo.callout(mo.md("Select your reflection answer to complete the lab."), kind="warn"),
|
||
)
|
||
|
||
_ref2_explanations = {
|
||
"ref2_a": (
|
||
"**Incorrect — too absolute.** Demographic parity is valid in some contexts, "
|
||
"particularly when base rates are approximately equal or when the historical "
|
||
"qualification gap is itself a product of discriminatory conditions. "
|
||
"The problem is applying it uncritically when base rates differ for legitimate reasons.",
|
||
"warn"
|
||
),
|
||
"ref2_b": (
|
||
"**Correct.** This is the key insight. Demographic parity equalizes the "
|
||
"outcome (selection rate) without conditioning on actual qualification. "
|
||
"If Group B has 40% qualified candidates and Group A has 10%, equal "
|
||
"selection rates require either: (a) lowering the bar for Group A "
|
||
"(accepting unqualified applicants) or (b) raising the bar for Group B "
|
||
"(rejecting qualified applicants). Both produce decisions that are "
|
||
"conditionally unfair given true qualification. Equalized odds avoids "
|
||
"this by conditioning on the true label — it guarantees that equally "
|
||
"qualified candidates are equally likely to be selected.",
|
||
"success"
|
||
),
|
||
"ref2_c": (
|
||
"**Incorrect.** The accuracy cost of demographic parity depends on how "
|
||
"different the base rates are and how tight the parity constraint is. "
|
||
"The Pareto frontier in Act II shows it can be 1–5pp in typical configurations. "
|
||
"It is not always >10pp.",
|
||
"warn"
|
||
),
|
||
"ref2_d": (
|
||
"**Incorrect.** Demographic parity does not violate GDPR — GDPR does not "
|
||
"prohibit all use of protected attributes; Article 9 allows processing for "
|
||
"specific purposes including combating discrimination. Many GDPR-compliant "
|
||
"fairness audits explicitly compute demographic parity metrics. The legal "
|
||
"challenge to DP is its substantive unfairness, not its data requirements.",
|
||
"warn"
|
||
),
|
||
}
|
||
|
||
_expl_r2, _kind_r2 = _ref2_explanations.get(
|
||
act2_reflection.value, ("No answer selected.", "info")
|
||
)
|
||
mo.callout(mo.md(_expl_r2), kind=_kind_r2)
|
||
return
|
||
|
||
|
||
# ── CELL 22: KEY TAKEAWAYS ────────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(mo):
|
||
mo.vstack([
|
||
mo.md("---"),
|
||
mo.md("## Key Takeaways"),
|
||
mo.callout(
|
||
mo.md("""
|
||
**1. Equal accuracy is not equal treatment (Chouldechova 2017).**
|
||
When base rates differ between groups, a calibrated model with a shared threshold
|
||
*structurally* produces unequal false positive and false negative rates. This is
|
||
not a training failure — it is a mathematical consequence of the base rate gap.
|
||
Auditing accuracy by group is a necessary but insufficient compliance check.
|
||
"""),
|
||
kind="info",
|
||
),
|
||
mo.callout(
|
||
mo.md("""
|
||
**2. Choosing a fairness criterion is a policy decision, not an engineering one.**
|
||
Demographic parity, equalized odds, and individual fairness are incompatible
|
||
when base rates differ. Each encodes a different moral theory of equal treatment.
|
||
Equalized odds is most defensible in hiring and lending under US law because it
|
||
conditions on actual qualification. The engineering job is to implement the chosen
|
||
criterion at minimum accuracy cost — not to choose it.
