The IBM PC — publishing an interface is not granting a right, and it did not matter

**proposals.md filed this under idea and the filing is right, but every previous entry that argued its way into idea got there by failing the moment test, and this one passes it.** The house definition of moment, fixed by eliza-1966 when it talked its way out of the kind and reused since by llama-weights-2023, is a date on which something visibly happened in public. unix-linux-1969-1991 failed on "in public" — a Usenet post read by a few hundred people. tcp-ip-1983 failed on the identical clause — a cutover announced to host administrators in a DoD newsletter. Neither excuse is available here. On 12 August 1981 IBM announced the Personal Computer at a press event at the Waldorf-Astoria in New York. There was a date, there was a public, and the public noticed. The clause that has done all the work in this canon does not touch this entry, so the case for idea has to be made positively, and made against a kind the entry genuinely qualifies for.

The positive case is the standard one and it lands cleanly. build.py defines idea as "a framing later work is built out of rather than argued about," and the house test — set by shannon-chess-1950, reused by dijkstra-1959, minimax-alpha-beta and printing-press-1440 — is that there is a mechanism in it you could implement. The mechanism here is not the 8088 and not the five expansion slots. It is a two-part decision, and both parts are artifacts:

1. Publish the interface completely — the bus, the memory map, the schematics, and the entire ROM BIOS as commented assembler source, in a manual anyone could buy. 2. Grant nothing — no licence, no covenant, no permission; keep the copyright on the implementation and enforce it.

That is implementable, it was implemented, and it has been implemented since by people who had never heard of Boca Raton. It is also, as section 3 grades, a mechanism whose two halves came apart — the first half worked past all expectation and the second half was void inside three years — which is exactly the sort of thing an idea entry is for and a moment entry cannot carry.

The negative case against moment is about what the date is worth. A moment's evidential content is its date: chatgpt-2022 and alphago-move-37-2016 are about what became true on a particular afternoon. Here the date is the least useful thing in the file. Nothing became possible on 12 August 1981. The 5150 was not the best personal computer on sale that morning and was not close; it entered a market that already had the Apple II, the TRS-80 and a shelf of CP/M machines, and it entered it fifth. Now list the dates a reading would actually cite from this entry — June 1982, March 1983, August 1983, early 1984, 1984, April 1987, September 1988, July 1994, 1994, December 2004 — and the shape is obvious. A moment that needs twenty-three years of sequel before it means anything is an idea wearing a date.

The year field is 1981 and the choice is deliberate. build.py reads the first four-digit number in that field, so this file marks the canon's chart at the decision rather than at any of its consequences. The consequences run to 2005 and section 3 grades each of them on its own date; putting the mark at, say, 1987 or 1994 would file the entry under its most dramatic year rather than its causal one.

descends_from is [hollerith-census-1890], and the edge is corporate and doctrinal rather than technical. No wire runs from a punched-card tabulator to an 8088. What runs is the firm and its operating doctrine. hollerith-census-1890 already carries the lineage on disk — the Tabulating Machine Company in 1896, merged into C-T-R in 1911, renamed International Business Machines by Thomas J. Watson Sr. in 1924 — and it carries the doctrine too, in the observation that Hollerith's machines were leased and not sold, which is why "the corporate relationship with IBM was continuous." That is the model this entry breaks. The PC was sold outright, through dealers, with a manual that told the buyer everything, and with no continuing relationship of any kind. Ninety-one years of corporate practice said the customer never owns the machine and never learns how it works; a twelve-month schedule in Boca Raton overturned both, and the overturning is the entry's entire content. That is descent that explains shape, which is the test.

**What is deliberately not claimed as an ancestor:**

• arpanet-1969 and tcp-ip-1983. The temptation is strong, because TCP/IP is the canon's other worked example of a published specification with no owner, and because tcp-ip-1983 is filed under idea on adjacent reasoning. But the ARPANET cutover is 1 January 1983 — seventeen months after this — so it cannot be an ancestor of it, and there is no evidence that RFC culture reached an IBM product division in Florida. They are siblings, they should be read together, and the differences are the interesting part: TCP/IP was published by people who wanted it copied, and this was published by people who did not.
• unix-linux-1969-1991. Downstream if anything. Linux's commercial life happened on the hardware this entry created, which makes the edge run the other way.
• llama-weights-2023. The entry this one mostly exists to be read against, and the relationship is analogy, not descent. Nobody at Meta reasoned from Boca Raton. Section 4 spends its length on keeping those two things apart, because collapsing them is the specific error this file is written to prevent.

