Fifteen ways a deep foundation job goes wrong, ranked by cost

A geotechnical boring log scores the ground inside the hole. The failures that actually put earth-support contractors out of business are usually about what sits next to it: the building three metres away, a legacy mine working nobody remembered, a plume upgradient, or a dewatering drawdown that reached further than anyone drew on the plan.

That gap is the reason we built a risk register instead of a risk score. This article publishes all fifteen hazards we screen for, the arithmetic behind each one, the public dataset that feeds it, and, for the ones we cannot yet fully score, the dataset we are still missing. You should be able to disagree with any number in here on the evidence, which is the point of printing them.

The hole is not the hazard

There is a quiet bias in most subsurface software. Tools tend to score what is convenient to score. Blow counts, refusal depth, and water table all arrive as tidy numbers on a log, so they get modelled carefully. Adjacent-structure settlement does not arrive as a number at all, so it tends to get a paragraph of boilerplate and no score.

The trouble is that the tidy hazards are mostly cheap. A caving hole costs you casing and a day. The untidy ones are the expensive category. Undermining a neighbour is a claim, an injunction, and occasionally a company. Ranking hazards by how measurable they are gets the priority order almost exactly backwards.

So the register is ordered by expected loss, meaning severity multiplied by frequency, and it deliberately includes high-loss hazards we can only partly score. Leaving them out would understate precisely the failures that cause the largest losses. A hazard with no data still appears, marked as pending, with the source it needs named.

The register

Fifteen hazards. Eleven of them are constructability hazards that cost job-days and feed the composite index. Four are durability hazards that cost service life and are reported on a separate axis. Weight is an expected-loss weight, not a confidence value and not a measure of how well we can compute it.

Two axes, and why corrosion is not in the index

Corrosion is a serious hazard. It is also the wrong kind of hazard to fold into a construction risk score, because it never costs a job-day. It costs you the element at year forty. Averaging a forty-year durability question together with a next-Tuesday constructability question produces a number that answers neither.

Corrosion, seismic and liquefaction, frost heave, and karst therefore sit on a durability axis, carry a weight of exactly zero, and are additionally excluded from the composite by an axis filter. Two independent mechanisms keep them out, which is deliberate: a weight can be edited by accident, a filter is harder to trip over. They are still scored, still shown, and still matter. They just do not move a number that is meant to price construction.

What the curves actually say

Every scoring curve is piecewise linear between published knots, and held flat outside the end knots. There is no fitted surface and no learned model in this layer. That is a design choice rather than a limitation: an estimator can reproduce any of these by hand, and a reviewer can argue with a knot in a way they cannot argue with a learned weight.

Groundwater is the clearest example. Water at grade is close to the worst case the curve admits, and the penalty falls away steeply through the first ten feet, then flattens.

Hole caving carries the more interesting piece of physics. The score starts from the blow count of the loosest sand layer, then gets multiplied by 0.45 if that layer sits above the water table. A layer counts as saturated if its top is within two feet of the water table, which is a capillary-fringe allowance rather than a rounding convenience.

The practical reading is blunt. Crossing the water table roughly doubles caving hazard for the same soil. That single multiplier is the difference between an open hole and a decision to case or run slurry, and it is why the water table matters twice in this register: once on its own line, and again through everything it saturates.

Dewatering: the choice that decides who else is affected

Dewatering method is usually discussed as a cost line. In hazard terms it does something more consequential: it sets how far your excavation reaches. The engine treats method as two separate multipliers, and they do different jobs.

The first scales drawdown severity, from 1.0 for no active dewatering up to 1.35 for deep wells. The second multiplies cut depth to give a zone of influence, from 1.5 times the depth up to 4 times for deep wells. The second is the one that surprises people. A twenty foot cut with deep wells has an eighty foot influence radius, which routinely puts a neighbour inside your scope who was never mentioned in the geotechnical report.

Note what the explorer does not do. It does not predict settlement in inches, and it does not tell you the wall will hold. It produces a screening number for the period before borings exist, when you are deciding whether to bid the job and what to carry against it. Every output carries the same instruction to confirm with site-specific borings, because that instruction is the honest one.

