Point it at a dataset and a metric. Ascent designs experiments, checks each idea against published research, runs them overnight on your own laptop, and keeps only the changes that survive three automatic integrity checks — typically 9 kept out of 218 tried. Validated across six domains.
Try enough ideas and one will look brilliant by luck alone. Careful researchers guard against that by hand — holding data back, checking the literature, refusing to keep something that only worked once. It is slow, unglamorous work, and at 2am nobody does it properly. That is not a character flaw. It is a workload problem, and it is the one nobody has automated.
A practitioner runs one to three experiments a day, and spends 38–45% of their time building scaffolding rather than testing ideas.A39
Under independent review, 42% of the best-known automated research system's experiments failed outright.A16
Plenty of tools run more experiments. None enforce the discipline that decides which results deserve to be believed.
You give it a dataset and the single number you want to improve — an accuracy, a forecast error, a risk-adjusted return. Setup takes about twenty minutes and needs no ML engineering.
It forms a hypothesis, checks it against published research, changes one thing, measures, then keeps or discards. Then again, for as long as you let it — on your laptop, on your own API key.
A result, the full record of how it was reached, and a package a skeptical colleague can re-run with a single command.
The rule: three automatic checks run before any result is recorded — that the data was split honestly, that the idea was grounded in real published work, and that the reasoning was written down. Fail one and the experiment is discarded. The system cannot override them, and neither can you.
There are good tools in this space. None of them do the part that decides whether a result is real.
| What it is | What it automates | What it leaves to you |
|---|---|---|
| AutoML (AutoGluon, H2O, DataRobot) | Hyperparameter search | The evaluation protocol, the literature, and whether the result means anything |
| AI scientists (paper-writing agents) | One-shot papers on their own problems | Your metric, your data, and sustained improvement |
| Experiment trackers (W&B, MLflow) | Recording what you did | Designing the experiments and judging them |
| A chatbot | Suggesting ideas | Verification — a tired human at midnight |
| Ascent | The scientific method itself | Steering, in one or two sentences a day |
A single day on two live campaigns, taken from the pack's day-in-the-life journey. Every figure below is projected from proof-of-concept timings and labelled as such.
This is the part every backtest gets wrong. Try enough ideas and the winner looks impressive even when nothing real is happening underneath. Statisticians have a correction for exactly this, and it needs one input almost nobody bothers to keep: how many things were actually tried, failures included. Ascent keeps that count on every experiment — which is what lets it publish the discounted number rather than the flattering one.
You do not have to take our word for it. Every kept result ships with the data splits, the reasoning, the full trial count and a one-command reproduction. Run it yourself; that is the product.
AutoML searches hyperparameters against an evaluation harness you built. Ascent builds the harness, grounds each idea in published work, and refuses experiments that fail integrity checks. The search is the easy half.
Training and evaluation data never leave your hardware. Your steering instructions do go to a hosted model API — we say so plainly rather than let a security review discover it. Bring-your-own-endpoint is on the roadmap.ROADMAP
Roughly $3–12 a day of your own model credits, plus $125/month for the layer that does the checking. You pay the compute bill directly, so it never becomes our margin.
No. Setup is a guided twenty-minute questionnaire, and the system supplies the evaluation method — including catching mistakes an ML engineer would have caught.
The trial-count-corrected acceptance gate, attestation packs, bring-your-own-endpoint and fingerprint refusal. They are marked ROADMAP everywhere on this site rather than implied as shipped.
The minimal loop drew 94,800 stars, then lost its maintainer. Of sixteen notable forks, none added rigor.A2
Turning off one check raised invalid experiments 42% and produced 3 leakage incidents. The discipline is worth its overhead.
Users pay their own compute bill directly, so it never touches our margin: 82.0% gross at 150 users.
Sixty-four documents covering research, strategy, product, architecture, validation and financials — every claim tagged to where it came from, and every gap named rather than hidden.