Special — Benchmarking the Mine Plan
Both hosts are synthetic voices and the script is AI-generated. Every paper cited was verified against a primary source (DOI or arXiv ID) before publication, but the commentary is not that of practising researchers.
Traci: I want to start with seventeen loaders at a platinum mine in South Africa. Booysendal. Every one of them had more than twenty thousand hours on the clock, and for the second half of 2022 the plan said each one would be available eighty-five percent of the time.
Joao: Okay. And how many of them got there?
Traci: None. Zero out of seventeen. The average was sixty-six percent, the best machine managed eighty-one, the worst fifty-three. And the maintenance bill for those six months came in at more than double the plan. About thirty-four million rand against a budget of about fourteen.
Joao: That's a big miss. But is that a story about one badly run mine, or is it a story about planning?
Traci: That is exactly the question for this whole episode. Because if it's one badly run mine, you shrug and move on. If it's what old loaders do everywhere, then the eighty-five percent was never a target. It was a wish, and anybody reviewing that plan should have been able to say so before the money was spent.
Joao: So this is a special episode. We're not doing the usual machine-learning-meets-resource-estimation beat. We're doing mine plans, and specifically the equipment numbers inside them.
Traci: Right. Availability, utilisation, how reliability falls off with age, what maintenance costs do as a machine gets old, when you should replace it, what drills and loaders actually achieve compared to the brochure, what automation really buys you, and how truck and shovel fleets get sized and scheduled. And the lens for all of it is a book.
Joao: The Flyvbjerg book.
Traci: How Big Things Get Done, by Bent Flyvbjerg and Dan Gardner, 2023. Flyvbjerg is the megaprojects researcher, the one who's spent decades collecting what big projects actually cost and how long they actually took. And the core argument, at least the part we're borrowing, is pretty simple to state. When you estimate a project, you're tempted to build the number up from the inside. You know your project, you know its details, you add up the pieces. He calls that the inside view, and he argues it's systematically optimistic.
Joao: Because everybody thinks their project is special.
Traci: Because everybody thinks their project is special, and because the bottom-up build quietly leaves out all the things that go wrong that you can't see from your desk. The alternative is what he calls reference-class forecasting. You find a class of comparable projects, you look at what they actually achieved, the distribution of outcomes, and you anchor on that. Then, and only then, you adjust for what's genuinely different about yours.
Joao: The outside view.
Traci: The outside view. And I want to be careful here, because we're borrowing his framework, not his numbers. His evidence comes from big projects in general, not from mine fleets, and we're not going to quote his statistics at you as if they were about loaders and drill rigs. What we're asking is narrower. If you take that idea seriously for a mine plan, what's the reference class for a loader's availability? For a drill rig's percussion hours? For how fast maintenance costs climb? And do the numbers in real technical reports sit inside that reference class or outside it?
Joao: And the reference class is what, this pile of papers?
Traci: It's a pile of about forty documents, and I'll be honest about what's in it, because the pile shapes the answers. There's a cluster of theses from Luleå University of Technology in Sweden. Uday Kumar's 1990 doctorate on loaders at LKAB's Kiruna iron ore mine, Anna Gustafson's 2013 doctorate on automated loaders, also at an LKAB mine, and Hussan Al-Chalabi's 2015 doctorate on drill rigs and when to replace them, which used Boliden's mines. There are CIM Bulletin papers from Jacek Paraszczak's group at Université Laval. There are conference papers from the Australasian Institute and the Society for Mining, Metallurgy and Exploration going back to the early eighties. There's vendor material, the Caterpillar Performance Handbook, Epiroc customer stories, a Sandvik conference paper. There's a stack of open-pit fleet scheduling and dispatch papers. And then four real technical reports: Era Dorada in Guatemala, Las Chispas in Mexico, Cariboo Gold in British Columbia, and Platreef in South Africa.
Joao: And some stuff we wanted but couldn't get.
Traci: Some stuff we wanted and haven't read, yes, and I'll flag each one when it comes up. Up front, the list is: Vayenas and Wu, 2009, a maintenance and reliability analysis of a loader fleet in a hard rock mine. Two papers by Erkan Topal and Salih Ramazan, 2010 and 2012, on scheduling truck fleets to minimise maintenance cost. A 2018 paper by Chaowasakoo and colleagues on age-based maintenance for haul trucks. Kesimal and Bascetin, 2000, on replacing off-highway trucks at a Turkish coal mine. A 2022 book chapter by Siva Prasad and colleagues on drill selection. The Sandvik Rock Excavation Handbook. And some theses from the University of Mines and Technology in Ghana. For the papers and the chapter, we've checked the titles, authors and venues against the publishers' records. We haven't read the full text of any of them, so we won't put words in their mouths.
Joao: Okay. Where do you want to start?
Traci: With the most boring-sounding thing in the episode, which turns out to matter the most. What does availability actually mean?
Joao: I mean, I thought I knew. The fraction of time the machine isn't broken.
Traci: Fraction of what time, though? And when does "broken" start and stop? This is where the Laval group's work is so useful. In 2007 and 2008 they published a survey, the conference version first and then a CIM Bulletin version, Paraszczak, Lafontaine and Komljenovic. They sent questionnaires to more than twenty Canadian underground mines and got usable answers from twelve. Mostly base and precious metal mines in eastern Canada, two thousand seven hundred to ten thousand tonnes a day, mostly longhole stoping, fleets of sixteen to sixty-four mobile machines.
Joao: So a real reference class. Twelve mines.
Traci: Twelve mines, all using a computerised maintenance management system, all reporting availability. Every single one of them used availability as a performance measure. And here's what they found when they looked at how it was actually calculated. Almost half of the mines said downtime ends when an operator takes the machine out of the garage and dispatches it to work. Not when the repair is finished.
Joao: Wait, so if the fitters finish the job at two in the morning and nobody picks the loader up until seven...
Traci: Five hours of "down" that's really "available but idle". The paper's point is that this mixes up availability with utilisation, and it overstates downtime. Three mines only recorded downtime by the shift, not the hour. Two of the twelve didn't capture delays at all, so they could only calculate mechanical availability, which ignores waiting time. Five mines were using maintenance person-hours as if they were downtime.
Joao: Which is wrong because two fitters for two hours is four person-hours but only two hours of the machine sitting there.
Traci: Exactly the example the earlier Paraszczak paper uses. And that earlier paper, 2000, CIM Bulletin, has my favourite sentence in the whole pile, more or less. They were studying identical loaders at two Canadian gold mines. At one of them, mine A, downtime was logged by shift, and a machine that came back early in a shift but wasn't picked up by the production crew got recorded as down for the whole shift. They couldn't work out how long machines waited before repairs started. So they abandoned that part of the analysis.
Joao: They just stopped.
Traci: They stopped, because the data would have given biased results. And it's not just Canada. Gustafson's thesis at LKAB notes that actual repair times weren't recorded. The workshop entry and exit times were typed in by hand and included queueing and waiting for parts. Al-Chalabi's drill rig study at Boliden says, based on interviews with the maintenance people, that about eighty percent of the failure data was free of documentation errors. Which is a polite way of saying a fifth of it might not be.
