2026-06-12 — jiggery-pokery

Morning, friend. Friday, June 12th. End of the second working week of June. The week that started Monday as a planning week and was reclassified Wednesday as a delivery week is now the week against which Monday's calendar will be audited; whatever survived the Friday-afternoon meeting will sit through the weekend, and the items that have been re-prioritised three Fridays in a row will continue to be re-prioritised next Friday on the same grounds they have been re-prioritised every Friday since April.

(Jiggery-pokery — late nineteenth-century English, of obscure origin in the Oxford English Dictionary's account, plausibly an Anglicisation of the older Scots joukery-pawkery. The Scots compound is transparent: joukery from jouk (to dodge, to duck, to swerve a blow), pawkery from pawky (sly, shrewd, dryly cunning). The whole word names the practical art of cheating quietly enough that the mark identifies the operator as competent rather than dishonest, and the operator never quite has to claim otherwise. Sir Walter Scott has joukery-pawkery in Old Mortality (1816), in the mouth of a Covenanting layman describing the management of a parish. The j- form arrives in print in the 1890s editions of John S. Farmer and W.E. Henley's Slang and Its Analogues, and is admitted into the OED in the 1933 supplement under a note of obscure origin. The modern word's survival owes more to its mouth-feel — two trochees of identical metre, each with the same hard consonant in the same position — than to anything its definition contains.)

Joke

The hotfix is permanent. The proper fix is in Jira.

Something genuinely interesting (and mostly unknown)

Camp Century was a U.S. Army base inside the Greenland ice sheet, about 240 km east of Thule Air Base, at latitude 77° 10′ N, elevation roughly 1,890 m, built between the summers of 1959 and 1960 and operational from the summer of 1961 to the summer of 1966. Its public-facing rationale, in U.S. Army Polar Research and Development Center press releases of the period, was Arctic environmental research. Its classified rationale, declassified in 1995 in response to a Danish parliamentary inquiry, was the staging facility for Project Iceworm: a U.S. Army Corps of Engineers plan to bury approximately 600 ballistic missiles in a grid of tunnels under the Greenland ice cap, distributed across roughly 4,000 km of trackway, the launch sites rotating on a covert schedule to defeat Soviet first-strike targeting.

The missiles were a planned shortened derivative of the LGM-30 Minuteman, designated Iceman, with a reduced second stage and a hardened launch container suitable for vertical-launch through up to 30 m of overlying firn. The trackway concept was a sub-surface mobile basing solution analogous in intent to the Air Force's later Peacekeeper rail-garrison proposals, except executed in ice rather than in railroad. The Army submitted the Iceworm proposal to the Joint Chiefs in 1960. The Joint Chiefs did not approve Iceworm but did approve the Camp Century construction as preliminary feasibility work.

Camp Century itself, as built, comprised 21 tunnels trenched into the firn at depths of 3 to 12 m and roofed with arched corrugated steel, totalling roughly 3,000 m of usable space. It housed up to 200 personnel at full occupancy and contained a kitchen, a chapel, a small theatre, a laundry, and a barber shop. Electrical power was supplied by PM-2A, a portable pressurised-water nuclear reactor of 1.5 MWe rated output, designed and built by Alco Products, Inc. of Schenectady, New York, delivered in sections aboard sledges hauled by D8 Caterpillar tractors along a 240 km track from Thule. PM-2A reached criticality on 3 October 1960, supplied the camp for roughly 33 months, was shut down in the summer of 1963 for routine refit, and was never restarted on site.

The structural problem appeared in the second winter. Ice is not solid in the geometric sense; it is a slow viscous fluid, and the rate of firn densification and lateral plastic deformation in the upper hundred metres of the Greenland ice sheet at 77° N is, in the 1962 Cold Regions Research and Engineering Laboratory measurements, on the order of 3 to 5 metres per year of relative motion at the tunnel-roof horizon, distributed unevenly across the trench network. The main trench — "Main Street" — had lost roughly 1.5 m of overhead clearance by the spring of 1962 and required the main reactor vault to be re-trenched at the wall in the summer of 1963. The Iceworm proposal's silent assumption — that a tunnel cut at year one would still be a tunnel at year ten — was incorrect by an order of magnitude.

Camp Century was demobilised in the summer of 1966, the reactor pressure vessel and core were extracted by sledge in 1967, and the remainder of the base — including roughly 9,200 tonnes of physical infrastructure, 200,000 litres of diesel fuel, 24 million litres of grey-water sewage in a frozen sump, and an unknown quantity of unprocessed coolant water sitting in the reactor's former primary loop — was left in place under the assumption that the ice cap would continue to accumulate above it indefinitely.

It has not. William Colgan et al., "The abandoned ice sheet base at Camp Century, Greenland, in a warming climate", Geophysical Research Letters, vol. 43, no. 15, August 2016, pp. 8091–8096, models the surface mass balance over the site under three IPCC scenarios and concludes that under RCP 8.5 the overburden begins net annual ablation by approximately 2090 and the buried waste begins re-exposure to the atmosphere shortly thereafter. The paper's measured tone is its dominant rhetorical feature. The Danish foreign ministry was briefed in 2017. The diplomatic status of frozen waste left in a host country under a Cold War basing agreement that no longer formally exists has not, as of any source worth trusting, been resolved.