Ids named in this file that do not exist in canon/ and are not being invented into the header: apple-iie-1983, macintosh-1984 and lego-mindstorms-1998 (all three proposed in proposals.md, none written), and compaq-1982, phoenix-bios-1984, wintel-1994, apple-v-franklin-1983, which are not proposed anywhere and are named here only as prose.

What it is

The schedule made every decision that is now called a philosophy

IBM's normal development cycle for a new machine ran four to five years through the usual committees. The PC did not get one. IBM chartered what Wikipedia's account calls "an independent internal working group" that "abandoned all traditional IBM methods," with a mandate to produce a prototype in thirty days and a shipping design in a year, and permission to grow to 150 people by the end of 1980. William C. Lowe got it approved; Philip Don Estridge ran it. The codename was Chess. Announcement to announcement, the whole machine took about twelve months.

Every celebrated decision falls straight out of that constraint:

• Off-the-shelf parts, from other companies. Wikipedia's account is explicit that the design used "entirely 'off-the-shelf' parts from third party manufacturers, rather than unique hardware designed by IBM," and gives a second, non-obvious reason alongside the schedule: market research found that "computer dealers were very interested in selling an IBM product, but they insisted the company use a design based on standard parts, not IBM-designed ones so that stores could perform their own repairs." The channel wanted a repairable box. IBM had never built one.
• An Intel 8088 at 4.77 MHz. Bought, not designed. Its 8-bit external bus — a compromise against the 8086 — let the team surround it with cheap support chips that already existed.
• An operating system from outside. IBM signed with Microsoft on 6 November 1980. Microsoft did not own an operating system either; it bought a non-exclusive licence to 86-DOS from Seattle Computer Products and delivered it to IBM as PC-DOS. IBM's negotiators, per the standard accounts, declined to pay a perpetual royalty and took a flat fee of roughly $250,000 — which is only half the story, because the other half is that they let Microsoft keep the right to license the same system to anybody else. That clause is the single most expensive sentence in the entry.

The machine shipped at $1,565 for a configuration with 16 KB of RAM, a Color Graphics Adapter, a keyboard, and no disk drives.

Estridge was killed on 2 August 1985, aboard Delta Air Lines Flight 191, which flew into a microburst on approach to Dallas–Fort Worth; 135 people died, including his wife Mary Ann. He was 48. The man who made the architectural decision was dead twenty months before IBM tried to reverse it, which is a fact worth carrying into the 1987 half of section 3.

The manual, which is the actual artifact

The entry's whole argument turns on what was in the book, so it is worth being exact rather than gesturing at "open specifications."

The IBM Personal Computer Technical Reference (August 1981) contains, as Appendix A, "ROM BIOS LISTINGS," pages A-1 through A-81, the complete source of the machine's firmware. This is not a paraphrase of somebody's summary; the appendix was read directly for this entry. Its contents page lists, by line number and ROM address: Equates and Data Areas (line 1), Power-on Self-test (line 198, E016), Boot Strap Loader (line 1355, E6F2), and then I/O support for asynchronous communications (1410, E729), the keyboard (1640, E82E), diskette (2255, EC59), printer (3007, EFD2), display (3119, F045) and cassette (4977, F859), followed by memory-size determination (4903), equipment determination (4933), the graphics character generator (5503), time-of-day (5642), print screen (5817), and "Notes for the BIOS Listing" (A-81).

What follows the contents page is an assembler listing in four columns — location, object bytes, line number, source — with IBM's own engineering comments left in. Line 1 reads $TITLE(ROM BIOS FOR IBM PERSONAL COMPUTER). Line 92 is annotated ; POINTER TO TAIL OF KEYBOARD BUFFER. Line 127 is BAD_DMA EQU 09H ; ATTEMPT TO DMA ACROSS 64K BOUNDARY. The listing runs to line 5,940, ending on page A-80 in the power-on reset vector at segment FFFFH with the release marker DB '04/24/81' — the date the ROM was frozen, three and a half months before the machine was announced.

Published alongside it: the system bus specification and memory map, complete and immediate on release, and the schematics.

