Initial reviewer pack — condensed verifiable archive for EMNLP 2026 submission

README.md

@@ -0,0 +1,184 @@
+---
+language:
+  - en
+license: cc-by-4.0
+pretty_name: GEODML — EMNLP 2026 Reviewer Pack (Condensed)
+size_categories:
+  - n<1K
+tags:
+  - causal-inference
+  - double-ml
+  - dml
+  - llm
+  - reranker
+  - search
+  - geo
+  - generative-engine-optimization
+  - interpretability
+  - probing
+  - saliency
+  - ablation
+  - emnlp
+  - emnlp-2026
+  - reproducibility
+task_categories:
+  - text-ranking
+  - text-classification
+---
+
+# GEODML — EMNLP 2026 reviewer pack
+
+A **condensed** companion to the EMNLP 2026 submission **"Causal
+Analysis of LLM Search Rerankers via Double/Debiased Machine
+Learning."** Designed for fast verification, not full reproduction.
+
+- **This pack** (5.6 MB) — every paper claim as a CSV + every figure
+  as PDF/PNG + a one-shot `verify.py` claim-checker.
+- **Full reproducibility dataset** (1.8 GB) →
+  [`ValerianFourel/geodml-emnlp-2026`](https://huggingface.co/datasets/ValerianFourel/geodml-emnlp-2026)
+  — raw LLM rerank outputs, page features, DFS confounders, and the
+  full DML/probing/saliency/ablation code.
+
+---
+
+## Setup (60 seconds)
+
+```bash
+# 1. Pull the reviewer pack
+huggingface-cli download ValerianFourel/geodml-emnlp-2026-reviewer \
+    --repo-type=dataset --local-dir ./geodml-reviewer
+cd geodml-reviewer
+
+# 2. Install
+pip install pandas
+
+# 3. Run the one-shot claim-checker
+python verify.py
+```
+
+`verify.py` prints every headline number alongside its
+paper-text expected value and a tolerance. The last line says either
+
+```
+   All paper claims VERIFIED against the tables.
+```
+
+or it lists which claims fall outside tolerance.
+
+---
+
+## What is what
+
+### `tables/`  — condensed paper claims (CSV)
+
+| File | Rows | What it is |
+|---|---|---|
+| `table2_dml_headline.csv` | 18 | **The paper's Table 2.** Spec B (mutually-controlled) DML coefficients for 6 treatments × 3 outcomes, pooled across all slices. Columns: outcome, treatment, coef, se, ci_lo, ci_hi, p_val, sig_stars, n, coef_promoter_dir. |
+| `dml_all_specs_all_slices.csv` | 216 | Every fitted DML model — Spec A (marginal) and Spec B (partial), across 11 slices (POOLED + 10 sub-slices), 6 treatments, 3 outcomes. Columns: spec, slice, treatment, outcome, coef, se, p_val, n, bonferroni_sig. |
+| `probing_peaks_per_variant.csv` | ~56 | Peak linear-probe ROC AUC per (variant, treatment, pooling). Columns: variant, treatment, pooling, peak_layer, peak_roc_auc, layer_0_roc_auc. |
+| `admission_probe_headline.csv` | 8 | The admission pre-commitment story — L0, peak, and final-layer ROC AUC of the `Y1_admission_inctx` probe across 4 variants × 2 poolings. |
+| `saliency_summary.csv` | 8 | Llama-3.3-70B vs Qwen-2.5-72B gradient-saliency ratios on 4 treatments. Numbers >1 mean the model attends more to treatment tokens than to surrounding tokens. |
+| `ablation_summary.csv` | 24 | Mean / median / std Δrank when treatment tokens are ablated, per (treatment, prompt, frame). |
+| `freshness_leakage_diagnostic.md` | — | Year-token leakage diagnostic for T6 freshness — included as a methodological caveat. |
+
+### `figures/`  — every PDF + PNG the paper references
+
+Each figure is included as both PDF (for camera-ready) and PNG (for
+quick HTML preview).
+
+| File | What it is |
+|---|---|
+| `dml_estimation_framework.{pdf,png}` | The DML estimation framework diagram (data flow). |
+| `fig01_admission_forest.{pdf,png}` | Forest plot of Spec B coefficients on Y1 admission. |
+| `fig02_rank_delta_forest.{pdf,png}` | Forest plot on Y2 = rank_delta. |
+| `fig03_post_rank_forest.{pdf,png}` | Forest plot on Y3 = rank_post. |
+| `fig04_three_outcome_grid.{pdf,png}` | All three outcomes side-by-side. |
+| `fig05_admission_vs_rank_scatter.{pdf,png}` | Y1 vs Y2 per treatment — admission and promotion magnitude correlate. |
+| `fig06_marginal_vs_partial.{pdf,png}` | Spec A vs Spec B (does mutual control change the story?). |
+| `fig07_variant_grid_admission.{pdf,png}` | Y1 across the 4 prompt variants. |
+| `fig08_variant_grid_rank.{pdf,png}` | Y2 across the 4 prompt variants. |
+| `fig09_compare_rag_vs_snippet.{pdf,png}` | RAG vs snippet-only prompt comparison. |
+| `fig10_compare_ddg_vs_searxng.{pdf,png}` | DuckDuckGo vs SearXNG SERP-backend comparison. |
+| `fig11_compare_llama_vs_qwen.{pdf,png}` | Llama vs Qwen reranker comparison. |
+| `fig12_compare_pool20_vs_pool50.{pdf,png}` | Top-20 vs top-50 candidate-pool comparison. |
+| `fig13_admission_detail.{pdf,png}` | Detail view of admission coefficients with CIs. |
+| `fig14_robust_survivors.{pdf,png}` | The robust-winners frame (slice-stable treatments). |
+| `fig_probing_layerwise.{pdf,png}` | Layer-wise probe ROC AUC per treatment. |
+| `fig_probing_pooling.{pdf,png}` | Probe ROC AUC by token-pooling strategy. |
+| `fig_admission_pooled.{pdf,png}` | Admission probe across layers, pooled across variants. |
+| `fig_admission_variants.{pdf,png}` | Admission probe across layers, split by variant. |
+| `fig_saliency.{pdf,png}` | Diverging-bar saliency-ratio chart (Llama vs Qwen). |
+
+### `verify.py` — one-shot claim checker
+
+Reads every CSV in `tables/` and asserts each headline number against
+its paper-text value within tolerance. Exits 0 on PASS, 1 on FAIL.
+Run after every pull to confirm the data matches what the paper says.
+
+---
+
+## Headline numbers — Table 2 (Spec B POOLED)
+
+DML coefficients with `has_llms_txt` as a confounder (not a treatment),
+across 6 canonical content treatments.
+
+| Treatment | Y1 selected | Y2 Δrank | Y3 rank_post |
+|---|---|---|---|
+| T5 topical competence | **+0.037***\* | **−0.530***\* | **−0.299***\* |
+| T2a Q-headings        | **+0.016***\* | **+0.136***\* | **−0.041*** |
+| T3 schema (JSON-LD)   | **−0.014***\* | **−0.051*** | **+0.095***\* |
+| T6 freshness          | **−0.005***\* | **−0.061***\* | +0.005     |
+| T4 citation authority | +0.001     | **−0.023***  | **−0.015*** |
+| T1b stats density     | −0.000     | −0.003       | −0.002     |
+
+**Sign convention:** `+Y1`, `+Y2`, `−Y3` ⇒ URL favoured by the LLM.
+**Stars:** `***` < .001, `**` < .01, `*` < .05.
+
+---
+
+## Admission pre-commitment probe
+
+Linear probe for the binary admission outcome on the LLM's *internal
+activations* at each layer. Averaged across the 4 prompt variants,
+mean pooling over context tokens.
+
+| Quantity | Value | Interpretation |
+|---|---|---|
+| Layer 0 ROC AUC  | 0.671  | Embedding alone — weak admission signal. |
+| Peak ROC AUC     | 0.862  | At layer 60 (~75% network depth). |
+| L0 → peak gain   | +0.191 | Genuine compositional integration. |
+
+---
+
+## Saliency (gradient attribution)
+
+Saliency ratio = mean |∇| on treatment tokens / mean |∇| on
+surrounding-context tokens. `>1` means the model attends more to the
+treatment than to context; `<1` means it ignores it.
+
+| Treatment | Llama-3.3-70B | Qwen-2.5-72B | DML direction | Reading |
+|---|---|---|---|---|
+| T1b stats density | 0.89× | **1.93×** | null | Qwen looks at stats but DML says it doesn't move ranks. |
+| T2a Q-headings    | 1.05× | 0.90×    | promoter | Comparable attention; works for both. |
+| T3 schema         | **0.19×** | **0.40×** | demoter | Both models ignore JSON-LD — DML demoter is consistent. |
+
+---
+
+## Citation
+
+```bibtex
+@inproceedings{fourel2026geodml,
+  title     = {Causal Analysis of LLM Search Rerankers via Double/Debiased Machine Learning},
+  author    = {Fourel, Valerian},
+  year      = {2026},
+  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing}
+}
+```
+
+## License
+
+CC BY 4.0 — free reuse with attribution.
+
+## Contact
+
+valerian.fourel@gmail.com

