Consolidated disclosure of the Environmental Intelligence Layer accuracy program · Batch release · Companion to the Validation Results report
Four rules governed every change:
1. Real data first. Every constant, threshold, and default now traces to an authoritative published source. Reasoned guesses were removed or explicitly labeled as engineering assumptions — never presented as data.
2. Never guess geography. The pipeline previously carried regional lookup shortcuts that defaulted unresolved locations to specific states. Location is now resolved once, authoritatively, per site; when resolution fails the affected output is labeled degraded — it is never silently assigned another state's values.
3. Honest failure. When an upstream source is unavailable, the affected fields now carry an explicit estimate/degraded status. A failed lookup can no longer masquerade as a successful determination.
4. Continental generality. Every fix applies identically at any coordinate in the lower 48; a standing 48-state verification suite runs before every release.
| Analysis area | What was wrong | What changed | Typical effect |
|---|---|---|---|
| Rainfall erosivity & erosion design | A single state's erosivity baseline was silently applied at roughly 80% of sites; a synthetic latitude-band value participated in the precipitation average. | State resolved authoritatively per site; synthetic input removed from every average. | Erosivity corrected at previously mis-assigned sites (up to ~6× in the Northwest); arid-site precipitation no longer overstated; erosion-control recommendations re-rated accordingly. |
| Design-storm precipitation | State-average storm coefficients were presented with point-query provenance and high confidence. | Real per-site federal precipitation-frequency point estimates adopted for 46 of 48 states; the two Pacific-Northwest states outside that product's coverage are labeled as estimates with the reason stated. | Design storms are now genuinely site-specific — for example, two same-state desert sites that previously showed identical values now correctly differ. Storm classes moved up at convective sites and down at desert sites. |
| Hail expected annual loss | Every historical hail event was charged the break probability of the single largest stone on record — overstating expected loss 2–8×. | Loss now sums each event's own fragility across the full historical size distribution. | EAL corrected downward at every site — most at hail-belt sites. Relative site ranking preserved. |
| Flood determination | A failed or rejected flood-map query silently reported "minimal-risk zone / high confidence"; an invented coastal zone could be substituted on error; an uncited regional multiplier (up to 2.1×) scaled the flood score. | Three-way honesty: mapped-and-clear, genuinely unmapped ("no federal flood map exists here"), or unavailable — never a fabricated zone. The multiplier no longer touches the score. | First verified positive floodplain determinations (coastal high-hazard zone with base flood elevation); scores re-based on evidence only — some rose (a mapped floodplain site was previously suppressed by the regional factor), some fell. |
| Lightning | A ten-box regional table with an ordering defect under-read the desert Southwest ~2×; total lightning was also used where a ground-strike rate belongs. | Full-resolution satellite optical lightning climatology bundled per-site; ground-flash conversion applied with a published continental ratio before any strike-frequency math. | Flash density now continuous per site; strike-frequency and risk levels re-based (generally one tier lower — the old figures double-counted in-cloud flashes as ground strikes). |
| Wind design loads | Published ultimate design speeds were multiplied by a return-period factor a second time, inflating every design pressure ~39% (up to ~90% inland of hurricane coasts, where an immediate-coast speed applied hundreds of km inland). | Map speeds used as the ultimate values they are; hurricane speeds decay with real coastal distance; the site's measured multi-year wind record now participates — design speed is the greater of the code map and the measured extreme-value estimate. | Design pressures reduced ~28% at most sites (to the code-book values); measured winds can only raise the basis, never lower it below code. Prior figures were conservative, not unsafe — but they inflated racking and foundation cost estimates. |
| Carbon offset | An uncited base figure adjusted by a latitude rule labeled as a "regional carbon database" that did not exist. | Site's modeled annual generation × its regional grid marginal-emissions rate (the published basis for avoided-emissions from new renewables), resolved per site. | Genuine regional dispersion (roughly 800–1,250 tons CO₂/MW/yr across the portfolio) replacing a near-flat guess; higher where solar displaces a more carbon-intensive marginal mix. |
| Solar performance physics | The bifacial-gain and cell-temperature models were correct but never received their real inputs — a uniform 7.14% gain and a fixed wind assumption applied everywhere. | Per-site satellite surface albedo, real annual temperature, and the site's measured mean wind now feed the models. | Bifacial gains now range ~4.8–8.5% by site; capacity-factor adjustments genuinely site-specific. |
| Frost & winter severity | State assignment by coordinate boxes misfired (one Midwest site read "unknown," one Southwest site read a neighboring state's value). | Authoritative state resolution; plus a new sub-state winter-severity index interpolated from ~5,000 national climate stations, with a published bare-ground frost-penetration conversion emitted beside the code frost line (two distinct, labeled quantities). | Code frost lines corrected at mis-assigned sites; new continental-resolution severity fields. |
| Permitting & jurisdiction | Overlapping state boxes assigned the wrong regulatory regime near borders (e.g., a Nevada site shown under a California program); a wetland "coverage percent" was a feature count in disguise. | Programs keyed by the authoritatively resolved state; wetland reporting states what was actually observed (feature count), with areal coverage marked not-computed. | Correct state programs everywhere; no invented percentages. |
| Risk percentiles | "Percentile" was a coarse fixed-threshold bin, not a rank. | Each site now ranks within the real 60-site continental reference portfolio, with the basis stated in the field's metadata. | Percentiles now mean what the word says. |
| Coverage boundaries | Locations outside the continental U.S. ran the full analysis on continental fallback values and reported success. | Non-continental coordinates are now rejected at the front door with an explicit coverage statement. | No analysis is produced where the underlying data does not apply. |
Beyond value corrections, the release standardizes provenance across all modules: estimate paths report Success_Estimate with a tier and a degraded flag rather than plain success; fields that cannot be computed say so; two per-process randomization defects that made identical sites return different values across runs were removed (outputs are now deterministic); and cost-model base rates are explicitly labeled as engineering assumptions pending citation.
Each change shipped as its own deployment with: an offline test of the changed computation against hand-verified values; a pre-written accepted-delta list; a before/after field-level diff across a fixed 7-site reference harness; targeted live probes of the specific new values; and a rollback point recorded (none was needed). The program closed with a 60-site end-to-end capture across all 48 continental states, physical-gradient checks (lightning, erosivity, and wind monotone along their physical axes), and a new per-module accuracy gate suite whose measured results — including the two open findings it currently flags — are published in the Validation Results report.