EVIDENCE | placement frontier
This page separates evidence by where it came from. Table 1 is computed by this server, live, on small 1024×1024 crops. Table 2 is full 10,980×10,980 resolution: LIVE rows were computed by this server, RECORDED rows were measured on a laptop and recorded in EVIDENCE.md — they are not recomputed here. Edge rows marked MEASURED were timed on a physical Raspberry Pi 4B (its power draw is still an estimate, not metered); rows marked SIMULATED are modeled from the ground baseline. No row used space-qualified hardware.
1 · Live small-crop runs — computed on this server
7 of 7 workloads measured| Workload | Dataset | DRF | Ground ₹/run | Edge ₹/run | Call | Break-even runs/day | Break-even ₹/GB | Edge timing |
|---|---|---|---|---|---|---|---|---|
| SatNOGS Payload Anomaly Proxy | satnogs | 7.6× | 0.00 | 164.38 | GROUND | 129864.2 | 64,948,068 | MEASURED Raspberry Pi 4B (measured) |
| Sentinel-2 NDVI Summary | GHRCE agricultural region (Nagpur) | 6314.6× | 1.68 | 164.38 | GROUND | 98.1 | 49,125 | MEASURED Raspberry Pi 4B (measured) |
| Cloud Mask | GHRCE agricultural region (Nagpur) | 5750.4× | 1.67 | 164.38 | GROUND | 98.2 | 49,126 | MEASURED Raspberry Pi 4B (measured) |
| Ship Detect | Mumbai/JNPT coastal water | 669.7× | 0.76 | 164.38 | GROUND | 217.1 | 108,586 | MEASURED Raspberry Pi 4B (measured) |
| Sentinel-2 Land-cover Classifier | GHRCE agricultural region (Nagpur) | 3681.8× | 1.67 | 164.38 | GROUND | 98.2 | 49,130 | MEASURED Raspberry Pi 4B (measured) |
| Sentinel-2 Lossless Recompress | GHRCE agricultural region (Nagpur) | 1.1× | 1.68 | 165.96 | GROUND | 1606.4 | 846,739 | MEASURED Raspberry Pi 4B (measured) |
| Sentinel-2 Quicklook Thumbnail | GHRCE agricultural region (Nagpur) | 16.6× | 1.67 | 164.48 | GROUND | 104.5 | 52,259 | MEASURED Raspberry Pi 4B (measured) |
At this crop size (~3 MB) the fixed edge hardware charge (₹164.38/run at 1 run/day) dwarfs any downlink saving, so ground wins for every workload. That is the honest small-payload answer, not a failure.
2 · Full-tile runs — LIVE rows computed here, RECORDED rows measured locally
10,980 × 10,980 px| Workload | Source | Input MB | DRF | Ground ₹/run | Edge ₹/run | Call | Break-even runs/day | Break-even ₹/GB |
|---|---|---|---|---|---|---|---|---|
| Cloud Mask | RECORDED | 384.0 | 653,113.83× | 192.07 | 164.38 | EDGE | 0.86 | 427.90 |
| Sentinel-2 NDVI Summary | RECORDED | 384.0 | 709,853.85× | 192.20 | 164.38 | EDGE | 0.86 | 427.56 |
| Quicklook Thumbnail | RECORDED | 384.0 | 16.48× | 192.09 | 176.03 | EDGE | 0.91 | 455.48 |
| Lossless Recompress | RECORDED | 384.0 | 1.05× | 192.41 | 346.61 | GROUND | 16.15 | 8,377.98 |
| Ship Detect (coastal, 1 band) | RECORDED | 180.4 | 69,580.68× | 90.20 | 164.38 | GROUND | 1.82 | 911.35 |
| Land-cover Classifier | RECORDED | 384.0 | NOT MEASURED — exceeded the local machine's memory at full-tile scale; needs a chunked rewrite or a larger host. | |||||
Edge wins on three workloads at full resolution — and the margin is small (₹16–28 per run). Edge cost is simulated (a fixed hardware charge plus electricity), not measured on real edge silicon.
What the frontier shows
Data-reduction factor alone does not decide placement. Ship Detect has the highest DRF measured (69,580×) and still recommends ground, because its input is only 180 MB: the downlink it saves (~₹90) never clears the flat ₹164.38 edge hardware charge. Edge starts to win when the input is large enough that avoided downlink exceeds that charge (about 329 MB at ₹500/GB) and the output is small. Lossless Recompress, with a ~1× DRF, stays on ground at any scale.
What this does not prove
Edge timing has been measured on one Raspberry Pi 4B only (one run set, power estimated, not metered) — no space-qualified hardware has been measured. The edge hardware cost and 1 run/day duty cycle are assumptions; the result flips with either. Inputs are two Sentinel-2 scenes (agricultural and coastal). There is no customer workload yet.