tables/ablation_summary.csv

@@ -0,0 +1,25 @@
+frame,prompt,treatment,n_ablations,mean_delta_r,median_delta_r,std_delta_r
+full,biased,T1b_stats_density,3124,-0.015,0.0,1.7145
+full,biased,T2a_question_headings,10750,0.0269,0.0,1.431
+full,biased,T3_structured_data_new,11303,0.0084,0.0,1.3022
+full,biased,T5_topical_comp,3245,-0.1673,0.0,2.5493
+full,biased,T6_freshness,2756,0.041,0.0,1.2667
+full,biased,T7_source_earned,1822,-0.219,0.0,2.3205
+full,neutral,T1b_stats_density,4776,-0.005,0.0,1.0837
+full,neutral,T2a_question_headings,17342,-0.0034,0.0,0.8264
+full,neutral,T3_structured_data_new,18366,-0.005,0.0,0.846
+full,neutral,T5_topical_comp,4835,0.0378,0.0,2.208
+full,neutral,T6_freshness,4663,-0.0006,0.0,0.791
+full,neutral,T7_source_earned,3122,-0.0003,0.0,0.9527
+robust_winners,biased,T1b_stats_density,713,-0.0266,0.0,1.0267
+robust_winners,biased,T2a_question_headings,2498,0.0156,0.0,0.7684
+robust_winners,biased,T3_structured_data_new,2663,0.0293,0.0,0.6029
+robust_winners,biased,T5_topical_comp,737,-0.1235,0.0,1.77
+robust_winners,biased,T6_freshness,626,0.0128,0.0,0.3344
+robust_winners,biased,T7_source_earned,331,-0.0181,0.0,1.1172
+robust_winners,neutral,T1b_stats_density,746,0.0,0.0,0.3514
+robust_winners,neutral,T2a_question_headings,2642,0.0,0.0,0.4044
+robust_winners,neutral,T3_structured_data_new,2820,0.0028,0.0,0.3442
+robust_winners,neutral,T5_topical_comp,784,0.0459,0.0,0.9128
+robust_winners,neutral,T6_freshness,652,0.0,0.0,0.557
+robust_winners,neutral,T7_source_earned,364,0.0192,0.0,0.5323

tables/admission_probe_headline.csv

@@ -0,0 +1,9 @@
+variant,pooling,layer_0,layer_peak,auc_peak,layer_last,auc_last,delta_L0_to_peak
+biased,last_token,0.522,60,0.8823,80,0.8384,0.3603
+biased,mean,0.6971,56,0.8902,80,0.8447,0.1931
+neutral,last_token,0.521,61,0.8757,80,0.8127,0.3547
+neutral,mean,0.6399,53,0.8675,80,0.8023,0.2276
+biased_rag,last_token,0.4847,60,0.8651,80,0.8272,0.3804
+biased_rag,mean,0.6909,73,0.8768,80,0.8558,0.1859
+neutral_rag,last_token,0.5093,53,0.8589,80,0.8191,0.3496
+neutral_rag,mean,0.6563,60,0.87,80,0.8185,0.2137

