acceptodds
Under review as a conference paper at ICLR 2027

Learned Enterprise Data Comprehension: Compression and Routing for Data Agents

Abstract

Structured-data agents in enterprise settings must reason over complex data environments whose relevant evidence is distributed across schemas, relationships, policies, and recurring business roles. Modern agentic systems often address this burden through reusable markdown-style memory or skill files that preserve previously discovered information for later queries, reducing the need to rediscover the same structure repeatedly. This is useful, but it obscures a natural division of labor: agents are well suited to semantic reasoning, while learned systems are well suited to predicting and organizing recurring structure. We introduce latent equivalence learning to bridge this gap. The framework separates persistent task-relevant identities from their dataset-relative realizations. In our realization, supporting and opposing evidence shape support-realized Gaussian prototypes that learn how those identities are expressed in a particular data environment, while soft-membership profiles retain distinctions lost under a hard assignment. A separate learned query-prototype system represents recurring evidential requirements and maps them through a learned compatibility function into the same persistent identity structure. This identity-factorized, query-conditioned routing materializes the relevant dataset-specific evidence for downstream reasoning, allowing the agent to operate over an already organized evidential state rather than reconstructing cross-schema structure at every query. On the Data Agent Benchmark, spanning 54 queries across 12 heterogeneous datasets, our full implementation achieves 94.67% dataset-macro stratified Pass@1 over five complete trials and 258/270 successful raw query attempts, compared with 55.51% for the benchmark's Claude Opus 4.6 reference agent, ranking first among 40 leaderboard entries at submission.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Data agents, Structured data reasoning, Representation learning, Metric learning, Query-conditioned routing

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_ads4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper targets data agents that answer natural-language questions over heterogeneous enterprise databases. It proposes “latent equivalence learning”: a finite inventory of persistent identities (business roles), whose dataset-specific realizations are learned as Gaussian prototypes from supporting and opposing evidence over evidence objects such as columns, record groups, and relations. A separate query-prototype model and a learned compatibility matrix map each query to activations over these identities. A bounded evidence view is then materialized and passed to an LLM Arbiter and a DuckDB-equipped agent. The theoretical part covers the induced equivalence relation, a membership pseudometric, and a margin–Lipschitz local stability certificate. On DAB (54 queries, 12 datasets), the full system reaches 94.67% stratified Pass@1, versus 84.13% for the same harness without the learned representation and 55.51% for the benchmark’s Opus 4.6 reference agent, ranking first on the leaderboard.

Critical strengths

The result on DAB is strong in absolute terms, and the paper is honest about some limitations, e.g., cost and the conservative verification policy. The intuition of separating reusable structural organization from query-time LLM reasoning is reasonable.

Critical weaknesses

Overall this reads as a strong engineering system wrapped in academic formalism rather than a research contribution that can be evaluated.

First, the parts that actually determine the system are never disclosed: how the identity inventory is defined, how the supporting and opposing evidence weights are constructed (the paper explicitly leaves this unspecified), what the feature map is, what data the prototype and query models are trained on, and the hyperparameters. The code is not released because the system is proprietary. What remains is a generic notation layer that a reader cannot reproduce, verify, or build upon.

Second, the theory adds little. Equivalence classes induced by a function, a pseudometric on softmax profiles, and a standard margin–Lipschitz certificate are all elementary, as the authors themselves acknowledge, and Proposition 3 is essentially a tautology. The method itself is close to semantic column-type detection and schema matching combined with prototypical networks and retrieval routing. Related work on schema linking in text-to-SQL, schema matching and entity resolution, semantic layers, and agent memory is missing, and the markdown memory and skill-file approach the paper argues against is never compared.

Third, the empirical evidence does not support the claims. Everything rests on a single 54-query benchmark with no train/test separation, so it is unclear whether the prototypes and query models were fit on the evaluation datasets. The leaderboard entry is marked as using a DAB-specific tuned prompt and hints, which seems to contradict the reproducibility statement, and there are multiple successive submissions on the same 54 queries. The only ablation (Core vs. full system) changes many things at once: the Arbiter, precomputed query interpretation and resolved-field artifacts, different prompts, and roughly twice the tokens and tool calls with deeper post-answer verification. The 10-point gain therefore cannot be attributed to the learned representation. The 39-point gain over the reference agent largely reflects a stronger model and a different harness. There are no component-level ablations, no compute-matched baseline, no evaluation on other benchmarks or unseen datasets, and no real enterprise data.