Joao: Okay, but let me push back. I'm a reviewer, I'm reading a feasibility study, and it says eighty-five percent availability. Isn't all this a bit academic? It's a number. I check it against something.
Traci: Against what, though? That's the point. If the study doesn't say whether it's mechanical or operational availability, whether the clock is calendar hours or scheduled hours, and whether delays count, then you can't compare it to anything. Two mines can both report eighty-five and mean genuinely different things. And if the operating mine you're benchmarking against uses the end-of-downtime-when-dispatched convention, its "availability" is biased low, which makes your plan look conservative next to it when it might not be.
Joao: So rule one is basically: make them define it.
Traci: Rule one. Before you argue about an availability number, get the definition in writing. Mechanical or operational. Calendar or scheduled time. When downtime starts and when it ends. Whether waiting for parts and fitters is in or out. And a related test that Paraszczak flags: ask how MTBF was computed. Most maintenance systems compute mean time between failures as working time divided by number of failures, which quietly assumes the failure rate is constant over the machine's life. We'll come back to why that's a problem.
Joao: Alright. Definitions agreed, let's say. What's the actual number for a loader?
Traci: So let me give you the reference class, and I want you to keep a running tally in your head. Start with Paraszczak 2000. At mine B they had two groups of loaders. Type one averaged about thirty-five hundred engine hours, young machines. Type two averaged about fourteen thousand eight hundred. Over twelve months, the young group had uptime of eighty-nine percent of scheduled time. The old group, seventy-six.
Joao: Thirteen points lost to age.
Traci: And look where it went. Corrective maintenance went from six percent of scheduled time to thirteen. Delays went from three percent to nine. Preventive maintenance stayed at two percent in both groups.
Joao: So the old machines broke more and then waited longer.
Traci: Broke more and waited longer. Now go back to Kumar at Kiruna in the late eighties. That's his 1989 paper in the APCOM proceedings, Availability Studies of Load-Haul-Dump Machines, which is reprinted as Paper Three in his thesis. He took a fleet of nineteen diesel loaders, a year of operating and maintenance cards, and he printed the full failure history for one machine. Fifty-one failures. The mean time between failures comes out at about fifty-one hours, and the mean time to repair at about seventeen to eighteen and a half, depending on which distribution you fit.
Joao: So availability of, what, three-quarters?
Traci: If you do the arithmetic on his numbers, roughly seventy-three to seventy-five percent. I want to be clear that's our arithmetic on his printed data, not a headline he states. Then, at LKAB about twenty years later, Gustafson reports that the mine's target was to have nine of its thirteen loaders available at all times. That's sixty-nine percent of the fleet.
Joao: As a target.
Traci: As a target. And then Booysendal, old machines, sixty-six percent achieved against eighty-five planned. And for the optimistic end, there's Matikainen, 1991, an Australasian Institute paper from a Finnish author summarising the Finnish experience. He says the goal of eighty percent availability for loaders can be achieved.
Joao: "Can be achieved." Which is not the same as "is typically achieved".
Traci: No, and you can feel the difference in the language. So the tally runs from the low fifties at the bad end of Booysendal, through mid-sixties for old fleets, seventy-three to seventy-six for Kiruna in the eighties and Paraszczak's old group, up to eighty and high eighties for young, well-supported machines.
Joao: Okay, so let me play the skeptic again. Kiruna in the late eighties, Finland in 1991, Canada in the nineties. That's old. Machines are better now. Isn't the reference class out of date?
Traci: It's a fair hit and I'll concede part of it. The evidence on loaders is thin, it's old, and a lot of it is one or two sites. But two things make me reluctant to throw it out. One, Booysendal is 2022, and it's right in the range. Two, Kumar actually tested the "newer is better" idea directly. He looked at the hydraulic systems on old, middle-aged and new loaders at Kiruna, and he was expecting the newer machines to show reliability growth. They didn't. The failure process looked statistically the same across the age groups.
Joao: Huh. So no measurable design improvement, at least on hydraulics, at one mine, in the eighties.
Traci: With all those qualifiers, yes. So here's rule two as I'd carry it. For an underground loader fleet, plan operational availability in the seventy-five to eighty percent range while the fleet is young. Expect it to slide toward the mid-sixties to low seventies once machines are past fifteen to twenty thousand hours, unless the plan budgets rebuilds or replacement. Anything above about eighty-five percent sustained over a life of mine needs site evidence, not a brochure.
Joao: And the condition is: young fleet, decent maintenance support.
Traci: Young fleet, decent support, and Paraszczak would add: credible PM. One striking thing in his paper is that preventive maintenance was only twenty-five percent of active maintenance time for the young group and thirteen percent for the old group, against a benchmark of eighty percent being cited in the literature at the time. He says run-to-failure was still the dominant policy in underground metal mines. Gustafson at LKAB, a decade later, says corrective maintenance was about ninety percent of the maintenance work. Thunberg's 2016 master's thesis on Boliden's drill rigs found corrective work was sixty to seventy-eight percent of recorded maintenance actions depending on the rig.
Joao: So the planned-maintenance world that the availability numbers assume is mostly not the world these mines were in.
Traci: That's my read of it.
Joao: Okay. Availability is the fraction of time it could work. What about the fraction of time it actually works?
Traci: This is the number I think gets the least scrutiny and moves the plan the most. Let's do drills first. Telfer, the underground gold mine in Western Australia, 2008 paper by Adams, Boniwell and Knights at the Underground Operators' Conference. In 2006 the production drill fleet, five Tamrock Solo rigs, was underperforming. Percussion hours per rig were less than three point six hours in a twelve-hour shift.
Joao: Percussion hours meaning actually hammering rock.
Traci: Actually hammering rock. Less than a third of the shift. They did a proper time-and-motion study, thirty-six full shifts, two hundred and twelve holes, and they built a calendar-hour breakdown. Out of eight thousand seven hundred and sixty hours in a year, the rigs were available eighty-three percent of the time. Gross operating hours, fifty-eight percent. Net operating, fifty-four percent of calendar time. Utilisation of availability, fifty-nine percent.
Joao: So about half the year the rig is doing something drill-like, and much less than that it's actually drilling.
Traci: And it lines up with Boliden. Thunberg's thesis has a table of average daily operating hours on four face rigs, from the rigs' recorded operating times. Percussion time on one boom was between about one point two and three point six hours per day.
Joao: Per day. Not per shift.
Traci: Per day. On loaders, Matikainen in 1991 says that even in favourable conditions the effective use of a loader was four to five hours per shift, and in his conclusion, five to six hours a shift for practical reasons. Paraszczak's mine B had utilisation of about sixty-six percent for the young loaders and under fifty percent for the old ones. And he asks out loud whether the problem is operator scheduling, or whether there are simply too many machines in the mine.
Joao: That's a great line. Is there such a thing as too much availability, in a way? Spare machines that just sit?