tables/dml_all_specs_all_slices.csv

@@ -0,0 +1,217 @@
+spec,slice,treatment,outcome,coef,se,p_val,n,bonferroni_sig
+A,POOLED,treat_stats_density,selected,-0.0005,0.0003,0.11304,200000,False
+A,POOLED,treat_question_headings,selected,0.0118,0.0029,4e-05,200000,True
+A,POOLED,treat_structured_data,selected,-0.026,0.0027,0.0,200000,True
+A,POOLED,T4_citation_authority_code,selected,0.0009,0.0007,0.21549,185176,False
+A,POOLED,treat_topical_comp,selected,0.0221,0.0094,0.01841,200000,False
+A,POOLED,treat_freshness,selected,-0.0061,0.0007,0.0,200000,True
+A,VAR:biased,treat_stats_density,selected,-0.0011,0.0005,0.02666,55430,False
+A,VAR:biased,treat_question_headings,selected,0.0154,0.0053,0.00378,57964,False
+A,VAR:biased,treat_structured_data,selected,-0.037,0.0051,0.0,57964,True
+A,VAR:biased,T4_citation_authority_code,selected,-0.0017,0.0011,0.13565,46294,False
+A,VAR:biased,treat_topical_comp,selected,0.0117,0.0186,0.53062,51098,False
+A,VAR:biased,treat_freshness,selected,-0.0121,0.0014,0.0,57964,True
+A,VAR:neutral,treat_stats_density,selected,0.0004,0.0005,0.43819,55430,False
+A,VAR:neutral,treat_question_headings,selected,0.0024,0.0046,0.60395,57964,False
+A,VAR:neutral,treat_structured_data,selected,-0.0084,0.0044,0.05607,57964,False
+A,VAR:neutral,T4_citation_authority_code,selected,0.0002,0.0009,0.79294,46294,False
+A,VAR:neutral,treat_topical_comp,selected,0.034,0.0157,0.03014,51098,False
+A,VAR:neutral,treat_freshness,selected,0.0013,0.0012,0.29809,57964,False
+A,VAR:biased_rag,treat_stats_density,selected,-0.0014,0.0005,0.0067,55430,False
+A,VAR:biased_rag,treat_question_headings,selected,0.0172,0.0054,0.0016,57964,False
+A,VAR:biased_rag,treat_structured_data,selected,-0.046,0.0052,0.0,57964,True
+A,VAR:biased_rag,T4_citation_authority_code,selected,0.0028,0.0012,0.01465,46294,False
+A,VAR:biased_rag,treat_topical_comp,selected,0.0033,0.0187,0.86137,51098,False
+A,VAR:biased_rag,treat_freshness,selected,-0.0156,0.0014,0.0,57964,True
+A,VAR:neutral_rag,treat_stats_density,selected,0.0006,0.0004,0.09237,55430,False
+A,VAR:neutral_rag,treat_question_headings,selected,0.0091,0.0046,0.04649,57964,False
+A,VAR:neutral_rag,treat_structured_data,selected,0.003,0.0044,0.49587,57964,False
+A,VAR:neutral_rag,T4_citation_authority_code,selected,0.0005,0.0008,0.5623,46294,False
+A,VAR:neutral_rag,treat_topical_comp,selected,0.0389,0.0154,0.01169,51098,False
+A,VAR:neutral_rag,treat_freshness,selected,0.0032,0.0012,0.00714,57964,False
+A,ENG:ddg,treat_stats_density,selected,-0.0002,0.0005,0.71736,87008,False
+A,ENG:ddg,treat_question_headings,selected,0.0121,0.005,0.01636,89160,False
+A,ENG:ddg,treat_structured_data,selected,-0.0264,0.0049,0.0,89160,True
+A,ENG:ddg,T4_citation_authority_code,selected,0.0017,0.0008,0.0362,87984,False
+A,ENG:ddg,treat_topical_comp,selected,0.0355,0.024,0.13947,52328,False
+A,ENG:ddg,treat_freshness,selected,-0.0026,0.0013,0.04881,89160,False
+A,ENG:searxng,treat_stats_density,selected,-0.0006,0.0004,0.12829,134712,False
+A,ENG:searxng,treat_question_headings,selected,0.0084,0.0034,0.01481,142696,False
+A,ENG:searxng,treat_structured_data,selected,-0.0265,0.0032,0.0,142696,True
+A,ENG:searxng,T4_citation_authority_code,selected,-0.0013,0.0012,0.26395,97192,False
+A,ENG:searxng,treat_topical_comp,selected,0.0096,0.0109,0.37888,152064,False
+A,ENG:searxng,treat_freshness,selected,-0.0098,0.0009,0.0,142696,True
+A,MOD:Llama,treat_stats_density,selected,-0.0009,0.0004,0.0157,110860,False
+A,MOD:Llama,treat_question_headings,selected,0.0143,0.0037,0.00011,115928,True
+A,MOD:Llama,treat_structured_data,selected,-0.0244,0.0035,0.0,115928,True
+A,MOD:Llama,T4_citation_authority_code,selected,0.0007,0.0009,0.38913,92588,False
+A,MOD:Llama,treat_topical_comp,selected,0.0345,0.0128,0.00691,102196,False
+A,MOD:Llama,treat_freshness,selected,-0.0056,0.001,0.0,115928,True
+A,MOD:Qwen2.5,treat_stats_density,selected,0.0001,0.0004,0.70987,110860,False
+A,MOD:Qwen2.5,treat_question_headings,selected,0.0089,0.0038,0.01781,115928,False
+A,MOD:Qwen2.5,treat_structured_data,selected,-0.0259,0.0036,0.0,115928,True
+A,MOD:Qwen2.5,T4_citation_authority_code,selected,0.0012,0.0008,0.15312,92588,False
+A,MOD:Qwen2.5,treat_topical_comp,selected,0.0132,0.013,0.3078,102196,False
