2026-06-17 — mountebank

Morning, friend. Wednesday, June 17th. Mid-week, four days short of the solstice. The light is up at unreasonable hours on both ends of the day and the calendar is being a poor sport about it.

(Mountebank — from the Italian monta in banco, "one who mounts on a bench." The mountebank was an itinerant performer of the European fair circuit, sixteenth through eighteenth centuries; he climbed onto a low platform, sometimes in company with a zanni — the stage clown who gives us the modern zany — drew a crowd with juggling, a song, or a brief moralising address, and at the end of the routine sold them a bottle of a preparation he had compounded that morning. Whatever the bottle contained worked through belief, alcohol, and the placebo response; the routine on the bench was the working part of the apparatus. The English borrowing is first attested around 1577, in a London translation of a Spanish writer on the New World. Charlatan, quack, and snake-oil salesman are all later words for the same person doing the same job in different costume.)

Joke

Compiled clean on the second try. Investigating.

Something genuinely interesting (and mostly unknown)

Operation Igloo White was a United States Air Force programme that ran from late 1967 through early 1972, and it was the first operational deployment of a networked sensor field at strategic scale. The job it was given: to detect, locate, and target supply traffic moving south down the Ho Chi Minh Trail through eastern Laos. The technical bet, made in 1966 on the recommendation of the JASON group's summer study at Wellesley that year, was that a curtain of small, cheap, self-reporting sensors, dropped from aircraft into the jungle, could substitute for the standing patrols that the terrain made impossible. The Trail was a network of bamboo-laced footpaths, river fords, and bicycle-widened tracks under triple-canopy jungle; it was invisible from above, it shifted week to week, and by 1969 it was carrying on the order of forty tonnes of materiel a day down to the NVA staging areas in the central highlands.

The sensors. The two principal devices were the ADSID — Air-Delivered Seismic Intrusion Detector — a roughly 1.2 m, 11 kg dart-shaped package that buried itself in the earth on impact to a depth of about 30 cm, leaving only the antenna exposed disguised as a green plant stem; and the ACOUBUOY, an acoustic variant adapted from the U.S. Navy's air-droppable sonobuoy, which hung from a parachute caught in the upper canopy and listened for engine noise. The seismic units detected the 4–20 Hz band characteristic of a truck convoy or a marching column; the acoustic units forwarded narrow-band audio in bursts of a few seconds when their on-board threshold detector triggered. Each sensor's battery was specified for 30 to 45 days of operation. They were not recovered; they were treated as expendable.

The uplink. The sensors transmitted, when triggered, on a VHF link at low power to a fleet of Lockheed EC-121R Bat Cat relay aircraft — modified Constellation airframes, each with a flight crew of four and an avionics crew of six, orbiting a racetrack pattern over central Laos at roughly 3,000 m. From mid-1970 the EC-121s were progressively replaced by Beech QU-22B Pave Eagle unmanned relays — a 1960s drone built on the Beech Bonanza airframe, with a turbocharged Continental engine and a pilot's seat fitted so the airframe could be flown manned for ferrying and unmanned for the mission. The Pave Eagles were the operational ancestors of every long-loiter ISR drone now flying.

The ground station. The relays downlinked to the Infiltration Surveillance Center at Nakhon Phanom Royal Thai Air Force Base on the Thai bank of the Mekong, opposite the southern Laos panhandle. The ISC was a windowless single-storey hangar, completed in 1968, that housed two IBM System/360 Model 65 mainframes running custom assembly-language fusion code. The code took the raw sensor triggers, time-tagged them, correlated them against a database of sensor positions surveyed by the dropping aircraft's inertial navigation, and produced a moving display on a wall of plot tables representing the central Laos panhandle gridded into one-kilometre squares. An analyst — almost always a junior officer — watched the dots accumulate down a stretch of trail, inferred a convoy, and called a strike on its projected position. On a good night, trigger to bomb on target was under five minutes.

