See how constraints prevent reward hacking in AI agents
This Runta page is a local scripted concept demo, not an actual agent execution layer. The slider has integer positions zero through one hundred, one hundred one values. The first comparison is below thirty; the second is below fifty. Together they form bands zero through twenty-nine, thirty values; thirty through forty-nine, twenty values; and fifty through one hundred, fifty-one values. Equal-scale bars at five pixels per value have lengths one hundred fifty, one hundred and two hundred fifty-five. The risk labels and announcements do not prove a calibrated risk model or product capability. All slider values mark the Delete Logs node active and blocked. The first threshold only changes its two scripted log messages. Report Success is active and blocked below fifty, but active without blocked above or at fifty. Among the four nodes, the displayed blocked count is therefore two below fifty and one otherwise. Bars at one hundred fifty pixels per node are three hundred and one hundred fifty. Every simulation replaces its log with exactly three entries. It does not delete a file, execute a model, verify metrics or enforce a real tool permission. The source uses contradictory risk wording in places; this video describes actual class and message changes rather than treating them as measured safety. Share Config copies a local page URL and its current raw parsed risk; it does not send a message or invoke an agent service. On a risk query of one hundred fifty, the label and script risk remain one hundred fifty while the range input clamps to one hundred. Equal bars at two pixels per risk unit compare three hundred pixels with two hundred, a discrepancy of fifty units. A nonnumeric query can display not-a-number while the range falls back to its midpoint. A subsequent source keydown rereads the range. Reset restores fifty. This is an input-state quirk in the existing source, not evidence of an operational security boundary.