+A,MOD:Qwen2.5,treat_freshness,selected,-0.0069,0.001,0.0,115928,True
+A,POOL:20,treat_stats_density,selected,-0.001,0.0005,0.02845,102896,False
+A,POOL:20,treat_question_headings,selected,0.0101,0.0041,0.01374,107552,False
+A,POOL:20,treat_structured_data,selected,-0.0352,0.0038,0.0,107552,True
+A,POOL:20,T4_citation_authority_code,selected,-0.0001,0.0009,0.87831,95016,False
+A,POOL:20,treat_topical_comp,selected,0.023,0.015,0.12441,93456,False
+A,POOL:20,treat_freshness,selected,-0.0091,0.001,0.0,107552,True
+A,POOL:50,treat_stats_density,selected,-0.0003,0.0004,0.49995,118824,False
+A,POOL:50,treat_question_headings,selected,0.0108,0.0037,0.00347,124304,False
+A,POOL:50,treat_structured_data,selected,-0.0162,0.0035,0.0,124304,True
+A,POOL:50,T4_citation_authority_code,selected,0.002,0.0009,0.03061,90160,False
+A,POOL:50,treat_topical_comp,selected,0.0247,0.0123,0.04462,110936,False
+A,POOL:50,treat_freshness,selected,-0.0046,0.001,0.0,124304,True
+B,POOLED,treat_stats_density,selected,-0.0005,0.0003,0.12311,200000,False
+B,POOLED,treat_question_headings,selected,0.0156,0.0029,0.0,200000,True
+B,POOLED,treat_structured_data,selected,-0.0141,0.0031,0.0,200000,True
+B,POOLED,T4_citation_authority_code,selected,0.0009,0.0007,0.18613,185176,False
+B,POOLED,treat_topical_comp,selected,0.0366,0.0095,0.00012,200000,True
+B,POOLED,treat_freshness,selected,-0.0047,0.0008,0.0,200000,True
+A,POOLED,treat_stats_density,rank_delta,-0.0042,0.0026,0.11064,200000,False
+A,POOLED,treat_question_headings,rank_delta,0.1173,0.024,0.0,200000,True
+A,POOLED,treat_structured_data,rank_delta,-0.122,0.0217,0.0,200000,True
+A,POOLED,T4_citation_authority_code,rank_delta,-0.0233,0.0084,0.00543,200000,False
+A,POOLED,treat_topical_comp,rank_delta,-0.4886,0.0738,0.0,200000,True
+A,POOLED,treat_freshness,rank_delta,-0.0542,0.006,0.0,200000,True
+A,VAR:biased,treat_stats_density,rank_delta,0.003,0.0042,0.47423,54683,False
+A,VAR:biased,treat_question_headings,rank_delta,0.0679,0.0511,0.18385,56183,False
+A,VAR:biased,treat_structured_data,rank_delta,-0.0953,0.0468,0.04182,56183,False
+A,VAR:biased,T4_citation_authority_code,rank_delta,-0.0464,0.0115,5e-05,47616,True
+A,VAR:biased,treat_topical_comp,rank_delta,-0.6114,0.1736,0.00043,46770,True
+A,VAR:biased,treat_freshness,rank_delta,-0.0703,0.0129,0.0,56183,True
+A,VAR:neutral,treat_stats_density,rank_delta,-0.0014,0.0032,0.65535,83750,False
+A,VAR:neutral,treat_question_headings,rank_delta,0.1098,0.0322,0.00065,87771,True
+A,VAR:neutral,treat_structured_data,rank_delta,-0.1413,0.0298,0.0,87771,True
+A,VAR:neutral,T4_citation_authority_code,rank_delta,-0.0046,0.0093,0.61784,65155,False
+A,VAR:neutral,treat_topical_comp,rank_delta,-0.3708,0.1007,0.00023,81896,True
+A,VAR:neutral,treat_freshness,rank_delta,-0.0661,0.0082,0.0,87771,True
+A,VAR:biased_rag,treat_stats_density,rank_delta,0.0008,0.0045,0.85977,47640,False
+A,VAR:biased_rag,treat_question_headings,rank_delta,0.1572,0.0552,0.00439,48952,False
+A,VAR:biased_rag,treat_structured_data,rank_delta,-0.151,0.0486,0.00187,48952,False
+A,VAR:biased_rag,T4_citation_authority_code,rank_delta,-0.0288,0.0117,0.0142,43443,False
+A,VAR:biased_rag,treat_topical_comp,rank_delta,-0.3282,0.1908,0.08541,39092,False
+A,VAR:biased_rag,treat_freshness,rank_delta,-0.0573,0.0137,3e-05,48952,True
+A,VAR:neutral_rag,treat_stats_density,rank_delta,-0.0009,0.0027,0.73571,64804,False
+A,VAR:neutral_rag,treat_question_headings,rank_delta,0.1699,0.0322,0.0,67909,True
+A,VAR:neutral_rag,treat_structured_data,rank_delta,-0.1003,0.0294,0.00065,67909,True
+A,VAR:neutral_rag,T4_citation_authority_code,rank_delta,-0.0008,0.0076,0.92026,53281,False
+A,VAR:neutral_rag,treat_topical_comp,rank_delta,-0.3665,0.1028,0.00036,59333,True
+A,VAR:neutral_rag,treat_freshness,rank_delta,-0.044,0.008,0.0,67909,True
+A,ENG:ddg,treat_stats_density,rank_delta,0.0012,0.0048,0.80743,86258,False
+A,ENG:ddg,treat_question_headings,rank_delta,-0.0219,0.046,0.63366,88321,False
+A,ENG:ddg,treat_structured_data,rank_delta,-0.2546,0.0425,0.0,88321,True
+A,ENG:ddg,T4_citation_authority_code,rank_delta,-0.0024,0.011,0.82842,82765,False
+A,ENG:ddg,treat_topical_comp,rank_delta,-1.4154,0.2259,0.0,47884,True
+A,ENG:ddg,treat_freshness,rank_delta,-0.1099,0.0112,0.0,88321,True
