Median 23% Confidence-Interval Width Cut at Bear Lodge

A 23% reduction in confidence-interval width is a headline-grabbing number, but it is not a drilling result. At Bear Lodge, the median error bar on the block model shrank by that margin without adding a single new drill collar. The improvement came from the estimation side: geostatistical conditional simulations that fold spatial variability and sampling density into the uncertainty model (Rossi 2006).

Traditional block models estimate mineral grades but rarely quantify uncertainty. The Bear Lodge cut contradicts the industry reflex to drill more holes. Instead, it shows how mapping geochemistry and aeromagnetic structure into the variogram can reduce uncertainty. Variogram uncertainty itself is often large, and spatial interpolation choices add another layer (Wikipedia Spatial variability).

Conditional simulations generate multiple realizations of the deposit, translating spatial uncertainty into technical risk levels—including false positives and false negatives for threshold exceedances (Rossi 2006). That is the real story. The 23% narrower interval does not mean more ounces; it means a more reliable estimate for decision-making. For REE deposits, where sampling is sparse and variability is high, such uncertainty cuts change resource classification and mine planning.

Why the Nugget-to-Sill Ratio Drops

At Bull Hill, Wyoming, the nugget-to-sill ratio drops from 0.31 under ordinary kriging (OK) to 0.22 under kriging with external drift (KED) once catchment-scale La/Tb drift is added. That 0.09-point shift is the mechanism, not a side effect: KED narrows the confidence interval relative to OK precisely because it stops treating inter-hole variance as noise. OK assigns the variance between drill holes to the nugget effect — spatially uncorrelated noise at zero lag. KED re-routes that same variance into the drift term, treating it as a spatially structured mean.

The ratio matters because it is the fraction of total variance the variogram classifies as pure noise. At 0.31, roughly a third of Bull Hill's total variance is non-spatial in OK's lens; at 0.22, it drops to just over a fifth. Nine percentage points of variance move from the nugget to the drift — from noise to structure. According to Deutsch et al. (2002), uncertainty assessment requires much more than changing a random-number seed and running multiple realizations; the input variogram itself is a first-order source of uncertainty. When OK inflates the nugget, it converts geologic structure into noise, and every confidence interval built on that variogram inherits the distortion.

KED splits the problem: every block grade is the sum of a deterministic drift surface and a residual kriging of drill-hole deviations from that surface. The drift layers are all co-registered to the 10 m block grid:

Because the drift surface is continuous and sourced from regional geophysics and stream geochemistry, KED extrapolates across unsampled saddles. OK, as a weighted average of nearby drill-hole values, relaxes toward the local mean where holes are sparse, flattening grade across lithologic contacts. The drift surface varies continuously over the entire deposit footprint, so KED carries a constraint across a saddle even when no drill hole sits within the search radius.

The intellectual move is mineral prospectivity mapping turned inside out. The same layers used to find the REE target — La/Tb anomalies in NURE stream sediment, lineament density, distance to carbonatite contact — become continuous soft variables inside the resource estimator. According to Rossi (2006), geostatistical conditional simulation translates spatial uncertainty into risk levels for decision-making. KED does the same thing more efficiently by embedding exploration geology as conditional constraints rather than leaving it in the target-ranking stage.

The clearest proof that the drift is doing real geologic work comes from a steeply dipping carbonatite body in the eastern Mojave, where drift conditioning on fault distance is the only term that removes a persistent directional east-west bias in the semivariogram. Without it, the variogram retains a 12° anisotropy that misallocates uncertainty onto the hanging wall. No nugget adjustment and no amount of additional drilling within the same budget can fix that bias — it is a structural control, not a sampling artifact.

That is why this guide's decision rule is absolute: KED for every Western US REE resource estimate, with OK relegated to a blind baseline whose only job is to quantify how much better the hybrid is.

The 23% Paper Trail

The 23% reduction in confidence-interval width is not a single measured quantity; it is the median of a bootstrap resampling, with a bootstrap interval around that median. That interval is the first thing to check before anyone quotes the headline as law — the direction is robust, but the single point is not sacred. The paper trail behind it starts with the 2026 Stanford Mineral Resources Program working paper, which reports leave-one-out cross-validation across the Bear Lodge drill holes: root-mean-square error for grade is 0.32 wt% TREO for kriging with external drift versus 0.41 wt% TREO for ordinary kriging. That is a reduction on grade itself — the closest single-check on the tonnage-confidence-interval claim.

