Stage-Audit: Auditable Source-Frontier Discovery for Cross-Wiki Tables

# Stage–Audit: Auditable Source-Frontier Discovery for Cross-Wiki Tables
Source: [https://arxiv.org/html/2605.20478](https://arxiv.org/html/2605.20478)
\\workshoptitle

AI Agents for Discovery in the Wild

###### Abstract

LLM\-curated tables can appear source\-grounded while containing unsupported rows: the curator may recall entries from parametric memory and retroactively attach page\-level citations that are not the actual source\. We study this hazard in*Seed2Frontier discovery*: the task of finding complement Wikipedia pages from a seed page to assemble a structured table\. Stage–Audit addresses it with disjoint curator\-auditor write rights, a row\-level source\-citation gate, and a 12\-check audit taxonomy over keys, schema, source roles, cardinality, and scope\. On a curated 51\-instance Seed2Frontier evaluation set spanning 15 top\-level domains, Stage–Audit improves source\-frontier precision over a vanilla LLM curator from 0\.356 to 0\.505 \(\+42%\+42\\%relative\) and F1 from 0\.334 to 0\.451 \(\+35%\+35\\%\), while maintaining explicit per\-row source traceability\. The vanilla\-LLM\-vs\-Stage–Audit comparison isolates the policy contribution rather than LLM\-based discovery in general\.

## 1Introduction

Take a question like “list transboundary rivers longer than 1000 km by continent\.” No single Wikipedia page carries the answer\. An LLM agent that wants to return a structured table, not a paragraph, has to walk a master list, the continent overview pages, and the individual river articles, then stitch them into rows that are keyed and source\-traceable\. We call this problem*Seed2Frontier discovery*: starting from a seed Wiki page, discover the source frontier needed to assemble a structured table\.

The technical hazard is unintentional ungroundedness\. An LLM curator can emit rows recalled from parametric memory and retroactively attach plausible page\-level citations; the row looks cited even when the citation is not the actual source of the row\. Prior work documents self\-preference and weak self\-correction at evaluation time\(Panicksseryet al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib21); Wataokaet al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib32); Barkanet al\.,[2025](https://arxiv.org/html/2605.20478#bib.bib3); Huanget al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib10); Kamoiet al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib13)\); here the failure happens earlier, when the reference table is built\.

We introduce*Stage–Audit*, a governance protocol for Seed2Frontier\. A curator can stage canonical rows only with an external source URL and locator; an auditor cannot edit the table and instead appends findings under a 12\-check taxonomy over row evidence, primary keys, schema, source roles, cardinality, and scope \(Figure[1](https://arxiv.org/html/2605.20478#S3.F1)\)\. Existing retrieval systems can attempt Seed2Frontier \(Section[2](https://arxiv.org/html/2605.20478#S2)\); our contribution is the governance overlay, not a new retrieval method\. A table can have individually plausible rows and still be unusable if its partition, key, or exhaustiveness claim is wrong, which is what our table\-level audit catches\.

We make three contributions\.\(1\) Task and curated set:a 51\-instance Seed2Frontier evaluation set spanning 15 top\-level domains, with Wikipedia seed pages, labelled complement pages, and primary\-key ground\-truth tables\.\(2\) Protocol:Stage–Audit defines role\-disjoint write rights, a source\-citation gate, a row\-witness property, and a 12\-check audit taxonomy for structured tables\.\(3\) Policy ablation:a four\-configuration comparison \(memory\-only, seed\-outlink, vanilla LLM curator, Stage–Audit\) shows that the source\-citation gate plus audit, not LLM\-based discovery alone, drives source\-frontier precision and F1\.

## 2Related Work

Source\-grounding and judge bias\.FActScore, ALCE, SAFE, and PaperTrail evaluate generated text after the fact\(Minet al\.,[2023](https://arxiv.org/html/2605.20478#bib.bib20); Gaoet al\.,[2023b](https://arxiv.org/html/2605.20478#bib.bib1); Weiet al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib33); Martin\-Boyleet al\.,[2026](https://arxiv.org/html/2605.20478#bib.bib18)\); LLM\-as\-judge work documents bias and self\-preference under same\-family setups\(Zhenget al\.,[2023](https://arxiv.org/html/2605.20478#bib.bib34); Panicksseryet al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib21); Wataokaet al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib32)\)\. Stage–Audit places the gate before a row enters the table, so the output is a structured table with keys and scope claims rather than a sequence of atomic facts\.

