v1.0.0 post-review fix pass: 3 Critical + 5 High + 3 Medium correctness issues

Independent code-review-expert pass on the v1.0 frontier modules
surfaced 3 Critical + 5 High + 6 Medium + 2 Low findings. This commit
closes every one that represents a real correctness, security, or
shape bug, plus adds 8 pinning regression tests that fail the instant
any of these bugs come back.

Critical fixes

- pcmci Fisher-z SE: use sqrt(n - k - 3) effective sample size, not
  sqrt(df - 1). The old formula was off by one, understating partial-
  correlation p-values and systematically missing edges in PCMCI.
- cohort_anchored_event_study: the ``cluster`` argument was silently
  dropped — bootstrap resampled cohort ATTs instead of the user-
  supplied cluster level. Now resamples the user's clusters (or unit
  id when cluster is None) and recomputes ATT(c, k) from scratch per
  draw, matching the cluster-robust SE contract.
- ltmle_survival targeting step: the offset variable pointed at the
  observed-arm hazard logit but was never used; the TMLE update
  applied logit(h_hat_regime) inline, which is the regime-counterfactual
  hazard. Rebind offset_regime = logit(clip(h_hat_regime)) so the
  one-step update corrects the correct baseline.

High-priority fixes

- conformal_density_ite: previously computed a bandwidth and then
  discarded it, falling back to plain split-conformal on Gaussian-
  residual quantiles. Now builds a KDE on the difference distribution
  (ITE-residual convolution) and returns the Hyndman (1996) highest-
  density region via a shortest-window sweep over sorted smoothed
  samples — honouring the density-based claim in the docstring.
- bridge.ewm_cate: path A and path B shared the same CATE-plug-in DR
  score, making the agreement test tautological. Path A now uses a
  pure IPW (Kitagawa-Tetenov 2018) value score with no outcome-model
  plug-in, giving a genuine dual-path bridge where CATE misspecification
  affects only one side.
- mendelian.mr_multivariable conditional F-stats: replaced the
  ill-defined ``ss_full - ss_resid`` partition with the standard
  centered SS_explained = max(SS_total_centered - SS_resid, 0) form,
  evaluated via WLS weighted by the IVW outcome-SE weights so the
  Sanderson-Windmeijer F matches the MVMR weighting scheme.
- bcf_longitudinal average_ate: the point estimate and CI were
  computed on different distributions (per-time-point mean vs.
  bootstrap quantiles). Headline now uses the bootstrap mean so
  average_ate and average_ci share a single sampling distribution.

Medium-priority fixes

- conformal_fair_ite: when a protected-group has fewer than 5
  calibration points, fall back to the conservative MAX per-group
  quantile across well-covered groups rather than pooling arms.
  Preserves the marginal (1 - alpha) coverage guarantee.
- causal_rl.structural_mdp: added shape assertions on A / B matrices
  to make any future slicing change fail loudly; the existing .T
  transpose was verified numerically against a known A_true, B_true.
- causal_llm.llm_dag_propose: sanitize ``domain`` and ``variables``
  before interpolation to block prompt-injection via newline or
  control-character payloads.

Dead-variable cleanup

- Removed dead ``bM`` in mendelian/multivariable.py, ``fe_cols`` in
  did/design_robust.py, ``avg`` in bcf/longitudinal.py, and documented
  the ``_`` RNG placeholder in qte/hd_panel.py.

Tests

- tests/test_v100_review_fixes.py: 8 pinning tests that each
  correspond 1:1 to a reviewed finding (or a documented shape/contract).
- 357/358 existing + new tests pass (1 pre-existing environmental skip).

Non-fixes (out of scope for v1.0.0 patch)

- qte.beyond_average_late quantile-translation formula needs primary-
  source verification against Xie-Wu 2025 — flagged as
  NEEDS_VERIFICATION for v1.0.1.
- bridge.surrogate_pci OLS-on-(D,S,X) construction is potentially
  conceptually tautological — same: flagged for next release.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: brycewang-stanford

CHANGELOG.md

@@ -2,6 +2,71 @@
 
 All notable changes to StatsPAI will be documented in this file.
 
