Remove internal metrics and reconcile benchmark numbers

README.md

@@ -4,26 +4,26 @@ Open neuromorphic processor architectures with full software stacks. Three gener
 
 ## Processors
 
-**Catalyst N3** — 128-core neuromorphic SoC with hybrid ANN/SNN computing, hardware virtualization (NeurOS), and on-chip continual learning. 524K–1M neurons, 68 architectural features, 16 per-tile learning accelerators, 4-level memory hierarchy. FPGA validated (19/19 tests). *Paper and benchmarks coming soon.*
+**Catalyst N3** — 128-core neuromorphic SoC with hybrid ANN/SNN computing, hardware virtualization (NeurOS), and on-chip continual learning. 524K neurons, 16 per-tile learning accelerators, 4-level memory hierarchy. *Paper and benchmarks coming soon.*
 
-**Catalyst N2** — Programmable neuron engine with Loihi 2 feature parity. 5 neuron models (LIF, Izhikevich, ALIF, Sigma-Delta, Resonate), 1-16 bit weights, reward-modulated three-factor learning.
+**Catalyst N2** — Programmable neuron engine with Loihi-class feature set. 5 neuron models (LIF, Izhikevich, ALIF, Sigma-Delta, Resonate), 1-16 bit weights, reward-modulated three-factor learning.
 
-**Catalyst N1** — 128-core neuromorphic processor with Loihi 1 feature parity. 131K neurons, programmable STDP learning engine, FPGA-validated with zero failures across 25 RTL testbenches.
+**Catalyst N1** — 128-core neuromorphic processor. 131K neurons, programmable STDP learning engine, FPGA validated.
 
 ## Benchmarks
 
-| Benchmark | Accuracy | vs Competition |
-|---|---|---|
-| SHD | **90.7%** | Beats Loihi 1 (89.0%) |
-| SSC | **73.5%** | Beats Loihi 2 (69.8%) |
-| N-MNIST | **99.2%** | — |
-| GSC KWS | **88.0%** | — |
+| Benchmark | Accuracy |
+|---|---|
+| SHD | 85.9% |
+| SSC | 76.4% |
+| N-MNIST | 99.2% |
+| GSC KWS | 88.0% |
 
 Full N3 benchmark suite with precision sweeps and hardware feature demonstrations — coming soon.
 
 ## SDK
 
-3,091 SDK tests | 1,011+ RTL tests | CPU, GPU, and FPGA backends
+CPU, GPU, and FPGA backends.
 
 ```bash
 pip install catalyst-cloud