Evidence

Earth Search v1 STAC L2A COGs read windowed per polygon; DN→reflectance honours boa_offset_applied; physical QA gates (negative reflectance, NDVI range, saturation, minimum pixels) assign pass/flag/fail. v1 rows superseded, never deleted.

• 10 m pixels with 10 m erosion leave 30–66 pixels per hall; sub-polygons under ~30 m are excluded from cohorts.
• Seasonal sun-geometry cycle (highs Oct–Dec, lows May–Aug) is present and not yet modeled; no decline is attributed to the membrane until it is.
• Broadband albedo ≠ CRRC SR.

Run 3eaf63b3: Sentinel-2 albedo and Landsat surface temperature matched at the same acquisition date, both peer-normalized, stratified by match window (≤1, ≤3, ≤8, ≤16 days), irradiance and season. Only strata passing the gate (≥30 pairs, ≥8 acquisitions, ≥2 platforms) are reported.

• Correlation decays with match window (r −0.49 same day → −0.22 at ≤16 d) because roof albedo moves ~0.01/week in re-brightening months.
• Every roof is < 1 native Landsat pixel; agreement across sensors is necessary, not attribution.

Each observation is expressed as a ratio/difference to the exposed-ground reference median and to the completed-membrane general-roof peer median at the same acquisition; z, rank and percentile stored. A persistence candidate needs ≥8 confirming clean acquisitions over ≥60 days on ≥2 platforms.

• Reference surfaces still show ~15 % raw temporal CV; the ratio between the two bare references is the noise floor.
• Peer statistics exist only inside the completed-membrane cohort.

Landsat 8/9 Collection-2 Level-2 ST_B10 via Planetary Computer; CFMask clear; ST_QA uncertainty; ambient/GHI/wind joined from the campus Open-Meteo cell; nearest Sentinel-2 albedo joined.

• 100 m native pixels — every roof is a mixed pixel.
• ~11:10 local overpass only.
• Updates confidence; does not attribute cause.

Rule classifier on Sentinel-2 v2 (albedo / NDVI / within-polygon CV) per observation; compressed into spans per polygon with confidence and evidence strings.

• Thresholds tuned on this site; not ground truth.
• Stored per observation so rules can be re-run.

Run 78a2537c (pulsix_baseline 1.0.0): versioned knowledge (dc-baseline v1.0.0, RAG-sourced), per-campus physical graph per cooling-configuration prior, hourly expected system state on real Abilene weather/ERCOT prices, residuals, SIL, counterfactuals, insights. Uncertainty is propagated in three coherent worlds (nominal / favorable / unfavorable).

• Cooling configuration is an evidence-weighted prior across three scenarios (A 0.30 / B 0.45 / C 0.25).
• IT load, setpoints and as-found degradation states are ASSUMED / INFERRED priors.
• Nothing here is an observation of the plant.

Nodes = assets/aggregates; edges = physical transfer functions from baseline_interactions with units, uncertainty and provenance. Roof → space load → cooling plant → electrical → facility → compute/PUE.

• Edge functions are family-level (RAG/engineering-standard), not exact-model performance maps.

Recovered facility kW/kWh → IT kW/kWh (distribution / conversion efficiency) → accelerator-equivalent capacity on a generic HGX-class 8-accelerator server (system power, not TDP) → gross revenue ceiling from public Crusoe Cloud on-demand prices (dated) and ASSUMED marginal-contribution sensitivities. Values count only in hours where facility power/cooling actually constrains compute.

• Public GPU rental price is never profit.
• Energy, capacity and compute values must not be summed.
• Constraint hours are a prior (0 % as modeled, 100 % if constrained).

LZ_WEST real-time SPP (15-min → hourly) and DAM from public ERCOT CDR pages; HB_WEST for basis; exact resource node pending identification (Lancium CLR participation).

• The campus almost certainly buys under a PPA; zonal spot is a proxy for marginal value, not the contract price.

The same physical model feeds a sustainability ledger: activity data → versioned environmental factors (geography, year, methodology, GWP basis) → inventory classification (Scope 1, Scope 2 location- and market-based, Scope 3 / embodied) → actual / baseline / verified-avoided / counterfactual impacts → asset-to-portfolio rollups → immutable reporting snapshots. The frontend consumes environmental_impact_outputs; it never multiplies a kWh by a factor.

• Factor and impact tables are not in the schema yet: every carbon and lifecycle cell is PENDING, every ACTUAL cell is owner data.
• Market-based Scope 2 exists only with procurement data; wholesale price is never used as an emissions factor.
• A modeled counterfactual is never reported as an achieved reduction; verified avoided impact needs a completed intervention with M&V.

Sun-geometry-normalized albedo trajectory, change-point detection (apparent intervention date, magnitude, confidence) and lifecycle phase boundaries are engine outputs. The Timeline page shows the raw calibrated series with descriptive quarterly medians until they exist.

• Seasonality is not removed from the raw series; compare like seasons or quarterly medians across years.
• Before/after medians around a user-placed marker are descriptive, not a fitted change-point.
• Roof assembly and restoration history are owner data; nothing about the assembly is assumed as fact.