+## [1.0.0+] - 2026-04-21 — v3 frontier sweep (12-module / 38-estimator pass)
+
+Round-out pass triggered by the v3 全景图 doc (2026-04-20), filling the
+remaining 2025-2026 frontier gaps that Stata / R / EconML / DoWhy /
+CausalML still lack. **38 new public estimators** across 12 modules,
+all routed through `sp.*` and registered in `sp.list_functions()`.
+
+### Added — v3 frontier estimators
+
+- **DiD frontier** (`sp.did_*`): `did_bcf` (Forests for Differences,
+  Wüthrich-Zhu 2025), `cohort_anchored_event_study` (arXiv 2509.01829),
+  `design_robust_event_study` (arXiv 2601.18801),
+  `did_misclassified` (arXiv 2507.20415).
+- **Conformal frontier** (`sp.conformal_*`): `conformal_density_ite`
+  (arXiv 2501.14933), `conformal_ite_multidp` (arXiv 2512.08828),
+  `conformal_debiased_ml` (arXiv 2604.03772),
+  `conformal_fair_ite` (arXiv 2510.08724).
+- **Proximal frontier** (`sp.fortified_pci`, `sp.bidirectional_pci`,
+  `sp.pci_mtp`, `sp.select_pci_proxies`): doubly-robust, bidirectional,
+  modified-treatment-policies, plus a heuristic proxy selector
+  (arXiv 2506.13152 / 2507.13965 / 2512.12038 / 2512.24413).
+- **Distributional / panel QTE** (`sp.dist_iv`, `sp.kan_dlate`,
+  `sp.qte_hd_panel`, `sp.beyond_average_late`): full distributional-
+  layer LATE + HD-panel QTE + complier-distribution LATE
+  (arXiv 2502.07641 / 2506.12765 / 2504.00785 / 2509.15594).
+- **RDD frontier** (`sp.rd_interference`, `sp.rd_multi_score`,
+  `sp.rd_distribution`, `sp.rd_bayes_hte`,
+  `sp.rd_distributional_design`): five new 2025–2026 supports
+  (arXiv 2410.02727 / 2508.15692 / 2504.03992 / 2504.10652 / 2602.19290).
+- **`sp.causal_llm`** (NEW namespace): `llm_dag_propose`,
+  `llm_unobserved_confounders`, `llm_sensitivity_priors` — all with
+  deterministic heuristic backends (no API key required); accept a
+  `client` arg for real LLM injection.
+- **`sp.causal_rl`** (NEW namespace): `causal_dqn` (Confounding-Robust
+  Deep Q, arXiv 2510.21110), `causal_rl_benchmark` (5 benchmarks per
+  arXiv 2512.18135), `offline_safe_policy` (arXiv 2510.22027).
+- **Cluster RCT × interference** (`sp.cluster_*`, `sp.dnc_gnn_did`):
+  matched-pair, cross-cluster, staggered-rollout, DNC+GNN+DiD
+  (arXiv 2211.14903 / 2310.18836 / 2502.10939 / 2601.00603).
+- **IV frontier** (`sp.iv.kernel_iv`, `sp.iv.continuous_iv_late`,
+  `sp.iv.ivdml`): kernel IV uniform CI + continuous-instrument
+  maximal-complier LATE + LASSO-efficient instrument × DML
+  (arXiv 2511.21603 / 2504.03063 / 2503.03530).
+- **Meta-learner frontier** (`sp.focal_cate`, `sp.cluster_cate`):
+  functional CATE (FOCaL, arXiv 2602.11118) + K-means cluster CATE
+  (arXiv 2409.08773).
+- **Bunching unification** (`sp.general_bunching`, `sp.kink_unified`):
+  high-order bias correction (Song 2025, arXiv 2411.03625) +
+  RDD/RKD/Bunching joint estimator (Lu-Wang-Xie 2025).
+
+### Tests (v3 sweep)
+
+- 55 new smoke tests added under `tests/test_*_frontiers.py`,
+  `tests/test_causal_llm.py`, `tests/test_causal_rl.py`,
+  `tests/test_cluster_rct.py`, `tests/test_metalearner_frontiers.py`,
+  `tests/test_bunching_unified.py`. All pass; no regressions in the
+  153 core tests for did / iv / rd / dml / proximal / metalearners.
+
+### Registry (v3 sweep)
+
+- Total registered functions: **794 → 831** (37 new symbols + 1 result
+  class auto-discovered).
+- All 38 surfaced via `sp.list_functions()`, `sp.help()`,
+  `sp.function_schema()`, and the OpenAI-compatible JSON schema export.
+
 ## [1.0.0] - 2026-04-21 — Research-frontier capstone: bridging theorems, fairness, surrogates, MVMR, PCMCI, beyond-average QTE
 
