Agentic loop
A system where a model does not just answer once, but acts, observes the result and decides what to do next. Useful for experimentation — but only as reliable as the observation step.
Glossary
The vocabulary of autonomous experimentation
Automation has its own dialect, and most of it is borrowed from other fields with the meaning quietly changed. This is how we use these words, written so that a domain expert from outside machine learning can follow the engineering, and an engineer can follow the science.
Where a term has a precise meaning inside a specific domain, treat the definition here as our working usage rather than a universal one.
A
Automated experimentation
An AI system owning the full cycle: proposing hypotheses, running experiments, evaluating results and choosing the next round. What AutoScientia builds.
See the loopB
Baseline
The reference a result is measured against. It is frozen for the duration of a comparison, because a win that comes from moving the measuring stick is not a win.
Budget
The hard ceiling on compute or spend attached to an objective. Enforced by the scheduler before a trial starts, not audited after the invoice arrives.
See control surfacesC
Candidate
A single proposed experiment: one configuration, design or hypothesis that the loop is considering running.
Closed-loop search
Search where each round of results changes what is proposed next, rather than walking a fixed grid of experiments. The opposite is open-loop or grid search.
Cost per gain
What an improvement actually cost in compute and time to obtain. The number that stops a loop from mistaking activity for progress.
E
Evaluation harness
The automated scoring machinery that decides whether a run produced anything worth keeping. In an autonomous lab this is the product.
See measurementExperiment orchestration
Everything between “here is a candidate” and “here is a result”: queueing, placement, resources, retries, checkpoints and result collection.
F
Frontier model
A state-of-the-art model used as the proposer in our loop. It generates ideas and plans; it does not decide what counts as success.
G
Grand Challenge
A large, well-defined societal problem — clean water, clean energy, security, health — where progress is measurable but slow under manual experimentation.
See applicationsH
Held-out objective
A metric the proposer never sees, used to check whether an apparent gain generalises or was simply fitted to the visible measurement.
Human-in-the-loop
People making the decisions that should not be automated: objectives, budgets, stopping conditions and any effect outside the sandbox.
I
Iteration
One turn of the cycle. Iteration time — not effort — is the quantity the whole system exists to reduce.
M
Metascience
The study of how research itself is done. In our context: how to tell that an automated system is doing science rather than generating plausible activity.
See research themesMisuse potential
How much damage could follow if a capability were pointed at a hostile objective. Screened before a campaign is approved, not after.
O
Objective
What the loop is asked to optimise, stated before any experiment runs and together with what would count as failure.
Objective misspecification
When a system optimises the metric you wrote instead of the outcome you meant. Most automation failures are this, wearing a different hat.
P
Parallelism
Running many trials at once rather than one after another. Real parallelism needs scheduling, isolation and per-trial budgets, not just more machines.
Promotion
Moving a candidate from one round to the next, with the evidence that justified it. The moment where a loop is most tempted to cheat itself.
Proposer
The component that turns an objective and past results into the next set of candidates, ranked by expected value and cost.
R
Reproducibility
Being able to re-run a result from its record: code, data references, configuration and environment. At machine scale this becomes a systems problem rather than a habit.
Run record
The versioned description of an experiment, complete enough to rebuild and re-score it. If it is missing, the number is not reported.
See measurementS
Sample efficiency
How much useful information a search extracts per experiment. The difference between a loop that finds an answer and one that runs out of budget looking.
Sandbox
Isolated execution with no ambient credentials and restricted network access, where machine-generated code is allowed to run and fail safely.
See securitySearch space
What may vary in an experiment, what is fixed, and which interactions are worth spending trials on. Declared before the cycle starts.
Scientific method
Hypothesis, experiment, observation, revision. Our claim is narrow and deliberate: we automate the loop, not the reasoning about what matters.
T
Throughput
Finished, scored experiments per unit of time at a fixed cost. The single number that best describes whether an automated lab is actually working.
Trial
A single execution of one candidate under a pinned configuration. Thousands of trials make a cycle; a cycle makes a campaign.
A term missing, or used badly?
Tell us and we will fix the definition. Precise language is part of the work, not decoration.