Cross\-page Wiki retrieval and citation\-aware generation\.SRAG, KARMA, WikiContradict, and InfoGather bear on multi\-page Wiki construction\(Linet al\.,[2025](https://arxiv.org/html/2605.20478#bib.bib24); Lu and Wang,[2025](https://arxiv.org/html/2605.20478#bib.bib25); Houet al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib26); Yakoutet al\.,[2012](https://arxiv.org/html/2605.20478#bib.bib29)\); HybridQA, OTT\-QA, and Open\-WikiTable evaluate QA over fixed multi\-page table corpora\(Chenet al\.,[2020](https://arxiv.org/html/2605.20478#bib.bib37),[2021](https://arxiv.org/html/2605.20478#bib.bib38); Kweonet al\.,[2023](https://arxiv.org/html/2605.20478#bib.bib39)\), whereas Seed2Frontier evaluates*discovery*of the source frontier itself\. Our vanilla\-LLM\-curator baseline stands in for unguarded LLM\-based retrieval of this family\. RARR retrofits citations to free\-text outputs and Self\-RAG learns to self\-cite during generation\(Gaoet al\.,[2023a](https://arxiv.org/html/2605.20478#bib.bib35); Asaiet al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib36)\); both place the citation step inside the generator and have no separate audit role over a structured table contract\. Stage–Audit instead enforces citation*before*a row is staged and adds a disjoint auditor whose findings cannot rewrite the curator’s table; SPADE and MAST motivate the table\-contract view\(Shankaret al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib31); Cemriet al\.,[2025](https://arxiv.org/html/2605.20478#bib.bib5)\), while generator–critic, debate, Self\-Refine, and Constitutional AI explore critique\-via\-extra\-calls without a row\-level source gate\(Madaanet al\.,[2023](https://arxiv.org/html/2605.20478#bib.bib16); Duet al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib7); Baiet al\.,[2022](https://arxiv.org/html/2605.20478#bib.bib2)\)\.

## 3Method

### 3\.1Stage–Audit Oversight Protocol

LetMMbe a model or agentic framework used to curate source\-grounded tables for user questions\. Some curated tables may later be used to evaluate or diagnose the same model family\. The threat is not adversarial deception but unintentional self\-reference:MMemits remembered rows and then attaches plausible page\-level citations\. Because intrinsic self\-correction is weak without external feedback\(Huanget al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib10); Kamoiet al\.,[2024](https://arxiv.org/html/2605.20478#bib.bib13)\), the protocol makes the source policy, not the model’s self\-critique, the load\-bearing mechanism\.

User query\.“List transboundary rivers≥1000\\geq 1000km by continent, with length and traversed countries\.”Seed\.en\.wikipedia\.org/wiki/Transboundary\_riverCurator \(stage locator\-backed rows\)\. Proposes 13 complement pages \(e\.g\.,*Senegal\_River*,*Amur*,*Ural\_\(river\)*\)\. Stages 31 rows; each row carriessource\_pageand a section/infobox locator\.Auditor \(append\-only findings\)\. ✓\\checkmarkSource\-citation gate: every row cites a page and locator\. △\\trianglescope, blocking:*Ural*crosses Europe/Asia; partition label ambiguous\. △\\trianglecardinality, blocking: 31 emitted vs\. 33 expected from frontier coverage\.Repair \(curator restages\)\.2 missing rows pulled from*List\_of\_transboundary\_rivers*;*Ural*continent set toEurope; Asia\.Audited table\.33 rows×\\times4 cols, primary key=*River*, every row→\\toWiki locator\.Human acceptance \(deployed\)\.Reviewer signs off on audited table;*ablated in our experiment*\(LLM auditor only\)\.restage

Figure 1:Stage–Audit walkthrough on a curated Seed2Frontier instance\. The curator proposes locator\-backed rows; the auditor appends findings under the 12\-check taxonomy and cannot edit the table; blocking findings trigger a repair\-restage loop back to the curator\. The dashed final step shows the deployed human\-acceptance role; our experiment ablates the human step and runs a single curator\-and\-auditor pass with no repair\-restage iteration, to isolate the single\-step value of the policy\.Design tenets\.Stage–Audit has four operational rules\.*Append\-only audit*: the auditor cannot edit the canonical table and records findings separately\.*Cited\-source\-first*: every emitted row must carry a source URL and locator before staging\.*Parametric memory is not evidence*: remembered facts may suggest where to search, but cannot justify a row\.*Two\-axis severity*: every finding is tagged by issue type \(factual / structural / scope\-related\) and severity \(blocking / hygiene / suggestion\), so repair can distinguish blockers from cosmetics\.