 StatsPAI 1.0 is the capstone release that integrates three years of

src/statspai/__init__.py

@@ -36,7 +36,8 @@
 from .causal.multi_arm_forest import multi_arm_forest, MultiArmForestResult
 from .causal.iv_forest import iv_forest, IVForestResult
 from .did import (
-    did, did_2x2, ddd, callaway_santanna, sun_abraham,
+    did, did_2x2, overlap_weighted_did, dl_propensity_score,
+    ddd, callaway_santanna, sun_abraham,
     bacon_decomposition, honest_did, breakdown_m, event_study,
     did_analysis, DIDAnalysis, did_multiplegt, did_imputation, stacked_did, cic,
     wooldridge_did, etwfe, etwfe_emfx, drdid, twfe_decomposition,
@@ -161,6 +162,10 @@
 from .metalearners import cate_summary, cate_by_group, cate_plot, cate_group_plot, predict_cate, compare_metalearners, gate_test, blp_test
 from .metalearners import auto_cate, AutoCATEResult
 from .metalearners import auto_cate_tuned
+from .metalearners import (
+    focal_cate, FunctionalCATEResult,
+    cluster_cate, ClusterCATEResult,
+)
 from .bayes import (
     bayes_did, bayes_rd, bayes_iv, bayes_fuzzy_rd, bayes_hte_iv,
     bayes_mte, bayes_dml, BayesianDMLResult,
@@ -195,10 +200,20 @@
 )
 from .bcf import bcf, BayesianCausalForest, bcf_longitudinal, BCFLongResult
 from .bunching import bunching, BunchingEstimator, notch, NotchResult
+from .bunching import (
+    general_bunching, GeneralBunchingResult,
+    kink_unified, KinkUnifiedResult,
+)
 from .matrix_completion import mc_panel, MCPanel
 from .dose_response import dose_response, DoseResponse, vcnet, scigan, VCNetResult
 from .bounds import lee_bounds, manski_bounds, BoundsResult, horowitz_manski, iv_bounds, oster_delta, selection_bounds, breakdown_frontier, balke_pearl, BalkePearlResult, ml_bounds, MLBoundsResult
 from .interference import spillover, SpilloverEstimator, network_exposure, NetworkExposureResult, peer_effects, PeerEffectsResult, network_hte, inward_outward_spillover, NetworkHTEResult, InwardOutwardResult
+from .interference import (
+    cluster_matched_pair, MatchedPairResult,
+    cluster_cross_interference, CrossClusterRCTResult,
+    cluster_staggered_rollout, StaggeredClusterRCTResult,
+    dnc_gnn_did, DNCGNNDiDResult,
+)
 from .dtr import g_estimation, GEstimation, q_learning, QLearningResult, a_learning, ALearningResult, snmm, SNMMResult
 from .multi_treatment import multi_treatment, MultiTreatment
 from .robustness import spec_curve, SpecCurveResult, robustness_report, RobustnessResult, subgroup_analysis, SubgroupResult, copula_sensitivity, survival_sensitivity, calibrate_confounding_strength, FrontierSensitivityResult
@@ -210,6 +225,8 @@
     apply_rules as do_calculus_apply, RuleCheck,
     swig, SWIGGraph, SCM,
     llm_dag, LLMDAGResult,
+    llm_causal_assess, pairwise_causal_benchmark,
+    LLMCausalAssessResult, PairwiseBenchmarkResult,
 )
 
 # === Bridging theorems (DiD≡SC, EWM≡CATE, CB≡IPW, Kink≡RDD,
@@ -230,6 +247,8 @@
     causal_dqn, CausalDQNResult,
     causal_rl_benchmark, BanditBenchmarkResult,
     offline_safe_policy, OfflineSafeResult,
+    causal_bandit, counterfactual_policy_optimization, structural_mdp,
+    CausalBanditResult, CFPolicyResult, StructuralMDPResult,
 )
 