Questions for authors

see weakness

Reviewer_oRNhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes latent equivalence learning for enterprise data agents. The method represents recurring semantic roles through dataset-specific Gaussian prototypes and uses a learned query-prototype compatibility mechanism to route relevant evidence to a downstream LLM agent. Evaluation on the Data Agent Benchmark shows that the proposed system achieves 94.67% stratified Pass@1, compared with 84.13% for the agentic baseline using the same reasoning model.

Critical strengths
  1. The proposed system achieves substantial improvements over a competitive agentic baseline using the same reasoning model. The evaluation across 12 heterogeneous datasets and five complete trials provides encouraging evidence of its practical effectiveness.
  2. The separation of persistent semantic identities, dataset-specific evidence representations, and query-conditioned routing provides a useful framework for organizing heterogeneous enterprise data. This design addresses an important challenge in structured-data agent systems.
Critical weaknesses
  1. The evaluation lacks component-level ablations that isolate the contributions of signed evidence, Gaussian prototypes, and query-conditioned routing. Furthermore, the proposed system uses substantially more tokens and tool calls than Core. Without controlling for inference resources and verification depth, it is difficult to determine how much of the improvement comes from the proposed representation.
  2. Important implementation details, including the construction of supporting and opposing evidence, the persistent identity inventory, and the training of the query-prototype compatibility mechanism, are insufficiently specified. The experiments also provide limited direct evidence that the learned identities and routing mechanism generalize across unseen data environments.
Questions for authors
  1. Can the authors provide component-level ablations and comparisons under matched token or tool-call budgets?
  2. How are the signed evidence weights and persistent identity inventory constructed in practice?
  3. How are the query prototypes and compatibility matrix trained, and what evidence supports their transfer to unseen datasets?
1 comment

Thank you for taking the time to review our manuscript! We appreciate your recognition of the improvement over Core. We agree our original experiments left questions about attribution and cost, and your review made clear where more evidence was needed.

Before addressing your points, we want to add an AI disclosure. We used AI to assist with adapting code and preparing review-specific diagnostics, checking results against our baseline. AI checked, but did not develop, our new theorems. These build on a planned theoretical extension originally omitted to keep our narrative focused. We hope they strengthen the mathematical basis of the architecture.

All references are to our revised manuscript.

  1. (Ablations and matched resources). We agree that EQ versus Core does not isolate signed evidence, Gaussian prototypes, or routing individually. Both use Opus 5, but differ in evidence, orchestration, and verification. Thus 94.67% versus 84.13% is a full-system result.

We added Flat Retrieval (Sections 4.1–4.2), a familiar RAG-style alternative combining MiniLM cosine similarity and word TF–IDF through reciprocal-rank fusion. Flat indexes the same source-bound records with table, field/relation, and source references, but receives no EQ memberships, plans, candidates, or answers. Both feed the same fresh Sonnet 4.6 Arbiter and Opus 5 agent with the same prompts, tools, source access, settings, and stopping rules. They share an 8,192-token evidence budget and limits of two million total tokens, 150 tool calls, and one final verification.

On the ten-question primary panel, usage was approximately 404k versus 388k tokens and 22.7 versus 21.3 tool calls for EQ and Flat. Yet EQ exposed 70.3% of reference-required evidence versus 31.7%, with higher recall on nine of ten cases. Neither agent saw the evaluator-authored references; recall remains exploratory. Across all 13 cases, we validate 12/13 versus 10/13 (Table 3). All attempts were retained. This examines the evidence interface under comparable resources, not every component. A reconstructed soft-versus-hard test also tied at 2/4. This directly observes your request to see under comparable token budget constraints, can we get reasonable improvement, and the answer is clearly yes. We also did this across multiple runs where the results remaining consistent at 12/13 vs 10/13.

