Quantistic

LPA Term Extraction Accuracy Benchmark

Automated extraction of fee terms, distribution waterfalls, and governance provisions from real-world private equity Limited Partnership Agreements.

April 2026 · 4 Funds · 271 Fields · 6 Term Categories

Results

We benchmarked our extraction pipeline against four real-world LPAs from large-cap private equity funds. Each document was processed with zero human intervention. Every extracted value was compared against the corresponding provision in the source LPA by a qualified analyst.

97.4%aggregate accuracy across
4 funds and 271 verified fields
Carlyle
100%
67 / 67
Blackstone
100%
68 / 68
ILPA Standard
97.1%
66 / 68
KKR
92.6%
63 / 68

Two of four funds achieved 100% accuracy with zero extraction errors. The remaining errors are concentrated in provisions that reference external agreements not included in the LPA document itself — a structural limitation, not an extraction failure.

Fund Coverage

The benchmark deliberately spans structurally diverse fund types to stress-test extraction across different LPA architectures.

FundStrategyStructure ComplexityAccuracy
Carlyle Partners VILarge-Cap BuyoutStandard — all terms in main body100.0%
Blackstone Capital Partners VLarge-Cap BuyoutHigh — fee terms in external Investment Advisory Agreement; dual waterfall structure100.0%
ILPA Standard Growth Fund IIGrowth EquityModerate — ILPA-template structure with bracketed agreed terms97.1%
KKR Millennium FundLarge-Cap BuyoutVery High — Management Agreement in external exhibit; tiered fee structure; multi-fund netting92.6%

Accuracy by Term Category

Every LPA is decomposed into six standardized term categories. Each category contains between 8 and 16 verified fields capturing rates, thresholds, trigger events, formulas, and governance provisions.

Management Fee
97.9%
Fee rates, bases, payment timing, commencement triggers, offset provisions, organizational expense caps
Carried Interest
95.3%
Waterfall type, carry rate, preferred return, catch-up, clawback, escrow, recycling, GP commitment
Key Person
100%
Named key persons, trigger events, thresholds, consequences, cure periods, reinstatement mechanisms
Investment Period
96.9%
Duration, start trigger, follow-on rights, extension provisions, early termination triggers
Fund Term
96.9%
Initial term, extensions, dissolution triggers, GP removal thresholds (for-cause and no-fault)
Organizational Expenses
100%
Expense caps, GP vs fund allocation, covered & excluded categories, fee offset policies, broken deal costs

Distribution Waterfall Validation

Waterfall classification is the highest-stakes extraction in LP fee verification. Misclassifying a European waterfall as American — or vice versa — fundamentally changes how carry is calculated and when the GP earns economics. Our pipeline correctly classified all four funds.

FundClassificationResultClassification Basis
Carlyle Partners VIEuropean CorrectTier 1 aggregates capital recovery across all realized investments
Blackstone Capital Partners VEuropean CorrectWritedown mechanism deems unrealized losses as dispositions at distribution time
ILPA Standard Growth Fund IIEuropean CorrectTier 1 returns capital from all realized investments plus aggregate unrealized losses
KKR Millennium FundEuropean CorrectWhole-of-fund economics with cross-fund netting across parallel structures

Why Waterfall Classification Is Hard

The binary European-vs-American distinction obscures significant structural complexity. Real LPAs rarely present a clean textbook waterfall. Our benchmark encountered three distinct challenges that most automated systems mishandle.

Blackstone Capital Partners V
Modified European — Deal-by-deal form, whole-of-fund economics

Blackstone structures distributions on a per-investment basis with a separate Current Income and Disposition Proceeds waterfall. On the surface, this resembles an American deal-by-deal structure. However, paragraph 4.3.4(a) introduces an automatic writedown mechanism: at the time of any distribution, all unrealized investments trading below cost are deemed disposed at fair market value. This forces unrealized losses into the Unrecouped Losses calculation, making the LP whole across the entire portfolio before carry flows. The preferred return in the Disposition Proceeds waterfall compounds across all disposed investments, not per-deal. Combined with end-of-fund-only clawback, this produces whole-of-fund economics despite deal-level distribution mechanics. Our system correctly identified these structural indicators and classified the waterfall as European.

ILPA Standard Growth Fund II
Standard European — Four-tier waterfall with partial catch-up

ILPA Standard presents a clean four-tier European waterfall: Tier 1 returns capital contributions across all realized investments plus aggregate unrealized losses and fund expenses; Tier 2 distributes the 8% preferred return; Tier 3 allocates 80% to the GP during catch-up until carry reaches 20% of net profits; Tier 4 splits residual 80/20 to LP and GP. The partial catch-up at 80% (rather than 100%) is a modified ILPA structure that some systems misinterpret as a separate tier or miscalculate the carry target. Our system correctly identified the catch-up rate and verified the mathematical consistency: the 80/20 catch-up delivers the GP to exactly 20% of total profits.

KKR Millennium Fund
European — Cross-fund netting with predecessor aggregation

KKR Millennium operates with cross-fund performance netting across parallel fund structures — an unusual provision where carry is calculated on aggregate performance across multiple vehicles. The investment period is event-driven, triggered by the termination of a predecessor fund's commitment period rather than the fund's own closing. Fee terms reside in a separate Management Agreement appended as an exhibit, requiring the system to correctly identify and extract from external agreements. Our system resolved the exhibit structure and classified the waterfall as European based on whole-of-fund capital recovery requirements.

