I'm especially interested in feedback on the modeling choices — prevalence curves, specificity sliders, and whether (and how) to add correlated errors or adversarial behavior.
If anyone works in health tech, security, or trust & safety, I'd also love to hear how this maps to your practical experience with low-prevalence detection.
If you test the simulator, I'd love to hear what surprised you most — especially if something behaved differently than you expected. Thanks in advance!
An interactive simulator for visualizing how false positives explode in mass screenings for low-prevalence problems. Relevant across diverse domains including security (e.g., Chat Control; iBorderCtrl), medicine (e.g., mammography; PSA testing), and so much more.
The purpose is to begin to more easily visualize how prevalence, sensitivity, specificity, investigative capacity, and strategic behavior interact to shape real-world outcomes.
A lot of what's changed isn't parenting instincts but the systems kids are embedded in. Traffic speed, liability fears, school policies, and surveillance culture all push toward zero-risk tolerance. When the background structure punishes normal exploration, parents behave accordingly. Freedom and safety aren't individual traits; they're properties of an environment.
One underappreciated challenge in geothermal is simply the distribution of usable rock formations — low-prevalence resources. Better drilling reduces the rarity penalty by making more sites viable. Similar structure to many screening problems: improve detection and you can shift what counts as "reachable."
Medieval life lacked precision measurement: no clocks, no KPIs, no micro-evaluation. Much of modern stress isn't workload but being quantified. That shift from 'work until the task is done' to 'work until the metric is satisfied' changes everything.
The directional‐ablation approach in Heretic is clever: by identifying residual “refusal directions” and ablating them, they shift the trade-off frontier for the model. In rare‐event screening terms: they’re effectively changing the detection threshold geometry rather than trying just to get better data. It resonates with how improving a test’s accuracy in low-prevalence settings often fails unless you address threshold + base rate.
So one thing missing from the excitement around this line of work: how little these worm effects generalize to mammals.
C. elegans has very unusual biology — direct soma→germline communication pathways, minimal nervous systems, and short generational cycles. Epigenetic inheritance is much easier to observe there than in mice or humans, where mechanisms differ and dilution across meiosis tends to erase these “marks.”
This means that, even if the PA14 avoidance effect replicates, it’s not evidence that humans inherit learned behaviours. It’s evidence that worms are an interesting edge-case system.
If an opaque actor can fabricate legal-ish complaints and pressure DNS providers into blocking a site, the system is wide open for abuse. Smaller services without legal teams would just fold.
Curious if others are seeing this kind of “shadow regulation” pop up more frequently elsewhere — especially in email filtering, CDN layers, and AI content moderation.
Grateful for Breyer's consistent excellent work tracking and opposing these proposals.
The Zombie proposal just keeps rising from the dead. The technical/mathematical objections don’t change.
I still haven’t seen a counter-argument stating why a mass-scanning architecture should be expected to work, given the base rates and error rates involved.
If anyone works in health tech, security, or trust & safety, I'd also love to hear how this maps to your practical experience with low-prevalence detection.
If you test the simulator, I'd love to hear what surprised you most — especially if something behaved differently than you expected. Thanks in advance!