Artifact lifecycle\.A staged table moves through four states \(*proposed*,*staged*,*audited*,*repaired*\), iterating until no blocking findings remain\. The curator proposes and stages only locator\-backed rows; the auditor appends findings but cannot edit canonical rows\. In deployment, a human acceptance step gates final sign\-off; this paper evaluates the LLM\-curator \+ LLM\-auditor sub\-pipeline only\.

Source\-gated acceptance and audit checks\.We say that an accepted row has a non\-parametric witness with respect toMMif every emitted rowRRhas an evidence locatorE​\(R\)E\(R\)such that the locator content supportsRRunder an extractorffthat operates only on that locator content\. Stage–Audit operationalizes this through a source\-citation gate: proposed rows without locators are rejected before audit\. The gate does not certify truth\. It only shifts the failure mode: a row is wrong only if the extractor misread the cited locator\. The artifact contractA=\(Q,K,C,P,S,E\)A=\(Q,K,C,P,S,E\)records the user query, primary\-key schema, columns, scope statement, source set, and evidence map\. Audit checks span row evidence, key integrity, schema conformance, cardinality, source\-role coverage, claim type, normalization, scope match, and temporal knowability \(checks: App\.[I](https://arxiv.org/html/2605.20478#A9); severity: App\.[F](https://arxiv.org/html/2605.20478#A6); proof sketch: App\.[C](https://arxiv.org/html/2605.20478#A3)\)\.*Factual*findings target row evidence;*structural*findings targetK,CK\{,\}C;*scope\-related*findings targetP,SP\{,\}S\. Figure[1](https://arxiv.org/html/2605.20478#S3.F1)shows two scope/cardinality findings the citation gate alone would miss\.

### 3\.2Seed2Frontier Discovery Procedure

Seed2Frontier takes a user question and a seed Wiki page, then asks for the complement pages needed to assemble a structured table\. Stage–Audit bounds search through a logged frontier discipline: every page inSSrecords the operator that admitted it \(seed table, sibling list, redirect, entity page, optional Wikidata count or rank check; template in Appendix[B](https://arxiv.org/html/2605.20478#A2)\)\. Audit treats missing required operators, unexplained pages, and mismatched source roles as scope failures\. Full pseudocode appears in Appendix[A](https://arxiv.org/html/2605.20478#A1)\. The procedure is a source\-discovery policy, not a new retrieval method; existing planners can attempt the same task, and our experiment isolates the value of the governance overlay\.

## 4Anchor Experiment: Seed2Frontier Discovery

Curated evaluation set\.We curate 51 Seed2Frontier instances spanning 15 top\-level domains \(geography, culture, science, politics, sports, business, technology, and others\)\. 35 instances are independently authored questions; 16 are programmatic derivatives produced by filter, top\-K, or year\-range modification of multi\-page parents \(Appendix[D](https://arxiv.org/html/2605.20478#A4)\)\. Each instance has a Wikipedia seed page, labelled complement pages \(45 of 51, mean 10\.1 per instance, range 1–35\), and a primary\-key ground\-truth table\.

Configurations\.Four conditions evaluated with the same model\.*Memory\-only*: one prompt for parametric\-memory rows, no source frontier produced\.*Seed\-outlink*: deterministic Wikipedia\-API enumeration of all outlinks of the seed; no LLM, no rows\.*Vanilla LLM curator*: a single LLM prompt asking for complement pages and rows, no source\-citation gate, no audit \(a proxy for an unguarded LLM\-based retrieval planner\)\.*Stage–Audit\-governed*: curator prompt with the Wikipedia discovery hint and source\-citation requirement, followed by an auditor prompt that appends findings under the 12\-check taxonomy\.

Setup\.We run the experiment withgpt\-5\.4atT=0T\{=\}0\.111Identifiersgpt\-5\.4andLlama 3\.3 70Bare API release tags, quoted verbatim from the API response\.We manually label the seed Wikipedia URL for each instance as the most natural starting page among the candidate URLs\. Two protocol steps are ablated: human acceptance \(Figure[1](https://arxiv.org/html/2605.20478#S3.F1), gray dashed\) and any repair\-restage iteration\. We run a single curator\-and\-auditor pass to match the single\-pass budget the vanilla and memory\-only baselines receive \(loop ablation in Appendix[J](https://arxiv.org/html/2605.20478#A10)\)\. Following WikiTabGen\-style scoring \(Appendix[E](https://arxiv.org/html/2605.20478#A5)\), we report recall, precision, and F1 over two axes: predicted Wikipedia pages against the labelled complement set, and emitted rows against the labelled primary keys\.