What was not published is the part everyone forgets, and it is a null. No licence accompanied any of it. No covenant not to sue, no patent grant, no permission of any kind. IBM printed the source of its firmware and retained every right it had in it. The word "open," applied to this document, collapses two separate facts — fully described and freely usable — and only the first one is true. Holding them apart is the entire value of this entry, and section 4 is about what goes wrong when they are collapsed.

What happened to the manual, in order

• June 1982. Columbia Data Products shipped the MPC 1600, generally credited as the first machine more or less compatible with the IBM PC — under a year after the announcement.
• November 1982 / March 1983. Compaq announced the Compaq Portable and shipped it four months later, essentially 100% compatible. Compaq is reported to have spent about $1 million on the clean-room procedure: Team A disassembled the ROM and wrote a functional specification of what each call did; Team B, screened to exclude anyone who had worked with the code or the architecture, implemented from that specification and nothing else. IBM never sued Compaq. The method was designed for exactly that eventuality and it worked.
• 1983. Compaq's first full year: 53,000 units and more than $111 million, at the time the largest first-year revenue of any company in American business history.
• 30 August 1983. The Third Circuit decided Apple Computer, Inc. v. Franklin Computer Corp., 714 F.2d 1240 — holding that a computer program in object code burned into a ROM is copyrightable subject matter, and that an operating system is no exception. Franklin had copied Apple's ROMs literally for the Ace 100. It was the first US appellate holding that firmware of this kind is protected, and it cut both ways at once: it made IBM's control point legally real, and it made the clean room the only lawful route around it.
• Early 1984. IBM sued Eagle Computer, Corona Data Systems and Handwell for infringing its BIOS copyright. With Franklin freshly decided against them, they settled, stopped, and reimplemented. IBM's control point worked every single time it was tested against literal copying. This is the fact that disappears from the popular telling.
• 1984. Phoenix Technologies began licensing a clean-room BIOS to anyone who wanted one; AMI and Award followed. Phoenix advertised its own methodology as the product: the engineer who wrote it had come from TMS9900 work, had never touched an Intel processor, and was given functional specifications only. From that point, building a compatible machine required a purchase order rather than an engineering department.

The asymmetry to state plainly, because it is the mechanism: publishing the specification did not transfer the implementation, and IBM was right that it did not. But it reduced the cost of re-implementing to something a startup could afford, because it did Team A's work for them. Working out what the BIOS does is the expensive half of a clean-room reconstruction, and IBM had already done it, in print, on eighty-one pages, with the comments left in. Copyright forbade copying the expression. Nothing forbade knowing the behaviour, and IBM had published the behaviour.

The attempt to close it again

April 1987. IBM announced the PS/2 line and, with it, the Micro Channel Architecture — a genuinely better bus than the one it replaced, 16- or 32-bit, with bus mastering and none of ISA's interrupt starvation. It was proprietary, and the licence terms were the news: 5% of the selling price of every Micro Channel machine, plus a retroactive 5% on every ISA-bus clone the licensee had already built and sold. The retroactive half asked the clone industry to pay IBM for machines built to a bus IBM had published and did not own. One contemporary estimate put the effect at about $150 on the price of a system. Bill Lowe, then running the Entry Systems Division, opened the year by remarking publicly that anyone who tried to reverse-engineer Micro Channel would need either very good engineers or very good lawyers, and settled in to wait for the licensees.

They did not come. Compaq had a backwards-compatible 32-bit bus of its own in development and took the design to eight other manufacturers. In September 1988 the "Gang of Nine" — Compaq, AST Research, Epson, Hewlett-Packard, NEC, Olivetti, Tandy, Wyse and Zenith Data Systems, holding 33.5% of the 1987 PC market between them — announced EISA, which extended ISA to 32 bits, accepted existing ISA cards, and licensed for a nominal fee. The Wall Street Journal called it the Revolt of the Clones.

EISA was not a triumph of engineering. It was slower than Micro Channel, it was awkward, and within a few years both were displaced by PCI. That is not the point. The point is that Micro Channel lost on terms while holding the technical advantage, and it lost to a bus whose main virtue was that using it did not involve IBM.