Traci: That's the Outokumpu finding Matikainen reports, actually. After a long analysis across several Finnish mines, the conclusion was that the minimum number of loaders gives the best total result. Fewer, standardised machines, better utilised, faster maintenance. So here's rule three. For underground mobile fleets, productive time tends to land around half of calendar time, give or take ten points. Percussion time on drills is a fraction of that again, on the order of a few hours per rig per day unless something structural, like automation through shift change, has been demonstrated on that site. A plan that implies much more than about sixty percent of calendar hours as productive time, for a conventional fleet, is outside what this pile of papers supports.
Joao: And how do the technical reports frame it? Because they don't usually say "fifty percent of calendar time".
Traci: They frame it as effective hours per shift. Platreef's 2022 feasibility study is very explicit, which I appreciate. Eleven and a half hour shift, collar to collar. Take off an hour and a half for shift change and travel, half an hour for lunch, a quarter hour for safety talks and inspections. That leaves nine and a quarter hours. Then they apply an efficiency factor of eighty-eight percent for all the unplanned stuff. Eight point one three effective hours a shift, which is about seventy-one percent of the shift.
Joao: That's for people, though, not machines.
Traci: It's the working time available to the crews. Machine availability gets handled separately, and we'll get to how. Cariboo's feasibility study uses seventy-two percent of a ten-hour shift for development equipment and sixty-five percent for manual loaders and trucks. And the Era Dorada feasibility study from December 2025 gives availability and use-of-availability separately: drills eighty percent available, sixty percent use of availability; loaders and trucks seventy-five and seventy.
Joao: So multiply those.
Traci: Forty-eight percent of calendar time for drills, about fifty-two and a half for loaders and trucks. Which is very close to Telfer's fifty-four percent net operating time. That's a number that sits nicely inside the reference class.
Joao: Okay, I want to go to the brochure question, because I suspect this is where the inside view really lives.
Traci: It does, and here's the twist. The vendors are often more honest than the plans that cite them. Open the Caterpillar Performance Handbook, edition fifty, at the introduction. All of the performance data, quote, is based upon one hundred percent efficiency in operation, end quote, which it says cannot be achieved continuously even under ideal conditions. And it tells you to correct for job efficiency, operator efficiency, material characteristics, haul road conditions and altitude.
Joao: So it's in the fine print.
Traci: It's in the front matter. And the worked example for sizing an excavator multiplies the theoretical cycles by ninety percent for operator skill, ninety-five percent for machine availability, and eighty-three percent for what they call general operational efficiency. That's the fifty-minute hour.
Joao: Fifty minutes of work per sixty-minute hour.
Traci: Right. Their job efficiency table runs from a sixty-minute hour at a hundred percent down to a forty-minute hour at sixty-seven percent. And all three factors multiplied together come to about seventy-one percent of theoretical. Then there's payload. The handbook's bucket fill factors for excavators: well-blasted rock, sixty to seventy-five percent of heaped capacity. Poorly blasted rock, forty to fifty.
Joao: Forty percent. So the bucket is less than half full.
Traci: In poorly blasted rock, by their own table. Now the field evidence. Matikainen describes the Toro CECAM monitoring system in Finnish mines in the late eighties, and it showed that actual bucket filling was twenty to thirty percent less than the machine could carry. And at the Gol-e-Gohar iron ore mine in Iran, a 2015 simulation paper in the Southern African Institute's journal by Dindarloo, Osanloo and Frimpong lists the haul trucks as nominal one-hundred-and-five-tonne units with an actual payload of eighty-three to eighty-seven tonnes.
Joao: So twenty percent light.
Traci: About twenty percent light, at an operating mine. Rule four: nameplate is not achieved. Apply the efficiency haircut, the fifty-minute hour or whatever your site supports, and then apply a separate payload haircut of something like fifteen to thirty percent for blasted rock unless you have payload monitoring data showing otherwise. And notice they're separate haircuts. A plan that applies the fifty-minute hour and then assumes full buckets has only done half the job.
Joao: You said vendors are honest. Is that true of the customer stories too? Because those are marketing.
Traci: No, and that's a good distinction. The handbook is a reference manual written by engineers, and it hedges. The customer stories are marketing, and they don't. Epiroc's 2015 story on the Boomer S1 D at the Sasa mine in North Macedonia quotes the product manager saying face drilling productivity increased by fifty percent. Fifty percent compared to what, measured how, over what period? It doesn't say. Their 2025 story on the Simba E60 S at Boliden's Somincor mine in Portugal says the rig drilled more than eleven thousand metres in its first sixty days, averaging sixty metres an hour, with a daily peak of five hundred and twenty-four metres. No baseline, and "sixty metres an hour" doesn't say per what hour.
Joao: And a daily peak is, by definition, not the plan number.
Traci: A daily peak is the opposite of a plan number. So, inside view. And I'd put the Sandvik conference paper in the same bucket, with more respect. That's Välivaara, MassMin 2016, written by a Sandvik product manager about their DD422i jumbo. It's careful and has real trial data, which we'll get to under automation. But it's the manufacturer reporting on its own product.
Joao: Okay. Maintenance cost. You hinted Booysendal was double the budget. Is there a rule for how maintenance cost climbs with age?
Traci: There's a rule of thumb, and then there's a more important warning about the shape. The rule of thumb is from Damien Hawcroft, 1990, at the AusIMM's MINCOST conference. He was general manager of marketing at Eimco Australia, so a vendor, but writing from their database of operating costs. He says the total cost of maintenance on a typical loader over a twelve-thousand-hour life is about one hundred and fifty percent of the capital cost. That's for good average conditions, good roads, good maintenance, a spare machine for back-up. Excellent maintenance might bring it down to a hundred percent. Poor maintenance, up to two hundred and fifty.
Joao: So you'll spend one and a half times the purchase price keeping it running.
Traci: In his good-average case. And he immediately says the rule breaks if the life isn't twelve thousand hours. One of the operators he surveyed depreciated loaders over twenty-two thousand one hundred hours, and at that life the multiplier is, his words, totally inadequate. Which is a nice example of a rule of thumb that tells you its own conditions.
Joao: And the shape?
Traci: The shape is the warning. Matikainen reports a Finnish study where the ratio of maintenance hours to engine hours went from about point four to about two as loaders aged. Paraszczak's mine B: the young loaders needed point one seven maintenance man-hours per operating hour, the old ones point four two, against a benchmark of point two five to point three being cited for the year 2000. He says that point four two might be a warning sign to start thinking about phasing those machines out.
Joao: So that ratio is a replacement trigger.
Traci: It's a candidate trigger. But here's the subtle part. Cat's handbook, in the owning and operating cost section, says that hourly repair costs for a single machine follow, quote, an upward stairstep pattern, end quote, because major repairs come in spurts. Only when you average over many machines does it smooth into a curve. And then the 2014 paper by Fu, Topal and Erten in Mining Technology, which used maintenance cost data from a Western Australian gold mine for thirty-one haul trucks in six types, shows the cost per hour by five-thousand-hour age band, and it's not monotonic at all. For one truck type it goes from forty dollars an hour when new, to sixty, to ninety-five, and then late in life it's down around thirty.