Table 1:Stage–Audit on the 51\-instance Seed2Frontier evaluation set with gpt\-5\.4 atT=0T\{=\}0\. Frontier metrics \(primary axis\) on the 45 instances with complement\-page labels; PK metrics \(secondary axis, see body\) on all 51\. Closed\-loop scoring: when the auditor returns no accepted rows or pages \(8/51 for rows, 4/51 for pages\), Stage–Audit emits 0 for that instance rather than passing curator output through\.Frontier comparison\.Stage–Audit improves source\-frontier precision \(0\.505 vs\. 0\.356\), F1 \(0\.451 vs\. 0\.334\), and complement recall \(0\.491 vs\. 0\.384\) over the vanilla LLM curator on the 45 instances with complement labels\. We compute paired\-bootstrap 95% CIs by resampling the 45 per\-instance Stage–Audit−\-vanilla differences with replacement \(5000 draws\) and taking the central 95% range of the resulting means:\[\+0\.05,\+0\.25\]\[\+0\.05,\+0\.25\]for precision \(two\-sided sign testp=0\.006p\{=\}0\.006\),\[\+0\.03,\+0\.22\]\[\+0\.03,\+0\.22\]for F1 \(p=0\.029p\{=\}0\.029\), and\[\+0\.02,\+0\.21\]\[\+0\.02,\+0\.21\]for recall\. The precision and F1 intervals lie entirely above zero, so the gains are robust to instance resampling\. The seed\-outlink\-vs\-vanilla gap \(precision 0\.020 vs\. 0\.356\) shows that LLM\-based discovery does most of the precision lift over deterministic outlink enumeration; Stage–Audit then adds the source\-citation gate and the audit step, raising precision a further \+0\.149 absolute \(\+42% relative\) while producing per\-row locator output the unguarded curator does not\.

PK rows \(secondary axis\)\.By construction Stage–Audit emits only source\-grounded rows, so unguarded baselines that emit every remembered row rank higher on PK overlap by design \(memory\-only0\.6190\.619\>\>vanilla0\.5880\.588\>\>Stage–Audit0\.5340\.534\); the primary axis is frontier quality with per\-row source traceability, and a memory\-only table cannot satisfy a source\-grounded consumer\.

Where the policy helps\.Stage–Audit’s frontier\-F1 gain over vanilla is largest on small\-frontier instances \(1–3 complement pages,\+0\.25\+0\.25,n=17n\{=\}17\); the gain shrinks on larger frontiers \(4–9 pages,\+0\.07\+0\.07,n=10n\{=\}10; 10\+ pages,\+0\.02\+0\.02,n=18n\{=\}18\)\. The largest per\-domain gains are in*culture*\(\+0\.32\+0\.32,n=7n\{=\}7\),*sports*\(\+0\.25\+0\.25,n=4n\{=\}4\), and*science*\(\+0\.18\+0\.18,n=5n\{=\}5\);n≤3n\{\\leq\}3buckets are deferred to the supplement\.

Audit\-finding density\.182 findings across 51 instances \(mean 3\.6, range 1–5; issue\-type×\\timesseverity in Figure[2](https://arxiv.org/html/2605.20478#A8.F2), Appendix[H](https://arxiv.org/html/2605.20478#A8)\); the auditor returned no accepted rows on 8/51 instances and no accepted complement pages on 4/51, contributing 0 to the Stage–Audit means under closed\-loop scoring\. Stage–Audit’s frontier averages 11\.4 pages per instance,54×54\\timessmaller than seed\-outlink’s 612\.3\.

Cross\-model\.Repeating the four\-configuration evaluation on Llama 3\.3 70B \(open\-source; Appendix[G](https://arxiv.org/html/2605.20478#A7)\) yields a small directional gain \(\+0\.008\+0\.008frontier F1 vs\. vanilla,\+0\.012\+0\.012precision,\+0\.033\+0\.033recall\); the policy contribution is most pronounced on stronger instruction\-following models\.

