docs(citations): fix 49 wrong attributions + add automated auditor

Systematic audit of all arXiv / NBER / DOI references in src/ and docs/
found 49 unique IDs attributed to the wrong authors — predominantly in
post-2024 citations where LLM-generated attributions diverged from
arXiv / NBER primary sources.

Build: new tools/audit_citations.py (stdlib-only, ~790 lines)

  - EXTRACT  regex over src/ + docs/ (arXiv IDs, NBER WP, DOIs)
  - VERIFY   batch arXiv API, NBER meta scrape, Crossref JSON, all
             cached on disk under tools/.citation_cache/
  - DIFF     unicode-aware last-name match, hyphen/apostrophe-tolerant,
             paragraph + bullet scope isolation, multi-citation-line
             narrowing, bibtex + markdown-table special cases
  - REPORT   audit_report.md with OK / MISMATCH / UNRESOLVED sections

Final audit: 298 citations, 298 OK, 0 mismatches, 0 unresolved.

Fix examples (metadata only, no numerical path touched):

  * 1905.12741  Veitch-Wang-Blei (2020) -> Veitch-Sridhar-Blei (2019)
  * 2111.07966  Tibshirani swap -> Brunskill; 2023 -> 2021
  * 2311.08527  Ghassami et al. 2024 -> Tran-Bibaut-Kallus 2023
  * 2503.11375  Shi-Athey -> Sun-Xie-Zhang
  * 2503.13323  Roth-Sant'Anna-Bilinski-Poe -> Baker-Callaway-
                Cunningham-Goodman-Bacon-Sant'Anna
  * 2506.17214  Qian-van der Laan -> Li-Qiu-Wang-van der Laan
  * 2510.21110  Fu-Zhou -> Li-Zhang-Bareinboim
  * 2510.26723  Ferman et al. -> Kato
  * 2511.14133  Agarwal-Shah -> Han-Shah
  * 2512.00566  Cattaneo-Jansson-Ma 2026 -> Cavaliere-Goncalves-
                Nielsen-Zanelli 2025
  * 2512.12038  Park-Ying -> Olivas-Martinez-Gilbert-Rotnitzky
  * NBER 34550  Borusyak-Hull-Jaravel -> Abadie-Angrist-Frandsen-
                Pischke (a different paper shares the "Harvesting DiD"
                title family; Borusyak-Hull-Jaravel are on w24997/w33236)
  * 10.1002/jae.3103  remove phantom "Kurz" — JAE paper has 4 authors
                      not 5

Verification: pytest on 19 touched test modules -> 188 passed / 0 failed.
Re-running tools/audit_citations.py confirms 298/298 OK.

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

Author: brycewang-stanford

.gitignore

@@ -179,3 +179,11 @@ test-notebooks/demo_causal_inference.ipynb
 # Claude Code session state (local only)
 .claude/
 CLAUDE.md
+
+# Citation auditor artefacts (tools/audit_citations.py)
+tools/.citation_cache/
+audit_report.md
+audit_report.json
+
+# Ad-hoc citation-audit scratch (not from the tool)
+.citation_audit/

CHANGELOG.md

@@ -752,13 +752,13 @@ surfaces through `sp.list_functions`, `sp.describe_function`, and
   [`synth/experimental_design.py`]
 
 - **`sp.rdrobust(..., bootstrap='rbc', n_boot=999, random_state=...)`**
-  — Cattaneo, Jansson & Ma (arXiv:2512.00566, 2026) robust-bias-corrected
+  — Cavaliere, Gonçalves, Nielsen & Zanelli (arXiv:2512.00566, 2025) robust-bias-corrected
   studentised percentile bootstrap.  Empirically delivers CIs ~3–15%
   shorter than the analytic robust CI without sacrificing coverage.
   New ``model_info['rbc_bootstrap']`` block exposes the CI, p-value,
   length-ratio, and effective replicate count.
 
