diff --git a/v2/crates/wifi-densepose-train/src/ablation.rs b/v2/crates/wifi-densepose-train/src/ablation.rs
new file mode 100644
index 00000000..c3f3e59e
--- /dev/null
+++ b/v2/crates/wifi-densepose-train/src/ablation.rs
@@ -0,0 +1,306 @@
+//! ADR-145 — Ablation evaluation harness with privacy-leakage + latency metrics.
+//!
+//! Runs the sensing pipeline under a matrix of feature combinations
+//! (CSI-only / CIR-only / CSI+CIR / +Doppler / +BFLD / +UWB) and binds a metric
+//! set — presence accuracy, localisation error, FP/FN, latency p50/p95,
+//! privacy-leakage (membership-inference), and cross-room degradation — so every
+//! pipeline change is measured, not guessed (ADR-145 §10/§14). The model runs
+//! themselves are external; this module owns the deterministic metric
+//! computation + the auto-report.
+
+use core::fmt::Write as _;
+
+/// One feature combination in the ablation matrix (ADR-145 §2).
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub enum FeatureSet {
+    /// CSI amplitude/phase only.
+    CsiOnly,
+    /// CIR taps only (ADR-134).
+    CirOnly,
+    /// CSI + CIR.
+    CsiCir,
+    /// CSI + CIR + passive Doppler.
+    CsiCirDoppler,
+    /// CSI + CIR + Doppler + BFLD privacy gate.
+    CsiCirDopplerBfld,
+    /// Full fusion including UWB range constraints (ADR-144; deferred until hw).
+    FullUwb,
+}
+
+impl FeatureSet {
+    /// The six-variant ablation matrix, runnable order.
+    pub const MATRIX: [FeatureSet; 6] = [
+        FeatureSet::CsiOnly,
+        FeatureSet::CirOnly,
+        FeatureSet::CsiCir,
+        FeatureSet::CsiCirDoppler,
+        FeatureSet::CsiCirDopplerBfld,
+        FeatureSet::FullUwb,
+    ];
+
+    /// Stable label for reports.
+    #[must_use]
+    pub fn label(self) -> &'static str {
+        match self {
+            Self::CsiOnly => "csi_only",
+            Self::CirOnly => "cir_only",
+            Self::CsiCir => "csi+cir",
+            Self::CsiCirDoppler => "csi+cir+doppler",
+            Self::CsiCirDopplerBfld => "csi+cir+doppler+bfld",
+            Self::FullUwb => "full+uwb",
+        }
+    }
+}
+
+/// `(p50, p95)` percentiles of a latency sample set (ms), nearest-rank.
+#[must_use]
+pub fn latency_percentiles_ms(samples_ms: &[f64]) -> (f64, f64) {
+    if samples_ms.is_empty() {
+        return (0.0, 0.0);
+    }
+    let mut s = samples_ms.to_vec();
+    s.sort_by(|a, b| a.partial_cmp(b).unwrap());
+    let pick = |q: f64| {
+        // Nearest-rank: ceil(q * n) - 1, clamped.
+        let rank = ((q * s.len() as f64).ceil() as usize).clamp(1, s.len()) - 1;
+        s[rank]
+    };
+    (pick(0.50), pick(0.95))
+}
+
+/// False-positive and false-negative rates from a confusion count.
+#[must_use]
+pub fn confusion_rates(tp: u64, fp: u64, tn: u64, fn_: u64) -> (f64, f64) {
+    let fp_rate = if fp + tn == 0 { 0.0 } else { fp as f64 / (fp + tn) as f64 };
+    let fn_rate = if fn_ + tp == 0 { 0.0 } else { fn_ as f64 / (fn_ + tp) as f64 };
+    (fp_rate, fn_rate)
+}
+
+/// Privacy-leakage score in [0, 1] via a membership-inference (MIA) proxy
+/// (ADR-145 §2): how separable are confidence scores of training-set *members*
+/// from *non-members*? Computed as `|AUC - 0.5| * 2` — 0.0 when the two score
+/// distributions are indistinguishable (no leakage), 1.0 when perfectly
+/// separable (an attacker can tell who was in the training set).
+#[must_use]
+pub fn membership_inference_leakage(member_scores: &[f64], nonmember_scores: &[f64]) -> f64 {
+    if member_scores.is_empty() || nonmember_scores.is_empty() {
+        return 0.0;
+    }
+    // AUC = P(member_score > nonmember_score) over all pairs (+ 0.5 for ties).
