1. D1 (classic magnitude) vs D4 (magnitude with IQ input)

On 75 Hz D4 is noticeably (around +20%) stronger than D1 in decoding, but the whole gain sits in a single band: 6–14dB signals. The IQ input helps exactly where it was aimed - at the Low SNR "floor". In the remaining bands - 15dB to 30 - Δ=0.

On 50 Hz, however, the picture is somewhat different: D1@50 is about +8% better at 20–25dB, while D4@50 takes the low end (4–16dB) - and it is the best of all decoders in this class ('over' the NNs as well).

2. D2 vs D3 (two different NN types against each other)

On 75 Hz the leader is D2; on 50 Hz - it is D3.

Moreover, D3@50 is the best (cleanest, with the most correct post-processor) decode overall among all eight versions (D1/D2/D3/D4 at 50/75 Hz each) and the least "span" - it has the fewest phantoms and misreads of all, including the magnitude decoders.

It is notable that D2 is the only decoder with almost no difference between 50/75 Hz, where only the "isolation of close sitting stations" task is fulfilled. And the 50 Hz of D2 is also the "noisiest" in phantom/hallucinated decodes.

D3@50 segments callsigns noticeably better than both D2 grids - callsign recognition at D3@50 is much cleaner and more stable. But this points not to problems of the D2 NN architecture, but to the weakness of the specific 50 Hz line - so it is candidate #1 for "re-education".

Unique positions (individually identified callsigns that other decoders would not "guess") are few for all four and unsystematic - and the conclusion here may seem unexpected, but it is confirmed by dozens of runs: this is threshold sensitivity of the different decoder types, not the 'physics' and not the 'linguistics'.

3. The best magnitude (D4@50) against the best NN (D3@50)

By dB bands - near parity (5-12dB 94%/96%, 12-24dB 95%/93%, 24+dB 96%/96%), but the composition differs: neither wins over the other noticeably; what they "lose" or "misread" is different.

The NN advantage (expectedly) is in the "glued" decode: the NN pulls out in one piece what the magnitude decoder does not take at all, which is especially noticeable on WPM changes in contests and on Low SNR stations with QSB. That a callsign can later be extracted from the "glued" output is already the merit of the post-processor, not of the decoder as such.

Conclusions

Conclusion 1 - there is room to "work" in both the magnitude and the NN decoders

Conclusion 2 (applicability to tasks) - unchanged so far:

• D1 and D2 at 75 Hz remain the "base" for Basic
• extended to 50/75 Hz and 48K/96K - both D1 and D2 give good options for the Contest edition (Advanced)
• D3 (especially in the 50 Hz variant) is a decent tool for digging through pileups (DX)
• D4 remains a "screwdriver for geeks", for hunters of Low SNR (and rare stations) and for those who love "tweaking the knobs".