Does It Tie Out? Towards Autonomous Legal Agents in Venture Capital

• Despite impressive legal-reasoning benchmark and contract-analysis capabilities [7] [5] [6] [10] [1], LLMs have not displaced this recurring manual workflow.Source: §1 · p. 1
• Ownership verification is a transaction bottleneck: lawyers manually reconcile thousands of pages under capped fees, while later-stage financing datarooms often reach tens of thousands of pages.The paper reports over 25,000 venture financings per year, circa 15,000 in the US.Source: §1 · p. 1
• Tie-out reconciles each capitalization-table position to signed agreements, board approvals, payment records, and amendments, producing verified ownership records with traceability and flags for discrepancies.Errors such as a missed cancellation or overlooked amendment can misstate ownership across many stakeholders.Source: §1 · p. 1, Figure 1 · p. 2
• Unlike single-document QA [4] [8] and clause extraction [9], tie-out requires cross-file reasoning, source-linked claims, and identical outputs across repeated runs; agentic LLMs struggle with verification and reproducibility [3].Relevant information may be scattered across dozens of files, and a single grant can depend on multiple documents and amendments.Source: §1 · pp. 1–2
• The paper characterizes capitalization tie-out’s complexity and identifies architectural requirements for its reliable automation.Source: §1 · p. 2

2.1 The Dataroom

• The dataroom holds a company’s legal and financial history in documents that can contain partial, overlapping, or conflicting ownership and transaction information.Documents range from contracts and notices to certificates and spreadsheets; a few categories, including board consents and SPAs, account for most documents, while specialized instruments form a long tail [12].Source: §2.1 · pp. 2–3, Figure 2 · p. 2, Table 1 · p. 3

2.2 The Capitalization Table

• The cap table consolidates outstanding securities, holders, and ownership percentages, typically in spreadsheet ledgers organized by share class.It can include common stock, preferred shares, options, and SAFEs.Source: §2.2 · p. 3
• A tie-out treats the cap table as a master claims ledger: each entry must trace to authoritative dataroom documents, and unsupported entries are flagged.The goal is to confirm support for each line, not reconstruct the company’s full legal history.Source: §2.2 · p. 3

2.3 The Tie-Out Problem

• Tie-out asks whether the reference cap table and the legal capitalization implied by dataroom documents coincide, and identifies where and why they diverge.The legal state is typically implicit: verification transforms check specific aspects of the two capitalization states rather than constructing a complete virtual cap table.Source: §2.3 · pp. 3–4
• The tie-out outcome records failed or unestablished constraints with an explicit, minimal evidence set for each anomaly.The evidence set comes from the dataroom and supports the relevant quantities—or the failure to determine them.Source: §2.3 · p. 4
• Tie-out anomalies include securities missing from the reference cap table, reference entries without sufficient documentation, and differences in terms for securities present in both.Term mismatches can concern vesting, acceleration, price per share, or liquidation preference.Source: §2.3 · p. 4
• A cap-table row can compress a long, branching lifecycle of grants, repricings, splits, exercises, expirations, and transfers that requires cross-referencing agreements across years.With 30 stakeholders, 5 security classes, and 200 documents, the paper describes thousands of distinct verification paths.Source: §2.3 · p. 5, Figure 3 · p. 5
• Real-world datarooms may contain concatenated, scanned, incomplete, or near-duplicate documents, while amendments, separate batches, missing references, and variant party names complicate identifying operative evidence.Drafts, executed copies, redlines, clean versions, and signature packets may coexist, and the operative document is not always obvious.Source: §2.3 · p. 5

Evidence scaling

• Securities increase sevenfold, from 184 to 1,292, while dataroom size roughly doubles between Seed and late Series B.At Seed, documents and securities are roughly 1:1; by Series B, one document may define hundreds of security issuances, requiring more granular extraction and verification.Source: §3 · p. 6, Figure 4 · p. 6

Shifting anomaly types

• Seed-stage anomalies center on missing information in unstructured records, while mature datarooms require checking historical transaction inconsistencies and completeness.Comp. B.2 surfaces nearly 2.5× the issues of Comp. S.1; Missing Information and Missing Consent/Approval remain top categories in later stages.Source: §3 · p. 6, Figure 5 · p. 6, Figure 6 · p. 7

