FlowCLD

Iterations

Edit the same model as the Canvas in spreadsheet form. Changes sync both ways — switch to Canvas to see them on the diagram.

Nodes

Name	Start	Retention	Floor	Ceiling	ƒ(x)	Formula	Sink	Sink formula	Source	Source formula

Edges

Source	Target	Correlation	Decay	Confidence	Delay

Run the two-phase simulation. The backend returns raw series; charts are built in the browser with Chart.js (hover, legend toggle, range trim) — same stack as Legacy Time Series.

Iterations

Available Connections in Graph (Not in Database):

Node 1	Node 2	Polarity	Strength	Action

OpenAI API Key (For Generate explanation feature):

Node-to-Node Connection

Node 1:

Node 2:

Polarity:

Strength:

Justification:

Output:

No output yet...

View All Connections

Filter by Node Name:

Node 1	Node 2	Polarity	Strength	Justification

Cycle Detection

CLD Signal Propagation Analysis

Advanced analysis tools for Causal Loop Diagram signal propagation, behavior classification, and system stability assessment.

Analysis Controls

Simulation Steps: Monte Carlo Runs:

Analysis Results

Ready to run analysis...

Behavior Classification

System Summary

Recommendations

Parameter Space

2D Stability Maps

Generates three heatmaps in parallel — one for each pair of parameters (decay, delay, retention).

Decay – Steps

Delay – Steps

Retention – Steps

3D Parameter Space

Sweeps retention × decay × delay. For each retention value, a full 2D stability map (decay vs. delay) is generated.

Initializing Python environment...

Retention – Steps

Decay – Steps

Delay – Steps

Iterations

Fan Chart

Monte Carlo uncertainty analysis — runs many simulations with perturbed edge weights based on confidence values, producing confidence bands per node.

Simulations

Noise

Parameter Optimization

Finds the best (retention, decay, delay) combination for your graph using a grid search followed by simulated annealing refinement. Returns top 5 configurations ranked by stability score.

Note: This runs ~200+ simulations and may take 1–2 minutes on the free backend.

Controlled Structural Edits

Stage 3

User guide

Using Controlled Edits

Controlled Edits is a guarded way to change model structure. Edits are collected into a proposed transaction, checked on a private copy of the Canvas graph, and applied to the shared model only after the complete transaction is accepted and you choose Apply to Canvas.

1. Choose an edit

Add or remove a node or edge. Complete the displayed fields, then select Queue edit. Nothing on Canvas changes while an edit is only queued.

2. Build one transaction

Queue related edits together, such as adding a node and its connecting edges. Later queued fields can reference nodes and edges proposed earlier in the same transaction.

3. Set preservation rules

Protect important nodes or edges and add causal paths that must remain reachable. These rules supplement the always-active anti-collapse constraints.

4. Check, review, apply

Check transaction validates the entire proposal. Review its reason, reward, graph-size summary, and preservation score. An accepted check exposes Apply to Canvas.

Edit reference

Add node	Sets name, starting value, retention, optional bounds, and optional value/source/sink formulas. A formula converter must use retention 0.
Remove node	Removes the node and its incident edges. The full transaction must still satisfy every safety constraint.
Add edge	Sets source, target, correlation, decay, confidence, delay, and functional form. New self-loops and duplicate edges are rejected.
Remove edge	Removes the selected directed connection, provided protected elements and required causal paths remain intact.

Always-active safeguards

Retain at least 50% of the original nodes.
Retain at least 25% of the original edges.
Do not increase graph fragmentation or create isolated nodes.
Do not introduce new self-loops.

Important behavior

A rejected transaction does not change Canvas. If Canvas changes after a successful check, the preview becomes stale and must be checked again. Applying is rollback-safe: if replay fails, the original model is restored.

Recommended workflow: protect critical nodes and paths first, queue the smallest coherent set of edits, check it, read any rejection reason, and apply only when the preview matches your intent.

New edit

Pending transaction 0

Safety constraints

Keep 50% of nodes Keep 25% of edges Prevent fragmentation No new self-loops

Protected nodes Protected edges

Required causal paths

Team Workspace

Stage 5

User guide

Using Team Workspace

Team Workspace lets role-based teams test changes against the model currently open on Canvas. It uses the same graph data and simulation logic. A normal session remains a preview until you select Apply final to Canvas or keep the final replay frame.

1. Prepare the model

Build or load a connected model on Canvas first. Clear node names and meaningful starting values make objectives and results easier to interpret.

2. Define each team

Start from a role preset or use Custom. Choose what node the team owns and which node defines its objective. Behavior objectives target a dynamic class; trajectory objectives target a sequence of values. An optional maximum spectral radius adds a whole-model linear-stability goal.