-- **`sp.fairness.evidence_without_injustice`** — Kwak & Pleasants
+- **`sp.fairness.evidence_without_injustice`** — Loi, Di Bello & Cangiotti
   (arXiv:2510.12822, 2025) counterfactual-fairness test that freezes
   admissible-evidence features at their factual values and tests
   whether predictions still change under ``do(A = a')``.  Returns a
@@ -774,7 +774,7 @@ surfaces through `sp.list_functions`, `sp.describe_function`, and
 
 ### Added — P1 frontier (4 methods, within-sprint week 2)
 
-- **`sp.harvest_did`** — Borusyak et al. MIT/NBER WP 34550 (2025)
+- **`sp.harvest_did`** — Abadie, Angrist, Frandsen & Pischke, NBER WP 34550 (2025)
   Harvesting DID + event-study framework: extracts every valid 2×2
   DID comparison from a staggered panel, combines them via
   inverse-variance weights, and reports event-study + pretrend Wald
@@ -881,23 +881,23 @@ parameter docs, examples, references, and tags.
 - `sp.kernel_iv` (top-level alias of `sp.iv.kernel_iv`) — Lob et al.
   (2025, arXiv:2511.21603) kernel IV regression with wild-bootstrap
   uniform confidence band over the structural function `h*(d)`.
-- `sp.continuous_iv_late` (top-level alias) — Xie et al. (2025,
+- `sp.continuous_iv_late` (top-level alias) — Zeng et al. (2025,
   arXiv:2504.03063) LATE on the maximal complier class for continuous
   instruments via quantile-bin Wald estimator. (Also fixed a summary
   formatting bug — see below.)
 
 **TMLE**
 
 - `sp.hal_tmle` + `sp.HALRegressor` / `sp.HALClassifier` — TMLE with
-  Highly Adaptive Lasso nuisance learners (Qian & van der Laan, 2025,
-  arXiv:2506.17214). Two variants: `"delta"` (plug HAL into standard
+  Highly Adaptive Lasso nuisance learners (Li, Qiu, Wang & van der
+  Laan, 2025, arXiv:2506.17214). Two variants: `"delta"` (plug HAL into standard
   TMLE) and `"projection"` (apply tangent-space shrinkage to the
   targeting epsilon). Recovers ATE within ~3% on n=400 with rich
   nuisance.
 
 **Synthetic Control**
 
-- `sp.synth_survival` — Synthetic Survival Control (Agarwal & Shah,
+- `sp.synth_survival` — Synthetic Survival Control (Han & Shah,
   2025, arXiv:2511.14133). Donor convex combination on the
   complementary log-log scale matches the treated arm's pre-treatment
   Kaplan-Meier, then projects forward and reports the survival gap
@@ -1065,9 +1065,10 @@ all routed through `sp.*` and registered in `sp.list_functions()`.
   `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).
+- **`sp.causal_rl`** (NEW namespace): `causal_dqn` (Li-Zhang-Bareinboim
+  confounding-robust Deep Q, arXiv 2510.21110), `causal_rl_benchmark`
+  (5 benchmarks per Cunha-Liu-French-Mian, arXiv 2512.18135),
+  `offline_safe_policy` (Chemingui et al., 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).

docs/guides/bridging_theorems.md

@@ -27,11 +27,11 @@ disagree beyond sampling error.
 
 | `kind`          | Theorem                                           | Reference                                                       |
 |-----------------|---------------------------------------------------|-----------------------------------------------------------------|
-| `did_sc`        | DiD ≡ Synthetic Control                           | Shi & Athey, arXiv:2503.11375 (2025)                            |
-| `ewm_cate`      | EWM ≡ CATE-max policy                             | Ferman et al., arXiv:2510.26723 (2025)                          |
-| `cb_ipw`        | Covariate Balancing ≡ IPW ≡ DR                    | Zhao & Percival, arXiv:2310.18563 v6 (2025)                     |
+| `did_sc`        | DiD ≡ Synthetic Control                           | Sun, Xie & Zhang, arXiv:2503.11375 (2025)                       |
+| `ewm_cate`      | EWM ≡ CATE-max policy                             | Kato, arXiv:2510.26723 (2025)                                   |
+| `cb_ipw`        | Covariate Balancing ≡ IPW ≡ DR                    | Słoczyński, Uysal & Wooldridge, arXiv:2310.18563 (2023)         |
 | `kink_rdd`      | Bunching ≡ Kink RDD first-order expansion         | Lu, Wang, Xie, arXiv:2404.09117 (2025)                          |
-| `dr_calib`      | Doubly-robust ≡ outcome + Riesz joint calibration | Zhang et al., arXiv:2411.02771 (2025)                           |
+| `dr_calib`      | Doubly-robust ≡ outcome + Riesz joint calibration | van der Laan, Luedtke & Carone, arXiv:2411.02771 (2024)         |
 | `surrogate_pci` | Long-term Surrogate Index ≡ PCI                   | Imbens, Kallus, Mao & Wang, JRSS-B 87(2) 2025; arXiv:2202.07234 |
 
