Analysis after the jump.
Angie has dropped from her initial rating of 96% to make the Top 3 to under 80% now. This roughly matches up with the judges comments and the fact that she was not in the top 2 this week. Candice, on the strength of two great performances, regained second place from Kree, though Kree herself gained 5 points. Amber is off her peak of 57%, down 4 points, which makes sense given that she was in the bottom 2.
The Top 3 Tracker in this round (nominally round 8, so that it matches up with previous Top 5 years) makes use of an averaging scheme to provide a less noisy picture of the contest. For historical context, here are the average round 6 to round 8 not-safe probabilities plotted against whether or not the person made the Top 3 (if the dot is at the top (1) then he or she did make the Top 3, and if it’s near the bottom (0) he or she did not. Some vertical jitter has been introduced to make the dots visible.).
Hover your cursor over the points to see which past contestants they correspond to. Now that we have only 5 contestants, a majority of them made the Top 3. The cutoff for making it is about 0.45. The only person to ever make it with a score above that was Syesha Mercado. Janelle is above this cutoff, meaning it would be surprising if she made it.
The cutoff for not making the Top 3 is about 0.35. The only people not to make it with a score better than that were Chris Daughtry, Jason Castro, and James Durbin. Amber is not below this cutoff, meaning that it would not be surprising if she were not to make it, though it would not be surprising if she made it either; 8 people have made the Top 3 with scores worse than hers.
The contest continues to look like the Top 3 will be Angie, Candice, and Kree. Amber and Janelle have failed to make significant gains on these three, and even if Angie continued to fall at her current pace, it would still not be enough for them to be favorites.
*Correction: This post originally had the labels for Kree and Candice swapped on the lead chart. The mistake has been corrected.