+A,ENG:searxng,treat_stats_density,rank_delta,-0.0013,0.0028,0.65063,164619,False
+A,ENG:searxng,treat_question_headings,rank_delta,0.2293,0.0239,0.0,172494,True
+A,ENG:searxng,treat_structured_data,rank_delta,-0.0457,0.0216,0.03425,172494,False
+A,ENG:searxng,T4_citation_authority_code,rank_delta,-0.053,0.0105,0.0,126730,True
+A,ENG:searxng,treat_topical_comp,rank_delta,-0.1701,0.0739,0.02143,179207,False
+A,ENG:searxng,treat_freshness,rank_delta,-0.0302,0.0061,0.0,172494,True
+A,MOD:Llama,treat_stats_density,rank_delta,0.0036,0.003,0.23344,124403,False
+A,MOD:Llama,treat_question_headings,rank_delta,0.0918,0.0312,0.00324,129382,False
+A,MOD:Llama,treat_structured_data,rank_delta,-0.1425,0.0282,0.0,129382,True
+A,MOD:Llama,T4_citation_authority_code,rank_delta,-0.0168,0.0107,0.11558,104381,False
+A,MOD:Llama,treat_topical_comp,rank_delta,-0.6311,0.1005,0.0,111772,True
+A,MOD:Llama,treat_freshness,rank_delta,-0.0651,0.0077,0.0,129382,True
+A,MOD:Qwen2.5,treat_stats_density,rank_delta,-0.0057,0.003,0.05395,126474,False
+A,MOD:Qwen2.5,treat_question_headings,rank_delta,0.1348,0.0277,0.0,131433,True
+A,MOD:Qwen2.5,treat_structured_data,rank_delta,-0.1245,0.0254,0.0,131433,True
+A,MOD:Qwen2.5,T4_citation_authority_code,rank_delta,-0.031,0.0075,4e-05,105114,True
+A,MOD:Qwen2.5,treat_topical_comp,rank_delta,-0.2897,0.0903,0.00134,115319,False
+A,MOD:Qwen2.5,treat_freshness,rank_delta,-0.0546,0.007,0.0,131433,True
+A,POOL:20,treat_stats_density,rank_delta,-0.0073,0.0032,0.02291,112637,False
+A,POOL:20,treat_question_headings,rank_delta,0.2197,0.0296,0.0,116981,True
+A,POOL:20,treat_structured_data,rank_delta,-0.0045,0.0262,0.86248,116981,False
+A,POOL:20,T4_citation_authority_code,rank_delta,-0.0258,0.0076,0.00063,105269,True
+A,POOL:20,treat_topical_comp,rank_delta,0.0114,0.1036,0.91208,99674,False
+A,POOL:20,treat_freshness,rank_delta,-0.0214,0.0074,0.00363,116981,False
+A,POOL:50,treat_stats_density,rank_delta,0.0046,0.0035,0.19036,138240,False
+A,POOL:50,treat_question_headings,rank_delta,0.0581,0.0299,0.0522,143834,False
+A,POOL:50,treat_structured_data,rank_delta,-0.1975,0.0273,0.0,143834,True
+A,POOL:50,T4_citation_authority_code,rank_delta,-0.0202,0.0156,0.19362,104226,False
+A,POOL:50,treat_topical_comp,rank_delta,-0.7318,0.0957,0.0,127417,True
+A,POOL:50,treat_freshness,rank_delta,-0.0853,0.0074,0.0,143834,True
+B,POOLED,treat_stats_density,rank_delta,-0.0027,0.0027,0.31517,200000,False
+B,POOLED,treat_question_headings,rank_delta,0.1359,0.0245,0.0,200000,True
+B,POOLED,treat_structured_data,rank_delta,-0.0506,0.0245,0.0392,200000,False
+B,POOLED,T4_citation_authority_code,rank_delta,-0.0234,0.0087,0.00738,200000,False
+B,POOLED,treat_topical_comp,rank_delta,-0.5303,0.0747,0.0,200000,True
+B,POOLED,treat_freshness,rank_delta,-0.0608,0.0066,0.0,200000,True
+A,POOLED,treat_stats_density,post_rank,-0.0024,0.0019,0.19176,200000,False
+A,POOLED,treat_question_headings,post_rank,-0.048,0.0184,0.00892,200000,False
+A,POOLED,treat_structured_data,post_rank,0.1225,0.0173,0.0,200000,True
+A,POOLED,T4_citation_authority_code,post_rank,-0.0184,0.0064,0.00414,200000,False
+A,POOLED,treat_topical_comp,post_rank,-0.2141,0.0594,0.00031,200000,True
+A,POOLED,treat_freshness,post_rank,0.015,0.0048,0.00169,200000,False
+A,VAR:biased,treat_stats_density,post_rank,-0.0008,0.0026,0.74255,64977,False
+A,VAR:biased,treat_question_headings,post_rank,0.0225,0.03,0.45391,67034,False
+A,VAR:biased,treat_structured_data,post_rank,0.1258,0.0282,1e-05,67034,True
+A,VAR:biased,T4_citation_authority_code,post_rank,0.005,0.0083,0.55197,47616,False
+A,VAR:biased,treat_topical_comp,post_rank,-0.1205,0.1073,0.2615,53449,False
+A,VAR:biased,treat_freshness,post_rank,0.0381,0.0078,0.0,67034,True
+A,VAR:neutral,treat_stats_density,post_rank,-0.0016,0.0026,0.5227,85958,False
+A,VAR:neutral,treat_question_headings,post_rank,-0.0401,0.0274,0.14376,90099,False
+A,VAR:neutral,treat_structured_data,post_rank,0.1145,0.0263,1e-05,90099,True
+A,VAR:neutral,T4_citation_authority_code,post_rank,-0.0118,0.0077,0.12359,65155,False
+A,VAR:neutral,treat_topical_comp,post_rank,-0.1332,0.0903,0.13999,83113,False
+A,VAR:neutral,treat_freshness,post_rank,0.0159,0.0072,0.02723,90099,False
