Good catch! Even if you assume they were talking about amplitude flatness instead of accuracy, neither the Fluke 5500 scope calibrator option nor the older Tek SG503 / SG5030 leveled sine wave generators have a flatness specification better than 1%.
I also found the leveled sinewave requirements in all service manuals, even the correct ones, are vague. Only a single number in percentage, "amplitude accuracy" is specified, and this is problematic - it makes the requirements much stricter without justifications.
For example, TBS1000C, an entry-level oscilloscope (100 MHz & 200 MHz), specifies "3% amplitude accuracy" and suggests a Fluke 5500A-SC calibrator. But even with its top-of-the-line option, Fluke 5520A-SC1100, it has just an amplitude accuracy of 2% at reference, and 1.5% flatness at 100 MHz, this is already a 3.5% error.
Similar problems exist for 3-series MDO's 4% requirements. 3-series MDO is a middle-level oscilloscope (1 GHz). Even when we use the Fluke 9500B officially recommended by the manual, equipped with the best, top-of-the-line 9530 output head, under the most ideal conditions (VSWR 1.2)... The error is 1.5% at reference and the flatness is 3.0% at 1 GHz, giving an error of 4.54%, 0.54% higher than the official "minimum requirement".
If the calibration lab doesn't use the top-of-the-line option, the error can even be 0.5% or 1% more in addition to that.
Another interesting observation is how the performance of leveled sinewave generators changed over time. A vintage Tektronix SG503 (1974, 250 kHz to 250 MHz) has 3% error at reference and 3% flatness at 100 MHz in all ranges, a total error of 6% (but it's only 4% at X1 range). Now take a look at the modern Fluke 5520A-SC1100, the reference accuracy is 1.5%, and at 100 MHz the flatness is 2%, so the absolute error is 3.5%.
SG504 (1974, 200 MHz - 1 GHz) has 3% error at reference, and 4% flatness across the spectrum, thus the absolute error is 7.12%. For a Fluke 5520A-SC1100, at 1 GHz the flatness is 5% - the absolute error is still 7% - the more things change the more they stay the same. The top-notch Fluke 9500B is better, 3.0% flatness at 1 GHz, 1.2 VSWR, the error is 4.54%.
Conclusions:
1. For leveled sinewave generators, 3.5% to 4% amplitude accuracy at 100 MHz is top-notch. However, even 7% is already pretty good.
2. The "minimum requirement" for leveled sinewave generators in oscilloscope service manuals is poorly specified and safe to ignore, read the datasheets of the recommended equipment instead. Further, if you're a hobbyist with a 20-year-old oscilloscope, and you happen to have access to a 7% sinewave generator, but the service manual calls for a 4% one, just ignore the requirements and do the calibration, your oscilloscope will probably still be better.