 Each bridge is importable at top level as `sp.bridge(kind="..." ...)` or

docs/guides/causal_mas.md

@@ -193,7 +193,7 @@ Explicit kwargs always win over the name heuristic.
 - arXiv:2509.00987 (2025/09).
   *Causal MAS: A Survey of Large Language Model Architectures for
   Discovery and Effect Estimation.*
-- Kiciman, E. & Sharma, A. (2025).
-  *Causal Reasoning with Large Language Models.*  arXiv:2402.11068.
+- Wan, G., Lu, Y., Wu, Y., Hu, M. & Li, S. (2024).
+  *Enhancing Causal Discovery with Large Language Models.*  arXiv:2402.11068.
 - Anthropic (2025).
   *Extended Thinking in Claude 4.5 / Opus 4.7.*  Technical report.

docs/guides/causal_rl_family.md

@@ -81,7 +81,7 @@ data will happily converge to a Q-function that is biased in the direction
 of the logging policy's mistakes.  `causal_dqn` explicitly caps how much
 the confounding can distort Q and returns a sensitivity range.
 
-Citation: Fu & Zhou (2025), arXiv:2510.21110.
+Citation: Li, Zhang & Bareinboim (2025), arXiv:2510.21110.
 
 ---
 
@@ -174,8 +174,8 @@ bench.regret_curves    # dict: algorithm → regret over time
 bench.best             # algorithm with lowest cumulative regret
 ```
 
-Reference: Zhou-Bareinboim (2025), arXiv:2512.18135 — the canonical
-unified causal-RL benchmark taxonomy.
+Reference: Cunha, Liu, French & Mian (2025), arXiv:2512.18135 — a
+canonical unified causal-RL benchmark taxonomy.
 
 ---
 

docs/guides/causal_text_family.md

@@ -169,7 +169,7 @@ and human labels exist, then divide:
 
 ## References
 
-- Veitch, V., Wang, Y., & Blei, D. M. (2020). "Adapting text embeddings for causal inference." *UAI*. [arXiv:1905.12741](https://arxiv.org/abs/1905.12741).
+- Veitch, V., Sridhar, D., & Blei, D. M. (2019). "Adapting text embeddings for causal inference." *UAI*. [arXiv:1905.12741](https://arxiv.org/abs/1905.12741).
 - Egami, N., Hinck, M., Stewart, B., & Wei, H. (2024). "Using imperfect surrogates for downstream inference: Design-based supervised learning for social science applications of large language models." *NeurIPS*. [arXiv:2306.04746](https://arxiv.org/abs/2306.04746).
 - Hausman, J., Abrevaya, J., & Scott-Morton, F. (1998). "Misclassification of the dependent variable in a discrete-response setting." *Journal of Econometrics*, 87, 239–269.
 - Aigner, D. J. (1973). "Regression with a binary independent variable subject to errors of observation." *Journal of Econometrics*, 1(1), 49–59.

docs/guides/choosing_iv_estimator.md

@@ -144,7 +144,7 @@ too restrictive, v1.2 ships two non-parametric alternatives:
 
 - `sp.kernel_iv` (Lob et al. 2025, arXiv:2511.21603) delivers a uniform
   confidence band over the whole `D`-grid, not just pointwise.
-- `sp.continuous_iv_late` (Xie et al. 2025, arXiv:2504.03063) identifies
+- `sp.continuous_iv_late` (Zeng et al. 2025, arXiv:2504.03063) identifies
   the LATE on the subpopulation most responsive to the instrument —
   the natural generalisation of Angrist-Imbens LATE beyond binary `Z`.
 

docs/guides/conformal_family.md

@@ -295,7 +295,7 @@ print(r.dose_curves)       # full curve per test row across the grid
 fertilizer application, time-in-app. Returns a point estimate + band
 at each `(test_row, dose)` pair in the grid.
 