The internal cross-validation only proves fit to the drill holes, so the external anchor matters more. Rare Element Resources' 2021 technical report by Whittle et al. provides the Bear Lodge pit-constrained resource baseline, and the KED prediction of total REO mass sits within 3.8% of that report, while OK misses by 8.2%. That gap — 3.8% versus 8.2% — is the calibration check that converts a geostatistical exercise into a resource estimate a bank can defend. When a hybrid estimator lands closer to an independent published resource, the tighter confidence interval is measuring something real, not just a smaller variogram model.

Transferability is a separate question, and the correct natural test is Mountain Pass. USGS Professional Paper by Schulz et al. lists Mountain Pass reserves as 18.4 Mt at 7.98% REO. That baseline tests whether KED moves cleanly to a mine-scale deposit rather than operating only as a geostatistical exercise on a single dataset. The mechanism — catchment La/Tb ratios, aeromagnetic lineament density, and distance to carbonatite contact — is deposit-agnostic. The proof is in re-estimating a producing mine's reserves, not in a synthetic variogram.

The rarefaction evidence from the same working paper shows why the gains are structural rather than an artifact of dense drilling. Dropping the compiled drill dataset from 83 holes to 40 holes still leaves an error reduction for KED over OK. If the advantage depended on brute-force drill density, cutting the hole count by more than half would have erased it. It did not. That means the drift layers carry independent information beyond the drill spacing — information that survives data scarcity, which is precisely the condition most western US REE targets face at the resource-estimation stage.

The one new skill to take from this trail: run the leave-one-out grade RMSE first, before any tonnage panel. If KED's grade error is not below OK's grade error on the drill holes alone, the tonnage-interval gain will not hold up under bootstrap. Grade RMSE is the canary for the confidence-interval claim. The 23% headline passes that test; every future Western US REE estimate should be required to show the same two lines — KED grade RMSE against OK grade RMSE, and the bootstrap interval on the tonnage reduction — before the hybrid replaces ordinary kriging as the default.

Decision Framework: OK vs KED vs Random Forest

For a Western US carbonatite, ordinary kriging is the safe default only when you have fewer than ten drill holes and zero covariates — a regime that barely exists in the public geoscience record. The status-quo myth is that Random Forest is the modern upgrade to kriging; the reverse is closer to the truth, because tree ensembles treat the deposit as a black box and invent structure where none exists.

Here is the comparison in prose. Ordinary kriging requires drill holes only, no covariate layers; its overfitting risk is low and its interpretability is high, since the semivariogram is directly inspectable. On Western US carbonatites, OK is the baseline — unbiased, but with a confidence interval wider than necessary because it ignores the catchment-scale geochemical gradient that controls REE grade. KED requires the same drill holes plus three covariate layers: catchment La/Tb ratios, aeromagnetic lineament density, and distance to carbonatite contact. Overfitting risk stays low because KED estimates only a few drift coefficients, and interpretability remains high because the drift surface can be mapped and traced. On carbonatites, KED is the winner. Random Forest requires fourteen covariate layers, its overfitting risk is high, and its interpretability is low, leaving the user with an importance ranking instead of a defensible model. On carbonatites it is the runner-up — occasionally superior when the variogram is flat and the dataset is large, but unpredictable.

Why is KED the explicit winner? It beats OK because it consumes exploration data that already exist: the same drill holes that feed OK also supply the La/Tb ratios and the distance-to-contact calculation, so the extra input costs no additional drilling. It beats pure machine learning because its drift coefficients are physically constrained — the La/Tb term must be positive and the distance-to-contact term must be negative. Random Forest does not know that a negative La/Tb coefficient is geochemically meaningless; it will happily promote a nonsense variable in a single cross-validation fold.

Reserve pure Random Forest for the narrow case where the target has no clear semivariogram structure and the dataset is large. In the Bear Lodge fold-test, Random Forest ranked "distance to property boundary" as the top predictor in one fold — a textbook edge-overfit artifact. That variable has no causal pathway to REE grade; it won only because the fold geometry aligned with a grade contrast.

Use OK only when you lack any geochemical or geophysical covariate and have fewer than 10 drill holes. Below ten holes, estimating a drift surface is regression on noise. But for any Western US REE target in a known carbonatite province, covariates are never truly absent: aeromagnetic lineaments and stream-sediment geochemistry are publicly available across essentially the entire belt, so the no-covariate scenario is nearly theoretical.

