Fitz Doud / Compute & energy / Analytical accounting note This is an analytical screen, not a hardware measurement or a full-model energy claim. 1. Sampling-node upper bound 64 channels; capacitance 4 pF each; full-amplitude voltage 2 V. Baseline: conventional resistive charge from zero to V, then reset to zero. The two steps dissipate C*V^2 total; 0.5*C*V^2 is stored before reset. E_cycle = 64 * 4e-12 F * (2 V)^2 = 1.024e-9 J = 1.024 nJ. Stored before reset = 0.512 nJ at full amplitude. Against the SEPARATE reported active-tile estimate of 1.51 microjoules: fraction = 1.024e-9 / 1.51e-6 * 100 = 0.06781457 percent. This ratio compares an analytical full-amplitude charge/reset bound with a separately reported engineering estimate. It is not recoverable savings: recovery circuit overhead, loss, timing and modified circuit behavior remain unmeasured. Real node amplitudes need not all reach full scale. 2. Matched ideal readout schedule (reported analytical screen) Workload: 56 resident 64x64 banks, first 64 outputs of a query projection. Dedicated: 2 banks active x 64 channels; 28 waves x 8 us = 224 us. Shared: 8 banks active x 8 channels; 7 waves x 32 us = 224 us. Dedicated active-tile subtotal = 56 * 2.58 uJ = 144.48 uJ. Shared active-tile subtotal = 56 * 1.51 uJ = 84.56 uJ. Difference = 59.92 uJ per slice. The rounded per-tile values come from the saved architecture-screen note, which attributes them to a separate circuit-study summary. The underlying final circuit deck was not independently audited for this website. The difference is only a nominal allowance for extra omitted shared costs. Neither number is complete accelerator energy. Costs of input drive, routing, holding outputs, waking banks, weight storage and final reduction must be counted. The ideal 224 us excludes those overheads and is neither a measured latency nor a demonstrated application deadline. 3. Projection CSV scope Selected rows from a saved Qwen2.5-Coder-7B numerical tile study. 56 tiles of 64 inputs by 64 outputs. Symmetric 8-bit input quantization. 6-bit tile weights. Calibration: 472 token occurrences from six prompts. Evaluation: 462 occurrences from six further short code prompts. Repeated token IDs may occur across the split. Not a representative benchmark. Reference: floating-point projection on dequantized shipped model weights. Error: RMS error / RMS reference projection before bias, expressed as percent. No added programming error in the selected rows. No end-to-end model score, training measurement, physical ADC measurement or silicon-energy result. Source artifact types: saved results.csv, sample_cap_recovery_screen.py, START-HERE.md and ARCHITECTURE-SCREEN.md. Original local paths and private receipts are intentionally excluded from this public note.