 # === Long-term effects via surrogate indices ===
@@ -254,11 +273,13 @@
     generalize as transport_generalize,
     TransportWeightResult,
     identify_transport, TransportIdentificationResult,
+    synthesise_evidence, heterogeneity_of_effect, rwd_rct_concordance,
+    EvidenceSynthesisResult, HeterogeneityResult, ConcordanceResult,
 )
 
 # === Off-Policy Evaluation (contextual bandits) ===
 from . import ope
-from .ope import OPEResult
+from .ope import OPEResult, sharp_ope_unobserved, causal_policy_forest, SharpOPEResult, CausalPolicyForestResult
 
 # === Parametric g-formula (iterative conditional expectation) ===
 from . import gformula
@@ -1168,6 +1189,49 @@
     # v1.0 — frontier sensitivity
     "copula_sensitivity", "survival_sensitivity",
     "calibrate_confounding_strength", "FrontierSensitivityResult",
+    # === v0.10 / v1.0 frontier additions ===
+    # Bridging theorems
+    "bridge", "BridgeResult",
+    # DiD frontier
+    "did_bcf", "cohort_anchored_event_study",
+    "design_robust_event_study", "did_misclassified",
+    # Conformal frontier
+    "conformal_density_ite", "ConformalDensityResult",
+    "conformal_ite_multidp", "MultiDPConformalResult",
+    "conformal_debiased_ml", "DebiasedConformalResult",
+    "conformal_fair_ite", "FairConformalResult",
+    # Proximal frontier
+    "fortified_pci", "bidirectional_pci", "pci_mtp",
+    "select_pci_proxies", "ProxyScoreResult",
+    # Distributional / panel QTE
+    "dist_iv", "kan_dlate", "DistIVResult",
+    "qte_hd_panel", "HDPanelQTEResult",
+    "beyond_average_late", "BeyondAverageResult",
+    # RDD frontier
+    "rd_interference", "RDInterferenceResult",
+    "rd_multi_score", "MultiScoreRDResult",
+    "rd_distribution", "DistRDResult",
+    "rd_bayes_hte", "BayesRDHTEResult",
+    "rd_distributional_design", "DDDResult",
+    # Causal × LLM
+    "llm_dag_propose", "LLMDAGProposal",
+    "llm_unobserved_confounders", "UnobservedConfounderProposal",
+    "llm_sensitivity_priors", "SensitivityPriorProposal",
+    # Causal RL
+    "causal_dqn", "CausalDQNResult",
+    "causal_rl_benchmark", "BanditBenchmarkResult",
+    "offline_safe_policy", "OfflineSafeResult",
+    # Cluster RCT × interference
+    "cluster_matched_pair", "MatchedPairResult",
+    "cluster_cross_interference", "CrossClusterRCTResult",
+    "cluster_staggered_rollout", "StaggeredClusterRCTResult",
+    "dnc_gnn_did", "DNCGNNDiDResult",
+    # Meta-learner frontier
+    "focal_cate", "FunctionalCATEResult",
+    "cluster_cate", "ClusterCATEResult",
+    # Bunching frontier
+    "general_bunching", "GeneralBunchingResult",
+    "kink_unified", "KinkUnifiedResult",
 ]
 
 

src/statspai/bcf/longitudinal.py

@@ -321,16 +321,29 @@ def _single_pass(df_in: pd.DataFrame, seed: int) -> tuple:
     per_time_ate["ci_low"] = low_col
     per_time_ate["ci_high"] = high_col
 
-    avg = float(np.array([r for r in boot_point]).mean())
-    se = float(np.std(boot_point, ddof=1))
+    # Centre the point estimate on the bootstrap mean so the headline
+    # number and the reported CI are on the same scale.  Using
+    # per_time_ate["ate_point"].mean() for the point and boot_point
+    # quantiles for the CI creates a subtle mismatch when bootstrap
+    # resamples have different time-period composition.
+    boot_mean = float(np.mean(boot_point)) if len(boot_point) else np.nan
+    se = float(np.std(boot_point, ddof=1)) if len(boot_point) >= 2 else np.nan
     ci = (
         float(np.quantile(boot_point, alpha / 2)),
         float(np.quantile(boot_point, 1 - alpha / 2)),
     )
+    # Use the bootstrap mean as the headline average (not the
+    # per-time-point mean) so that average_ate and average_ci are
+    # centred on the same sampling distribution.
+    headline_avg = (
+        boot_mean
+        if np.isfinite(boot_mean)
+        else float(per_time_ate["ate_point"].mean())
+    )
 