The mine that was not a mine

Legacy underground workings are a genuine high-loss hazard, so we search the USGS Mineral Resources Data System within eight kilometres of the site. Eight kilometres sounds generous until you consider that a worked underground deposit undermines a wide area, and subsidence above one continues for decades.

The interesting part is not the search radius. It is the scoring discipline, because a naive implementation of this produces alarming nonsense. Records in MRDS carry a name, a development status, and a commodity code, and those three fields disagree with each other constantly.

So the ladder reads the commodity code first, the development status second, and the name last. A record named as a mine whose commodity code is SDG, sand and gravel, is a sand pit. It scores 18, not 88, because you cannot fall into a hole that was never dug underground. A processing facility scores 3, on the reasoning that nobody extracted ground there at all; it is a contamination question, not a subsidence one. A worked deposit that really is underground, or that carries a void-prone commodity such as coal or a base metal, scores 88.

Two more rules keep the number honest. Distance decay is applied per record, so a worked shaft a quarter mile away retains 92 percent of its severity while one five miles out retains 10 percent. And the roll-up takes the worst single record rather than a sum, because one close underground working dominates a cluster of distant pits. A density bump exists, capped at ten points, but it counts only genuine void candidates. Twenty gravel pits cannot add up to a mine.

A measurement can raise the hazard, never lower it

Where satellite radar interferometry is available, we fold measured ground displacement into the adjacency and mine-void rows. The fusion rule is one line and it is worth stating precisely.

fused = max(priorScore, subsidenceSeverity)

Maximum, not mean. If the ground is measurably moving, that raises the hazard. If it is measurably stable, that does not lower it, because a quiet two-year radar record is not evidence that a void is safe. It is evidence that it has not collapsed yet. Averaging the two would let a reassuring measurement cancel a real structural concern, which is exactly the failure mode a screening tool must not have.

The same instinct shows up elsewhere in the system. Inferred evidence can only widen an uncertainty band, never narrow it. Rules that can only move one direction are much harder to accidentally misuse than rules that average.

A rule, not a fitted number

The natural-obstruction row carries a twelve point bump when the local bedrock is crystalline, meaning igneous or non-carbonate metamorphic. The mechanism is ordinary geology: crystalline rock weathers to corestones, and glacial transport seeds those corestones through the drift as boulders. A rig meets them as false refusal.

Two things about that bump are worth saying out loud. It is a geological rule, not a coefficient fitted to a dataset, and we label it as such rather than dressing it in false precision. And it is gated on there being till present at all, because corestones live in the drift. Bare rock with no overburden gets no bump. A boulder hazard on a site with no soil to hold boulders would be a nonsense the arithmetic should refuse to produce.

Pending is an answer

Four rows in the register are not fully scored. Adjacency and base stability need your excavation and wall scope before they can finish. Man-made obstructions and uncontrolled fill are scored today from soil survey made-land detection, which is a weak proxy for what we actually want, which is historical mapping and utility records.

Those rows still appear, marked pending or partial, with the dataset they need named. This matters more than it sounds. A register that silently drops what it cannot compute looks more complete than it is, and it teaches the reader that the missing hazards do not exist. A register that says we cannot score this yet, and here is what it would take is less tidy and considerably more useful.

How the index composes

The composite is a weight-normalised mean over the constructability rows that actually have data:

index = sum(weight_i x score_i) / sum(weight_i)   over scored and partial constructability rows

Normalising by the weights that participated, rather than by all eleven, is what lets a site with four scored rows and a site with eleven sit on the same 0 to 100 scale. The alternative, dividing by the full weight set, would quietly reward sites we know less about with lower scores, which is the opposite of what a screening tool should do.

The index reports alongside it how many rows it rested on. A score of 52 from eleven scored rows and a score of 52 from four are not the same claim, and the interface should never let those look alike.

What this does not do

Why publish the thresholds

A screening number that cannot be interrogated is worth very little to a general contractor deciding whether to carry contingency. If we tell you a site scores 68 and cannot tell you which rows drove it, from what data, on what curve, then we have given you a mood rather than information.

So the curves are published, the weights are published, the data sources are named, and the gaps are named too. If you think the groundwater curve is too steep through the first ten feet, or that uncontrolled fill deserves more than a weight of 0.5, that is a conversation we would rather have on the numbers.