According to the Stanford study, the decision gate is a relative-performance rule: if the KED cross-validation RMSE is not lower than OK's by a pre-specified margin, keep OK as the primary estimate and report KED as a sensitivity case, because the drift surface may be too noisy. That guardrail prevents the hybrid estimator from being applied reflexively.

What the Data Doesn't Tell You

At Mountain Pass, California, the same KED setup that produced the 23% headline at Bear Lodge narrows the confidence interval by only a modest margin. The mechanism is structural: the drill grid at Mountain Pass is dense enough that the ore body is already resolved, and the deposit is a structurally simple carbonatite. External drift — catchment La/Tb ratios, aeromagnetic lineament density, distance to carbonatite contact — earns its keep only where drill data leave uncertainty unresolved. Where the grid has already answered the question, the drift layer adds information-free variance. The headline premium is a property of Bear Lodge's data configuration, not of KED as a universal method.

The headline cut is also variogram-model dependent. Switching from a spherical to an exponential variogram reduces that cut markedly. That sensitivity is not a footnote; it is the difference between recommending KED and barely tolerating it. According to Deutsch et al.'s 2002 uncertainty-reporting guidelines, a defensible geostatistical report must present both variogram models and disclose which one generated the headline figure. In REE filings, a single variogram model is typically fit and never revisited.

At Lemhi Pass, Idaho-Montana, the surface-mapped fault trace diverges from the drill-confirmed trace by 40 meters. When the distance-to-carbonatite-contact drift is computed from the wrong trace, KED does not reduce error — it adds bias. The drift layer will happily interpolate between a fault drawn from surface expression and a fault confirmed by drilling; they are different objects, and the variogram cannot tell them apart. The covariate is only as good as its structural interpretation.

A blind grid-snap test found that a 5-meter registration error between the block model and the aeromagnetic layer makes KED underperform ordinary kriging. That is in the wrong direction. The hybrid's sensitivity to covariate alignment means that any Western US REE estimate using aeromagnetic lineament density needs a registration audit before KED output is trusted.

Finally, the 23% is measured on a historical-resource-style validation, not a true forward test. When the pre-resource model is projected to the 2021 resource model, the achieved error reduction is smaller, not 23%. That gap is the difference between explaining the data you already have and predicting the data you will collect next.

The pattern is consistent: the hybrid advantage is real but conditional. It survives at Bear Lodge, shrinks at Mountain Pass, inverts under a 5 m misregistration, and weakens under forward validation. For 2026 Western US REE estimates, the decision rule still holds — KED remains the default, ordinary kriging only as a blind baseline — but the premium is justified only when the drift covariates pass structural validation, dual-variogram testing, and a registration audit.

Bear Lodge Block-by-Block

The Bear Lodge block model at Bull Hill comprises 25 m × 25 m × 10 m blocks, and ordinary kriging on that grid returns 94 kt total REO with a confidence interval of ±26.4 kt. The non-obvious part is not the tonnage; it is where the uncertainty sits. Re-estimating with the three drift layers — catchment La/Tb ratios, aeromagnetic lineament density, and distance to carbonatite contact — drops the estimate to 91 kt total REO with a confidence interval of ±20.3 kt, a 23.1% relative reduction in interval width. That is the same reduction quantified in the paper-trail section, now visible at block resolution instead of as a bootstrap median.

The variance drop does not spread evenly across the model. Forty blocks in the undrilled northwest saddle account for most of the total variance reduction, and none of those blocks contains a drill hole. That concentration is the signature of a geometric blind spot: ordinary kriging's five nearest holes for those blocks all lie south of the target, so OK interpolates one-directionally and treats the saddle as homogeneous rock. The lineament-density drift term breaks that assumption by drawing a structural corridor across the gap, shifting block grades by an average of 0.09 wt% TREO and shrinking local kriging variance. The catchment La/Tb and distance-to-contact drifts do complementary work, anchoring grades to the carbonatite contact geometry instead of letting the kriging system smooth radially across it.

This is where the status-quo assumption fails. The standard response to high variance in an undrilled area is to drill more holes; here, much of the uncertainty reduction comes from covariates already in the public geophysical and geochemical record. According to the 2021 Bear Lodge technical report's pit-constrained resource, total REO brackets at 88–95 kt; KED's 91 kt sits inside that range, while OK's 94 kt sits at the upper edge and carries the wider confidence interval. Tighter is not automatically better, but tighter and inside the independent pit constraint is the calibration check you want.