3. Limit its authority

Permissions determine which parameters or structural edits a team may use. Budgets and move costs discourage excessive changes. Objective scope restricts actions to the objective node and adjacent edges.

4. Choose an approach

Greedy baseline picks the best immediate move. Epsilon-greedy learns a transparent value for each action family while gradually reducing random exploration. Learned actor-critic estimates longer-term value from graph features.

Color roles: Green and Blue are cooperative stabilizers. Red and Black are bounded adversaries. A disruptor still targets a healthy state in the form because the reward engine reverses that score. Presets fill editable settings; they do not bypass permissions, costs, budgets, or anti-collapse checks. Autonomous turns search bounded parameter changes and permitted structural motifs, including mediated paths, balancing links, and safe edge removal. Controlled Edits remains available when you want to queue an exact transaction yourself.

Learned-agent settings

Practice rounds	How many training episodes the selected learner receives. Start with 10–30.
Moves per round	Maximum learner decisions in each practice episode. Start with 5–8.
Learning window	How many future steps contribute to each learning update. Five is a practical default.
Frozen opponents	How non-learning teams behave during practice: hold, act randomly, or use a fixed greedy policy.
Planning depth	How far the learned agent looks ahead during the final session. Use 1 first; 2–3 take progressively longer.
Epsilon-greedy	Initial exploration is the chance of trying a random legal move during practice. Minimum exploration is its floor, and decay reduces exploration after each update. Evaluation always uses zero exploration.
Goal safeguard	Final learned moves are checked with the canonical simulator. If a stabilizer stalls before an explicit spectral target, the bounded planner supplies the next permitted retention or decay step; it cannot bypass team permissions.
Comparison seeds	Runs random, greedy, and learned policies on matching scenarios. Three seeds is useful for a first comparison.
Maximum spectral radius	Optional whole-model limit from the notebook transition matrix. Values below 1 are linearly stable; the reward still keeps activity, transmission, and move-cost safeguards.
Structural budget	Counts individual node and edge edits. Inserting a node between two existing nodes uses four edits: add node, remove old edge, and add two replacement edges.
Balance current model	Builds a bounded retention/decay plan for a connected model. Two-node balancing applies directly; larger-model plans stay in preview until you review the change table and select Apply final to Canvas.

Reading the results

Return: the accumulated objective score; higher is better for the defined team objective.
Learning curve: return by practice round. Some noise is normal.
Comparison: seed-matched averages for random, greedy, and learned decisions.
Exploitability proxy: how much a team could gain by switching policy. Lower is better, but it is not proof of equilibrium.

Replay and keep changes

Use Apply final to Canvas to commit the result directly. Use Replay on Canvas to inspect each move first. Exit restores the graph that was open before replay, while Keep final commits the last frame.

Recommended first run: 10 practice rounds, 5 moves, depth 1, no comparison seeds. Once that succeeds, use 30 rounds and 3 comparison seeds.

Balance the current model

Choose a whole-model spectral limit and which bounded parameters may change. The planner prioritizes influential nodes and edges, preserves structure and polarity, and checks the result with the full notebook simulation.

Maximum spectral radius

Adjust retention Adjust decay

Safety bounds

Minimum retention Maximum decay

Teams 0

Role preset Click the selected color again to return to Custom.

Name Orientation Owned node Objective node Objective Desired behavior Max spectral radius (optional) Weight Score preset

Agent settings

Future weight Parameter move cost Structural move cost Move budget Structural budget Minimum live nodes

Permissions

Team	Orientation	Objective	Weight	Permissions

Move order 0

Team Move type

Parameter Target Mode Value

Agent approach

Approach

Simulation steps Agent turns

Session result

Team scores

Team	Initial	Final	Cumulative

Move log

#	Team	Action	Status	Shared score	Reason

Researcher Access Management

Grant or revoke researcher access for users by email. Only accounts that have already signed up can be granted access.

Loading...

Diagram Generator

Describe a system in plain English and generate a causal loop diagram directly on the canvas — nodes, links, and polarities are laid out automatically.

Model Choose any provider — you'll need an API key for the one you pick.

API Key Kept only for this browser tab and sent only to the selected provider.

Description

Try:

Replace current diagram Uncheck to add the generated nodes to your existing model.

Nodes

Edges

Available Connections in Graph (Not in Database):

Node-to-Node Connection

Output:

View All Connections

Cycle Detection

CLD Signal Propagation Analysis

Analysis Controls

Analysis Results

Behavior Classification

System Summary

Recommendations

Parameter Space

2D Stability Maps

3D Parameter Space

2D slice preview

Fan Chart

Parameter Optimization

Controlled Structural Edits

New edit

Pending transaction 0

Team Workspace

Balance the current model

Teams 0

Move order 0

Agent approach

Session result

Balance preview

Team scores

Learning result

Move log

Researcher Access Management

Diagram Generator