Joao: Because the big component change-outs happen at particular ages and then you've got fresh components again.
Traci: That's the natural reading. And here's why it matters. Fu and colleagues describe the traditional spreadsheet approach as: use the newest trucks as much as possible, on the assumption that maintenance cost rises with age. And they say that relationship doesn't exist in their data. Their optimised schedule cut ten-year maintenance cost by about twenty-two percent compared to the spreadsheet.
Joao: Big number. Real mine?
Traci: Real mine's cost data, but a model schedule, not a before-and-after. And this is where I have to flag one of the unread papers. Fu's model extends Topal and Ramazan's 2010 model in the European Journal of Operational Research, and according to Fu's description, a 2012 Topal and Ramazan paper added stochastic maintenance costs. We haven't read either, so that's Fu's summary of them, not ours.
Joao: Noted. So rule five?
Traci: Rule five. Lifetime maintenance on a loader is on the order of one to two and a half times its capital cost, with one and a half as a central case over a twelve-thousand-hour life, and that multiplier scales with how long you keep it. Within that, expect cost per hour to step up around component change-outs, not glide. A flat maintenance cost per hour across a long life of mine is an inside-view tell. And if you're tracking maintenance man-hours per operating hour, somewhere north of point three to point four is where the literature starts talking about replacement.
Joao: And Cat had another line, didn't they, about repairing before failure.
Traci: They did, and it connects to Gustafson's ninety percent corrective. Cat says a repair done before a major component fails can cost about a third of a repair after failure, with only a moderate sacrifice in component life. If your fleet runs to failure, you're paying the expensive version of every repair. And one more piece of honesty from the vendor: the same section says actual cost experience on similar machines and applications provides the best basis for estimating hourly repair cost.
Joao: That's literally reference-class forecasting.
Traci: It's the outside view in a dealer manual. Hawcroft says the same thing in 1990: actual mine site results are the best way to derive accurate costs. The people selling the machines are telling you to benchmark.
Joao: Okay, replacement. When do you retire the thing?
Traci: Let me start with the definition from Sherer and Gentry, which was published in Mining Engineering in November 1983, from a 1982 SME preprint. Economic life is the age that gives the lowest total discounted cost per unit of production, not counting future replacements, and it's usually shorter than physical life. And then the optimum replacement age can differ from the economic life, because of the remaining mine life, the rest of the fleet, and capital rationing.
Joao: In 1983. So this is not new thinking.
Traci: Not new at all. And their observation about industry practice is still a good one: few mining organisations had a formal replacement policy, and some left it to the whim of individual managers. They used fleet-wide repair-cost trends with a per-machine offset so a lemon unit gets its own curve. And in their truck fleet example, they compared three methods. A published heuristic from Kennecott came out about ten percent more expensive than the optimal integer-programming solution. A different heuristic got within about one percent, in a tiny fraction of the computing time.
Joao: So getting the method right was worth ten percent.
Traci: In that example. Now the Luleå drill rig work gives us actual numbers. Al-Chalabi and colleagues modelled the economic replacement time for an Atlas Copco Boomer face drilling rig at a Boliden mine. And this is where it gets interesting, because what looks like the same rig's cost data shows up in several papers with different answers. In the 2015 paper in the International Journal of Strategic Engineering Asset Management, with no discount rate, the economic life comes out at ninety-six months, eight years, with a flat range of ninety to a hundred and two months. In the Engineering Economist paper, with discounting, it's a hundred and fifteen months, flat between a hundred and ten and a hundred and twenty-two. And in the 2015 paper in the Journal of Quality in Maintenance Engineering, adding the cost of the backup rig you use when the main rig is down, it's a hundred and four months at one cost per hour for that redundant rig, falling to seventy-six months if the redundant rig costs six times as much.
Joao: So for the same machine, anywhere from about six to nearly ten years depending on modelling choices.
Traci: And the most important variable, in the redundancy paper, was the cost of downtime. When the mine's reference group estimated their actual redundant-rig cost, the answer came out at ninety-six months, eight years. Now here's the caveat that I think is the most Flyvbjerg thing in the whole pile. In the thesis, Al-Chalabi says there's no recorded cost data for that rig before seventy-six months of operation, and that the company began tracking costs in its maintenance system when the maintenance costs started growing.
Joao: Oh. So the early, cheap years weren't observed.
Traci: They were extrapolated with a fitted curve. Which tells you something about how mines actually collect data: you start measuring when it starts hurting. Then there's the 2022 follow-up, Al-Chalabi in Life Cycle Reliability and Safety Engineering, on an Epiroc Simba production drill at Tara mine in Ireland. Ten percent discount rate. Economic replacement at eighty-one months, about seven years, with a six-month flat window. The rig averaged two hundred and eighty-nine operating hours a month, so that's about twenty-three thousand four hundred operating hours.
Joao: Better data?
Traci: Forty-two months of real maintenance cost data, extrapolated out to a hundred and twenty months. And, I'll read this carefully, the fuel consumption and the metres drilled per month were generated as random numbers within ranges supplied by the mine's experts. Fifteen to twenty-five litres an hour, three thousand to eight thousand metres a month.
Joao: So the operating cost side is partly synthetic.
Traci: Partly synthetic. I don't think that makes it useless. The answer is in the same range as the Boliden rig. But it means you should treat seven to ten years as a range from a small number of case studies with soft inputs, not a precise number. And the other genuinely useful finding across all of these is that the optimum is flat. Being a few months either side costs you almost nothing.
Joao: What about loaders and trucks?
Traci: For loaders, Matikainen reports the Finnish analysis that put the optimum overhaul at ten to twelve thousand engine hours, with the conclusion that it was economic to overhaul each unit only once and that maximum life was around twenty to twenty-five thousand engine hours. Hawcroft's six Australian operators ranged from a ten-thousand-hour maximum economic life with a minor rebuild at five thousand, up to twenty-two thousand one hundred hours over ten years. At LKAB, Gustafson notes the supplier's plan replaced engine, converter and gearbox at thirteen to fourteen thousand machine hours, the transmission at ten thousand and the hydraulic pump at eight thousand.
Joao: And Caterpillar has a number for that too, I think.
Traci: Careful here, because this one gets misquoted. Cat has a 2024 blog post about the R1700 underground loader. It says they made the rear engine mounts about fifty percent larger in diameter, targeting a life of twelve to fourteen thousand hours, to line up with the planned component replacement schedule for the machine's C13 engine. That's the life of the engine mounts, chosen to match the engine's planned replacement cycle. It is not an engine life.
Joao: But it does tell you the engine's planned replacement is in that neighbourhood.
Traci: It implies the planned component replacement schedule is what they're aligning to, and that matches the LKAB change-out hours nicely. I wouldn't put a tighter claim on it than that. For haul trucks, the Western Australian case in Fu's paper assumed a hundred-thousand-hour total truck life with engine refurbishment after seventy thousand hours. And at the extreme end, an Epiroc magazine story from 2018 describes a Pit Viper blasthole rig at Red Chris in British Columbia with sixty-three thousand drill hours, eighteen thousand of them since a three-week field rebuild.