Key finding: All four funds in the benchmark use European (whole-of-fund) distribution waterfalls, consistent with the industry trend toward LP-protective structures in large-cap buyout and growth equity funds. Our system achieved 100% waterfall classification accuracy across all four funds, including the structurally complex Blackstone modified European and KKR cross-fund netting structures.

Waterfall Component Accuracy

Beyond the binary classification, the pipeline extracts and verifies the full waterfall structure — every tier, every threshold, every exception.

Waterfall ComponentFields VerifiedAccuracy
Waterfall type classification4 / 4 funds100%
Carry rate extraction4 / 4 funds100%
Preferred return rate3 / 4 funds100%*
Catch-up rate and structure3 / 4 funds100%*
Clawback existence & guarantee type4 / 4 funds100%
Full tier-by-tier waterfall capture4 / 4 funds100%

* 3/4 funds had preferred return and catch-up terms directly stated in the LPA. The fourth fund defines these in a separate Management Agreement not included in the LPA document. The system correctly returned null for fields defined in external documents rather than hallucinating values.

Fee Structure Verification

Management fee extraction requires identifying the fee rate, calculation base, payment timing, commencement trigger, and offset provisions — each of which can differ between the investment period and post-investment period.

Fee ComponentAcross 4 FundsNotes
Fee rate (investment period)100%Rates ranged from 0.75% to 2.00%; correctly distinguished from post-investment rates
Fee rate (post-investment)100%Step-down rates correctly identified; fee base transitions captured
Fee base classification100%Committed capital, invested capital, and net invested capital correctly distinguished
Payment timing100%Quarterly-in-advance and quarterly-in-arrears correctly identified from notice provisions
Fee offset and acceleration100%Offset provisions, fee income credits, and termination-triggered acceleration correctly captured
Organizational expense caps100%Both fixed-dollar caps and percentage-based formulas correctly extracted
Fee commencement handling: The benchmark included funds with four different fee commencement triggers — first closing, first investment, a predecessor fund event, and the fund's effective date. The system correctly classified each trigger type without defaulting to industry-standard assumptions.

Key Person & Governance Provisions

Key person and fund governance provisions were extracted at 100% accuracy across all four funds. The system correctly handled structurally distinct governance architectures.

ProvisionStructures EncounteredAccuracy
Key person identificationNamed individuals, group-based triggers, dual-clause thresholds100%
Trigger consequenceAutomatic suspension, LP vote to terminate, cure periods100%
GP removal thresholdsFor-cause (50–66.7%), no-fault (75%), single-threshold structures100%
Fund term & extensions8–11 year terms, 1–2 year extensions with tiered consent requirements96.9%

Methodology

Benchmark Protocol

  • Four real-world LPAs from active private equity funds were processed with zero human intervention — no pre-labeling, no prompt tuning per document, no manual corrections.
  • Each LPA was parsed, structurally analyzed, and extracted using the same pipeline configuration. No fund-specific customization was applied.
  • Extracted values were compared field-by-field against the corresponding LPA provisions by a domain analyst. Each field was scored as correct or incorrect with no partial credit.
  • Fields correctly returned as null (when the provision exists in an external agreement or is genuinely absent from the LPA) are scored as correct. The system is not penalized for acknowledging what it cannot determine.

What Counts as a Field

Each of the six term categories contains a fixed schema of 8–16 fields. Fields include numeric values (rates, thresholds, durations), categorical classifications (waterfall type, fee base type), boolean indicators (clawback existence, recycling permitted), and structured objects (distribution waterfall tiers, key person lists, early termination triggers). A total of 271 fields were verified across the four funds.

Scoring Rules

  • Numeric values must match within 0.1% tolerance (e.g., 0.02 vs 0.0201 is correct).
  • Categorical values must match exactly (e.g., "european" vs "american" — no partial credit).
  • Null values are correct when the provision is genuinely absent from the LPA or defined in an external document. Null is incorrect when the provision exists in the LPA and should have been extracted.
  • Structured objects (waterfall tiers, key person lists) are scored as present/absent — the object must exist and contain substantively correct content.

Document Diversity

The four benchmark LPAs were selected to represent the structural diversity encountered in large-cap PE fund documentation. They include funds with all fee terms in the main LPA body, funds with fee terms defined in external exhibits, funds with ILPA-template structures using bracketed agreed terms, and funds with multi-entity parallel structures requiring cross-reference resolution. Document lengths ranged from 87 to 141 pages.

Limitations

This benchmark measures extraction accuracy on the LPA document as provided. It does not measure:

  • Extraction from side letters, capital call notices, quarterly reports, or other post-LPA documents.
  • Cross-document verification (e.g., comparing extracted LPA terms against actual GP charges).
  • Accuracy on venture capital, real estate, infrastructure, or credit fund LPAs — this benchmark covers buyout and growth equity funds only.
  • Extraction of provisions defined entirely in external agreements (e.g., a Management Agreement appended as an exhibit but not included in the uploaded document). In these cases, the system correctly returns null rather than hallucinating values.

The five errors in the benchmark (across 271 fields) fall into two categories: provisions defined in external agreements that were not included in the document (3 errors), and a definitions section where the document parser did not capture the complete text for two defined terms (2 errors). No errors resulted from incorrect interpretation of provisions that were present in the document.

© 2026 QuantisticAI, Inc. · quantistic.ai

Benchmark conducted April 2026. Results reflect the production extraction pipeline with no per-document customization.