## 5Conclusion

Stage–Audit governs Seed2Frontier with disjoint curator\-auditor write rights, a row\-level source\-citation gate, and a 12\-check audit taxonomy\. On 51 instances using gpt\-5\.4 it improves source\-frontier precision over an unguarded LLM curator from 0\.356 to 0\.505 \(paired 95% CI\[\+0\.05,\+0\.25\]\[\+0\.05,\+0\.25\],p=0\.006p\{=\}0\.006\) and F1 from 0\.334 to 0\.451\. The present evaluation ablates the human\-acceptance step and the repair\-restage loop, and covers one closed\-source and one open\-source model\. Within those scope limits, Stage–Audit serves as a source\-grounding integrity overlay for cross\-Wiki table construction\.

## References

- A\. Asai, Z\. Wu, Y\. Wang, A\. Sil, and H\. Hajishirzi \(2024\)Self\-RAG: learning to retrieve, generate, and critique through self\-reflection\.International Conference on Learning Representations \(ICLR\)\.External Links:2310\.11511Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- Y\. Bai, S\. Kadavath, S\. Kundu, A\. Askell, J\. Kernion,et al\.\(2022\)Constitutional ai: harmlessness from ai feedback\.arXiv preprint arXiv:2212\.08073\.External Links:2212\.08073Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- C\. O\. Barkan, S\. Black, and O\. Sourbut \(2025\)Do large language models know what they are capable of?\.arXiv preprint arXiv:2512\.24661\.External Links:2512\.24661Cited by:[§1](https://arxiv.org/html/2605.20478#S1.p2.1)\.
- M\. Cemri, M\. Z\. Pan, S\. Yang, L\. A\. Agrawal, B\. Chopra, R\. Tiwari, K\. Keutzer, A\. Parameswaran, D\. Klein, K\. Ramchandran, M\. Zaharia, J\. E\. Gonzalez, and I\. Stoica \(2025\)Why do multi\-agent llm systems fail?\.arXiv preprint arXiv:2503\.13657\.External Links:2503\.13657Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- W\. Chen, M\. Chang, E\. Schlinger, W\. Y\. Wang, and W\. W\. Cohen \(2021\)Open question answering over tables and text\.InInternational Conference on Learning Representations \(ICLR\),External Links:2010\.10439Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- W\. Chen, H\. Zha, Z\. Chen, W\. Xiong, H\. Wang, and W\. Y\. Wang \(2020\)HybridQA: a dataset of multi\-hop question answering over tabular and textual data\.InFindings of the Association for Computational Linguistics: EMNLP,External Links:2004\.07347Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- Y\. Du, S\. Li, A\. Torralba, J\. B\. Tenenbaum, and I\. Mordatch \(2024\)Improving factuality and reasoning in language models through multiagent debate\.InInternational Conference on Machine Learning,External Links:2305\.14325Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- L\. Gao, Z\. Dai, P\. Pasupat, A\. Chen, A\. T\. Chaganty, Y\. Fan, V\. Y\. Zhao, N\. Lao, H\. Lee, D\. Juan, and K\. Guu \(2023a\)RARR: researching and revising what language models say, using language models\.Annual Meeting of the Association for Computational Linguistics \(ACL\)\.External Links:2210\.08726Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- T\. Gao, H\. Yen, J\. Yu, and D\. Chen \(2023b\)Enabling large language models to generate text with citations\.arXiv preprint arXiv:2305\.14627\.External Links:2305\.14627Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p1.1)\.
- Y\. Hou, A\. Pascale, J\. Carnerero\-Cano, T\. T\. Tchrakian, R\. Marinescu, E\. Daly, I\. Padhi, and P\. Sattigeri \(2024\)WikiContradict: a benchmark for evaluating llms on real\-world knowledge conflicts from wikipedia\.InAdvances in Neural Information Processing Systems,External Links:2406\.13805Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- J\. Huang, X\. Chen, S\. Mishra, H\. S\. Zheng, A\. W\. Yu, X\. Song, and D\. Zhou \(2024\)Large language models cannot self\-correct reasoning yet\.arXiv preprint arXiv:2310\.01798\.External Links:2310\.01798Cited by:[§1](https://arxiv.org/html/2605.20478#S1.p2.1),[§3\.1](https://arxiv.org/html/2605.20478#S3.SS1.p1.2)\.