Verification workload

• Verification workload nearly triples, from approximately 2,700 steps at Seed to nearly 8,000 at Series B, and each check becomes more cognitively demanding.The step count serves as a practical proxy for the total family of required verification transforms; the review can take tens of hours.Source: §3 · p. 7, Figure 7 · p. 7

4 Towards Tie-Out Automation

• Tie-out’s core challenge is bridging unstructured legal documents and rigid verification constraints; the section contrasts lazy agentic reasoning with Equall’s eager world-model approach.Source: §4 · p. 7

4.1 The Agentic Paradigm: Lazy Construction via RAG

• Lazy RAG agents retrieve and reason on demand for each cap-table query, but struggle to prove global absence and maintain long document lineages.The agent formulates queries, retrieves dataroom chunks, extracts values with supporting evidence, and compares them with the reference cap table; retrieval failure can be mistaken for missing documentation, while errors compound across dependent documents.Source: §4.1 · pp. 7–8

4.2 The Equall Paradigm: Eager Construction of a World Model

• Equall eagerly extracts source-linked stakeholders, securities, and atomic terms, then organizes them into temporal event nodes linked to the prior issuance or agreement they modify.The resulting Event Graph represents the company’s lifecycle as a reusable asset for tie-out and adjacent legal tasks.Source: §4.2 · p. 8
• Equall uses LLM-based extraction and event synthesis to build the graph, then applies deterministic logic to aggregate events into the virtual cap-table state.For a stakeholder’s share count, the query adds issuances and transfers in, subtracts transfers out, and applies corporate-action adjustments before comparison with the reference table.Source: §4.2 · p. 8

5 Experiments

• Evaluation covers four Seed-to-Series-B datarooms, with professional ground-truth flags and correctness requiring both anomaly type and supporting evidence to match.The harmonized categories include terms discrepancies, missing documentation, and securities or stakeholders missing from the cap table.Source: §5 · p. 9

5.1 Results

• Equall achieves an average F1 of 85%, versus 42% for agentic + structured representations and 29% for pure agentic.Figure 8 reports precision, recall, and F1 across harmonized flag categories.Source: Figure 8 · p. 9, §5.1 · p. 9
• Agentic baselines handle local inconsistent-terms comparisons better, but collapse on missing-documentation and missing-from-cap-table anomalies requiring global reasoning.Equall’s precomputed knowledge graph turns these reasoning chains into graph queries; the evaluated approaches include pure agentic, agentic over structured representations, and Equall’s eager Event Graph.Source: §5.1 · pp. 9–10, §5 · p. 9
• Equall offers a 22× speed advantage per check by shifting reasoning into upfront Event Graph construction and making verification deterministic graph traversals.This eager construction trades higher upfront investment for inexpensive checks; Figure 9 compares speed on a 300-document dataroom.Source: §5.1 · p. 10, Figure 9 · p. 10
• Agentic F1 falls from 55% to 28% as dataroom complexity grows, while manual tie-out time rises from 5 hours at Seed to nearly 27 hours at Series B.Equall-assisted review scales from 64m to 300m, with reported efficiency gains rising from roughly 79% at Seed to 81.5% at Series B.Source: §5.1 · p. 10, Figure 10 · p. 11, Figure 11 · p. 11
• Equall’s task-agnostic Event Graph models lifecycle events such as issuances, amendments, and transfers, and its strong performance indicates utility beyond tie-out.The paper presents the graph as a foundational substrate for downstream legal applications relying on the same historical ground truth.Source: §5.1 · p. 10

Towards Autonomous Tie-Out Agents

• Autonomous tie-out agents require objective rewards, scalable synthetic curricula, and a robust world model, with RL-driven training proposed to improve complex lineage-task reliability.The proposed curriculum injects known error patterns into validated clean datarooms; the Event Graph supplies structured temporal memory where raw-text retrieval struggles.Source: §6 · p. 11

Generalized Representation as a Foundation for Legal Intelligence

• Equall’s Event Graph represents legally operative events as temporally ordered transitions grounded in primary evidence, supporting deterministic verification and a generalizable foundation for applied legal reasoning.The conclusion attributes stable performance under combinatorial complexity to verification through deterministic queries over the model’s well-defined state space.Source: §6 · p. 11