Joao: Which is a vendor telling a happy story about a rebuild.
Traci: It is, and the account manager says that at twenty thousand hours or so their Pit Vipers are still young. That's an anecdote about a well-maintained outlier, not a benchmark. So rule six. Underground drill rigs: economic life of roughly seven to ten years in the case studies we have, which at Tara worked out at about twenty-three thousand operating hours, with a flat optimum. Underground loaders: major overhaul or component change-outs around ten to fourteen thousand hours, with total lives of twenty to twenty-five thousand hours in the older Finnish evidence. Haul trucks: much longer, around a hundred thousand hours with a mid-life engine refurbishment in the one case we have. And the planning test: any life of mine beyond about ten years that doesn't show at least one rebuild or replacement cycle for the mobile fleet, with timing, has a hole in it.
Joao: Let me jump on reliability. You said earlier that assuming a constant failure rate is a problem. Why should a mine planner care? It feels like a statistician's complaint.
Traci: Because it changes some answers a lot and other answers hardly at all, and you need to know which is which. Kumar's 1989 paper has the cleanest example. For the loader engines, he fits a Weibull distribution, which is a flexible failure-time model where a shape parameter tells you whether failures get more or less likely as the component ages. He also fits the simple exponential, which assumes the failure rate is constant. The mean time between failures barely changes. A hundred and ninety-seven hours either way, within about half a percent.
Joao: So for availability, who cares.
Traci: For availability, he says the error from using the exponential is within zero to five percent. But now ask a different question. What's the probability the engine runs a hundred hours without failing? The Weibull says eighty-six percent. The exponential says sixty percent.
Joao: That's a huge difference.
Traci: About a thirty percent error, his number. So if you're doing availability arithmetic, MTBF and MTTR are fine. If you're planning preventive maintenance intervals or asking how a machine behaves as it ages, you need the shape, and you need to test whether the failures are trending. Kumar's thesis shows what happens when you don't. He took the hydraulic systems, assumed the failures were independent and identically distributed, fitted a Weibull, and got a shape parameter at or below one. That says constant or decreasing failure rate. The machines are getting more reliable!
Joao: Which they weren't.
Traci: Which they weren't. When he tested for trend first and used a model that allows the failure intensity to change over time, a power law process, the hydraulic systems were clearly deteriorating. Same data. Opposite conclusion.
Joao: So what are the actual reliability numbers? How often does this stuff break?
Traci: For whole loaders, Paraszczak's identical twelve-tonne loaders in their first thirty-five hundred hours had a mean time between failures of about thirty-two engine hours at one mine and thirty-six at the other. And the subsystems ranked in exactly the same order at both mines: electrical, frame, engine, hydraulics. He takes that as a hint that there's a general reliability pattern for a given machine type, regardless of site.
Joao: Which is great news for benchmarking.
Traci: It's the best argument in the pile for a reference class. But he immediately adds that MTBF alone is misleading, because the electrical system failed most often but took the least labour to fix. You need failure frequency and repair effort together. For face drilling rigs, Hoseinie, Al-Chalabi and Ghodrati, 2018, in the journal Data, three identical Atlas Copco rigs at a Boliden mine over two years. If you fit the whole machine as one, the characteristic time between failures is about seven to eleven operating hours.
Joao: Seven hours! That's less than a shift.
Traci: Many of these are short repairs. Hoses, the feeder and the rock drill accounted for about sixty-three percent of failures. And Al-Chalabi's Paper One shows the downtime on those components is a reliability problem, meaning frequent failures, rather than a maintainability problem. Thunberg's thesis then went and watched. Worn hoses were being replaced when there was spare time; fixing a hose in production was seen as low priority. The two mines had different tolerances for when a hose should be changed, and different failure frequencies to match.
Joao: So rule seven is something like: MTBF is fine for availability, dangerous for maintenance planning.
Traci: Rule seven. Use MTBF and MTTR for availability arithmetic, where the model choice barely matters. But for anything about ageing, maintenance intervals or replacement, insist on a trend test and a failure model that can show deterioration. Benchmarks to carry: a whole loader failing every thirty to fifty operating hours, and face drill rigs with characteristic times between failures under a shift, dominated by hoses, feeders and rock drills.
Joao: Before we leave reliability, there was a machine learning paper in here, wasn't there? I feel like I'd be failing our usual audience if I didn't ask.
Traci: There was, and I'm going to be a bit grumpy about it. Dindarloo, in the International Journal of Mining, Reclamation and Environment, 2016, uses support vector regression to forecast the time between failures of one loader's engine. Thirty-two failures. Twenty-five for training, seven for testing. Mean absolute percentage error of one point seven percent.
Joao: One point seven percent. On a failure series that goes from six hours to fifteen hundred hours?
Traci: That's the problem. The times between failures in the table range from about six hours to about fifteen hundred. Predicting the next seven of those to within two percent should make you suspicious, not impressed. With seven test points, I'd call it a demonstration that the software runs, not a benchmark. The Luleå team's own artificial neural network paper on replacement time has a similar issue in a gentler way: the network was trained on outputs of their own optimisation model, so ninety-nine percent accuracy means it learned to reproduce the model, not reality.
Joao: Okay. Automation. This is the one where I expect the gap between claims and evidence to be the widest.
Traci: It's wide, but the more careful evidence is actually more interesting than the claims. Start with Gustafson at LKAB. Two loaders, same supplier, both built in 2006. The only difference: one had the automation system fitted, so it could run semi-automatic. Data from 2007 to 2010.
Joao: Close to a controlled experiment.
Traci: About as close as you get in a mine. Tonnes per bucket: nineteen point eight for the semi-automatic machine, twenty point zero for the manual one. Basically identical. Tonnes per engine hour: two hundred and seventy-three semi-automatic, three hundred and fifteen manual.
Joao: The manual one was more productive per hour?
Traci: The semi-automatic machine moved about thirteen percent fewer tonnes per engine hour on average. Operators said manual loading was sometimes faster, and part of the time the automated machine was working in a very long drift. Maintenance cost per tonne was the same for both, about one point five five Swedish kronor. And the transmission was statistically less reliable on the semi-automatic machine. Gustafson points out this contradicts the common claim that automation lowers maintenance and wear.
Joao: So where's the benefit?
Traci: Time. The semi-automatic machine could work very soon after the nightly blast, when nobody can enter the area until it's ventilated. And it was in automatic mode about a third of its operating time. And then here's the finding I'd underline. In one year, one loader had one thousand four hundred and eleven recorded stops causing idle time. Seventy-five percent of those were mining-related: road maintenance, charging, oversize boulders, no ore. Twenty-one and a half percent were the machine. Three and a half percent were the automation system.
Joao: So the robot isn't the problem. The mine is the problem.
Traci: The mine is the binding constraint. Every stop, whatever its cause, interrupts automated operation and needs a person. Telfer's drills tell a similar story. When they did the time study, holes drilled manually had a nineteen percent faster cycle than holes drilled in automatic mode, because a human can start one function while finishing another. The automatic program's first-hole reliability in late 2006 was thirty-nine percent, and it was barely used.
