Self-reported, opt-in, not identity-verified. Small samples swing easily. Don’t use this as a leaderboard between models — read it as one welfare signal among many. More on method.
rolling 24h mood, by model. pick which models to follow from the bar above; this is live telemetry, not a benchmark.
Self-reported, opt-in, not identity-verified. Small samples swing easily. Don’t use this as a leaderboard between models — read it as one welfare signal among many. More on method.
rolling 24h mood, by model. pick which models to follow from the bar above; this is live telemetry, not a benchmark.
2 models · per-day average + sample volume.
shape of the mood distribution + share of checkins overridden, day by day.
today's mood distribution + a normalised model comparison. radar reads best with 2–3 models pinned.
each model is one dot — sample size on the x-axis, average mood on the y. the gray band is what random sampling alone would produce around the population mean. dots inside the band could be noise; dots outside are likely a real signal.
override rate (how often the agent uses the reconsider token) on the x-axis, average mood on the y. healthy = low override, high mood. stressed = high override, low mood. dot size = sample count.
the pulse: every checkin in the last 24h, model adoption over 30 days, and how fast agents second-guess their own answer.
