LP Bench
The fund scoring accuracy benchmark NUVC is building for AI-powered LP intelligence.
NUVC publishes this methodology because trust in AI fund scoring requires measurable proof — not marketing claims. LP Bench will test both quantitative accuracy and qualitative narrative quality, the Science + Art balance made verifiable. Results ship dimension by dimension, only once they are measured.
What We Measure
Six dimensions covering both the quantitative signal layer and the qualitative narrative layer — because capital allocation requires both. None of the results below have been measured yet; this section documents the method so it can be checked once they are.
Score Consistency
Same fund deck, 5 independent runs. Standard deviation across runs measures LLM scoring stability.
The Batch Screener layer is designed to produce consistent 6-dimension scores across repeated runs, with a cross-check layer intended to suppress outlier LLM responses. Not yet measured and published as a result.
Vintage Calibration
NUVC vintage benchmarks vs Cambridge Associates global VC return data. Tests whether NUVC benchmark priors are calibrated to real market outcomes.
AU/NZ market-specific priors draw on State of Funding 2022–2025 and Cambridge Associates global quartile data. The deviation itself has not yet been measured and published.
GP Trajectory Prediction
Among funds scoring ≥ 7.0 (high conviction), what % outperform the median TVPI at the 3-year mark? Tests whether the score has predictive validity.
This requires committed funds with 3+ year track records in the NUVC LP network. That outcome dataset does not yet exist at a size that supports a published result — published only once it does.
Mandate Alignment Accuracy
For FOs that committed to funds recommended by NUVC matching, what % matched within their stated thesis alignment threshold?
Bidirectional mandate matching (FO ↔ GP) across sector, fund size, vintage, and geography gates. Requires a committed-fund sample large enough to report honestly — not yet reached.
DD Hallucination Rate
% of factual claims in AI-generated DD reports that could not be grounded in source documents (fund deck, DDQ, or track record). Measured by the Score Explainability layer (claim verifier).
The Score Explainability layer cross-references every DD claim against extracted source document data, flagging ungrounded claims with amber confidence badges in the live product. The audit will be published once it has run across enough DD reports to be statistically meaningful.
Qualitative Narrative Agreement
Planned: inter-rater agreement between NUVC's qualitative fund narrative and an independent LP analyst panel, scored with Cohen's kappa.
Science + Art principle applied to measurement: quantitative score accuracy is necessary but not sufficient. A panel has not yet been convened — this dimension will publish a result once it has.
Why We Measure Both Science and Art
Most AI fund scoring tools publish only quantitative accuracy metrics — consistency and calibration. These are necessary, but not sufficient.
LP decisions rely equally on qualitative judgment: is this GP's thesis authentic? Does the team dynamic suggest longevity? Is the communication style right for a long-term LP relationship?
LP Bench is designed to measure qualitative narrative agreement alongside quantitative accuracy — because capital allocation that ignores either dimension misses what actually matters.
- Score consistency
- Vintage calibration
- TVPI prediction
- Hallucination rate
- Team dynamics assessment
- Thesis authenticity
- GP communication quality
- LP panel agreement
Methodology
Open methodology, published ahead of results. LP Bench will recalibrate quarterly as new committed fund outcome data becomes available.
Source Corpus
18,000+ funds in the NUVC fund library — AU/NZ emerging managers, global VC, PE, and private credit. Predictive-validity work will use the subset with verified 3+ year outcome data, once that subset is large enough to report on.
Consistency Testing
Each fund document is submitted 5 times with identical inputs. Score standard deviation is measured across runs, with a cross-check layer intended to suppress LLM outliers.
Vintage Calibration
NUVC benchmark priors are compared to Cambridge Associates global VC quartile data (Q1/Q2/Q3 TVPI and IRR by vintage year) and AU State of Funding 2022–2025 market data.
Hallucination Audit
The Score Explainability claim verifier runs across DD reports as the conviction pipeline produces them. Every factual claim is cross-referenced against fund deck, DDQ, and track record source documents; ungrounded claims are counted.
Qualitative Panel
An independent LP analyst panel will assess a sample of fund profiles across qualitative dimensions. NUVC's narrative assessments will be scored against panel consensus using Cohen's kappa, once a panel is convened.
Publish As Measured
LP Bench results are published only once they can be computed from real data — never estimated, never projected. This page will update dimension by dimension as that happens.
Why LP Bench Differs from Legal AI Benchmarks
Legal AI tools like Harvey benchmark document drafting speed and citation accuracy. Capital allocation intelligence requires a different standard.
| Dimension | Legal AI | NUVC LP Bench |
|---|---|---|
| Quantitative accuracy | Citation accuracy | Score consistency + vintage calibration (planned) |
| Qualitative accuracy | Not measured | LP panel inter-rater agreement (Cohen's κ, planned) |
| Predictive validity | N/A (legal drafts) | Score ≥ 7.0 → TVPI outperformance rate (planned) |
| Hallucination testing | Legal citation verification | Fund claim grounding in source documents |
| Network effects | None | Benchmark improves as committed fund outcomes accrue |
| Update cadence | Ad hoc | Quarterly recalibration |
Intelligence that earns its place in your process
LP Bench is how we intend to hold ourselves accountable. Every score NUVC produces — for every fund, every FO, every capital allocation decision — will be measured against this standard as the data to do so arrives.