When to use each paradigm, what it measures, and live demonstrations of how bias, low power, and missing counterbalancing turn a clean study into a confidently wrong one. Every section links to a one-click template you can clone and adapt.
Seven well-established measurement contracts. Pick the one that matches your research question, not the one that looks coolest.
When to use: When you want a defensible ranking of items — designs, stimuli, logos, candidate options — free of order effects and anchoring bias.
What it measures: Pairwise/quad votes aggregated into ELO + Bradley-Terry rankings with standard errors and a publishable-threshold gate.
References
When to use: Whenever the dependent variable is speed: Stroop, Simon, flanker, go/no-go, lexical decision, primed classification.
What it measures: Trial-by-trial RT + accuracy, per-condition means, timing-audit grade (Reimers & Stewart 2015) on every published study.
References
When to use: Match-to-sample within vs cross-category trials – tests whether a linguistic or perceptual category boundary sharpens discrimination.
What it measures: Accuracy, RT, cross-minus-within advantage; ΔE-controlled CIE Lab stimuli via the lab-generator so luminance never confounds the result.
References
When to use: When the DV is a continuous error — participant drags a stimulus until it matches a reference. Müller-Lyer, Ebbinghaus, Ponzo and other length / size illusions.
What it measures: Settled value vs reference, per-trial illusion magnitude in pixels, per-participant adjustment trajectory, group means with normative-range overlay.
References
When to use: Product or landing-page tests where each participant sees only one variant. Random assignment with counterbalancing, per-variant KPIs, and a built-in χ² test.
What it measures: Conversion rate, click-through, time-on-task per variant; confidence intervals and effect size.
References
When to use: Attitudes, demographics, PROs. Drop-in question library for Likert, binary, numeric, free-text, with piped logic and jump rules.
What it measures: Per-item descriptives, Cronbach α, item-total correlations, exportable to SPSS/R/CSV.
References
When to use: Factorial moral / judgement studies: short scenarios with manipulated factors, followed by attitude or decision items. AI-assisted scenario generation + TTS narration.
What it measures: Full factorial designs with attention checks; item-level scores and between-condition contrasts.
References
These are not »try a paradigm« demos – those live on /demo. These are demonstrations of why each methodological decision matters: drag a contaminant into a clean study and watch the result break.
Methodological demo
Ranking is intact · ρ = 1.00
The measured order matches the truth — bias is small enough that ELO recovers cleanly.
Truth (ground-truth quality)
Measured (ELO leaderboard)
How SciBLIND prevents this: every blind-comparison study runs Latin-square counterbalancing on item ordering by default, with the matchmaking layer guaranteeing each item appears in each slot equally often. The leaderboard you publish is the ρ ≈ 1.0 column on the left, not the ρ < 0.5 column you can drag yourself into above. · 200 simulated comparisons · seed 42
Sample-size × effect-size × α calculator with a live curve and SciBLIND-default warning thresholds — so you size your study before you run it, not after.
Animated walkthroughs of published studies that failed replication, with the methodological flaw highlighted and the SciBLIND-built equivalent shown side-by-side.
Latin-square counterbalancing as a grid; ELO matchmaking depth as a saturation heatmap; staircase convergence as an animated trace. The inner workings participants never see.
Interactive d / OR / r explorers with worked examples mapped to SciBLIND paradigms. Translate a literature claim into a sample-size budget in two clicks.
Each kind ships when it can do its category justice. The alternative – four half-built demos – would be scientific theatre, and that's the opposite of what /learn is for.
Short, plain-English notes on the concepts researchers ask about most. No equations hidden; no jargon surfaced.
Jitter is dominated by display refresh and event-loop latency. With ≥60 Hz and our timing audit, differences ≥25 ms reproduce cleanly; sub-10 ms effects need lab equipment.
Read articleHand-picked hex codes give uneven perceptual distances. CIE Lab with fixed L* and uniform ΔE between swatches removes luminance confounds from any color discrimination study.
Read articleHit rate and false-alarm rate → d′ (sensitivity) and c (bias). Separates »good at the task« from »trigger-happy.« SciBLIND computes both on every DISCRIMINATION study.
Read articlePick a target accuracy (e.g. 75%). The n-up/n-down rule converges to the matching stimulus level in ~30 trials. Our builder exposes this as a single slider.
Read articleLatin-square randomization neutralises stimulus-order confounds across participants. Blocked randomization preserves within-condition learning. The builder picks sensibly by default.
Read articleEcological momentary assessment done right: schedules, prompt channels (push / Telegram / WhatsApp / email), compliance benchmarks, and how SciBLIND turns a 2-week diary study into one builder step.
Read articleHow completion codes work, why SciBLIND mints them server-side, and how to wire them into your MTurk HIT, Prolific study, or Bakker project. Covers fixed / per-session / JWT-embedded modes.
Read articleSkip the scripting. Describe your research question; the AI configurator pre-fills a valid study. You review every suggestion before participants see it.