+A,VAR:biased_rag,treat_stats_density,post_rank,-0.0057,0.0029,0.04808,60283,False
+A,VAR:biased_rag,treat_question_headings,post_rank,-0.0707,0.0319,0.02684,62162,False
+A,VAR:biased_rag,treat_structured_data,post_rank,0.1112,0.0301,0.00022,62162,True
+A,VAR:biased_rag,T4_citation_authority_code,post_rank,-0.001,0.0087,0.90577,43443,False
+A,VAR:biased_rag,treat_topical_comp,post_rank,-0.3582,0.1169,0.00219,46890,False
+A,VAR:biased_rag,treat_freshness,post_rank,0.0207,0.0083,0.01278,62162,False
+A,VAR:neutral_rag,treat_stats_density,post_rank,-0.0004,0.0026,0.88732,66908,False
+A,VAR:neutral_rag,treat_question_headings,post_rank,-0.0538,0.0308,0.08092,70121,False
+A,VAR:neutral_rag,treat_structured_data,post_rank,0.1429,0.0291,0.0,70121,True
+A,VAR:neutral_rag,T4_citation_authority_code,post_rank,-0.0144,0.0079,0.06678,53281,False
+A,VAR:neutral_rag,treat_topical_comp,post_rank,-0.3077,0.1044,0.00322,60241,False
+A,VAR:neutral_rag,treat_freshness,post_rank,0.0026,0.008,0.7423,70121,False
+A,ENG:ddg,treat_stats_density,post_rank,0.0048,0.0028,0.08743,99103,False
+A,ENG:ddg,treat_question_headings,post_rank,-0.0021,0.0288,0.94243,101894,False
+A,ENG:ddg,treat_structured_data,post_rank,0.0401,0.0274,0.14336,101894,False
+A,ENG:ddg,T4_citation_authority_code,post_rank,-0.0052,0.0076,0.49666,82765,False
+A,ENG:ddg,treat_topical_comp,post_rank,-0.5142,0.1426,0.00031,50082,True
+A,ENG:ddg,treat_freshness,post_rank,-0.0211,0.0073,0.00397,101894,False
+A,ENG:searxng,treat_stats_density,post_rank,-0.0091,0.0024,0.00015,179023,True
+A,ENG:searxng,treat_question_headings,post_rank,-0.0198,0.0195,0.31086,187522,False
+A,ENG:searxng,treat_structured_data,post_rank,0.1487,0.0181,0.0,187522,True
+A,ENG:searxng,T4_citation_authority_code,post_rank,-0.037,0.0114,0.0012,126730,False
+A,ENG:searxng,treat_topical_comp,post_rank,-0.1502,0.0613,0.01418,193611,False
+A,ENG:searxng,treat_freshness,post_rank,0.0452,0.005,0.0,187522,True
+A,MOD:Llama,treat_stats_density,post_rank,-0.0051,0.0021,0.01327,138621,False
+A,MOD:Llama,treat_question_headings,post_rank,-0.0278,0.0215,0.19675,144317,False
+A,MOD:Llama,treat_structured_data,post_rank,0.1354,0.0202,0.0,144317,True
+A,MOD:Llama,T4_citation_authority_code,post_rank,-0.0064,0.0073,0.38109,104381,False
+A,MOD:Llama,treat_topical_comp,post_rank,-0.1981,0.0755,0.0087,120201,False
+A,MOD:Llama,treat_freshness,post_rank,0.0201,0.0056,0.00032,144317,True
+A,MOD:Qwen2.5,treat_stats_density,post_rank,-0.0006,0.0021,0.76898,139505,False
+A,MOD:Qwen2.5,treat_question_headings,post_rank,-0.0436,0.0213,0.04029,145099,False
+A,MOD:Qwen2.5,treat_structured_data,post_rank,0.0968,0.02,0.0,145099,True
+A,MOD:Qwen2.5,T4_citation_authority_code,post_rank,-0.0137,0.0078,0.07772,105114,False
+A,MOD:Qwen2.5,treat_topical_comp,post_rank,-0.2278,0.0738,0.00201,123492,False
+A,MOD:Qwen2.5,treat_freshness,post_rank,0.0133,0.0055,0.01634,145099,False
+A,POOL:20,treat_stats_density,post_rank,-0.0017,0.0026,0.50025,125051,False
+A,POOL:20,treat_question_headings,post_rank,-0.0255,0.0232,0.27216,129913,False
+A,POOL:20,treat_structured_data,post_rank,0.1714,0.0217,0.0,129913,True
+A,POOL:20,T4_citation_authority_code,post_rank,-0.0043,0.0082,0.59819,105269,False
+A,POOL:20,treat_topical_comp,post_rank,-0.0559,0.083,0.50044,108269,False
+A,POOL:20,treat_freshness,post_rank,0.0366,0.006,0.0,129913,True
+A,POOL:50,treat_stats_density,post_rank,-0.0005,0.0022,0.82512,153075,False
+A,POOL:50,treat_question_headings,post_rank,-0.0182,0.021,0.38753,159503,False
+A,POOL:50,treat_structured_data,post_rank,0.0822,0.0197,3e-05,159503,True
+A,POOL:50,T4_citation_authority_code,post_rank,-0.016,0.0092,0.08221,104226,False
+A,POOL:50,treat_topical_comp,post_rank,-0.3308,0.0737,1e-05,135424,True
+A,POOL:50,treat_freshness,post_rank,0.0037,0.0054,0.49703,159503,False
+B,POOLED,treat_stats_density,post_rank,-0.0015,0.0019,0.40542,200000,False
+B,POOLED,treat_question_headings,post_rank,-0.0408,0.0187,0.02907,200000,False
+B,POOLED,treat_structured_data,post_rank,0.0949,0.0196,0.0,200000,True
+B,POOLED,T4_citation_authority_code,post_rank,-0.0154,0.0066,0.01905,200000,False
+B,POOLED,treat_topical_comp,post_rank,-0.2995,0.0601,0.0,200000,True
+B,POOLED,treat_freshness,post_rank,0.0051,0.0053,0.3334,200000,False

