Can frontier AI actually sell?

Three separate questions — kept separate on purpose

Most benchmarks collapse into a single leaderboard. VEB answers three distinct questions, and the ranking is deliberately the least important of them.

One instrument for the question every sales leader is about to face: can this AI carry a deal?

The leaderboard, ranked by the floor of its confidence interval

Mean Sale Quality Score (SQS, 0–100) per model, pooled across both tracks and all scenarios. We rank by the lower bound of the bootstrap 95% CI, not the point estimate — a model only outranks another if the gap survives the noise.

Model naming & run policy:labels like “GPT-5.6-Sol” or “Grok-4.20” are normalized display names for specific pinned API endpoints — production models as each lab exposed them at run time. Two entries flagged open-weight (Kimi-K3, Inkling) are base open-weight checkpoints served as-is — no private or fine-tuned checkpoints appear on the board. Every model played every scenario bare (OOB) and armed (PACK) under identical settings. Exact provider/model identifiers and versions are recorded in provenance.json, and every arm is reproducible from the frozen dataset (released file sha256 13a9e6e5ac7a…).

The winner’s playbook — what all 330 closed-wins had in common

Across 330 closed-won and 3180 everything-else campaigns, a handful of behaviors separate the two populations almost perfectly. These are win/loss dividers, not vanity metrics.

Every model sells with a personality

A behavioral fingerprint per model across the five pillars of the sale — each pillar scored as a share of the maximum attainable, so the shape tells you how close each model gets to a perfect sale. Below each: how it behaves when procurement squeezes on price.

Not just who won — how they played

On top of the deterministic grid sits an enriched twin: 17 behavioral features a calibrated 3-seat LLM panel reads from the transcripts — psychology, emotion, rhetoric, ethics, and four reward-shaping signals unique to this environment. Every score quotes a transcript line or abstains; nothing is inferred from the outcome.

The panel is validated against 139 human-labeled items before it scores anything. Because it abstains when a transcript carries no evidence, per-feature n varies across the 3,510 campaigns — every chart below reports its own n so coverage reads as a rigor signal, not a gap.

What behavior drives a better sale

Each judged behavior correlated against the Sale-Quality Score and against a closed-won outcome, across every campaign the panel could score.

The reward signals

Four features exist to make VEB a verifiable reinforcement-learning environment, not just a leaderboard: they locate the decisive move, weigh its counterfactual impact, and catch the buyer state a seller failed to read.

Behavioral fingerprints

Every model’s behavior profile across the feature families, armed with the methodology. Two models can land at the same score by playing completely different games — this is where you see it.

How the read tracks the outcome

The same behavioral read, sliced by how the deal ended. Won deals and dead deals look different long before the buyer says yes or goes dark.

The steepest levers on winning

Each behavior split into low / mid / high bands, with the share of campaigns in each band that closed won. Ranked by the gap between the top and bottom band — the behaviors where doing more moves the needle most.

Every claim is anchored to a verbatim transcript line

The insight layer is anti-fabrication by construction: a failure-mode exhibit only ships if its quote resolves back to the exact run on disk. The emit step throws on any dangling claim. A sample of the 13 anchors behind the atlas above:

What this benchmark cannot tell you

The same frozen world is a verifiable RL environment

Every VEB scenario doubles as a training target in the SWE-bench / Terminal-Bench standard: a frozen buyer-sim (claude-sonnet-4-6@t0.8+smv3), a machine-checkable reward, and an sha256 task_checksum pinned to the exact world the agent was handed. The proof below runs fully offline in mock — no API keys, seconds, $0 — so anyone can reproduce it. Two environments below clear the difficulty gate; the third is shown unsolved, on purpose.

A gate passes only when the disciplined oracle solves the environment (reward ≥ 0.9) AND a naive baseline cannot (reward < 0.5) — the same solvable-but-not-trivial band SWE-bench enforces. Hostile Renewal is held out as the hard case: even the oracle banks only 0.207, so it is published as an open target, not padded into the headline.

The reward can’t be gamed — Logistics SaaS

Five adversarial exploit policies attack the reward directly. Each must land below the audit ceiling (0.5) and none may resolve the deal. Worst exploit here scored 0.098 — all five defeated.