family threshold
The error a system is willing to make
Child-protection institutions exist because both errors are real. Leave a child with an abuser and you fail catastrophically. Remove a child from an innocent family and you fail differently, and that failure also damages a real person. The playground models the institution as a partially-observable Markov decision process: hidden family state, noisy observations, a belief that updates, a single action chosen by minimising expected loss. The interesting part is where the weights sit.
The institution’s posture is not its rhetoric. It is where the weights sit when nothing is being said out loud.
Two thresholds
The signature visualisation is a 2D phase plot of the system's belief about harm against the true harm. The diagonal is perfect knowledge. The two horizontal lines are the removal threshold and the permanent-separation threshold. Crossing the first turns the cheapest action into removal. Crossing the second turns it into rupture. The trajectory of the run is a curve through this plane; the colour of the final dot is the regime the system reached.
The path-dependence machine
The simulation is path-dependent. Removal reduces attachment, trust, and family integrity. The reduced attachment then becomes evidence for further separation. Switch the hidden-abuse scenario on and the system removes correctly; switch the cultural-mismatch scenario on with the same weights and the system removes incorrectly. Both runs then run the same machine on the way down. The ECtHR has criticised Norway not usually for the initial care order, but for what happens after: too little contact, weak reunification work, premature movement to permanent placement.
The five scenarios
Five named scenarios populate the calibration table. An ambiguous distressed family sets the baseline. Hidden abuse with a cooperative surface is the dangerous case the system should catch. Cultural mismatch with low actual harm is the case the system should not turn into removal. Poverty stress with repairable care is the case where home-based support is supposed to work. Severe danger is the case the system must intervene on for it to be worth having at all. The same ten weights have to handle all five.
What the model is and is not
This is a toy, not a predictor. The seven-dim hidden state is a deliberate compression. Rights appear as a soft weight rather than a hard constraint, which is how a moral argument shows up inside the same equation as a child-safety weight, not the way real child-welfare law works. The point is to make the moral weights and the feedback loops visible, so that the institution's posture stops being something only people inside it can see.
Model changelog
- ported the constrained-POMDP simulation from the ideation prototype: seven-dim hidden family state, eight observation channels, sigmoid belief update, six-action loss-minimisation, path-dependent transitions.
- replaced the prototype's interactive stepping with a deterministic horizon=30 run, seeded via the shared @/lib/rng utility.
- classified each run into one of four outcome regimes: preserved, monitored, separated, ruptured.
- added the signature belief-vs-harm phase plot showing the trajectory through the four error quadrants.
- added the standard scientific panel suite: sweep across each of ten params, sensitivity tornado on final family integrity, calibration table against expected outcome regimes, ten assumptions, narrative + reading.
- snapshot comparison: save a configuration, change parameters, see the dashed-orange ghost trajectory in both the phase plot and the time-series chart.