What makes a light-curve embedding good? Separate template-matching from genuine per-source period reading.
Probe each frozen encoder to regress log-period, then split the result two ways. Overall R² looks impressive — but it conflates two very different skills. A model can score high just by recognizing the class and predicting that class's mean period (template-matching), without ever reading the period off the individual star. So we also measure ρ_within: the within-class rank correlation between predicted and true period, with the class effect removed. That is the part that requires actually reading absolute time.
Eight encoders + a classical Lomb–Scargle extractor on 25,565 ZTF variables. Long blue bar with a short purple bar = high R² that is mostly template-matching. Only Lomb–Scargle has both bars long. Switch to the MLP probe: R² rises (sharper templates) but ρ_within stays ≤ 0.31.
Every self-supervised encoder lands here: R² of 0.5–0.7, but ρ_within ≤ 0.26. They identify the class and predict its mean period; they don't read the individual source.
Switching to a nonlinear probe raises R² (e.g. BiGRU 0.66 → 0.77) while ρ_within barely moves (≤ 0.31). The period information isn't hiding nonlinearly in the embedding — it largely isn't there.
A direct Lomb–Scargle extractor reaches ρ_within = 0.78 on the same light curves. The within-source period is there to be read; the encoders are under-reading it.