+    let mut wins = 0.0;
+    let total = (member_scores.len() * nonmember_scores.len()) as f64;
+    for &m in member_scores {
+        for &n in nonmember_scores {
+            if m > n {
+                wins += 1.0;
+            } else if (m - n).abs() < f64::EPSILON {
+                wins += 0.5;
+            }
+        }
+    }
+    let auc = wins / total;
+    ((auc - 0.5).abs() * 2.0).clamp(0.0, 1.0)
+}
+
+/// The metric bundle for one ablation variant (ADR-145 §2).
+#[derive(Debug, Clone)]
+pub struct AblationMetrics {
+    /// Which feature combination was evaluated.
+    pub feature_set: FeatureSet,
+    /// Presence-detection accuracy in [0, 1].
+    pub presence_accuracy: f64,
+    /// Mean localisation error (m).
+    pub localization_err_m: f64,
+    /// False-positive rate.
+    pub fp_rate: f64,
+    /// False-negative rate.
+    pub fn_rate: f64,
+    /// Latency 50th percentile (ms).
+    pub latency_p50_ms: f64,
+    /// Latency 95th percentile (ms).
+    pub latency_p95_ms: f64,
+    /// Privacy leakage in [0, 1] (MIA proxy; lower is better).
+    pub privacy_leakage: f64,
+    /// Cross-room accuracy degradation (room_A_acc - room_B_acc), >= 0.
+    pub cross_room_degradation: f64,
+}
+
+/// Raw per-variant inputs from a pipeline run; metrics are derived
+/// deterministically from these.
+#[derive(Debug, Clone)]
+pub struct VariantRun {
+    /// Feature combination evaluated.
+    pub feature_set: FeatureSet,
+    /// Confusion counts (tp, fp, tn, fn).
+    pub confusion: (u64, u64, u64, u64),
+    /// Mean localisation error (m).
+    pub localization_err_m: f64,
+    /// Per-frame latency samples (ms).
+    pub latency_samples_ms: Vec<f64>,
+    /// Member/non-member confidence scores for the MIA proxy.
+    pub member_scores: Vec<f64>,
+    /// Non-member confidence scores.
+    pub nonmember_scores: Vec<f64>,
+    /// Accuracy in the calibration room and a held-out room.
+    pub room_a_accuracy: f64,
+    /// Held-out room accuracy.
+    pub room_b_accuracy: f64,
+}
+
+impl AblationMetrics {
+    /// Derive the metric bundle from a raw variant run.
+    #[must_use]
+    pub fn from_run(run: &VariantRun) -> Self {
+        let (tp, fp, tn, fn_) = run.confusion;
+        let (fp_rate, fn_rate) = confusion_rates(tp, fp, tn, fn_);
+        let total = (tp + fp + tn + fn_).max(1);
+        let presence_accuracy = (tp + tn) as f64 / total as f64;
+        let (p50, p95) = latency_percentiles_ms(&run.latency_samples_ms);
+        Self {
+            feature_set: run.feature_set,
+            presence_accuracy,
+            localization_err_m: run.localization_err_m,
+            fp_rate,
+            fn_rate,
+            latency_p50_ms: p50,
+            latency_p95_ms: p95,
+            privacy_leakage: membership_inference_leakage(&run.member_scores, &run.nonmember_scores),
+            cross_room_degradation: (run.room_a_accuracy - run.room_b_accuracy).max(0.0),
+        }
+    }
+}
+
+/// An ablation report over the variant matrix (ADR-145 auto-report).
+#[derive(Debug, Clone, Default)]
+pub struct AblationReport {
+    /// Per-variant metrics in evaluation order.
+    pub rows: Vec<AblationMetrics>,
+}
+
+impl AblationReport {
+    /// Build from a set of variant runs.
+    #[must_use]
+    pub fn from_runs(runs: &[VariantRun]) -> Self {
+        Self { rows: runs.iter().map(AblationMetrics::from_run).collect() }
+    }
+
+    /// Look up a variant's metrics.
+    #[must_use]
+    pub fn get(&self, fs: FeatureSet) -> Option<&AblationMetrics> {
+        self.rows.iter().find(|m| m.feature_set == fs)
+    }
+
+    /// Acceptance check (ADR-145 / ADR-136 AC): does CSI+CIR beat CSI-only on at
+    /// least `min_wins` of {presence accuracy ↑, localisation error ↓, p95 latency ↓}?
+    #[must_use]
+    pub fn csi_cir_beats_csi_only(&self, min_wins: usize) -> bool {
+        let (Some(a), Some(b)) = (self.get(FeatureSet::CsiOnly), self.get(FeatureSet::CsiCir)) else {
+            return false;
+        };
+        let wins = [
+            b.presence_accuracy > a.presence_accuracy,
+            b.localization_err_m < a.localization_err_m,
+            b.latency_p95_ms <= a.latency_p95_ms,
+        ]
+        .iter()
+        .filter(|w| **w)
+        .count();
+        wins >= min_wins
+    }
+
+    /// Deterministic markdown report (stable column/row order).