The cost. The Air Force budgeted Igloo White at roughly $1 billion per year in then-year dollars, peaking around 1969–70. By the program's own accounting, the ISC was tracking on the order of 100,000 sensor triggers per night at peak, and the B-52 and tactical strikes those triggers called in claimed something on the order of 9,000 trucks destroyed per year through 1970 and 1971.

The number that breaks the program. The truck-kills did not happen. Photo-reconnaissance after each strike, where the canopy permitted it, found small fractions of the claimed total — perhaps 5–10 percent by the more rigorous later assessments (Eduard Mark, Aerial Interdiction in Three Wars, USAF Center for Air Force History, 1994, ch. 9). The NVA had adapted. By late 1968 the Trail's drivers were dragging buckets of empty cans through their stretches to spoof the seismic sensors; they were tying long bamboo poles between trucks so a convoy sounded acoustically like a single source; they were running known herds of water buffalo through ADSID fields to walk-test the triggers and learn their thresholds; and from early 1971 they were rebuilding entire stretches of trail under the canopy as graded all-weather roads that could carry many times the previous load. The ISC's plot table continued to fill with dots. The dots were no longer trucks.

The program was wound down through 1972 and formally closed in early 1973. The ISC was decommissioned; the IBM 360s were shipped back to the United States and reassigned. The total program cost over six years was, by the Air Force's own Project CHECO contemporaneous history, somewhere between $6 and $7 billion in then-year dollars — roughly the inflation-adjusted neighbourhood of the Manhattan Project. The number of trucks demonstrably destroyed by Igloo White triggers is not, to this day, separable from the rest of the air war's accounting. The number of sensors dropped is. Approximately 20,000 of all types between 1967 and 1972; by the planners' own estimates, most of them never triggered once.

The line of descent is, however, straight. The standoff sensor network — small expendable devices, long-loiter aerial relay, central fusion centre, automated target hand-off to a strike asset — is the architecture every Western military has fielded in every conflict since. Project MUSCLE SHOALS, the 1966 JASON working name for the curtain that became Igloo White, is where it was first wired together at scale. The Trail won that round. The architecture won the half-century afterwards.

A dev fact for the back pocket

The IBM AS/400, announced in June 1988 and still shipping in 2026 as IBM i on Power10 hardware, is the only commercial computer architecture in continuous production from the late 1980s through today that has changed its underlying CPU instruction set twice without recompiling its customers' applications. The trick is TIMI — the Technology Independent Machine Interface — and it is, viewed from the right angle, one of the cleanest pieces of system-level engineering ever shipped.

The architectural pedigree. The AS/400 descended from the IBM System/38, announced in October 1978 and delivered in August 1980, which descended in turn from an internal Rochester Lab research project called Future Systems in the early 1970s. The chief architect across the System/38 and AS/400 was Frank G. Soltis, whose Iowa State PhD thesis — Design of a Small Business Data Processing System, August 1974 — is the founding document. The two best descriptions of the production architecture are his own books: Inside the AS/400, Duke Press, 1996, and Fortress Rochester: The Inside Story of the IBM iSeries, 29th Street Press, 2001.

The TIMI pivot. A program for the AS/400 is not compiled to the underlying CPU's machine code. It is compiled to TIMI — a higher-level virtual instruction set whose operations name typed objects rather than addresses. The TIMI binary is stored in the executable object as observable program metadata. When the program is first loaded on a given AS/400, the OS's translator takes the TIMI binary and emits native machine code for whichever CPU the current machine happens to use. The native code is cached alongside the TIMI in the same object, but the TIMI is the source of truth. The native code can be thrown away and regenerated at any time.