- R\. Kamoi, Y\. Zhang, N\. Zhang, J\. Han, and R\. Zhang \(2024\)When can llms actually correct their own mistakes? a critical survey of self\-correction of llms\.Transactions of the Association for Computational Linguistics\.External Links:2406\.01297Cited by:[§1](https://arxiv.org/html/2605.20478#S1.p2.1),[§3\.1](https://arxiv.org/html/2605.20478#S3.SS1.p1.2)\.
- S\. Kweon, Y\. Kwon, S\. Cho, Y\. Jo, and E\. Choi \(2023\)Open\-WikiTable: dataset for open domain question answering with complex reasoning over table\.InFindings of the Association for Computational Linguistics: ACL,External Links:2305\.07288Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- T\. Lin, Y\. Zhu, Y\. Luo, and N\. Tang \(2025\)SRAG: structured retrieval\-augmented generation for multi\-entity question answering over wikipedia graph\.arXiv preprint arXiv:2503\.01346\.External Links:2503\.01346Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- Y\. Lu and J\. Wang \(2025\)KARMA: leveraging multi\-agent llms for automated knowledge graph enrichment\.arXiv preprint arXiv:2502\.06472\.External Links:2502\.06472Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- A\. Madaan, N\. Tandon, P\. Gupta, S\. Hallinan, L\. Gao, S\. Wiegreffe, U\. Alon, N\. Dziri, S\. Prabhumoye, Y\. Yang,et al\.\(2023\)Self\-refine: iterative refinement with self\-feedback\.arXiv preprint arXiv:2303\.17651\.External Links:2303\.17651Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- A\. Martin\-Boyle, C\. A\. C\. Leckey, M\. C\. Brown, and H\. Kaur \(2026\)PaperTrail: a claim\-evidence interface for grounding provenance in llm\-based scholarly q&a\.arXiv preprint arXiv:2602\.21045\.External Links:2602\.21045Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p1.1)\.
- S\. Min, K\. Krishna, X\. Lyu, M\. Lewis, W\. Yih, P\. W\. Koh, M\. Iyyer, L\. Zettlemoyer, and H\. Hajishirzi \(2023\)FActScore: fine\-grained atomic evaluation of factual precision in long form text generation\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,External Links:2305\.14251Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p1.1)\.
- A\. Panickssery, S\. R\. Bowman, and S\. Feng \(2024\)LLM evaluators recognize and favor their own generations\.InAdvances in Neural Information Processing Systems,External Links:2404\.13076Cited by:[§1](https://arxiv.org/html/2605.20478#S1.p2.1),[§2](https://arxiv.org/html/2605.20478#S2.p1.1)\.
- S\. Shankar, H\. Li, P\. Asawa, M\. Hulsebos, Y\. Lin, J\. D\. Zamfirescu\-Pereira, H\. Chase, W\. Fu\-Hinthorn, A\. G\. Parameswaran, and E\. Wu \(2024\)SPADE: synthesizing data quality assertions for large language model pipelines\.Proceedings of the VLDB Endowment17\(12\)\.External Links:2401\.03038Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- K\. Wataoka, T\. Takahashi, and R\. Ri \(2024\)Self\-preference bias in llm\-as\-a\-judge\.arXiv preprint arXiv:2410\.21819\.External Links:2410\.21819Cited by:[§1](https://arxiv.org/html/2605.20478#S1.p2.1),[§2](https://arxiv.org/html/2605.20478#S2.p1.1)\.
- J\. Wei, C\. Yang, X\. Song, Y\. Lu, N\. Hu, J\. Huang, D\. Tran, D\. Peng, R\. Liu, D\. Huang, C\. Du, and Q\. V\. Le \(2024\)Long\-form factuality in large language models\.InAdvances in Neural Information Processing Systems,External Links:2403\.18802Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p1.1)\.
- M\. Yakout, K\. Ganjam, K\. Chakrabarti, and S\. Chaudhuri \(2012\)InfoGather: entity augmentation and attribute discovery by holistic matching with web tables\.InProceedings of the ACM SIGMOD International Conference on Management of Data,pp\. 97–108\.External Links:[Document](https://dx.doi.org/10.1145/2213836.2213848)Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p2.1)\.
- L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. P\. Xing, H\. Zhang, J\. E\. Gonzalez, and I\. Stoica \(2023\)Judging llm\-as\-a\-judge with mt\-bench and chatbot arena\.InAdvances in Neural Information Processing Systems Datasets and Benchmarks Track,External Links:2306\.05685Cited by:[§2](https://arxiv.org/html/2605.20478#S2.p1.1)\.

