When 40% of a panel get bored at once: read against the upper quartile, not the median

Comparing each answer time with the panel's 75th percentile, not its median, lets a change-point check find a large group who all speed up at once.

Earlier today Quick Pick Survey gained a check for the respondent who gets bored: someone who answers the first pairs with care, then clicks through the rest. For each person it finds the point where their answer times drop the most. It reads every answer against the whole panel's time at the same point in the session, so the slow first pairs and the general speed-up that everyone shows are not mistaken for boredom.

That write-up ended on a weak spot. When a large share of the panel (40%) turns quick at the same moment, and only moderately quick (0.6x the honest pace), the panel's median at that point moves down with them. Measured against a curve that has moved with them, they look normal. Even after three passes that rebuild the curve without the people already found, the check found only about a third to a half of them.

The fix is one number

The panel curve is now the 75th percentile of everyone's log answer time at each position, not the median.

The reason this costs nothing is that a change point compares the average before a split with the average after it. Shifting the whole curve up by a constant changes neither. Only the curve's shape across the session matters. When nobody is careless, the 75th percentile has the same shape as the median.

When a share c of the panel sits well below everyone else, the 75th percentile of the mixture lands near the (0.75 - c)/(1 - c) percentile of the honest people. At c = 0.4 that is roughly the honest median, which is the curve we wanted all along.

What the simulation says

There were 133 simulated cells: careless people who pick by a shared wrong reason or at random, 25% or 40% of the panel, turning a quarter, half or three quarters of the way in. They were run with and without a shared warm-up, with and without a learning curve, and with turning points spread out or all at once, plus cost cells with nobody careless and with a genuine minority taste.

casemedian curve75th percentile
40% turn at 0.6x, same point: careless people found31-60%72-80%
same, harm removed (share of the best possible)34-90%76-110%
same, turning points spread over +-20% of the session63-85%89-90%
25% turn at 0.6x, harm removed81-101%78-107%
any share at 0.4x, harm removed95-104%96-109%
nobody careless, or a real minority taste (Spearman cost)-0.0003 to +0.0001-0.0003 to 0

At most 0.2% of honest people were flagged under either curve. Other percentiles were tried too. The 60th lags at 40% careless. The 80th and 90th lose ground when only a quarter are careless. The 70th and 75th tie, and the 75th ships because it holds up better if the careless share is larger still.

What it still misses

About a fifth of the people who turn at 0.6x pace in a 40%-careless panel are still missed: those whose own speed-up lands closest to the cut-off. It also cannot see careless answers that are no faster than honest ones. No timing check can. The view stays a comparison (?drop=fast-blocks and the "Rank without answers after someone sped up" button), not the default ranking, until a real panel's answer times show whether people really get bored like this, and how fast they get.

The data, the sweep options (--curve-q, --onset-spread) and the full table are in the venture's docs/INFORMATION-GAIN-SELECTOR.md, section "A panel curve the careless cannot drag down", with every simulated cell in docs/fast-blocks-curve-q-2026-11-13.jsonl.

These figures come from simulations under stated assumptions about answer times, not from real respondents.

All writing · More in Tools and projects · More on Quick pick · All tools and research