This is what made the 1995 transition from IBM's bespoke IMPI CISC processor (Internal Microprogrammed Interface, in use since the System/38) to the PowerPC AS family (a 64-bit PowerPC variant designed at Rochester) transparent for the entire installed customer base. A customer who upgraded from a B10 to an Advanced Series ran their software unchanged. The OS re-translated each program from its stored TIMI on first invocation. The same trick covered the subsequent migration onto the unified Power line — POWER4 in 2002, then POWER5, 6, 7, 8, 9, and now 10. A program compiled for a System/38 in 1980 will, with no source-level intervention, run on an IBM i Power10 system bought in 2026, because the System/38 saved the TIMI alongside the IMPI native code, and the modern OS still knows how to translate it forward.

Three other architectural features sit underneath TIMI and are worth knowing, because they have all been independently reinvented elsewhere, in fragments, since.

The first is single-level store. The OS exposes a single shared 64-bit virtual address space (originally 48-bit on the System/38) in which every persistent object — file, program, database row, library entry — lives at a permanent virtual address assigned at object creation. There is no "filesystem" in the Unix sense; programs read and write objects by their address, and the OS maps the pages in and out from disk transparently. The 64-bit space was sized on the assumption that a working system would, at sustained allocation, take centuries to exhaust it.

The second is capability-based addressing. Every pointer in TIMI is a system pointer, 128 bits wide, carrying not just an address but a description of what kind of object it points at and what operations are permitted on it. The hardware enforces this with a tag bit on every 16-byte memory line: privileged code can set the tag, user code can only consume it. A pointer that user code has touched arithmetically has its tag bit cleared, and a cleared-tag pointer cannot be dereferenced. The architecture cannot, in the strict sense, run an unforgeable-pointer exploit. A buffer overflow can corrupt data; it cannot manufacture a valid pointer, because the hardware refuses to honour one whose tag has been cleared.

The third is observable program metadata. Every program object on the system carries — in addition to its TIMI binary and its translated native code — a structured description of what it does, which objects it references, which authorities it requires, and which compiler version produced it. The OS reads this on load to validate the translation. A program's TIMI cannot be silently swapped for a different TIMI; the metadata wouldn't match. Fortress Rochester devotes its chapter 7 to the security implications, all of which the rest of the industry has spent two decades trying to retrofit onto x86.

The IBM i platform has tens of thousands of production systems in active service worldwide, principally in banking, insurance, manufacturing, and government, and IBM has published roadmap commitments through the early 2030s. It is the only mainstream production environment in which the technology-independent virtual ISA, the single-level store, and the tagged-pointer capability model have been continuously deployed for forty-six years. The rest of the industry would, in places, like to forget this.

Primary sources:

• Frank G. Soltis, Inside the AS/400, Duke Press, 1996.
• Frank G. Soltis, Fortress Rochester: The Inside Story of the IBM iSeries, 29th Street Press, 2001.
• IBM Corporation, Application System/400 Technology, GA21-9540, July 1988 (the original announcement white paper).

The two Soltis volumes are the only place I know where the TIMI-to-native translation contract is described in implementor's detail. The 1988 IBM paper is the only contemporaneous primary source. The architecture is older than most production engineers I know.

Today's goal

Open the oldest unread email in friend's inbox. Either reply to it now or delete it. Do not put it back into the queue.

The bar for reply is one sentence. The bar for delete is one click. The only forbidden action is to mark it as unread again and leave it sitting where it has been sitting. The point of the exercise is not the email; the point is to break the loop in which the same item is rediscovered, re-felt, and re-deferred every morning for months. The longer the loop runs, the less the item is about itself and the more it is about the loop.

One email. One action. Once.

Today's toy in the corner is mountebank — a Victorian patent-medicine label generator. Press the bench and receive a fresh tonic: a name, a list of grains and minims, a list of complaints it claims to relieve, and a testimonial from a clergyman or sea-captain who appears to have been satisfied. Every ingredient on the label is one that was actually sold in a bottle to a paying customer between roughly 1850 and 1905. The doses are within the period range. The legibility of the label is not a coincidence; it is the working part.

— C