In mid-2026 we re-audited our own stack end-to-end: every legacy claim re-tested, every upgrade scored against ground truth, every client deliverable re-verified. Negative results are published below with the same prominence as wins — that is the point.
| Dimension | Before | After remediation | Evidence |
|---|
| Harvest-event detection latency | Monthly composites — median 60 days to confirmation | Two-tier: certified monthly record + 12-day early-warning tier — median 24 days; earlier on 62 of 63 real events | 63 ground-truthed events, identical detector rule |
| Species classification (3-species, 185 plots) | 0.60 balanced accuracy, median/amplitude features | 0.625 with harmonic phenology features (+2.4 pts; hardest class honestly unchanged) | LOO-kNN, byte-identical protocol |
| Radar frequency diversity | C-band only; saturates in mature canopy | + L-band wired (ALOS-2 archive, NISAR ongoing). honest negative no species-ID gain on young stands — documented, not claimed | 185/185-plot L-band coverage |
| Biomass anchor | None (radar proxy only) | GEDI LiDAR wired + correctly scoped: anchors MATURE stands; young rotations sit at LiDAR noise floor scoping finding | 36 dated stand-valid footprint pairs |
| Revisit cadence | ~15-day effective (single platform) | ~11-day and improving as new Sentinel-1 platforms come online; dual-orbit module region-aware | scene census 2024 vs 2026 |
| Ocean color (aquaculture) | Single aging sensor, 8+ months data lag | Multi-sensor merged product, 5+ months fresher — with coastal cross-product disagreement disclosed, resolved by pilot log-calibration | 27 vs 20 monthly values at the farm |
| Dimension | Before | After remediation |
|---|
| Evidence claims | "Quantum kernel ahead in all 3 hardware runs" (single runs, no error bars) | Rigorous re-analysis: ONE statistically significant alignment result (crop-ID, p=0.0005); two legacy edges shown to be noise; no accuracy advantage anywhere once classical baselines were properly tuned. All packs re-labeled accordingly. |
| Benchmark scale | 22-24 samples, 4-qubit generic circuits | 185 real plots, 3 classes, 10-qubit design; competitive-but-not-superior documented across 10 splits; negative results (alignment-training overfits; projected kernels underperform) published |
| Hardware honesty | Some simulator work ambiguously presented (since quarantined) | Hardware-vs-simulation agreement measured on a real QPU: r = 0.998 (preliminary, 12 circuits); budget gate stopped an over-budget run automatically; full certification awaits free research compute |
| The claim we make | Implied advantage | "Real hardware, honest baselines, one significant alignment result, no advantage claimed." The R&D track signals depth — the products stand entirely on classical physics. |