The ending, with numbers

• 1983. IBM held about 25% of sales of personal computers priced between $1,000 and $10,000. A widely repeated secondary figure puts IBM's share at 63% in 1984 falling to 38% by 1987; the source found for it was a single blog and it could not be confirmed here, so it is reported as unverified and the direction of travel is what should be relied on.
• 1988. Gartner Group estimated the public was buying 1.5 clones for every IBM PC.
• First half of 1994. IDC put Compaq at 10.4% of the world PC market and IBM at 8.5%. Compaq had announced in September 1993 that it intended to displace IBM as the world's largest PC seller by 1996; it took under a year.
• 7 December 2004, completed 3 May 2005. Lenovo bought IBM's Personal Computing Division for $1.25 billion in cash and stock plus about $0.5 billion in assumed liabilities — $1.75 billion in total, with IBM taking an 18.9% stake in Lenovo. IBM created the category and left it twenty-three years later.

And, running underneath all of it, the part that actually matters. The platform did not become unowned. Intel held the instruction set and the fabs. Microsoft held the operating system and, more to the point, the OEM contracts: from 1988 it wrote per-processor licences, under which a manufacturer paid Microsoft for a copy of DOS on every machine of a covered type whether or not DOS was installed on it, with discounts of up to 40% for agreeing. The arithmetic is the whole trick — shipping a competitor's operating system meant the OEM paid for two. The Department of Justice sued and settled the same day; the consent decree of 15 July 1994 barred the practice. It took a federal antitrust action, thirteen years after the Waldorf-Astoria, to loosen a control point that IBM had handed over in a contract clause in 1980.

Why a reading would cite it

The concrete cases, in the order a reading is likely to need them.

1. Concentration — against "open weights decentralize the field." This is the entry's main job and the reason it earns a place. The IBM PC is the best-documented case in computing of an interface being opened and the resulting market becoming more concentrated, not less — one layer up. The shape to cite: publishing an interface commoditizes the layer the interface describes and moves the rent to whichever adjacent layer still has a control point. IBM opened the box; boxes became a commodity with commodity margins, which is precisely why Compaq could win the volume war and then lose it to Dell. The rent went to x86 and to DOS/Windows — the two control points IBM had created for itself and given away for schedule reasons in 1980 — and one of them needed a consent decree to touch. When a reading sees an open-weights release and reaches for "power dispersed," this entry supplies the question that has to come first: dispersed to whom, and what did not get commoditized? In August 2026 the honest candidates for the layer that did not get commoditized are leading-edge fabrication capacity, advanced packaging and high-bandwidth memory supply, and the dominant accelerator instruction set together with its software stack — which is the closest present-day analogue of what x86 was, and for exactly the same reason: everybody's code is already compiled for it.

2. Concentration — an incumbent trying to re-close a standard it opened. Micro Channel is the worked example, and it comes with terms, an opponent, a date and an outcome. Cite it when a lab that published something moves to restrict it: new terms attached to a previously permissive release, a licence with a user-count or revenue trigger, retroactive conditions, an "acceptable use" clause that arrives after the artifact. The 1987 base rate is not encouraging for the incumbent. IBM had the brand, the sales force, the enterprise channel, the largest share in the market, and the technically superior product, and it still could not make anyone pay. What beat it was that the old interface still worked and nobody was obliged to stop using it. The operative question for any modern instance is the same one: is the previously published version still runnable by the people who have it? If it is, the restriction is a request. llama-weights-2023 holds the modern instance of this with the clock running in days rather than years.

3. Capabilities and the public square — against "the leader announced it, so the leader has it." The 5150 was mediocre hardware announced by the fifth entrant into an existing market, and it took that market because it was documented, assemblable from parts anyone could buy, repairable by the dealer, and carried a name a corporate purchasing department could sign off on. Cite when an announcement is being scored on the artifact rather than on who can build against it — a smaller model with a documented interface and permissive terms is often the one to watch, and this entry is the historical reason to believe that.

4. Work and the economy — where the value lands versus where it is created. IBM invented the category, sold 750,000 machines in 1983 alone, and captured almost none of the economics of the twenty years that followed. Cite when a filing shows an enormous capability investment by a firm that does not hold a layer with a moat, and when a valuation is being justified by having been first.

The honest limit of the analogy, stated here rather than buried in section 4. IBM's decisions played out in a competitive hardware market with a deep parts supply chain, essentially no regulator until 1994, and a product whose marginal cost was dominated by components. Any of those could differ enough today to break the mapping — most obviously, the capital intensity of frontier training has no 1981 counterpart, and neither does the fact that the expensive artifact in question can be copied at zero cost once it exists. This entry is a base rate for how published interfaces behave, not a forecast. A reading that uses it as a forecast is doing the thing section 4 warns about.