  1. (Signed evidence and identity inventory). Appendix A, Table 1 now specifies 548 evidence-identity slots across ten families and 103 query slots; Patents omits finance (488 evidence identities). These are recurring roles, not DAB answers. Column, column-pair, row, and dataset objects retain source references and missing-feature masks. Features enter a learned 128-dimensional projector.

Our construction uses confidence/reliability signals and negative signals to distinguish supporting from opposing evidence. Section 3.4 shows how support builds precision-weighted prototypes and attraction losses; opposition penalizes proximity to incompatible identities. We cannot disclose our proprietary weights, but the mathematical uses of these signals are now explicit. Appendix C.4 derives a conditional loss-to-margin result connecting them to assignment separation.

We also agree that elementary equivalence relations or stability certificates alone do not justify the architecture. Theorem 1 characterizes the coarsest decision-preserving partition and bounds regret from merging distinguishable contexts. Theorem 2 identifies when hardening soft role profiles discards decision-relevant information. Full proofs appear in Appendix D. These are conditional guarantees, not benchmark accuracy claims.

  1. (Query prototypes and reuse). Section 3.5 describes a learned query projector and 103 query-prototype slots representing recurring evidential requirements. Query memberships activate shared persistent identity coordinates; each dataset then supplies its own realization of those identities through columns, relations, and other evidence. At inference the learned parameters are fixed, while the evidence view changes by request and database.

This is the sense in which query-side structure transfers: a request for a customer identifier activates the same recurring role even when different organizations use different schemas. Query requirements can be reused because their identity coordinates remain shared, while evidence-side realizations are learned locally. We did not design a frozen-model transfer test: our intended setting is learning from an organization’s data before querying it. We agree that unseen-database transfer without adaptation requires a separate experiment, and our conclusion does not claim it. Instead, our system was built around the concept that enterprise data that is incoming will remain consistent in type and structure, these signals do transfer in our system.

Thank you again for your suggestions. They prompted a stronger comparison and clarified the scope of our theory and experiments.

Reviewer_YUzmOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a framework for database comprehension for query answering agents interacting with a database. Dataset evidence (e.g. column names) are (soft-)assigned to persistent semantic identities through support-realized Gaussian prototypes. A separate query-prototype system maps natural-language requests to activations over these identities, which guide the construction of a budget-constrained evidence view. An LLM Arbiter interprets this evidence for a downstream agent that queries the database and performs calculations. On the Data Agent Benchmark, the full system achieves 94.67% dataset-macro Pass@1, compared with 84.13% for the authors’ Core system using the same downstream model.

Critical strengths
  1. The paper addresses a relevant problem. Repeatedly discovering schema relationships and interpreting heterogeneous data can impose substantial overhead on data agents. Mapping queries to evidence via dataset specific prototypes is an interesting direction, which to my knowledge had not been explored before. The reported improvement over the baseline is substantial.
Critical weaknesses
  1. The method and its training procedure are insufficiently specified to understand or gain insights from the evaluated system. The missing details concern central algorithmic choices:
  • Identity inventory CC: Its contents, size, and construction procedure are not specified.
  • Evidence objects: How evidence objects are extracted, their granularity and context are unclear.
  • Signed evidence ωD±\omega_D^\pm: The source and construction of supporting/opposing weights used in the experiments are not clear.
  • Evidence-side training: Episode construction, target-label generation, support/held-out separation, and the choice of interpolation coefficient αc,D\alpha_{c,D} are unspecified. The figure indicates that the system is trained separately for each new dataset, but it is nowhere specified.
  • Query-side training: The training queries, supervision, objective, and number of query prototypes are unspecified, as is how the encoder, prototypes, and compatibility matrix are jointly trained.
  • Training/evaluation separation: The training datasets and query overlap with DAB are unclear, as is which parameters are updated per database versus kept frozen.
  • Agent harness and Core comparison: Exact prompts, orchestration, evidence interfaces, and what replaces EQ’s routed evidence in Core are insufficiently documented.
Questions for authors

While the full system might be proprietary, could you share more fundamental details about your method, especially how your system is trained, what is chosen versus learned, and what parameters are frozen at what point?