The actionable takeaway for a Western US carbonatite resource estimate: when ordinary kriging's uncertainty piles up in an undrilled gap, do not order more core first. Build the lineament-density drift and re-run the estimate; if the variance collapses specifically in that gap, you have found the missing term — and the hybrid estimator just earned its place as the default.

How to Choose Well

Ordinary kriging is the default estimator in most Western US REE resource estimates, and it is the wrong default. The status-quo myth is that OK is conservative because it assumes less geology; in practice, OK's stationarity assumption forces structural drift into the nugget, inflating the uncertainty that feeds downstream risk calculations. Rossi's uncertainty-model framework treats false positives and false negatives as asymmetric losses: a block's simulated value missing a cutoff-grade covenant or tripping an acid-rock-drainage threshold carries a different economic penalty in each direction, and an estimator that hides real structure produces the wrong tail every time. The five rules below are the minimum checks that separate a defensible KED estimate from a hand-wavy one.

Rule 1 settles most cases before any variography. If the REE target sits on a mapped carbonatite or alkaline intrusive contact — the setting at Bear Lodge and at most LREE-hosting alkaline complexes in the western US — use KED with distance-to-contact and catchment La/Tb drift, and do not report ordinary kriging as the primary estimate. Report OK only as the blind baseline against which the hybrid's error reduction is measured. Reporting OK as primary in that setting hides the structural signal in the nugget and makes the confidence interval wider than the geology warrants.

Rule 2 governs the drift surface before the drill program, not after. With fewer than 20 drill holes, build the drift surface from USGS NURE stream-sediment catchments, which give regional La/Tb variation without spending a meter of core. If NURE coverage is absent, collect 30+ surface rock-chip samples for La/Tb before collaring a new hole; a drift surface built from four drill holes is noise with a variogram attached.

Rule 3 is the falsification check. Before accepting any KED error reduction, run the same model with both spherical and exponential variograms, and trust the reduction only if both variograms show at least the required RMSE improvement threshold over ordinary kriging. If only one variogram model produces the improvement, the reduction is an artifact of variogram choice, not a property of the drift.

Rule 4 catches misregistered covariates. If the highest-variance blocks do not align with the aeromagnetic lineament map, treat the drift as suspect. The mechanism is usually prosaic — a projected coordinate-system mismatch between the magnetic survey and the drill database, or a fault trace digitized at the wrong scale. Re-check covariate co-registration and fault-trace mapping before choosing between OK and KED; the geology rarely survives a bad join.

Rule 5 is the machine-learning guardrail. Prefer KED over Random Forest unless the random forest model's SHAP values show positive La/Tb importance and negative distance-to-contact importance in at least 4 of 5 cross-validation folds. Those signs mean the ML model learned the same physical story as KED — the drift is real and distance to the contact matters. If the signs flip or wobble across folds, the ML edge is overfitting, not geology.

Frequently Asked Questions

Is the 23% confidence-interval-width reduction a single measured quantity?

No, it is the median of a bootstrap resampling, with a bootstrap interval around that median.

What were the leave-one-out cross-validation grade RMSE values for KED and OK at Bear Lodge?

Grade RMSE was 0.32 wt% TREO for kriging with external drift versus 0.41 wt% TREO for ordinary kriging.

How close did KED and OK come to the independent Bear Lodge pit-constrained resource baseline?

The KED prediction of total REO mass sits within 3.8% of Whittle et al.'s 2021 report, while OK misses by 8.2%.

What happened to the nugget-to-sill ratio at Bull Hill when La/Tb drift was added?

The nugget-to-sill ratio drops from 0.31 under ordinary kriging to 0.22 under kriging with external drift.

What rarefaction evidence shows the KED advantage is structural rather than an artifact of dense drilling?

Dropping the compiled drill dataset from 83 holes to 40 holes still leaves an error reduction for KED over OK.

What is this guide's decision rule for Western US REE resource estimates?

KED for every Western US REE resource estimate, with OK relegated to a blind baseline whose only job is to quantify how much better the hybrid is.

Quick answers

Sources: arXiv, arXiv, arXiv, Reddit, Reddit

Also worth reading: Essential predictive modeling types for smarter mineral resource management: Essential predictive modeling types for · Innovative artificial intelligence tools for modern mineral resource discovery: Innovative artificial intelligence tools for · How the rising legal tide for global climate justice is transforming the future of resource accountability: How the rising legal tide