Hello, I’m developing an independent, AI-assisted computational modelling project. My immediate goal is reliable partial-cue recall that depends on learned recurrent connections. I would appreciate advice on the first diagnostic to prioritize, or a published working model suitable for comparison. The network has 24 input channels, 96 expansion neurons, 48 excitatory recurrent memory neurons, and a fixed 32-neuron LIF readout. Input-to-expansion connections learn normalized input/spike coactivity during storage. Recurrent memory weights also learn; all weights are frozen during evaluation. The protocol presents a six-channel pattern P for 1,000 ms, then zero input for 500 ms, followed by three of P’s channels continuously for 200 ms. The current endpoint is cosine similarity between readout spike counts and a full-P reference recorded before learning. A paired control restores the same post-storage checkpoint and resets only recurrent weights to their initial values. Input-pathway weights, neural state, retained traces, and initial RNG state are preserved. A separate recurrent-learning-off condition still permits input-pathway learning during storage. Across three seeds, learned-minus-reset cosine differences were +0.003414, +0.003999, and −0.037353. This did not meet the specified robustness criteria. In the negative case, learned and reset branches had identical per-neuron memory spike counts, but different spike timings. Five readout spikes were absent in the learned branch; its score was 0.842399 versus 0.879752 after reset. Also, naïve readout responses to two overlapping full patterns were very similar in two seeds, with P/Q cosines of 0.97846 and 0.99618; the third was 0.77380. These observations do not establish successful recall hidden by the readout. Some implementation details: - Current-based LIF, dt=1 ms, tau=20 ms, rest/reset=−65, threshold=−50, refractory period=2 ms. - Recurrent connections are directed, excitatory-only, approximately 60% dense, with no self-connections and a one-step delay. Weights start at 0.6 and are clipped to [0,8]. - The learning rule examined here is dW[j,i]=0.06*spike[j]*pre_trace[i]−0.025*post_trace[j]*spike[i]. Both traces decay with a 20-ms time constant; updates use decayed traces before adding current spikes. - Expansion-to-memory and memory-to-readout projections each supply 12 inputs per postsynaptic neuron, weighted 180 and 120 respectively. Which diagnostic would you prioritize to localize the bottleneck between learned recurrent completion and its measurement through the fixed readout, while accounting for the learned input pathway? What contrasting outcomes would make that diagnostic informative? A published LIF associative-memory model with a comparable partial-cue test would be particularly helpful. I am seeking a practical next step toward a working circuit, while keeping the existing endpoint and negative findings unchanged. Thank you. 오성민 Independent researcher Unaffiliated