2 comments

Thank you for taking the time to review our manuscript! We appreciate your interest in our separation of persistent identities and dataset-relative realizations. We agree that our original paper did not explain the learning procedure sufficiently. Some details were omitted to keep the narrative focused, but your review made clear they were essential.

Before addressing your points, we want to add an AI disclosure. We used AI to help adapt code and prepare review-specific diagnostics, checking results against our baseline. AI checked, but did not develop, our "new" theorems. We agree that the paper needed stronger results that were not merely cosmetic formalism. The theorems that we did add were direct build-offs of what we had planned for our theoretical extension, we had originally omitted them to not take away too much from the narrative, but we hope they help substantiate stronger our architecture.

We expanded Appendix A to distinguish configured, learned, and computed components, with further details in Sections 3.4–3.5. All references are to our revised manuscript.

  1. (Identity inventory and evidence construction). Appendix A, Table 1 now specifies 548 evidence-identity slots in ten families and 103 query slots. The families are semantic, role, structural, behavioral, temporal, operator, filter, finance, quality, and pairwise. Patents omits finance (488 evidence identities). The slots define recurring roles, not benchmark answers.

Evidence construction produces column, column-pair, row, and dataset objects while preserving source identities and missing-feature masks. Features enter a learned 128-dimensional projector. Our code supports confidence weighting, aliases, constraints, and negatives. Section 3.4 explains their use in prototype construction and losses. We cannot disclose our weights.

  1. (Evidence-side learning). Section 3.4 now explains the precision-weighted support calculation, prior/support interpolation, and episodic objective. The episode builder uses support count, confidence, consistency, and prior confidence, with prior fallback.

Notably, inspected checkpoints contain trained projectors, but stored prototype slots remain initialized priors with memory updates disabled. We therefore cannot claim every stored center was jointly optimized. Appendix A distinguishes this from support-derived computation and discloses incomplete historical episode targets, splits, and interpolation settings.

  1. (Query-side learning and materialization). The query side has a learned projector and 103 query-prototype slots with fixed initialized priors in the inspected bank. Routing combines available prototype evidence with query signals, source availability, and structured requirements. Section 3.5 analyzes matrix–softmax compatibility. We also want to observe that we did not perform a classical train/test split by design, in these enterprise systems our belief was companies want to train on their own data and that incoming data would be different, but not so different it was functionally another dataset.

We also describe materialization explicitly: activation, seeding, conditioned expansion, and compaction construct a bounded view retaining source references. The Arbiter derives field bindings, while the agent retains the request and source access for execution (Section 3.5 and Appendix E).

  1. (Training/evaluation access and comparison with Core). Table 2 separates fitting, prompt development, and runtime. Evidence-side fitting uses the database tables being served. During development, benchmark questions, descriptions, and hints informed interpretation, formatting, and verification rules; gold answers and validator outcomes were not supplied to runtime. The full-benchmark score uses five selected complete runs and benchmark-accepted replacements, not every development attempt. We do not present this as a held-out test.

Appendix A now clarifies Core's lexical retrieval, graph propagation, deterministic probes, orchestration, and critic policy. Because Core and EQ differ in several respects, 94.67% versus 84.13% remains a full-system result. We also added Flat Retrieval (Sections 4.1–4.2) with the same downstream Arbiter, agent, prompts, tools, source access, and resource caps. On ten primary questions, EQ exposes 70.3% of reference-required evidence versus 31.7% for Flat. Across 13 cases, the unchanged validator gives 10/13 versus 9/13; our separate author audit gives 12/13 versus 10/13. This tests the evidence interface more directly, though not every component separately.

Finally, we do not claim frozen-model transfer to unseen datasets. Our intended setting is database-specific adaptation, where an organization learns from its own data before querying it. The revised conclusion makes that scope explicit.

Your suggestions prompted us to clarify the implemented system, its evidence, and its limitations. Thank you again for your time!

I thank the authors for the additional explanations that were added in the paper. Unfortunately, some questions regarding the training remain unresolved.