What it got right, and what it got wrong

Section 3 is required for fiction and prediction and welcome anywhere. It is worth having here because there are three graded claims on the table, one of them IBM's, one of them IBM's again six years later, and one of them the proposal's.

Claim 1 — IBM's, 12 August 1981

The claim is never written in these words anywhere, but it is legible off the artifact and off the litigation: publish the interface completely, keep the copyright on the implementation, and we will get an ecosystem without losing the machine. Due date: continuously, from the first competitor.

First half — right, and understated by roughly an order of magnitude. Wikipedia gives IBM's own business plan as 220,000 machines over three years, and reports that sales exceeded expectations "by as much as 800%," with IBM at one point shipping 40,000 PCs a month and selling more than 750,000 machines in 1983. By 1984 IBM's PC revenue was $4 billion, more than twice Apple's; a 1985 Fortune survey found 56% of American companies with personal computers were using PCs against 16% using Apple. The ecosystem arrived, and it arrived because Appendix A let a stranger with an oscilloscope write a working device driver without asking IBM for anything.

Second half — wrong, and the manner of the wrongness is the entry's content. The trap here is that the copyright held. This is the part that gets dropped. IBM sued Eagle, Corona and Handwell in early 1984 and they folded; Apple v. Franklin, decided the previous August, had just established that a ROM operating system is copyrightable. The control point IBM selected was legally sound, judicially confirmed and practically effective against the thing it described. It simply did not describe the thing IBM needed controlled. Copyright forbids copying an expression; compatibility requires only knowing a behaviour — and IBM had published the behaviour, in full, deliberately, as a selling point. Graded: wrong within ten months of shipping (Columbia, June 1982) and wholly void within three years (Phoenix, 1984).

The general form, which is what a reading should carry away: a control point is worth exactly what it controls, and no more. Publishing a complete description of a behaviour is a decision about the behaviour, not about the text. IBM chose an enforceable right over the wrong object, and an enforceable right over the wrong object feels identical to the right one until somebody tests it.

Claim 2 — IBM's, April 1987

We can restore the boundary. Due within about two years, by IBM's own implied schedule.

Wrong, unambiguously, and faster than the first claim. Seventeen months after the PS/2 launch, nine companies holding a third of the market had published a competing bus; within a few more years the industry had walked past both to PCI. The failure mode is the instructive part: Micro Channel lost on price, not on engineering. It was the better bus. The market bought the worse one because the worse one came without IBM attached, and because the retroactive royalty asked firms to pay for the past as the price of the future.

What the 1987 attempt actually demonstrates is a transition of ownership that nobody signed: openness had stopped being IBM's policy and become the market's property. In 1981 IBM chose to publish. By 1987 IBM could not choose to stop, because the thing it had published was in a hundred thousand engineers' heads and in every card in every slot in every office in the country. That transition — from a thing the owner grants to a thing the owner cannot revoke — is the single most transferable finding in this file, and it is the one place where the analogy to weights is exact rather than approximate. llama-weights-2023 holds the modern instance with a seven-day clock where this one took six years.

Claim 3 — the proposal's own, canon/proposals.md, 17 August 2026

> the IBM PC, and the open architecture that produced the clone industry IBM > did not want. Cite when: an open interface decides who can build on a > capability, and the platform owner loses the platform — the closest historical > precedent for the open-weights argument.

Graded in four pieces, because it is right, wrong, misleading and half-right in turn — and the entry is worth more for saying so than for repeating it.

• "the clone industry IBM did not want" — right, and the strongest clause in it. The lawsuits are the evidence. A company that files copyright suits against three manufacturers in one year is not a company practising generosity, and the commercial record supports the proposal exactly.
• "open architecture" — wrong as stated, and the error is the one this file exists to correct. Nothing was licensed, granted or opened in any sense a 2026 reader means by the word. IBM published a description and asserted every right it possessed over the implementation. Reading "open architecture" onto a 1981 document imports a meaning that did not exist and produces precisely the confusion section 4 has to spend its length on.
• "the platform owner loses the platform" — misleading in the way that matters most. The platform was not lost. It was transferred, to two firms whose control points IBM had personally created — one in a contract clause in November 1980, one in a parts order — and one of the two was under a federal consent decree within thirteen years. "Loses the platform" invites the reading that opening an interface dissolves ownership. It does not dissolve it. It relocates it, and usually upward, to a layer with worse competition than the one that was opened.
• "the closest historical precedent for the open-weights argument" — right about the closeness, wrong about which layer maps to which. It is the closest precedent available and the proposal deserves credit for finding it. But the naive mapping runs backwards, and section 4 takes that apart.