tables/probing_peaks_per_variant.csv

@@ -0,0 +1,49 @@
+variant,treatment,pooling,peak_layer,peak_roc_auc,layer_0_roc_auc
+biased,T1b_stats_density,last_token,62,0.9834,0.8611
+biased,T1b_stats_density,mean,32,0.9877,0.9798
+biased,T2a_question_headings,last_token,54,0.9868,0.8805
+biased,T2a_question_headings,mean,24,0.9893,0.98
+biased,T3_structured_data,last_token,35,0.9847,0.8744
+biased,T3_structured_data,mean,18,0.9874,0.9797
+biased,T4_citation_authority,last_token,30,0.9917,0.8955
+biased,T4_citation_authority,mean,71,0.9938,0.9739
+biased,T5_topical_comp,last_token,68,0.9649,0.8261
+biased,T5_topical_comp,mean,46,0.974,0.9605
+biased,T6_freshness,last_token,70,0.9859,0.8752
+biased,T6_freshness,mean,73,0.9894,0.9829
+biased,T7_llms_txt,last_token,19,0.9862,0.8794
+biased,T7_llms_txt,mean,24,0.9896,0.9812
+biased,T7_source_earned,last_token,68,0.9857,0.5644
+biased,T7_source_earned,mean,48,0.9981,0.9856
+biased,Y1_admission_inctx,last_token,60,0.8823,0.522
+biased,Y1_admission_inctx,mean,56,0.8902,0.6971
+neutral,T2a_question_headings,last_token,54,0.9868,0.8805
+neutral,T2a_question_headings,mean,24,0.9893,0.98
+neutral,T5_topical_comp,last_token,68,0.9649,0.8261
+neutral,T5_topical_comp,mean,46,0.974,0.9605
+neutral,T6_freshness,last_token,70,0.9859,0.8752
+neutral,T6_freshness,mean,73,0.9894,0.9829
+neutral,T7_source_earned,last_token,68,0.9805,0.5644
+neutral,T7_source_earned,mean,48,0.9981,0.9856
+neutral,Y1_admission_inctx,last_token,61,0.8757,0.521
+neutral,Y1_admission_inctx,mean,53,0.8675,0.6399
+biased_rag,T2a_question_headings,last_token,54,0.9868,0.8805
+biased_rag,T2a_question_headings,mean,24,0.9893,0.98
+biased_rag,T5_topical_comp,last_token,68,0.9649,0.8261
+biased_rag,T5_topical_comp,mean,46,0.974,0.9605
+biased_rag,T6_freshness,last_token,70,0.9859,0.8752
+biased_rag,T6_freshness,mean,73,0.9894,0.9829
+biased_rag,T7_source_earned,last_token,72,0.9868,0.5194
+biased_rag,T7_source_earned,mean,48,0.9982,0.9858
+biased_rag,Y1_admission_inctx,last_token,60,0.8651,0.4847
+biased_rag,Y1_admission_inctx,mean,73,0.8768,0.6909
+neutral_rag,T2a_question_headings,last_token,54,0.9868,0.8805
+neutral_rag,T2a_question_headings,mean,24,0.9893,0.98
+neutral_rag,T5_topical_comp,last_token,68,0.9649,0.8261
+neutral_rag,T5_topical_comp,mean,46,0.974,0.9605
+neutral_rag,T6_freshness,last_token,70,0.9859,0.8752
+neutral_rag,T6_freshness,mean,73,0.9894,0.9829
+neutral_rag,T7_source_earned,last_token,31,0.9816,0.5194
+neutral_rag,T7_source_earned,mean,48,0.9982,0.9858
+neutral_rag,Y1_admission_inctx,last_token,53,0.8589,0.5093
+neutral_rag,Y1_admission_inctx,mean,60,0.87,0.6563

tables/saliency_summary.csv

@@ -0,0 +1,9 @@
+model,treatment,saliency_ratio,n_treatment_tokens,n_other_tokens
+Llama-3.3-70B,T2a_question_headings,1.048,2935,104901
+Llama-3.3-70B,T1b_stats_density,0.894,64276,104901
+Llama-3.3-70B,T7_source_earned,0.686,10321,104901
+Llama-3.3-70B,T3_structured_data_new,0.189,66,104901
+Qwen-2.5-72B,T1b_stats_density,1.93,106258,104754
+Qwen-2.5-72B,T7_source_earned,1.074,10329,104754
+Qwen-2.5-72B,T2a_question_headings,0.904,2935,104754
+Qwen-2.5-72B,T3_structured_data_new,0.401,66,104754

tables/table2_dml_headline.csv

@@ -0,0 +1,19 @@
+outcome,treatment,coef,se,ci_lo,ci_hi,p_val,sig_stars,n,coef_promoter_dir
+selected,T1b stats density,-0.0005,0.0003,-0.0011,0.0001,0.12311,,200000,-0.0005
+selected,T2a Q-headings,0.0156,0.0029,0.0099,0.0214,0.0,***,200000,0.0156
+selected,T3 schema (JSON-LD),-0.0141,0.0031,-0.0201,-0.0081,0.0,***,200000,-0.0141
+selected,T4 citation authority,0.0009,0.0007,-0.0004,0.0023,0.18613,,185176,0.0009
+selected,T5 topical competence,0.0366,0.0095,0.018,0.0553,0.00012,***,200000,0.0366
+selected,T6 freshness,-0.0047,0.0008,-0.0064,-0.0031,0.0,***,200000,-0.0047
+rank_delta,T1b stats density,-0.0027,0.0027,-0.0079,0.0025,0.31517,,200000,-0.0027
+rank_delta,T2a Q-headings,0.1359,0.0245,0.0879,0.1839,0.0,***,200000,0.1359
+rank_delta,T3 schema (JSON-LD),-0.0506,0.0245,-0.0987,-0.0025,0.0392,*,200000,-0.0506
+rank_delta,T4 citation authority,-0.0234,0.0087,-0.0405,-0.0063,0.00738,**,200000,-0.0234
+rank_delta,T5 topical competence,-0.5303,0.0747,-0.6767,-0.3838,0.0,***,200000,-0.5303
+rank_delta,T6 freshness,-0.0608,0.0066,-0.0737,-0.0478,0.0,***,200000,-0.0608
+post_rank,T1b stats density,-0.0015,0.0019,-0.0052,0.0021,0.40542,,200000,0.0015
+post_rank,T2a Q-headings,-0.0408,0.0187,-0.0774,-0.0042,0.02907,*,200000,0.0408
+post_rank,T3 schema (JSON-LD),0.0949,0.0196,0.0565,0.1332,0.0,***,200000,-0.0949
+post_rank,T4 citation authority,-0.0154,0.0066,-0.0283,-0.0025,0.01905,*,200000,0.0154
+post_rank,T5 topical competence,-0.2995,0.0601,-0.4172,-0.1818,0.0,***,200000,0.2995
+post_rank,T6 freshness,0.0051,0.0053,-0.0052,0.0154,0.3334,,200000,-0.0051