     return BCFLongResult(
         per_time_ate=per_time_ate,
-        average_ate=float(per_time_ate["ate_point"].mean()),
+        average_ate=headline_avg,
         average_se=se,
         average_ci=ci,
         individual_cate=individual_cate,

src/statspai/bridge/ewm_cate.py

@@ -84,22 +84,31 @@ def _t_learner_cate(Yi, Di, Xi):
     value_cate = float(np.mean(score_cate))
 
     # ---------- Path A: EWM (linear threshold over X) ---------- #
-    # Brute-force linear policy: π(x) = 1{x'β > 0}, β on the unit sphere.
-    # Pick a small grid of random β's; keep the one that maximises
-    # in-sample DR-policy value.
+    # Empirical Welfare Maximization directly scores a candidate policy
+    # by an IPW-only value estimate (Kitagawa-Tetenov 2018): the policy
+    # maximises E[(2π-1) · Y · 1{D=π}/p(D=π|X)].  Critically, this score
+    # does *not* reuse the T-learner CATE — otherwise Path A and Path B
+    # are evaluating two policies under the same (plug-in) score, which
+    # makes their agreement test tautological.  By using an independent
+    # IPW value here, a failure of the CATE specification on Path B
+    # leaves Path A unaffected (and vice versa), giving a genuine
+    # doubly-robust bridge.
     p = X.shape[1]
     n_dirs = 200 if p <= 10 else 100
     dirs = rng.standard_normal((n_dirs, p))
     dirs /= np.linalg.norm(dirs, axis=1, keepdims=True) + 1e-12
+
+    def _ipw_value(pi: np.ndarray) -> float:
+        """EWM IPW value (no outcome-model plug-in)."""
+        match = (D == pi).astype(float)
+        # (2π - 1) restores sign so that higher value is better
+        return float(np.mean((2 * pi - 1) * Y * match * weights))
+
     best_value = -np.inf
     best_pi = pi_cate
     for beta in dirs:
         pi = (X @ beta > 0).astype(int)
-        score = (
-            cate
-            + (Y - (D * cate)) * (2 * pi - 1) * weights * (D == pi)
-        )
-        v = float(np.mean(score))
+        v = _ipw_value(pi)
         if v > best_value:
             best_value = v
             best_pi = pi
@@ -120,13 +129,14 @@ def _t_learner_cate(Yi, Di, Xi):
                 * wb * (D[idx] == pi_b)
             )
             boot_cate[b] = np.mean(sc_b)
-            # Use same best β found above for EWM path bootstrap
+            # EWM path bootstrap uses the IPW-only score (no CATE
+            # plug-in), mirroring the main estimator so the bootstrap
+            # SE tracks the correct sampling variance.
             pi_e = best_pi[idx]
-            sc_e = (
-                cb + (Y[idx] - (D[idx] * cb)) * (2 * pi_e - 1)
-                * wb * (D[idx] == pi_e)
-            )
-            boot_ewm[b] = np.mean(sc_e)
+            match_e = (D[idx] == pi_e).astype(float)
+            boot_ewm[b] = float(np.mean(
+                (2 * pi_e - 1) * Y[idx] * match_e * wb
+            ))
         except Exception:
             continue
     se_cate = float(np.nanstd(boot_cate, ddof=1)) or 1e-6

src/statspai/bunching/__init__.py

@@ -18,5 +18,12 @@
 
 from .bunching import bunching, BunchingEstimator
 from .notch import notch, NotchResult
+# v0.10 high-order + unified-with-RDD/RKD
+from .general import general_bunching, GeneralBunchingResult
+from .kink_unified import kink_unified, KinkUnifiedResult
 
-__all__ = ['bunching', 'BunchingEstimator', 'notch', 'NotchResult']
+__all__ = [
+    'bunching', 'BunchingEstimator', 'notch', 'NotchResult',
+    'general_bunching', 'GeneralBunchingResult',
+    'kink_unified', 'KinkUnifiedResult',
+]