+    #[must_use]
+    pub fn to_markdown(&self) -> String {
+        let mut s = String::new();
+        let _ = writeln!(
+            s,
+            "| variant | presence_acc | loc_err_m | fp | fn | p50_ms | p95_ms | privacy_leak | xroom_degr |"
+        );
+        let _ = writeln!(s, "|---|---|---|---|---|---|---|---|---|");
+        for m in &self.rows {
+            let _ = writeln!(
+                s,
+                "| {} | {:.3} | {:.3} | {:.3} | {:.3} | {:.2} | {:.2} | {:.3} | {:.3} |",
+                m.feature_set.label(),
+                m.presence_accuracy,
+                m.localization_err_m,
+                m.fp_rate,
+                m.fn_rate,
+                m.latency_p50_ms,
+                m.latency_p95_ms,
+                m.privacy_leakage,
+                m.cross_room_degradation,
+            );
+        }
+        s
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn latency_percentiles_nearest_rank() {
+        let s: Vec<f64> = (1..=100).map(|i| i as f64).collect();
+        let (p50, p95) = latency_percentiles_ms(&s);
+        assert!((p50 - 50.0).abs() < 1e-9);
+        assert!((p95 - 95.0).abs() < 1e-9);
+        assert_eq!(latency_percentiles_ms(&[]), (0.0, 0.0));
+    }
+
+    #[test]
+    fn confusion_rates_basic() {
+        let (fp_rate, fn_rate) = confusion_rates(80, 10, 90, 20);
+        assert!((fp_rate - 0.1).abs() < 1e-9); // 10 / (10+90)
+        assert!((fn_rate - 0.2).abs() < 1e-9); // 20 / (20+80)
+    }
+
+    #[test]
+    fn mia_leakage_zero_when_indistinguishable_high_when_separable() {
+        // Identical distributions → ~no leakage.
+        let same = vec![0.5, 0.6, 0.7];
+        assert!(membership_inference_leakage(&same, &same) < 1e-9);
+        // Perfectly separable → leakage 1.0.
+        let members = vec![0.9, 0.95, 0.99];
+        let nonmembers = vec![0.1, 0.2, 0.3];
+        assert!((membership_inference_leakage(&members, &nonmembers) - 1.0).abs() < 1e-9);
+    }
+
+    #[test]
+    fn csi_cir_beats_csi_only_acceptance() {
+        let csi_only = VariantRun {
+            feature_set: FeatureSet::CsiOnly,
+            confusion: (70, 15, 70, 30), // acc 0.756
+            localization_err_m: 0.40,
+            latency_samples_ms: vec![10.0; 10],
+            member_scores: vec![0.5],
+            nonmember_scores: vec![0.5],
+            room_a_accuracy: 0.8,
+            room_b_accuracy: 0.6,
+        };
+        let csi_cir = VariantRun {
+            feature_set: FeatureSet::CsiCir,
+            confusion: (88, 6, 90, 12), // acc 0.908
+            localization_err_m: 0.22,
+            latency_samples_ms: vec![11.0; 10],
+            member_scores: vec![0.5],
+            nonmember_scores: vec![0.5],
+            room_a_accuracy: 0.85,
+            room_b_accuracy: 0.80,
+        };
+        let runs = [csi_only, csi_cir];
+        let report = AblationReport::from_runs(&runs);
+        // CSI+CIR wins on presence accuracy + localisation error (2 of 3).
+        assert!(report.csi_cir_beats_csi_only(2));
+        let md = report.to_markdown();
+        assert!(md.contains("csi_only") && md.contains("csi+cir"));
+        // Deterministic: same input → byte-identical report.
+        assert_eq!(md, AblationReport::from_runs(&runs).to_markdown());
+    }
+
+    #[test]
+    fn matrix_has_six_variants() {
+        assert_eq!(FeatureSet::MATRIX.len(), 6);
+    }
+}
diff --git a/v2/crates/wifi-densepose-train/src/lib.rs b/v2/crates/wifi-densepose-train/src/lib.rs
index 8534e2e9..0b72852c 100644
--- a/v2/crates/wifi-densepose-train/src/lib.rs
+++ b/v2/crates/wifi-densepose-train/src/lib.rs
@@ -67,6 +67,9 @@ pub mod metrics;
 pub mod model;
 #[cfg(feature = "tch-backend")]
 pub mod proof;
+
+/// ADR-145 — ablation evaluation harness (feature matrix + privacy/latency metrics).
+pub mod ablation;
 #[cfg(feature = "tch-backend")]
 pub mod trainer;