FOUNDING SEAT

Technical Co-Founder

Four founding seats

In person, San Francisco.

Founding Research Scientist, Training
Owns the policy: training on decisions and their measured outcomes. This page describes this seat, the technical co-founder.
Founding Research Scientist, Evaluation
Owns the measure: simulated companies with known ground truth, and the grading of decisions.
Founding Research Engineer
Owns the system: data generation, training and serving at scale.
Co-founder
Owns capital, partnerships and the first companies Volt operates.

For the other three seats, write to founders@govoltaire.com.

About Voltaire

Intelligence is becoming abundant. Companies still haven't converted it into meaningful gains in productivity and revenue.

That’s largely because the decisions that make or break a company — which outcomes to pursue and in what order — are still made by instinct and analogy.

Voltaire Labs trains Volt, a decision model that can run a company: it decides which outcomes the whole company should own, directs people and agents toward them, and learns from every result.

In its end state, the company runs itself. Agents do the work; Volt holds the intent and the record of what worked; humans own the company.

The companies of the future will run on a model.

The race to build it is a race to solve a handful of open problems, and Voltaire intends to lead it.

V1

V1 is built. It takes one outcome, ARR in twelve months, and runs the whole loop.

It rebuilds ARR from the billing ledger to the dollar and forecasts every driver with no change. It generates and values decisions, and chooses the set that fits the company's hours and budgets. It turns them into tasks for people and agents, each waiting on an approval, then measures the results against a holdout and re-ranks.

It runs on three simulated companies whose true effects we planted, so the engine can be graded against the answer.

Building it showed exactly how much is still open. Its next test is a real company.

The research frontier

The full decision layer - what a company should own - does not exist yet.

Building it means solving seven open problems. Each one below says what v1 does about it, and what remains.

  1. 01

    Outcome discovery

    From the company's state, its history and its market, generate the outcomes worth owning, each with a target, a deadline, its evidence, its value, a confidence level. Detectors yield anomalies and candidate moves; an outcome is a target over a horizon, composed from those signals and from what the company wants. Whatever ranks the candidates, the top estimate tends to be the most overestimated one: the optimizer's curse.

    V1

    Candidates come from a library of decisions per business model, and from detectors that fire when a driver drifts clearly and materially against the same months a year earlier. Each estimate is shrunk toward a skeptical base rate, to counter that curse.

    Still open

    V1 ranks decisions under one outcome we chose: ARR in twelve months. Generating the outcomes themselves is untouched, and the detectors do not yet travel beyond the companies they were written on.

  2. 02

    The objective

    What the company maximizes, over which horizon, at what tolerance for risk, under which constraints, and whose judgment prevails when its people disagree. This must be inferred from the data and from what people adopt, defer and reject. Written policy alone already defeats frontier models, which passed 25 percent of tau-Knowledge tasks at the first attempt in March 2026; unwritten objectives are harder still.

    V1

    The objective is set, not inferred: expected ARR in twelve months, within each owner's hours, the budgets, the prerequisites, how often a customer is contacted and how much a person can approve in a week.

    Still open

    Horizon, risk tolerance and disagreement between people are not modelled. Every approval, edit and rejection is logged; nothing yet learns the objective from them.

  3. 03

    Prediction

    A model of the company that can be run forward under any move and return a distribution of outcomes: a driver tree whose elasticities are estimated from the company's own history, with priors drawn from comparable companies. The space of moves is combinatorial; by contrast, outcome data is scarce. A 200-person company resolves too few outcomes a year to fit a rich model, and no public benchmark measures the effect of one action on one company's figures.

    V1

    Each driver gets a damped-trend forecast. Correlated Monte Carlo carries the drivers through the company's formulas to ARR. Effects combine reference-class priors, a skeptical base rate and the company's own comparisons.

    Still open

    In rolling backtests, the 80 percent ranges on drivers hold about 80 percent of the time. Well-calibrated parts do not yet compose into a well-calibrated whole.

  4. 04

    The portfolio

    The system must arbitrate between outcomes and decide how much to explore against how much to exploit. A portfolio of overlapping actions confounds itself, making allocation and measurement a single problem. No system yet arbitrates above the level of a single outcome.

    V1

    The chooser searches every subset up to 20 candidates and uses local search above that. Each decision left out carries the constraint that excluded it, and each choice is re-run on perturbed inputs and labelled a clear choice or a close call. For paid acquisition, a dedicated solver equalizes the cost of the next subscriber across channels, up to what a subscriber is worth.

    Still open

    The chooser maximizes expected ARR and puts no price on what a decision would teach. Exploration against exploitation is still unsolved.

  5. 05

    Calibration

    A figure attached to a decision means something only if its event is defined, if the figure is recorded at approval before the action runs, and if it is scored when the outcome lands. The live odds keep moving as data and other moves change them; the record of what was predicted does not.

    V1

    The Record keeps every prediction as it stood at approval, with the runners-up, and resolves it against what was measured. Track records run per class of action: counts first, then rates, then a Brier score against the base rate.