|
||
"""),
|
||
kind="info",
|
||
),
|
||
])
|
||
return
|
||
|
||
|
||
# ── CELL 23: LEDGER SAVE + HUD ────────────────────────────────────────────────
|
||
@app.cell(hide_code=True)
|
||
def _(
|
||
mo,
|
||
ledger,
|
||
COLORS,
|
||
context_toggle,
|
||
act1_prediction,
|
||
act2_prediction,
|
||
_a1_correct,
|
||
_a2_correct,
|
||
_final_eo_gap,
|
||
_final_accuracy_pct,
|
||
_audit_cost_k,
|
||
_reg_risk,
|
||
_disparate_impact_triggered,
|
||
_criterion,
|
||
):
|
||
# ── Safe defaults for unresolved reactive values ─────────────────────────
|
||
_ctx_val = context_toggle.value if context_toggle.value else "cloud"
|
||
_a1_pred = act1_prediction.value if act1_prediction.value else "none"
|
||
_a2_pred = act2_prediction.value if act2_prediction.value else "none"
|
||
_a1_ok = bool(_a1_correct) if _a1_correct is not None else False
|
||
_a2_ok = bool(_a2_correct) if _a2_correct is not None else False
|
||
_gap_val = float(_final_eo_gap) if _final_eo_gap is not None else 0.0
|
||
_acc_val = float(_final_accuracy_pct) if _final_accuracy_pct is not None else 84.0
|
||
_cost_val = float(_audit_cost_k) if _audit_cost_k is not None else 0.0
|
||
_risk_val = str(_reg_risk) if _reg_risk is not None else "low"
|
||
_hit_val = bool(_disparate_impact_triggered) if _disparate_impact_triggered is not None else False
|
||
_crit_val = str(_criterion) if _criterion is not None else "accuracy"
|
||
|
||
ledger.save(
|
||
chapter=15,
|
||
design={
|
||
"context": _ctx_val,
|
||
"fairness_criterion": _crit_val,
|
||
"equalized_odds_gap": round(_gap_val, 2),
|
||
"audit_cost_k": round(_cost_val, 2),
|
||
"act1_prediction": _a1_pred,
|
||
"act1_correct": _a1_ok,
|
||
"act2_result": round(_gap_val, 2),
|
||
"act2_decision": _crit_val,
|
||
"constraint_hit": _hit_val,
|
||
"regulatory_risk": _risk_val,
|
||
},
|
||
)
|
||
|
||
# ── HUD footer ─────────────────────────────────────────────────────────
|
||
_hud_color = COLORS["Cloud"] if _ctx_val == "cloud" else COLORS["Mobile"]
|
||
_ctx_display = "Cloud H100" if _ctx_val == "cloud" else "Mobile NPU"
|
||
_a1_display = "Correct" if _a1_ok else ("Incorrect" if _a1_pred != "none" else "—")
|
||
_a2_display = "Correct" if _a2_ok else ("Incorrect" if _a2_pred != "none" else "—")
|
||
_hit_color = "#f87171" if _hit_val else "#4ade80"
|
||
_risk_color = {"low": "#4ade80", "medium": "#fbbf24", "high": "#f87171"}.get(_risk_val, "#94a3b8")
|
||
_budget_ok = _cost_val <= 50.0
|
||
_cost_color = "#4ade80" if _budget_ok else "#f87171"
|
||
|
||
mo.Html(f"""
|
||
<div style="display: flex; gap: 22px; align-items: center; flex-wrap: wrap;
|
||
padding: 14px 24px; background: #0f172a;
|
||
border-radius: 12px; margin-top: 32px; font-size: 0.8rem;
|
||
border: 1px solid #1e293b; font-family: 'SF Mono', monospace;">
|
||
<span style="color: #475569; font-weight: 700; letter-spacing: 0.06em;">LAB 15</span>
|
||
<span>
|
||
<span style="color: #475569;">CONTEXT </span>
|
||
<span style="color: {_hud_color}; font-weight: 700;">{_ctx_display}</span>
|
||
</span>
|
||
<span>
|
||
<span style="color: #475569;">CRITERION </span>
|
||
<span style="color: #a5b4fc; font-weight: 600;">{_crit_val}</span>
|
||
</span>
|
||
<span>
|
||
<span style="color: #475569;">ACT I </span>
|
||
<span style="color: {'#4ade80' if _a1_ok else '#f87171'};">{_a1_display}</span>
|
||
</span>
|
||
<span>
|
||
<span style="color: #475569;">ACT II </span>
|
||
<span style="color: {'#4ade80' if _a2_ok else '#f87171'};">{_a2_display}</span>
|
||
</span>
|
||
<span>
|
||
<span style="color: #475569;">EO GAP </span>
|
||
<span style="color: {_hit_color};">{_gap_val:.1f}pp</span>
|
||
</span>
|
||
<span>
|
||
<span style="color: #475569;">AUDIT COST </span>
|
||
<span style="color: {_cost_color};">${_cost_val:.0f}K/yr</span>
|
||
</span>
|
||
<span>
|
||
<span style="color: #475569;">REG RISK </span>
|
||
<span style="color: {_risk_color}; font-weight: 700;">{_risk_val.upper()}</span>
|
||
</span>
|
||
<span>
|
||
<span style="color: #475569;">DISPARATE IMPACT </span>
|
||
<span style="color: {_hit_color};">{'YES' if _hit_val else 'No'}</span>
|
||
</span>
|
||
<span>
|
||
<span style="color: #475569;">LEDGER </span>
|
||
<span style="color: #4ade80;">ch15 saved</span>
|
||
</span>
|
||
</div>
|
||
""")
|
||
return
|
||
|
||
|
||
if __name__ == "__main__":
|
||
app.run()
|