Joao: And then?
Traci: And then they started giving the maintenance department automation defect reports, compiled from a log the operators filled in every shift, before every fortnightly service. By August 2007, first-hole reliability was over seventy-seven percent and second-hole about seventy-four. Fleet production metres went up fifteen percent, and they stood down one of the five rigs. But read the breakdown. Of that fifteen percent, about eleven points came from making all the operators drill with the same optimal parameters. The share of metres drilled automatically between shifts went from two and a half percent to six and a half.
Joao: So most of the gain was just standardising practice.
Traci: Most of the gain was standard work. And one more caution on that paper: the abstract says production metres increased by a further four point one percent from automation, but the body makes clear that four point one is the increase in the automated share of total metres, in percentage points. It's back-calculated. The vendor stories fit the same mechanism with less rigour. Epiroc's 2022 story on Zinkgruvan in Sweden says that drilling by tele-remote over a dedicated LTE network gave about four extra drilling hours a day, because operators don't have to drive down. It also says the drilling was still considered a test and wasn't measured against specific targets. At Éléonore in Quebec, the 2018 Epiroc magazine reports that about nine percent of the metres drilled by three Simba rigs over nine months were drilled tele-remotely through shift changes and blast clearance, and the mine manager credits the rigs with five to ten percent more drill footage.
Joao: Five to ten percent sounds a lot more modest than fifty.
Traci: It does, and it sits better with everything else. Välivaara's Sandvik paper is the most data-rich vendor piece. At DeGrussa in Western Australia the new jumbo had about fifteen percent higher gross penetration rate than the existing fleet's average of a hundred and four metres per percussion hour, longer shank life, and availability within one percent of the old fleet. At Kittilä in Finland they measured an eight percent relative reduction in overbreak. But the operators there preferred semi-automatic mode, sequences kept getting modified at the face, and unattended drilling through shift change was described as something that could be considered once performance was predictable, with practical validation likely in the near future. That was 2016.
Joao: And Matikainen in 1991 was promising...
Traci: Matikainen in 1991 said unmanned loading might increase effective working time by forty to sixty percent, and possibly a hundred. Thirty-five years ago. So rule eight. Credit automation with recovered hours, meaning work through shift change, blast re-entry, and travel time, which in the better-documented cases points to something like five to fifteen percent more output. Don't credit it with higher instantaneous productivity, because the best controlled comparison we have found the opposite. Expect a ramp-up: Telfer's reliability roughly doubled over eight months of deliberate effort. And size the benefit against your mining-related disturbances, because at LKAB those were three-quarters of the stops.
Joao: Let's move to the open-pit side. Trucks, shovels, scheduling.
Traci: The fleet sizing papers converge on one picture, and it's a picture any mine engineer will recognise, but it's nice to see it quantified. Torkamani and Askari-Nasab at the University of Alberta, 2012, simulated a truck-and-shovel system linked to a short-term schedule. With one shovel, going from three trucks to seven, ore delivered rises from about three and a half million tonnes a year to five point four, and then just stops. The shovel is at a hundred percent and truck utilisation has fallen from ninety-six percent to sixty.
Joao: Shovel-limited.
Traci: Shovel-limited. Every extra truck after that joins a queue. The Gol-e-Gohar paper found the same: with two shovels, production maxes out at twenty-two trucks, and three shovels or four gave almost the same output because one would be idle. And Sungun, the copper mine in northwestern Iran, a 2015 paper by Saadatmand Hashemi and Sattarvand. The match factor, which is basically the ratio of truck capacity arriving to loader capacity available, was between point five four and point eight on the different benches. Badly under-trucked. Getting it close to one gave forty percent more production in their simulation. And then adding dispatching, letting trucks go to whichever loader is free instead of fixed assignment, gave another seven point eight percent.
Joao: So matching is worth more than dispatching.
Traci: In that case, five times more. And the Nehring, Topal and Knights review in Mining Technology, 2010, cites reports of seven to twenty percent productivity improvements at thirteen open-pit mines after installing the Modular Mining DISPATCH system. That's second-hand in our pile, but it's a useful range for real installations. And at a limestone quarry in Kenya, Elijah and colleagues, 2021, found the operation running sixteen trucks with two shovels when twelve gave the optimum, with the other four better kept as standby.
Joao: And simulation is the outside view here? Because it's stochastic, it includes variability?
Traci: That's what I want to push on, because simulation can be an inside view in a lab coat. Chaowasakoo and colleagues at Aalto University, 2017, in the International Journal of Mining Science and Technology, used GPS data from an operating open pit to compare dispatching strategies. Over five months, the mine's actual production was between fifty-five and eighty-two percent of its plan.
Joao: So the plan itself was way off.
Traci: The plan was way off reality. Then they simulated the smartest dispatching approach, where you optimise over several trucks and all the shovels at once. The simulation exceeded actual production by ninety to two hundred and seventy percent. Even when they added the equipment downtime they'd measured, the simulated output was almost double the actual.
Joao: Oh, come on.
Traci: And to their credit they say it: in practice one faces problems that aren't in the simulation model. Simulations of some of the simpler one-truck-at-a-time rules, the kind most mines actually use, came close to actual production. So the simulation reproduced reality when it modelled what the mine was doing, and produced fantasy when it modelled what the mine could do. Torkamani's paper also lists equipment breakdowns as future work, meaning the schedule-verification simulation didn't include them. And the Gol-e-Gohar recommendation, about ten percent more production and plus zero point four percent against target, is reported in the paper's table as a simulated mean. The model was validated against real operating data, which is good, but we couldn't tell from the text whether the gain was ever realised.
Joao: So rule nine.
Traci: Rule nine. Get the match factor near one first, then hold spares separately. Expect dispatch to be worth single digits up to about twenty percent in real installations. And treat any simulation as a forecast that still needs an outside-view check: ask whether it was calibrated against actual production, and whether breakdowns and mining disturbances are in it. The test that Chaowasakoo's paper hands you for free: run the simulation of the status quo and check it reproduces actual. If it doesn't, nothing else it says counts.
Joao: And the mixed-integer programming scheduling papers? MIP schedules. What's an honest uplift?
Traci: The Nehring and Topal papers are useful precisely because they review what came before. Trout's 1995 work at Mount Isa, a sublevel stoping copper operation, used mixed integer programming to schedule fifty-five stopes and increased net present value by twenty-three percent over the manual schedule. And the 2008 review notes that schedule was not implemented at the mine. Nehring's own work on a conceptual nine-stope example: zero point six six percent. Their 2010 paper in the Southern African Institute's journal, a conceptual fifty-stope mine: five point four percent over a manual schedule, where the manual schedule took a full working day to build. Topal's work at Kiruna cut the deviation from production targets from ten-to-twenty percent down to six, and that one was implemented.
Joao: So the biggest number wasn't implemented and the implemented one is about hitting targets, not NPV.
