[jmxpearson]

Makes NMDA part of the shipped model.

Operating point: G_NMDA = 60, with the fast auditory→E weight lowered to B_SCALE = 10.

Sweep (5 seeds, intro notes on), across g_NMDA / B_SCALE:

setting	duration	extra spikes	latency
30 / 10	56 ms	1.56	33.9 ms
40 / 10	57 ms	1.59	33.8 ms
60 / 10	63 ms	1.66	33.2 ms
40 / 12	59 ms	1.76	—
60 / 12	64 ms	1.82	—
80 / 12	66 ms	1.89	—

60 / 10 lengthens the response and weakens it; at B_SCALE 12, spikes rise above the no-NMDA model.

Ten-seed means vs the current model (no NMDA):

	before	after	bird
duration	45.7 ms	60.6 ± 3.0 ms	90 ± 43
extra spikes / responder	1.79	1.66	~0.8–1.0
peak change	44 Hz	35 Hz	~20
onset latency	31.5 ms	33.9 ± 1.8 ms	23 ± 12 (≤ 35)
reversed/trained	1.175	1.156	> 1

Caveats:

• Latency is now the metric nearest its bound; some evaluation seeds would exceed 35 ms (the test seed gives 32.5).
• Charge ratio: the NMDA:AMPA charge ratio at threshold is ~3.3, on the high side; 40 / 10 (~2.2) is the conservative alternative at 57 ms.

Brian2:

• Implementation: NMDA is a fifth auditory→E synapse object, with per-synapse saturating gating summed through B, plus the Mg²⁺-blocked conductance term on E, using the same mV map as numpy. The guard is removed.
• Parity: numpy vs Brian2 E rates are 33.5 vs 32.2 Hz without NMDA and 39.8 vs 37.9 Hz with it.

Artifacts: retrained ven_model.npz, and a pure-defaults run reproduces it byte-for-byte. Pins (3-SD tolerances), README, MODEL.md, design notes and MANIFEST are updated. 207 tests pass; ruff clean.

finchsim: still pinned to an older spiking-ven and model; upgrading it is a separate PR.

🤖 Generated with Claude Code