## Appendix ASeed2Frontier Discovery Algorithm

Algorithm 1Seed2Frontier discovery procedure \(curator’s propose step in Figure[1](https://arxiv.org/html/2605.20478#S3.F1)\)1:natural\-language table query

QQ, seed Wiki page

ss
2:candidate source set

SS
3:

P←InferScope​\(Q\)P\\leftarrow\\textsc\{InferScope\}\(Q\)⊳\\trianglerightpartition/scope statement extracted fromQQ\(e\.g\., “by continent”⇒\\Rightarrowcontinents enumeration\)

4:

S←\{s\}S\\leftarrow\\\{s\\\}with source role and retrieval metadata

5:ifa single page covers all partitions in

PPthen

6:return

SS
7:endif

8:expand through sibling list pages, redirects, and list/table links

9:foreach under\-covered partition or tail entity in

PPdo

10:fetch the relevant entity page and add supporting pages to

SS
11:endfor

12:if

QQimplies a count, rank, or partition\-size constraintthen

13:run a Wikidata SPARQL count or rank sanity check

14:endif

15:log each URL, locator, retrieval time, and content hash for audit

16:return

SS

## Appendix BWikidata Sanity\-Check Template

When a user question implies a count, rank, membership, or partition\-size constraint, Wikidata can provide a weak sanity check on the source frontier\. The query is not the row witness: accepted rows still require source locators inEE\. In practice, the curator records the property and class choices, the query time, and the returned count or rank signal so an auditor can replay whether the frontier is plausibly complete\.

```
# Q_CLASS, P_FILTER, Q_FILTER are placeholders for instance-specific
# Wikidata classes and properties; substitute before issuing the query.
PREFIX wd: <http://www.wikidata.org/entity/>
PREFIX wdt: <http://www.wikidata.org/prop/direct/>
SELECT (COUNT(DISTINCT ?entity) AS ?n) WHERE {
  ?entity wdt:P31/wdt:P279* wd:Q_CLASS .
  ?entity wdt:P_FILTER wd:Q_FILTER .
}
```

## Appendix CProof Sketch for the Non\-Parametric Witness Claim

The source\-citation gate gives a simple invariant over staged rows\. LetRAR\_\{A\}be the rows in artifactAAand letBlock​\(A\)\\mathrm\{Block\}\(A\)be its unresolved blocking findings\. For rowrr, define the operational acceptance predicate:

RowOK​\(r,A\)≡∃e∈E​\(r\):e∈S∧supportsf​\(e,r\)\.\\mathrm\{RowOK\}\(r,A\)\\equiv\\exists e\\in E\(r\):e\\in S\\wedge\\mathrm\{supports\}\_\{f\}\(e,r\)\.The acceptance condition is then:

Accept​\(A\)≡∀r∈RA,RowOK​\(r,A\)∧Block​\(A\)=∅\.\\mathrm\{Accept\}\(A\)\\equiv\\forall r\\in R\_\{A\},\\ \\mathrm\{RowOK\}\(r,A\)\\wedge\\mathrm\{Block\}\(A\)=\\emptyset\.Heresupportsf​\(e,r\)\\mathrm\{supports\}\_\{f\}\(e,r\)means that extractorffreads locator contentee, rather than model memory, and recovers support for the row\. For any accepted rowrr, the staging rule requires such an evidence locator in the logged source setSS\. The audit rule then permits validation only from the locator content, not from the model’s remembered facts\. By induction over staged row updates, every row that passes the gate has such a witness, modulo extractor faithfulness and locator drift\.

## Appendix DCurated Evaluation Set

We curate 51 Seed2Frontier instances spanning 15 top\-level domains, with the largest concentrations in geography \(13 instances\), culture \(8\), science \(6\), politics \(4\), sports \(4\), and business \(3\); the remaining nine domains \(technology, religion, food, education, environment, health, history, law, society\) contribute one to three instances each\. Each instance is paired with a Wikipedia seed page, the labelled set of cited Wikipedia URLs \(mean 10\.1 pages per instance, median 6, range 1–35\), and a primary\-key ground\-truth table \(mean 68\.2 rows per instance, median 24, range 3–526\)\. We manually label the seed page for each instance as the most natural starting Wikipedia page for the query; the remaining labelled Wikipedia URLs form the complement set\.

## Appendix EScoring Notes

Complement\-page recall compares predicted source pages to the labelled complement set\. Source\-frontier precision and F1 use the same page set after Wikipedia\-title normalization\. Cardinality error is the absolute difference between emitted row count and labelled row count\. Table\-extraction F1 is computed over normalized primary\-key overlap when a configuration emits rows\.

