Methods

The method is part of the question.

How an experiment is designed, how its evidence is evaluated, and how the next experiment is chosen are themselves methodological decisions. Lab Loop makes them explicit and swappable, grounded in the theory of empirical research, not locked to one approach.

Method choice belongs outside the result.

The research question determines the method, not the other way around. Design, evaluation and search can change as the programme evolves.

Method coverage

Five methodological families, each grounded in a distinct theoretical tradition, compose into the research loop.

Experimental design

How trials are structured

The design determines what a result can teach. Factorial and fractional designs (Fisher) isolate main effects and interactions; response surface methodology (Box & Wilson) maps optima; optimal design theory (Kiefer) allocates trials efficiently under a model.

Full factorial Fractional factorial Response surface D-optimal Latin hypercube
Hypothesis evaluation

How evidence is assessed

A result only counts as evidence against a method for evaluating it. Frequentist testing (Neyman & Pearson) controls error rates; Bayesian updating (Laplace, Jeffreys) revises belief quantitatively; likelihood methods compare support across hypotheses directly.

Significance testing Power analysis Bayes factors Posterior updating Likelihood ratios
Search strategies

How the next experiment is chosen

When experiments are expensive, the choice of the next trial is itself a decision under uncertainty. Model-based search (Bayesian optimization) builds a surrogate and exploits it; direct search (Nelder & Mead, CMA-ES) needs no model; random and quasi-random baselines establish what structured search is worth.

Bayesian optimization Evolution strategies Direct search Random baseline Multi-fidelity
Multi-objective exploration

When objectives compete

Real research questions rarely have a single objective. Pareto-based methods (Edgeworth, Pareto; Deb) return a non-dominated set rather than compressing objectives into one arbitrary score; scalarization and preference elicitation let a decision-maker steer the search without pre-committing to weights.

Pareto frontier Scalarization Preference-guided Hypervolume
Sequential decisions

Under uncertainty and budget

Research runs under a finite budget of trials, time and cost. Multi-armed bandit theory (Thompson, Gittins, Lai & Robbins) formalizes the explore-exploit tradeoff; active learning chooses experiments that most reduce uncertainty; optimal stopping decides when enough evidence has accumulated.

Bandits (UCB) Thompson sampling Active learning Optimal stopping
Reproducibility & variance

Making evidence count

A result that cannot be reproduced cannot inform a decision. Replication and blocking (Fisher) separate signal from noise; variance estimation and resampling (Efron) quantify uncertainty without distributional assumptions; randomization guards against confounding the factor under test with uncontrolled variation.

Replication Blocking Randomization Bootstrap Variance reporting

One loop interface

The research programme remains stable while the method at each stage changes according to the structure of the question.

01

Programme

Goal, constraints, budget and the active question.

02

Design

A trial configuration produced by the chosen design and search method.

03

Evaluate

Evidence assessed by the chosen evaluation method, with variance reported.

04

Decide

Accept, reject or revise, and choose the next experiment under the remaining budget.