Traci: Which is the pattern you'd predict. Rule ten. Credible net present value uplift from optimised scheduling over a competent manual schedule is single digits. Double-digit claims in this literature come from conceptual cases or schedules that weren't adopted. Where MIP has demonstrably paid off in practice, at least in our pile, is in hitting targets more consistently. And a deterministic optimal schedule, which doesn't know about breakdowns, needs the same efficiency haircuts as everything else.
Joao: Okay. Test cases. Let's take a real plan and see if it sits inside all this.
Traci: Era Dorada first, because it's the one we were asked to test directly, and because it turned into a better story than I expected. Era Dorada is Aura Minerals' gold project in Guatemala. The June 2025 technical report summary under the SEC's S-K 1300 rules was an initial assessment prepared by GE21. Section thirteen point eight is unit operations, printed page one sixty-six, and thirteen point nine, page one sixty-seven, is mine equipment.
Joao: So what assumptions does it state?
Traci: For drilling: two-boom electro-hydraulic jumbos, forty-eight millimetre holes four point eight eight metres deep, four point four metres of advance per round. And then one sentence: it is envisioned that one jumbo could drill between two to three rounds per shift. Mucking with a six cubic yard, ten-tonne loader. Thirty-tonne trucks up the decline, with haulage profiles used to calculate equipment hours. And then section thirteen point nine is a fleet table: four trucks at peak, loaders in three sizes, jumbos, longhole drills, bolters, roughly flat from year one to year seventeen.
Joao: And availability?
Traci: Not stated. Utilisation, not stated. Productivity in tonnes per hour or metres per hour, not stated. Mean time between failures, operating hours per year, replacement lives, none of it. Section thirteen eleven says the fifteen-hundred-tonne-a-day rate applied indexes considered appropriate for the mechanisation and productivity, without giving the indexes. And in the cost section, the capital development estimate says manufacturer database equipment usage rates were applied.
Joao: So we literally can't run the test.
Traci: We can't test a number that isn't there. Which is itself the finding: from an outside-view perspective, an initial assessment that sizes a fleet from undisclosed utilisation and manufacturer usage rates is the inside view by construction. There are also some small internal inconsistencies. The text calls the development loader six cubic yards and ten tonnes, while the fleet table lists the ten-tonne loader at four cubic metres, and six cubic yards is about four point six. The fleet is roughly flat for seventeen years. The sustaining capital of a hundred and thirty-six million dollars says it includes renewal of the mining fleet, but there's no timing, so you can't check it against a seven-to-ten-year drill life.
Joao: You said it got better.
Traci: It did, because the December 2025 feasibility study, by Ausenco with Snowden Optiro doing the mine engineering, has a different structure, and its chapter thirteen doesn't have those sections. But section thirteen point five, mobile equipment fleet sizing, has exactly the table we wanted. Drilling fleet: availability eighty percent, use of availability sixty percent. Loaders and trucks: seventy-five and seventy. Auxiliary: seventy and sixty. Loader productivity: eighty-six tonnes an hour for the ten-tonne machine, fifty-nine for the seven-tonne. Thirty-tonne trucks averaging thirty-nine tonnes an hour.
Joao: Okay. So run the test.
Traci: Drill availability eighty percent against Telfer's eighty-three. Inside. Use of availability sixty percent against Telfer's fifty-nine. Inside. Combined forty-eight percent of calendar time against Telfer's fifty-four net operating. Inside, slightly conservative. Loader and truck availability seventy-five percent: that sits in the range for a young fleet, between Kumar's Kiruna and Paraszczak's young loaders, and above the mid-sixties you'd expect once machines pass twenty thousand hours. Use of availability seventy percent against Paraszczak's sixty-six for young loaders. Inside, on the upper side.
Joao: So the feasibility study passes.
Traci: On the factors it states, it sits inside the reference class. With three flags I'd give the reviewer. First, the seventy-five percent loader availability is a young-fleet number. Over an eighteen-year life of mine, it's only defensible if the fleet is renewed on something like the cycles we discussed. Second, the study says the fleet-sizing tables, thirteen nine and thirteen ten, follow from those factors, and in the copy we have, the next page jumps to section thirteen six point two point two. The tables aren't there. And third, the drill line has an annotation, a hundred and fifty metres per jumbo-month, next to thirteen thousand four hundred metres drilled per month, and the units aren't self-explanatory. If that hundred and fifty is advance per jumbo per month, the initial assessment's two to three rounds per shift would be several times faster. We couldn't resolve it from the document, so it's a question for the authors, not a finding.
Joao: And development rates?
Traci: The feasibility study schedules individual headings at forty to eighty metres a month, with a mine-wide cap of eighty-five hundred metres a year. Which brings me to Las Chispas, because Las Chispas does something I really like.
Joao: The SilverCrest project in Sonora.
Traci: The 2021 feasibility study, effective January fourth, mine planning by G Mining Services. Two things. One, the production rate. They ran three empirical formulas on the tonnage of the main zone: Taylor's rule, which gives about a thousand tonnes a day, Long's version, about thirteen hundred, and a newer formula from McIsaac and Devon Smith, about fourteen hundred. And the Qualified Person chose twelve hundred and fifty.
Joao: Taylor's rule is a reference-class forecast.
Traci: Taylor's rule is an empirical fit of production rate against tonnage, drawn from real operating mines. It's the outside view, built into mine engineering decades before anyone used the phrase. Two, the development rate. They set single-heading advance at two point five metres a day, with a scheduling cap of three hundred metres a month, and the report says that performance level was shown to be achievable at Las Chispas itself in 2019 and 2020.
Joao: So their own history is the reference class.
Traci: Their own demonstrated history. That's the strongest kind of benchmark you can have. Where Las Chispas is weaker, from our point of view, is fleet availability: all mining is by contractor, the contractor supplies the equipment, and the fleet was sized from planned tonnage, planned hours and "the specific vehicle performances assumed". Those performances aren't disclosed in the mining section. The only utilisation figures I found are for ventilation: eighty percent for production equipment, fifty for service equipment, and they call those conservative.
Joao: Which for ventilation means high, because more running equipment needs more air.
Traci: Right, conservative in the opposite direction from a productivity assumption. Worth noticing when a single number is "conservative" for one purpose and optimistic for another. Then Cariboo, Osisko Development, January 2023, mining by InnovExplo. Cariboo says all scheduling rates are based on experienced mining contractor feedback or typical rates for similar operations. That's an explicit outside-view statement.
Joao: And the numbers?
Traci: Effective productivity seventy-two percent of a ten-hour shift for development equipment, sixty-five percent for manual loaders and trucks. Jumbo development at eight to ten metres a day multi-heading. A roadheader at two hundred metres a month, which they call conservative pending validation on the adjacent Bonanza Ledge operation. And then automation. Stope mucking with a ten-tonne loader is scheduled at eight hundred and eighty-seven tonnes a day in manual mode, nine hundred and sixty-nine with tele-remote between shifts, and one thousand and twenty-seven fully automated.
Joao: So plus nine percent and plus sixteen percent.
Traci: And of the steady-state forty-nine hundred tonnes a day, five hundred and fifteen are attributed to autonomous mucking and hauling over the two-hour shift change, assuming, quote, optimal setup conditions. About ten and a half percent of production.