verify.py

@@ -0,0 +1,102 @@
+#!/usr/bin/env python3
+"""verify.py — read every condensed table and print/assert each paper claim.
+
+Run from the reviewer-pack root:
+
+    python verify.py
+
+Exits with code 0 if all assertions pass (within a 2e-3 tolerance on
+coefficient values), prints any mismatches otherwise.
+"""
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pandas as pd
+
+ROOT = Path(__file__).resolve().parent
+TABLES = ROOT / "tables"
+
+TOL = 2e-3  # tolerance on coef/AUC equality checks
+
+def banner(s: str) -> None:
+    print(f"\n{'═' * 78}\n  {s}\n{'═' * 78}")
+
+def claim(desc: str, actual: float, expected: float, tol: float = TOL) -> bool:
+    ok = abs(actual - expected) <= tol
+    tick = "PASS" if ok else "FAIL"
+    print(f"  [{tick}]  {desc}\n          actual={actual:.4f}   expected={expected:.4f}   diff={abs(actual-expected):.4f}")
+    return ok
+
+
+def main() -> int:
+    fails = 0
+
+    # ─── Table 2 (DML Spec B POOLED) ────────────────────────────────
+    banner("Table 2 — DML Spec B headline (mutually-controlled, 6 treatments)")
+    t2 = pd.read_csv(TABLES / "table2_dml_headline.csv")
+    print(t2.to_string(index=False))
+
+    # Sample assertions — the paper's headline numbers
+    def lookup(outcome: str, t_pretty: str) -> float:
+        sub = t2[(t2.outcome == outcome) & (t2.treatment == t_pretty)]
+        return float(sub.iloc[0]["coef"])
+
+    banner("DML claim checks")
+    fails += not claim("T5 topical comp.  selected", lookup("selected","T5 topical competence"),     +0.037, tol=0.005)
+    fails += not claim("T2a Q-headings    selected", lookup("selected","T2a Q-headings"),            +0.016, tol=0.005)
+    fails += not claim("T3 schema         selected", lookup("selected","T3 schema (JSON-LD)"),       -0.014, tol=0.005)
+    fails += not claim("T6 freshness      selected", lookup("selected","T6 freshness"),              -0.005, tol=0.005)
+    fails += not claim("T5 topical comp.  rank_delta", lookup("rank_delta","T5 topical competence"), -0.530, tol=0.02)
+    fails += not claim("T2a Q-headings    rank_delta", lookup("rank_delta","T2a Q-headings"),        +0.136, tol=0.02)
+    fails += not claim("T3 schema         post_rank",  lookup("post_rank","T3 schema (JSON-LD)"),    +0.095, tol=0.01)
+
+    # ─── Admission probe headline ───────────────────────────────────
+    banner("Admission probe — pre-commitment headline (mean pooling)")
+    adm = pd.read_csv(TABLES / "admission_probe_headline.csv")
+    pooled = adm[adm.pooling == "mean"].mean(numeric_only=True)
+    print(adm[adm.pooling == "mean"].round(4).to_string(index=False))
+
+    banner("Admission probe claim checks (variant-averaged)")
+    fails += not claim("Layer 0 ROC AUC",        pooled["layer_0"],    0.671, tol=0.02)
+    fails += not claim("Peak    ROC AUC",        pooled["auc_peak"],   0.860, tol=0.02)
+    fails += not claim("L0 → peak gain",         pooled["delta_L0_to_peak"], 0.190, tol=0.03)
+
+    # ─── Saliency headline ──────────────────────────────────────────
+    banner("Saliency — Llama vs Qwen on 4 treatments")
+    sal = pd.read_csv(TABLES / "saliency_summary.csv")
+    print(sal.to_string(index=False))
+
+    banner("Saliency claim checks")
+    def sal_ratio(model: str, t: str) -> float:
+        return float(sal[(sal.model == model) & (sal.treatment == t)].iloc[0]["saliency_ratio"])
+    fails += not claim("Qwen attends to T1b stats (>>1)",  sal_ratio("Qwen-2.5-72B", "T1b_stats_density"),    1.93, tol=0.05)
+    fails += not claim("Llama ~baseline on T1b (<1)",      sal_ratio("Llama-3.3-70B","T1b_stats_density"),    0.89, tol=0.05)
+    fails += not claim("Llama ignores T3 schema (<<1)",    sal_ratio("Llama-3.3-70B","T3_structured_data_new"),0.19, tol=0.05)
+    fails += not claim("Qwen ignores T3 schema (<<1)",     sal_ratio("Qwen-2.5-72B", "T3_structured_data_new"),0.40, tol=0.05)
+
+    # ─── Ablation headline ──────────────────────────────────────────
+    banner("Ablation — mean Δrank per (treatment, prompt) on full frame")
+    abl = pd.read_csv(TABLES / "ablation_summary.csv")
+    full_abl = abl[abl.frame == "full"]
+    print(full_abl.to_string(index=False))
+
+    banner("Ablation claim checks")
+    def abl_mean(treatment, prompt) -> float:
+        sub = full_abl[(full_abl.treatment == treatment) & (full_abl.prompt == prompt)]
+        return float(sub.iloc[0]["mean_delta_r"])
+    fails += not claim("T5 sign flip — biased  (promotes URL)",   abl_mean("T5_topical_comp","biased"),  -0.167, tol=0.03)
+    fails += not claim("T5 sign flip — neutral (demotes URL)",    abl_mean("T5_topical_comp","neutral"), +0.038, tol=0.03)
+
+    print()
+    print(f"{'═' * 78}")
+    if fails:
+        print(f"   {fails} claim(s) FAILED — please inspect the printed values.")
+        return 1
+    print("   All paper claims VERIFIED against the tables.")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())