05 — Evidence

Measured, not promised.
And clearly labelled.

Measured workloads, implementation results and external evaluation for the NeuroCore architecture — with the boundary of each claim stated alongside it.

142 µs adaptation<0.8% forgetting96.5% zero-shot7nm – 45nm flows
Measured hardware result

Directly recorded telemetry on FPGA evaluation hardware.

Implementation-flow validation

Verified RTL synthesis and place-and-route flow.

External evaluation

Verified with an external partner engineering team.

06 — Records

Four verified results

What was run, on what, and what it produced.

Measured hardware result

Real-time on-chip adaptation

Sensory pattern perturbation and error minimisation

142 µs
Adaptation latency
Model
Event-driven spiking recurrent network with hardware-resident e-prop
Platform
Single NeuroCore tile on Xilinx Artix-7 / Zynq-7000 FPGA
Execution
Single-core hardware evaluation running standalone firmware
Throughput
Instantaneous forward-only update cycle
Power and thermal
30–50 °C passive profile, near-zero idle clock dissipation
Energy
Derived estimate: sub-microjoule per adaptation event at active power
Methodology
Direct serial telemetry timestamp recording over UART

Pre-adaptation error rate of 99% reduced to 0% in 142 microseconds upon onset of a localised error signal.

Measured on FPGA emulation hardware. Physical ASIC silicon will achieve lower latency and power depending on process node and PDK selection.

Measured hardware result

Sequential lifelong task retention

50 sequential benchmark classification tasks

<0.8%
Degradation across 50 tasks
Model
Dual-tier consolidation with local plasticity and structural memory replay
Platform
Single NeuroCore tile on a Xilinx FPGA evaluation platform
Execution
Continuous sequential task switching without global re-initialisation
Throughput
Continuous streaming sequence evaluation
Power and thermal
Ambient passive air cooling
Energy
Minimal incremental overhead during consolidation cycles
Methodology
End-to-end task accuracy across all 50 sequential datasets

Task 1 accuracy sustained with under 0.8% degradation after learning 50 consecutive tasks, overcoming catastrophic forgetting.

Demonstrated on standardised synthetic and acoustic temporal benchmark sequences.

Measured hardware result

Zero-shot cross-modal transfer

Acoustic audio to RF pulse-strobe domain transfer

96.5%
Zero-shot accuracy
Model
Abstract temporal compositionality in recurrent state space
Platform
Single NeuroCore tile on a Xilinx FPGA evaluation platform
Execution
Direct sensory swap without gradient updates in the target domain
Latency
0.0 ms — no adaptation latency required
Power and thermal
Identical to baseline active inference
Energy
Zero additional training energy in the target domain
Methodology
Pre-trained acoustic weights evaluated against synthetic RF radar pulse inputs

96.5% immediate classification accuracy in the RF pulse domain using weights learned exclusively in the acoustic domain.

Relies on structural isomorphism between source and target temporal signal domains.

Implementation-flow validation

Standard-cell physical flow synthesis

Full-chip synthesis, placement and timing closure

0.60 mm²
Die area at 7nm FinFET
Model
NeuroCore synthesizable SystemVerilog / Verilog RTL
Platform
Commercial EDA flow across 7nm FinFET, 16/28nm and 45nm CMOS PDKs
Execution
Multi-corner timing analysis across PVT corners
Frequency
Up to 1.2 GHz nominal target at 7nm FinFET
Power and thermal
Standard-cell dynamic and leakage power extraction
Methodology
Synopsys and Cadence implementation scripts with SDC constraints
Result
0.60 mm² at 7nm · 0.85 mm² at 16/28nm · 1.32 mm² at 45nm

Verified RTL-to-GDSII physical implementation flow across four process nodes, with die area closure at each.

Represents verified physical flow validation, not manufactured or packaged ASIC silicon.

07 — External evaluation

Bharat Electronics Limited

An independent team ran it.

NeuroCore underwent an external hardware evaluation with Bharat Electronics Limited, covering FPGA programming, inference execution, status monitoring and reference-output verification.

Stage 1
FPGA programming

Bitstream loading and hardware initialisation on the target evaluation carrier.

Stage 2
Inference execution

Active execution of representative signal filtering and temporal patterns.

Stage 3
Status monitoring

Telemetry verification across hardware interfaces and event queues.

Stage 4
Reference verification

Output comparison against golden architectural reference vectors.

Scope notice: the external evaluation established functional execution on FPGA hardware. It does not constitute certification or a production deployment endorsement.

08 — Serial telemetry

Verbatim UART capture

Raw output from the bench.

Select a recorded run to inspect the telemetry captured during physical FPGA evaluation of the single-core implementation.

UART · 115200 8N1

[BOOT]single NeuroCore tile · Artix-7 · bitstream 0x4A21 loaded

[INIT]resident SRAM weights mapped · 0 external memory cycles

[ERR ]sensory pattern perturbation detected · baseline error 99%

[UPD ]forward-only e-prop update issued across local synapses

[DONE]error 99% → 0% · elapsed 142 µs · peak temp 34.2 °C

Verbatim serial capture · single-tile FPGA evaluation

Request benchmark replication packages.

Test vectors, bitstream configuration guides and simulation harnesses are provided to prospective licensees under an evaluation agreement.