    Still open

    Planted effects land inside the 80 percent ranges 87 percent of the time. So the ranges run wide, and the prior behind them was tuned on those same effects. That is a test of the machinery. The real score needs real outcomes, resolved at twelve months.

  6. 06

    Attribution

    Separating what Voltaire caused from what the team would have done regardless: a company with 2,300 customers and 3 percent annual churn cannot prove a 25 percent reduction in churn: at 80 percent power, the test requires about 7,100 customers per arm. The outcomes not chosen have no control group at all, unless what was declined is kept as a set of counterfactuals and the selection itself is graded on regret.

    V1

    A shared holdout assigned by a stable hash, and a resolver that reads only events, never a prediction. A jackknife gives an interval for any reader. Over 150 seeded runs of a separate noisy test company, those intervals cover the planted truth at about their stated rate; the three simulated companies replay fixed figures, so they cannot test an interval. The runners-up are recorded, so the chosen set can be graded on regret.

    Still open

    A decision's ARR is still its modelled effect times its measured share, because a direct read from the ledger was too noisy to show. Small companies still cannot power most tests.

  7. 07

    Reflexivity

    An engine that acts on a company alters the data it then uses to decide; an account that the team flags and then saves reads as a false alarm. Each outcome needs rules for when it is complete, when to retarget it and when to abandon it. Adopting or retiring an outcome is a class of decision with no evidence base yet.

    V1

    The loop re-reads, re-values and re-ranks on the company's cadence. A stop is always a proposal to a person. Autonomy is earned per class of action: 60 approvals with few edits for internal work, 300 for anything a customer sees, and a new model resets the count.

    Still open

    Reads assume a decision's full effect from the day it lands, so a slow winner can draw a false stop. Retargeting and retiring an outcome have no rules yet.

The foundations

Beneath the deciding layer sit four problems the field is working on now.

  1. 01

    Simulation

    Evaluating the engine on work that spans weeks or months requires a simulated company, with its people and its systems modelled from the company's own data and traces. A run can then be rewound, forked on a single change and played forward again, without the simulated world contradicting itself.

    V1

    Three simulated companies, each 24 months of event-level history and three months forward, with every decision's true effect planted, so the engine can be graded against the answer.

    Still open

    Rewinding and forking a run, and modelling a company from its own traces, are not built.

  2. 02

    Memory

    What the engine learns about a company must be stored, retrieved and compacted over years: its facts, its preferences, its exceptions and the decisions already taken, kept current as records land and as people change their minds.

    V1

    An append-only, as-of event store: re-reading a window adds nothing, and any run can be replayed as of any date.

    Still open

    It holds records, not preferences or exceptions. Nothing is compacted, and no knowledge yet persists beyond the events.

  3. 03

    Self-improvement

    Loops that turn traces into verified changes to the engine's skills, memory, code and policies, judged by an evaluator the modifier cannot touch, so that an improvement in one place cannot degrade another.

    Still open

    Not started. V1 updates its beliefs about decisions; it does not change its own code, skills or policies.

  4. 04

    Credit

    Scoring and credit across trajectories that run for months, where several actions, several agents, and the team's own work touch the same figures. The profit and loss arrives too late, and too coarse, to serve as a training signal on its own.

    V1

    Reads per driver against the shared holdout, with intervals.

    Still open

    Decisions on the same driver confound each other's reads. Scoring across long, overlapping trajectories remains open.

The split

This is a founding seat, in person, in San Francisco.

The founder owns the market: positioning, design partners and investors. HEC Paris Grande Ecole, X-HEC Masters, four startups, one exit, four years at Logitech.

You own the system, and you own the architecture. The research agenda, the product, the raise and the order in which we take the market are decided together.

The bar

  • You have built a system that made decisions with money or customers at stake.

  • You have founded or run a business with its own profit and loss.

  • You treat a prediction as a claim with an event, a horizon, and a score, and you distrust a ranking until you know how you selected its winner.

  • You draw the line between what a model may draft and what code must decide, and you can defend where you drew it.

  • You write the code. The company model, the ledger, the gate: whatever you design, you can build alone before anyone joins you.

Strong signals: harness, post-training, or applied AI work at a lab or on an applied AI team; published research, especially on agent memory or continual learning.

Design choices already made

Voltaire builds its own model for decisions, and every scored call trains it. In v1 that model is explicit: forecasts, priors and a chooser written as code, so every figure can be traced. Language models read, draft and plan; their output never touches a confidence, a sizing or a rank. Agents read and workflows write. The company's history is bitemporal, so that no prediction can see its own future. The engine does not depend on any one language model. It is written in TypeScript as pure functions, and it runs in the browser as easily as on a server. Storage, scheduling, syncs, accounts and how the engine meets its first real company are not decided. The rest is yours to decide.

The deal

Co-founder seat. Double-digit equity on a standard four-year vest. A strong and opinionated voice in where the company goes. The seat is equity-first: no cash compensation until the funding round closes.

If you want to tell us how you’d tackle the seven open problems above, let’s talk.