  • Appendix A, Table 1 now specifies 548 evidence-identity slots in ten families and 103 query slots. However, it is unclear to me where they come from. These counts are also reported for something called the "remaining-27 export". It is never explained what that means.
  • Regarding evidence-side training, it is still unclear how evidence–identity pairs are obtained. Are these labels generated by rules, an LLM, or another procedure? Also, Table 1 in Appendix A says: "...negative-generation settings, and active per-step terms are not recoverable from the exports." What exports does this refer to? Why is the full setup not recoverable?
  • Regarding query-side training, Appendix A explicitly warns that require_query_examples does not establish that benchmark questions were excluded from fitting. This is not further explained and raises concerns about test leakage.

Importantly, Appendix A says: "Table 1 distinguishes the mathematical formulation, the recovered training implementation, and the state actually available for the matched diagnostic." What do the authors mean by recovered training implementation? Is the full training procedure not recoverable?

Overall, the revisions contain several unexplained artefacts and raise concerns about whether the full training setup is recoverable and reproducible. I therefore maintain my score of 1.

Reviewer_gDiAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes learning dataset-specific realizations of persistent task-relevant identities for enterprise data agents. Supporting and opposing evidence shape Gaussian prototypes and soft membership profiles. A separate query-prototype model and learned compatibility matrix route queries to a bounded evidence view, which an Arbiter and tool-using agent interpret and execute against. The paper derives conditional assignment-stability bounds and reports 94.67% dataset-macro Pass@1 on DAB, versus 84.13% for a Core system using the same Opus 5 reasoner, over five trials of 54 queries.

Critical strengths

The same-reasoner Core comparison is more informative than the weaker reference-agent comparison: raw successes increase from 236/270 to 258/270, with improvements on several datasets (Section 4 and A ppendix A.1). The explicit distinction between dataset-macro and raw accuracy is also helpful. The paper distinguishes hard assignment from soft memberships and limits its stability certificate to perturbations under a fixed prototype bank. Appendix B provides a valid margin/Lipschitz proof structure and explicitly separates assignment stability from semantic correctness. The resource reporting also usefully reveals the cost of the full system rather than presenting quality alone.

Critical weaknesses

The same-reasoner Core comparison supports an end-to-end improvement. However, it removes representation and routing together, and EQ uses approximately 47.1M versus 24.1M tokens and 1,635 versus 885 tool calls per trial (A.2). Extra computation may be part of how the method succeeds; this does not invalidate the quality comparison. It leaves open whether a simpler evidence-indexing/retrieval method with comparable inference resources would achieve similar quality. Repeated trials on 54 questions establish repeatability, not generalization of a frozen identity/query model to new environments. Such transfer evidence would strengthen the broader reuse claim, but the immediate priorities are an assessable learning protocol and one informative alternative, rather than an exhaustive component-ablation program.

Questions for authors

Please specify the three operational stages: identity/evidence construction and its supervision; query-side training targets and data splits; and evidence-view materialization. For each, identify what benchmark data, queries, labels, and feedback are accessible during fitting, selection, and inference. A worked example and executable-level pseudocode would help. This information would allow assessment of training/evaluation validity even if code and some implementation details remain proprietary.

1 comment
AuthorsOpenReview ↗

Attempting to resolve your concerns

Thank you for taking the time to review our manuscript! Several concerns you raised were ones we anticipated. Before addressing them, we want to add an AI disclosure. We had removed some ablations to keep the narrative focused. We also used AI to assist with adapting code and preparing review-specific analyses, and checked results against our baseline. AI checked, but did not develop, our clarified theorems. We agree that the paper needed stronger theorems that were not just cosmetic formalism.

Your review identifies two priorities: a clearer protocol and a compute-matched retrieval alternative. We wanted to clearly address these in Appendix A and Sections 4.1–4.2. All references are to the revised manuscript.

  1. We agree our original Core comparison did not answer whether simpler retrieval could perform similarly with comparable resources. We added Flat Retrieval, a familiar RAG-style alternative combining MiniLM cosine similarity and word TF–IDF through deterministic reciprocal-rank fusion. Flat indexes the same source-bound records, retaining table, field/relation, and source references, but receives no EQ memberships, plans, selections, or answers.

