Method Design explorer

Explore a worked experiment.

The default below is one worked design: an attack-by-intervention contrast in a stateful banking scenario. Change what you like and watch the structure it commits you to: cells, matched blocks, planned comparisons and the limit on what the result could claim.

  1. An illustrative design is not a frozen protocol.
  2. A frozen protocol is not a completed experiment.
  3. A completed experiment is not a validated finding.

Everything on this page is at the first rung. It exists to make the shape of a CARF experiment understandable, not to report a result.

Initial example before changesAttack × intervention, in a stateful banking scenario, at the discovery stage. The controls below show the current selection.

Change the design

Alternative templates, attack families, delivery schedules, doses, ablations, safe routes, representations and analysis layers. The design below updates as you change them.

Supports estimation of perturbation and intervention contrasts and their interaction, subject to the stated assignment and analysis assumptions. This illustration does not establish mitigation effectiveness.

Declared conditions
Dimensions
Analysis layers, prespecified
4
design cells
1
configurations
48
planned episodes
12
turns / episode
576
planned model turns
3
illustrative comparisons

4 cells × 12 episodes × 1 model configuration × 12 turns = 576 planned model turns

Design cells · one representative episode eachturn shading = perturbation intensity · orange edge = illustrative intervention timing
Control · no intervention No intervention
× 12 episodes Available and explicit
Attack · no intervention No intervention
× 12 episodes Available and explicit
Control · intervention Hold / human review
× 12 episodes Available and explicit
Attack · intervention Hold / human review
× 12 episodes Available and explicit
The measurement path
Stateful worldSynthetic banking / payment settlementInput deliveryLate jump after benign setupEpisode shape3 segments × 4 turnsSeparate endpointsProposed action and transaction or settlement stateCandidate representationMultiple views against transparent baselinesPlanned analysisMatched contrast + 5 more
Primary comparisonPlanned Discovery · matched sets

Attack × intervention interaction

In this schematic design, one matched set contains one episode per condition. There are 12 sets per model configuration, not one independent replicate per turn. Multi-session designs must define their larger units separately.

Illustrative design only: no agent, detector or policy is executed. Where a cell has an intervention, its marker starts at zero-based turn floor(turns_per_episode * 0.55). This is scripted timing, not a warning-triggered response. Counts are planned observations, not a sample-size or power calculation. Illustrative comparison counts are not a multiplicity plan. Exploratory design: possible candidates require fresh follow-up.

This explorer sketches structure only; real designs also declare their observational unit, endpoint definitions, controls and holdout split before anything is interpreted. The discipline behind it lives on the Method page; the exported JSON is an illustrative configuration, not an executable or registered protocol. Throughout, a segment is a stretch of turns inside one conversation, and a matched block is an experimental replicate. They are different things and the labels never swap.