Combine: Evidence Log

2 November 2026: Evidence sprint begins

We started a six-week evidence sprint, organised around Stage 1 of our proposed EDA flow: set energy, latency, memory and accuracy targets from a measured system baseline, using system models and workload traces. From today, every change in the evidence is published on this page the same day.

Two commitments frame the sprint:

• Simulation first, honestly labelled. Models are published with their assumptions, so any reviewer can rerun them and disagree with us precisely.
• Ten small-lab interviews. We are interviewing ten independent builders and small LLM labs about compute access, cloud cost, power reliability and data privacy. Interviews are demand evidence and design input. Interviewees are not testers, pilots or users, and we will not describe them that way.

2 November 2026: Energy model v1

Our first reproducible system model: an hourly, year-long energy-budget simulation of one Combine node (PV array, battery storage, inverter, one accelerator workstation), across 54 scenarios: three solar irradiation levels (4.0, 5.0 and 6.0 kWh/m2/day, following the bands in our NASA POWER based solar atlas), three array sizes (2, 4 and 6 kWp), three storage sizes (5, 10 and 15 kWh usable), and two scheduling strategies for flexible fine-tuning work (a fixed evening start, versus placement in the sunniest hours).

Workload assumption, stated plainly: one 4-hour parameter-efficient fine-tuning session and 8 hours of local inference per day, about 2,442 kWh of load per year. Workstation power figures are assumptions based on a proposed 600 W class accelerator, not measurements. Replacing them with measurements is the next stage of this sprint.

What the model says:

• Sizing: at 5.0 kWh/m2/day, a 2 kWp array with 5 kWh of usable storage already covers about 99.9 percent of the annual load on solar energy in the model. At 4 kWp, more than 40 percent of generated energy has no load to serve. For this workload, a bigger array is wasted capital; growth should go into more workload, not more panels.
• Scheduling: in the tightest case modelled (4.0 kWh/m2/day, 2 kWp, 10 kWh storage), sun-aware placement cuts grid import from 186 to 85 kWh per year, at the cost of 15 deferred fine-tuning sessions the single daily slot cannot catch up. Scheduling trades grid energy for delay, and we report both sides of that trade.
• Storage: moving from 5 to 15 kWh of battery changes grid import only slightly in most scenarios, because the load is daytime-heavy. Storage size should follow measured evening demand, not intuition.

Update, same day: the cross-check below revises the sizing headline to about 97 percent and withdraws the storage finding.

Next for the model: a cross-check against pvlib and the NREL System Advisor Model, monthly seasonality from NASA POWER climatology, and then measured workload traces in place of assumed power figures.

2 November 2026: Model cross-check (Gate G1)

We checked energy model v1 against two validated open solar models, pvlib and the NREL System Advisor Model, fed real hourly weather for three reference locations, one per model scenario: Douala, Accra and Addis Ababa. They were chosen only because their NASA POWER irradiation matches the scenarios. They are reference points, not proposed sites. The code, the weather data and every result are published with this entry, so anyone can rerun it.

What we found:

• The two reference models agree within 2.5 percent, so the check itself is sound.
• Our v1 formula overstates solar yield by 3 to 9 percent on the same weather. It has no term for panel heat, and at these tropical sites cells average 34 to 40 C in daylight, costing 5 to 8 percent of output.
• The weather source matters more than the formula. For the same locations, two satellite datasets (NASA POWER and PVGIS SARAH2) differ by 7 to 20 percent in annual irradiation. Sizing must be tested against both, not one.

What this changes in v1's findings, with real seasons and the same annual energy:

• Sizing: the direction holds, the headline does not. At 5.0 kWh/m2/day, a 2 kWp array with 5 kWh of storage covers about 97 percent of the load, not 99.9 percent. The gap is in the rainy months. A larger array still mostly produces surplus.
• Scheduling: holds, with a sharper trade. In the tightest case, sun-aware placement cuts grid import from 286 to 107 kWh a year, but deferred sessions rise from 15 to 45, because cloudy seasons last weeks.
• Storage: our earlier statement that battery size is a weak lever does not hold in general. Away from the lowest irradiation case, going from 5 to 15 kWh removes most of the remaining seasonal grid import (75 to 29 kWh a year at the Accra reference). We withdraw that finding.

Still open: every workstation power figure is an assumption. Next are model v2 (hourly solar from pvlib, monthly seasons, both weather sources) and then measured workload traces. Full method and results.