## Appendix FSeverity Taxonomy Reference

Issue types are*factual*\(a source\-supported value is wrong or unsupported\),*structural*\(schema, primary key, type, duplicate, or artifact\-contract problem\), and*scope\-related*\(row\-universe, partition, rank, cardinality, or exhaustiveness problem\)\. Severity levels are*blocking*,*hygiene*, and*suggestion*\. The crossing is deliberate: a scope\-related blocker can make a table unusable even if every emitted row is factually correct, while a structural hygiene issue can be repairable without changing the row set\.

## Appendix GCross\-Model Generalization

We pick one closed\-source model \(gpt\-5\.4\) and one open\-source model \(Llama 3\.3 70B\) as the two model familiesMM, to test whether the policy contribution generalizes across release types and training pipelines\.

Table 2:Vanilla LLM curator vs\. Stage–Audit on the same 51\-instance Seed2Frontier evaluation set with picked seeds, across one closed\-source and one open\-source model\. Frontier metrics on the 45 instances with complement labels\.Δ\\Deltais Stage–Audit minus vanilla; sign\-testppon per\-instance F1 deltas\.Stage–Audit’s frontier\-F1 improvement over the vanilla LLM curator is large and paired\-significant on gpt\-5\.4 \(\+0\.118\+0\.118, two\-sided sign testp=0\.029p\{=\}0\.029\) and small but directionally consistent on Llama 3\.3 70B \(\+0\.008\+0\.008F1, with positive deltas on all three frontier metrics\)\. The pattern is consistent with the policy contribution being most effective on top of strong\-instruction\-following models, where the curator emits citation\-tagged rows the auditor can then filter cleanly\.

## Appendix HSeverity Taxonomy in the 182 Auditor Findings

![Refer to caption](https://arxiv.org/html/2605.20478v1/x1.png)Figure 2:Severity taxonomy applied to the 182 Stage–Audit auditor findings on the 51\-instance set with gpt\-5\.4\. \(a\) Issue\-type×\\timesseverity heatmap: factual findings are dominated by the source\-citation gate \(n=40n\{=\}40hygiene, 13 blocking\) confirming the gate fires on essentially every instance; scope\-related findings are spread across blocking, hygiene, and suggestion, reflecting the harder\-to\-automate cardinality and partition checks\. \(b\) Severity composition stratified by labelled complement\-page count: every frontier\-size bucket produces blocking findings, and the smallest\-frontier bucket carries the highest blocking share, consistent with the Section[4](https://arxiv.org/html/2605.20478#S4)finding that the gate provides the most lift on small\-frontier instances\. Issue\-type assignments use a conservative keyword classifier over the auditor’s free\-textchecklabels; the small “Other” bucket reflects labels that did not match the keyword set\.
## Appendix IAudit Check Enumeration

The 12\-check audit taxonomy used in Section[3](https://arxiv.org/html/2605.20478#S3)\. Each check has a stable identifier; the auditor assigns one identifier per finding\.

Factual\.F1row\-evidence \(locator content supports row\);F2source\-URL well\-formed and reachable;F3locator format and section/table reference valid\.

Structural\.S1primary\-key uniqueness;S2primary\-key non\-null;S3column type conformance;S4column completeness against the artifact contract\.

Scope\-related\.P1cardinality match against query expectation;P2partition coverage;P3source\-role coverage \(each page tagged with its admitting operator\);P4temporal knowability;P5normalization consistency across rows\.

## Appendix JRepair\-Loop Ablation

The body experiment caps Stage–Audit at a single curator\-and\-auditor pass to match the per\-instance call budget of the unguarded baselines\. To check whether additional repair\-restage iterations would change the picture, we run gpt\-5\.4 withk=0,…,5k\{=\}0,\\dots,5repair iterations on a 15\-instance domain\-stratified subset \(one instance per top\-level domain\)\. Each iteration afterk=0k\{=\}0feeds the auditor’s findings back to the curator with a restage prompt and re\-audits the new output; an instance converges if the auditor returns no blocking findings\.

Table 3:Stage–Audit performance vs\. repair\-loop countkkon a 15\-instance domain\-stratified subset \(gpt\-5\.4 atT=0T\{=\}0\);k=0k\{=\}0is the body experiment\. PK\-F1 gains\+0\.02\+0\.02atk=1k\{=\}1and plateaus\.Stage–Audit’s frontier metrics are stable throughk=3k\{=\}3and show no improvement past the initial pass; PK\-F1 receives a small one\-shot lift atk=1k\{=\}1and then plateaus\. Productive iteration would require either external retrieval or auditor signals naming candidate replacement pages, both deferred to future work\.