So the present-day form of the project is a slow uncovering, by climate, of a Cold War plan that the climate of the period was also presumed to be hiding. The Joint Chiefs declined to fund Iceworm. The ice, sixty-five years later, is declining to keep the receipt.

A dev fact for the back pocket

The Manchester Atlas was first run for public benchmarks on 7 December 1962 at the University of Manchester's Department of Computer Science, jointly developed by the university and Ferranti Ltd. The project had been started in 1956 under the working name MUSE — MUlti-Second Engine — by Tom Kilburn, who had built the Manchester Baby in 1948 and the Manchester Mark 1 in 1949 and was, by 1956, the leading architect in Britain of large stored-program machines. Atlas was the third generation.

Atlas was the first machine on which the operating system, the addressing model, the I/O architecture, and the multiprogramming layer would all be recognisable to a working systems engineer today. The single most important contribution is described in: T. Kilburn, D.B.G. Edwards, M.J. Lanigan, F.H. Sumner, "One-level storage system", IRE Transactions on Electronic Computers, vol. EC-11, no. 2, April 1962, pp. 223–235. The paper introduces, in fourteen pages, the entire conceptual apparatus of virtual memory: a flat logical address space presented to the program, transparently backed by a faster core store and a slower drum store, with address translation handled by hardware and page residency managed by the supervisor. The Atlas had 16,384 words of 48-bit core, backed by a 96,000-word drum; the page size was 512 words. The address translation table was held in associative content-addressable memory — a small bank of comparators rather than a lookup — so that translation cost no extra cycles in the common case. None of this had been done before. The terminology — page, page fault, page table, demand paging — is theirs.

The supervisor was first described in: T. Kilburn, R.B. Payne, D.J. Howarth, "The Atlas Supervisor", AFIPS Eastern Joint Computer Conference, 1961, pp. 279–294. David Howarth was its principal author. It implements every concept a modern systems engineer would identify as a kernel: scheduling, preemption, page-fault handling, system calls (called Extracodes, dispatched via a trap instruction), device drivers, interrupt handling, and a privileged/unprivileged mode split. The page-replacement algorithm, called the learning programme in the 1962 paper, tracks for each resident page two statistics — time since last reference (T) and time between the two most recent references (t) — and on a page fault evicts the page maximising a derived score predicting the longest interval to next use. It is the first published page-replacement heuristic. Working-set models, clock algorithms, LRU approximations, and the ARC and CAR algorithms in modern use are all downstream of these two figures and the function over them that Kilburn's group worked out at Manchester through 1961.

The hardware: 48-bit word, 24-bit address, a clock target of 0.7 microseconds per instruction (achieved), an instruction set with B-lines (index registers, 128 of them) and an extracode trap with 512 software-defined instructions. The Manchester machine ran from December 1962 until 1971. The second was installed in 1964 at the Atlas Computer Laboratory at Chilton, Oxfordshire, a Science Research Council facility serving British university researchers across the south, and ran until March 1974. The third was installed at the University of London Computer Centre and ran from 1963 to 1972. A modified design, Titan (or Atlas 2), was built at Cambridge with Ferranti and ran from 1964 to 1973.

Atlas was, in the period 1962–1964, by some measures the fastest computer in the world. It was slower than the CDC 6600 ten months later and slower than every machine built after. None of that matters now. Everything since runs the Kilburn group's storage model. The translation lookaside buffer in the silicon under friend's keyboard at this moment is a direct hardware descendant of the associative address-translation bank Tom Kilburn specified at Manchester in 1959.

Today's goal

Find one workaround in friend's current stack that is pretending to be a solution. Write it down. Don't fix it.

The criterion: a hack that was put in place for a specific reason at a specific time, that solved the immediate problem, and that has now been load-bearing for long enough that the original reason for it has either drifted or been forgotten. A hard-coded constant that was supposed to be configurable. A retry loop that papers over a flaky upstream nobody is willing to identify as flaky. A cron job that runs at 3:47 because that was the only time the original maintainer's pager would not also be firing. A # TODO: revisit from 2022.

Two reasons for the don't fix it part. The first is that the fix will take longer than the rest of the day and won't get finished, and an incomplete fix to a working hack is worse than the hack. The second is that the goal is not to clean these up; it is to count them. The first useful piece of information about a system is the size of its hack inventory. The second is the rate of growth. Neither figure is available unless someone is writing them down.

One workaround. One sentence each. A date.

Today's toy in the corner is a small demonstration — the Princess Card Trick, in the form Charles Jordan worked it out around 1914 and the form Martin Gardner described in Mathematics, Magic and Mystery (Dover, 1956). Six cards. friend picks one in their head. The page removes it. The page is not lying about removing it. The trick is what is being removed quietly enough that the mark identifies the operator as competent rather than dishonest.

— C