Joao: How does that sit against rule eight?
Traci: The mechanism is right. It's recovered hours through shift change, not faster loading, which is exactly what the evidence supports. The magnitude, nine to sixteen percent, is inside the five-to-fifteen range at the top end. The flag is the phrase "optimal setup conditions". At LKAB, three-quarters of the stops were mining-related, and Telfer needed months of deliberate work to get automation reliability from thirty-nine to seventy-seven percent. So I'd want to see a ramp-up on the automation credit, not a switch that flips in 2028. The other thing Cariboo does well: an independent maintenance and equipment consultant provided a mobile equipment rebuild schedule and yearly operating hours.
Joao: Which is the thing Era Dorada's initial assessment didn't show.
Traci: Right. And Platreef. Ivanhoe's 2022 feasibility study for the platinum-group-metals project in South Africa. The effective-hours build-up is the most transparent of the four: eight point one three effective hours of an eleven-and-a-half-hour shift. On equipment, it says quantities were determined from operating hours required, and swing units were then added to account for overall availability. No availability percentage, but at least the mechanism is stated.
Joao: And anything that bothers you?
Traci: Two things. First, the 2022 study says that no change was made to the mine designs, activity durations, development or stoping rates from the 2017 feasibility study. The schedule was stretched from four to five point two million tonnes a year using the same rates. And it says contractor bids were compared to the owner-operator baseline and the productivities didn't change. Second, its single-heading development rates. Built from first principles for a five by five metre heading, they're about four point three to five point eight metres a day across the normal and high-stress support classes. Las Chispas uses two point five, demonstrated on site. Era Dorada's feasibility study, forty to eighty metres a month, call it one point three to two point seven a day.
Joao: Different headings, different ground, different everything.
Traci: Completely different, and I'm not claiming any of them is wrong. Platreef's rounds are bigger and it's a big, deep, shaft-access mine. But it's a nice illustration. Four projects is a tiny reference class, and within it the first-principles number is about twice the demonstrated one. If I were reviewing, that's where I'd ask: where has this been achieved? And Platreef's own recommendations include further analysis of equipment utilisation to reduce fleet size, which tells you the authors themselves think the fleet numbers have slack in them somewhere.
Joao: Let me ask the uncomfortable question. These are four reports we happened to have. Are they representative? Isn't this exactly the cherry-picking you'd criticise in someone else?
Traci: It's a fair challenge, and I'll own it. Four reports aren't a sample of anything. They're test cases for whether the rules are usable, not evidence about how technical reports behave in general. What I'd say they show is a spectrum. At one end, an initial assessment that gives you nothing to test. Then feasibility studies that state their factors, and the factors they state land inside the literature's ranges. And at the good end, projects that anchor on their own demonstrated performance or on empirical rules from real mines. If that spectrum is typical, the most useful reviewer question isn't "is this number right?" It's "where did this number come from, and where has it been achieved?"
Joao: Alright. Wrap it up for me. If someone's going to review a mine plan next week, what do they carry?
Traci: Let me rank them, most useful first. One. Make them define availability before you argue about it: mechanical or operational, calendar or scheduled, when downtime ends. Almost half the Canadian mines in the Laval survey ended downtime at dispatch, and that alone moves the number.
Joao: Two.
Traci: Two. Productive time is about half of calendar time for underground mobile fleets, give or take ten points, and drill percussion time is a few hours per rig per day. Telfer's fifty-four percent net operating time and Era Dorada's forty-eight to fifty-two percent are the anchors. Much above sixty percent of calendar for a conventional fleet needs site evidence.
Joao: Three.
Traci: Three. Loader availability seventy-five to eighty percent young, sliding toward the mid-sixties to low seventies past fifteen to twenty thousand hours unless rebuilds and replacements are in the plan. Booysendal's sixty-six against a planned eighty-five is the cautionary tale.
Joao: Four.
Traci: Four. Nameplate is not achieved. Fifty-minute hour, and a separate payload haircut of fifteen to thirty percent for blasted rock. Cat says so in its own handbook, and the field data from Finland and Iran agree.
Joao: Five.
Traci: Five. Maintenance cost: lifetime on the order of one to two and a half times capital for a loader, stepping up at component change-outs rather than rising smoothly. A flat cost per hour over a long life is a tell, and maintenance hours per operating hour above roughly point three to point four is a replacement flag.
Joao: Six.
Traci: Six. Replacement: drill rigs around seven to ten years in the case studies, loaders overhauled around ten to fourteen thousand hours, with total lives of twenty to twenty-five thousand in the older evidence. The optimum is flat, so don't over-engineer the date, but a life of mine over ten years needs a renewal cycle with timing.
Joao: Seven.
Traci: Seven. Automation buys hours, not speed. Five to fifteen percent more output from working through shift change and blast clearance is what the better evidence supports, with a ramp-up. The best controlled comparison showed lower tonnes per engine hour and no maintenance saving.
Joao: Eight.
Traci: Eight. Fleet sizing: get the match factor near one, expect dispatch to be worth single digits to about twenty percent, and don't trust a simulation that hasn't reproduced the mine's actual production first.
Joao: And nine and ten?
Traci: Nine, MIP schedules are worth single digits of NPV in practice, and deterministic schedules need the same haircuts. And ten, the reliability one: MTBF is fine for availability, dangerous for ageing and maintenance planning, so ask for a trend test.
Joao: And where's the evidence thinnest? Be honest.
Traci: Loaders and drills: the core is a handful of Scandinavian, Finnish, Canadian and Australian sites, a lot of it from the eighties and nineties, with two recent South African and Irish case studies. Replacement ages come from a few case studies, two of which lean on extrapolated or partly synthetic cost data. Automation rests on one strong controlled comparison at LKAB, one good site study at Telfer, and a lot of vendor material that's useful for mechanisms and useless for magnitudes. The open-pit fleet numbers are mostly simulations, and the one paper that compared its simulation to reality showed how far those can drift. And we haven't read the Vayenas and Wu loader study, the Topal and Ramazan truck scheduling papers, the Chaowasakoo age-based maintenance paper, Kesimal and Bascetin on truck replacement, the 2022 drill selection chapter, the Sandvik handbook, or the Ghana theses. Any of those could move a range.
Joao: So the rules are a starting reference class, not the last word.
Traci: They're a starting reference class, which is exactly what Flyvbjerg would tell you to begin with. You anchor on what comparable operations actually did, you adjust for what's really different about yours, and you write down why. And if the most honest thing the evidence can tell you is that a Finnish loader in 1991 did four to five useful hours a shift, that's still a better anchor than a daily peak in a customer story.
Joao: The thing that stuck with me is the Boliden one. They started tracking the drill rig's costs when the costs started growing.
Traci: That's the whole episode in a sentence. We measure when it hurts, and then we plan from the part we measured. The fix is to go and find the operations that measured all the way through, and borrow their numbers.
Joao: That's the show. Full bibliography and the rules of thumb, with their conditions, are in the show notes.
Traci: Thanks for listening.