-Citation: Kim, Jeong, Barber, Lee (arXiv:2407.03094, 2024).
+Citation: Schröder et al. (arXiv:2407.03094, 2024).
 
 ---
 

docs/guides/harvest_did.md

@@ -1,6 +1,6 @@
 # Harvesting DID + Event-Study Designs
 
-> Borusyak, Hull & Jaravel (MIT/NBER WP 34550, 2025).
+> Abadie, Angrist, Frandsen & Pischke (NBER WP 34550, 2025).
 
 ## 1. What "harvesting" means
 
@@ -112,11 +112,11 @@ boot = bootstrap(df, stat, n_boot=500, cluster='unit', random_state=0)
 
 ## 7. References
 
-- Borusyak, Hull & Jaravel (MIT/NBER WP 34550, 2025).
-  "Harvesting Differences-in-Differences and Event-Study Designs."
+- Abadie, Angrist, Frandsen & Pischke (NBER WP 34550, 2025).
+  "Harvesting Differences-in-Differences and Event-Study Evidence."
 - Callaway & Sant'Anna (2021).  "Difference-in-Differences with
   multiple time periods."  *JoE* 225.
-- Roth, Sant'Anna, Bilinski & Poe (2025).
+- Baker, Callaway, Cunningham, Goodman-Bacon & Sant'Anna (2025).
   "Difference-in-Differences Designs: A Practitioner's Guide."
   arXiv:2503.13323.
 

docs/guides/interference_family.md

@@ -149,7 +149,7 @@ first-order nuisance misspecification. Requires a **pre-computed
 scalar neighbor exposure summary** — you control the mapping (share of
 treated neighbors is the usual choice). Needs `n ≥ 10 * n_folds`.
 
-Citation: Parmigiani et al. (arXiv:2509.18484, 2025).
+Citation: Wu & Yuan (arXiv:2509.18484, 2025).
 
 ### `sp.inward_outward_spillover` — directed networks
 
@@ -172,7 +172,7 @@ print(r.inward_effect, r.outward_effect, r.ratio_in_out)
 influence flow the same way both directions?"  A `ratio_in_out` far
 from 1 is the signal of directional asymmetry.
 
-Citation: Li, Ratkovic et al. (arXiv:2506.06615, 2025).
+Citation: Fang, Airoldi & Forastiere (arXiv:2506.06615, 2025).
 
 ---
 
@@ -218,9 +218,9 @@ print(r.spillover_effect, r.spillover_se)
 adjacent and leakage is plausible. The neighbor treatment share must
 be pre-computed from user-supplied adjacency.
 
-Reference: Ding et al. (2025), arXiv:2310.18836.
+Reference: Leung (2023), arXiv:2310.18836.
 
-### `sp.cluster_staggered_rollout` — Zhou et al. (2025)
+### `sp.cluster_staggered_rollout` — Chen & Li (2025)
 
 Cluster RCT with **staggered** treatment adoption — different clusters
 turn on at different calendar times. Produces a dynamic event-study
@@ -242,9 +242,9 @@ print(r.event_study)      # rel_time, att, ci_low, ci_high
 For individual-level staggered DiD without clustering, use
 `sp.callaway_santanna` instead.
 
-Reference: Zhou et al. (2025), arXiv:2502.10939.
+Reference: Chen & Li (2025), arXiv:2502.10939.
 
-### `sp.dnc_gnn_did` — double negative controls + GNN (Zhao et al. 2026)
+### `sp.dnc_gnn_did` — double negative controls + GNN (Zhang, Fu & Wang 2026)
 
 When you suspect **unmeasured** network confounding (e.g. latent
 homophily), combine double negative controls with a GNN embedding of
@@ -268,7 +268,7 @@ rooted in network structure (position, centrality, latent communities).
 Requires at least one valid negative-control outcome and one valid
 negative-control exposure — chosen from domain knowledge.
 
-Reference: Zhao et al. (2026), arXiv:2601.00603.
+Reference: Zhang, Fu & Wang (2026), arXiv:2601.00603.
 
 ---