VMQ: Venture Market Quality

VMQ deliberately separates the thing that predicts from the thing that explains, because conflating them is how rating systems become unfalsifiable.

An LLM analyst pipeline produces the rating and the reasoning: the dimension scores, the comparables, the flags, the upgrade and downgrade conditions. Separately, a quantitative model makes a real forward prediction scored against outcomes. Because the two are measured apart, the explanation cannot quietly absorb the prediction's errors.

The rating is the composite's percentile across the whole combined population, as one global ranking with exclusive bands. L5 means the top one per cent, L4 the top ten, L3 the top quarter and the beginning of investment grade, L2 the top half, L1 the rest. A top rating is genuinely scarce, the way AAA is, rather than being handed out whenever a market looks good in isolation.

One global ranking is a deliberate choice. The reference markets are synthetic, authored by a language model, so there is no principled reason to rank sub-groups separately. They are pooled and ranked together, and the real predictive signal is validated elsewhere.

The predictive model is trained and validated against real private-market data rather than against the synthetic labels. For a market defined by keywords and a vintage year, the cohort is the seed companies founded in the four years up to that vintage which had raised seed by it. The outcome counts as positive if the company reached Series D or beyond, an IPO, or an acquisition above $250M, within five years of the vintage.

The features are chosen with leakage discipline: category age, seed momentum and seed median, all computable at the decision point. It would be easy to build a model that looks excellent and is quietly reading the future; the restriction exists to make that impossible rather than merely unlikely.

A rating on its own is a number. Around the core sits the work an investor actually does: reassessing the whole population as of today, with forward predictions driving the ordering and a movers diff against the stored baseline; discovering missing markets through an LLM sweep gated by an expert review queue; generating value, momentum and quality lists for any segment.

Those lists feed directly into the portfolio construction and capital allocation engines, and investment-thesis memos are written in a split-screen panel beside the evidence that produced them.

Experts can challenge a rating, and those challenges are collected rather than discarded. Over time that becomes an evaluation benchmark: a record of where informed humans disagreed with the system, and who turned out to be right.

That is the part most rating systems lack. Without it, calibration is an assertion.

A Vela Research and Oxford collaboration, built on the Vela design system, with a Kronos growth forecast supplying the trajectory each rating is set against. Cloud credentials are resolved by path and never read or printed, and every warehouse query is capped on bytes billed, because a research tool that can run up an unbounded bill is not finished.