Both feed the same fresh Sonnet 4.6 Arbiter and Opus 5 agent, with the same prompts, tools, source access, settings, and stopping rules. They share an 8,192-token evidence budget and limits of two million total tokens, 150 tool calls, and one final verification. Across ten primary questions, EQ uses approximately 404k tokens and 22.7 tool calls per attempt versus 388k and 21.3 for Flat. EQ exposes 70.3% of reference-required evidence versus 31.7%, with higher recall on nine questions. Neither agent saw these evaluator-authored references, which lack complete independent review, so recall is exploratory.

Final-answer differences are clear, EQ yields 12/13 while Flat yields 10/13. Clearly, on complex questions EQ is a strong method, but RAG is not without value. However, we do want to point our we deliberately constrained the budget and were still able to beat this system on relevant difficult questions (although we did lose 1 question that we had previously gotten on the full DAB due to this loss of budget, see our revised paper). All launched attempts were retained. We regard this as a diagnostic whose clearest result is evidence exposure.

  1. Appendix A and Table 1 distinguish configured, learned, and computed quantities.

First, evidence construction produces column, column-pair, row, and dataset objects, retaining references and missing-feature masks. The export has 548 evidence-identity slots in ten families and 103 query slots; Patents omits finance. Features enter a learned 128-dimensional projector. Recovered code supports confidence weighting, aliases, constraints, and negatives (Section 3.4).

Second, the query side uses a learned projector and query-prototype slots with fixed initialized priors in the inspected bank. Routing combines prototype evidence with query signals, source availability, and structured requirements. Section 3.5 models matrix–softmax compatibility; Appendix A distinguishes this from recovered implementation.

Third, materialization activates identities, seeds and expands a source-referenced slice, then compacts it under a budget. The Arbiter derives interpretation and field bindings; the agent retains the request and source access for execution and verification. Section 3.5, the example in Section 3.1, and Appendix E illustrate this.

  1. (Training and eval) Table 2 distinguishes fitting, prompt development, and runtime. Evidence-side fitting uses the served database's tables. During development, benchmark questions, descriptions, and supplied hints informed refinements to interpretation, formatting, and verification rules; gold answers and validator outputs were not runtime inputs. The full-benchmark score uses five selected complete runs and benchmark-accepted replacements, not every development attempt. The matched diagnostic instead retained every launched attempt.

Additionally, we exposed as many of the systems as we could without disclosing core proprietary information. This was done in Appendix A and should make clear any missing artifact information or clarification that you need that the new diagnostics don't establish themselves.

  1. (Scope) We agree that repeated trials do not establish frozen-model transfer to unseen environments. Our setting is database-specific adaptation: learning from an organization's own data. The primary diagnostic uses archived EQ selections and CRM uses checkpoint reconstruction; neither is a protected holdout. The conclusion makes this explicit. Section 3.1 and Appendix D give conditional decision-preservation results, which do not replace transfer experiments.

We appreciate your suggestions; they prompted a comparison we should have included and clarified the limits of our claims. Thank you again for your time!

Comments on the paper

2 comments
AuthorsOpenReview ↗

A quick clarification

We also want to observe that in some small sections inside of the paper you will see we added "author adjudication" for our improved ablation as requested by our reviewers. This improved the score from 10/13 and 9/13 to 12/13 and 10/13 (this is present in Section 4 and the Appendix A). This was deliberate. The answer was correct but there was a formatting issue on the validator. We did not change the validator as it was coded by DAB. Our own validator accepted this character mismatch and it was marked as correct (the answers and traces were identical to a point where it was common sense to view them as perfectly equivalent e.g. invisible characters and bizarre tabbing formalities).

We can remove or further clarify this in the paper later pending your request (we figure clarification but we don't want to confuse our reader, we just wanted to see all of your opinions before we made the official decision).

Thanks, The Authors

Dear Reviewers,

The author-reviewer discussion phase is ending soon. If you have not done so yet, please read the author rebuttals and respond.

Please recall that, as a reviewer, you have agreed to engage in discussion with the authors to clarify and improve the paper. Failure to do so may result in penalties on your own submissions, as per the official ICLR policy. Given the high volume of submissions, your participation is crucial.

Thank you for your time and help.

Best regards,

Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.