Fitting Annex K to WhatsApp's returned JPEGs leaves a 37% mean residual

A quantization table normally identifies the encoder and its quality setting. Most software takes the Annex K reference table from the JPEG standard, scales it by one number and writes the result into the file, so reading the table back usually tells you what produced the image ^[2]. That inference only holds when the fit is close. For WhatsApp's standard send, the per-coefficient ratios against scaled Annex K span about six to one ^[12], the average step size corresponds to something near quality 65, and the roll-off toward high frequencies is far steeper than Annex K's ^[5].

A hardcoded table is still a build fingerprint. WhatsApp HD is a different table again, mean 12.2 against 35.6 for the standard send ^[10][3], roughly 2.9 times finer per coefficient ^[11].

Payload bits live in the 8x8 DCT blocks JPEG is built from. When a platform resamples the image, the block grid shifts, the decoder starts reading across boundaries, and the data is destroyed ^[17]. Every catastrophic failure in the measurements landed at roughly a 50% bit error rate, which is what guessing produces ^[18].

Instagram upscaled a 1080-wide upload to 1440 and padded a 4:3 image to square, so there is no size small enough to be safe ^[13]. WhatsApp caps the long edge, and a portrait photo came back 1200x1600 ^[14]. It also applies the EXIF orientation flag to the pixels and strips the tag: a 3840x2160 file sent with orientation=6 returned as 2160x3840 with no orientation tag ^[15]. For an ordinary photo that is correct and invisible; for a watermark or a forensic pixel comparison, a silent 90-degree rotation is total loss ^[16].

All of this is four production upload paths measured on a single handset over a month ^[1]. The dev.to post does not include app build numbers or the dates of the sends ^[22]. Byte-identical output across dozens of files shows the table is compiled into the build that answered, not that the next release keeps it. For the tables and the survival numbers to transfer to your pipeline, you need the same app builds and the same send mode, and standard send and HD are already two different tables inside one app ^[10].