Commonly misused as

Not required for idea — build.py reserves this section for limit — but included because this entry's entire function is to serve as an analogy, and the analogy is habitually run in reverse. An entry that supplies a comparison without supplying its failure modes is a liability.

1. "Open beats closed, so open weights win." What happened is narrower and considerably stranger than that. The published-interface machine took the volume market; the closed machine — Apple, which published nothing and licensed nothing and sued Franklin for trying — kept a smaller, durable, high-margin position and is still in the business. IBM is not. And the volume market's winners were not the party that did the opening: IBM opened and lost, Compaq copied and led for a decade before losing to a firm with a better supply chain, and the two firms that never opened anything took the profits for twenty years. The transferable claim is about interfaces, not about firms: a published interface with many implementations beats a private interface with one, at the layer the interface describes. It says nothing whatever about whether the publisher wins, and the historical answer to that question is no.

2. "Open weights are the IBM PC's open architecture." The layers do not match, and the mismatch flips the conclusion. IBM published the specification and kept the implementation: Appendix A described a behaviour that a competitor still had to rebuild from scratch, in a screened room, for a million dollars. Weights are the implementation. Releasing them is far more open than anything IBM did — it hands over the exact artifact Compaq spent a year and a million dollars avoiding the need to reconstruct.

The structure in this era that is actually IBM-PC-shaped is not the weights. It is the published API surface: an HTTP interface with a documented request and response schema, reimplemented by dozens of providers and by every local inference server, so that a client written against one runs unmodified against another. No licence, no permission, no owner, and enormous switching-cost relief for everyone downstream. That is the ISA bus of this decade, and it arrived by the same route — one vendor documented its interface to get an ecosystem, and then could not take the documentation back. A reading that wants this entry's lesson should look there first, and should immediately notice that the same relocation applies: commoditizing the model interface moves the rent toward whoever holds the silicon underneath.

3. "IBM gave the PC away." IBM gave nothing away. It published a manual, asserted copyright over the code inside it, sued three companies for infringing that copyright, and won every time. The failure was object selection, not generosity. Cite this against any claim that a legal or contractual control over an AI artifact is sufficient — a licence term, a usage policy, a terms-of-service clause, an export designation — without first asking the two questions IBM did not ask in 1981: what specifically does this forbid, and what would a competitor actually need to do that this does not forbid? IBM's answer to the second, had anyone asked it, was one sentence long: read the manual, then write your own.

4. "Estridge's team believed in openness." They believed in the calendar. The dealer channel wanted standard parts because dealers wanted to do their own repairs. The parts were bought because designing parts takes years and they had twelve months. The operating system was licensed from outside because writing one takes years. Every decision now read as architectural principle was a schedule concession made by people who expected to sell 220,000 machines. The general point is one this canon can use often, and not only about 1981: an architecture that later reads as principle is frequently the fossil of a deadline. It is worth remembering when a lab's release posture is being read as a philosophy rather than as a competitive position taken under time pressure — and worth remembering in the other direction too, since the posture outlives the pressure that produced it and becomes, as it did here, something its author cannot revoke.

Sources

Marked by what was actually read, in the house convention.

• **IBM, IBM Personal Computer Technical Reference, Appendix A: "ROM BIOS Listings" (August 1981), pages A-1 to A-81. Read directly, as rendered PDF pages, via the copy hosted at retrocmp.de/fdd/general/IBM-ROM-BIOS_Aug1981.pdf. The primary source for everything in this file about what IBM actually published: the contents list and its line numbers and ROM addresses, the four-column assembler format, the retained engineering comments, $TITLE(ROM BIOS FOR IBM PERSONAL COMPUTER) at line 1, and the listing's end at line 5,940 with DB '04/24/81' as the release marker in the power-on reset vector. A part-number discrepancy worth recording: bitsavers' directory index names the August 1981 Technical Reference 6025008, while minuszerodegrees and several other archives name it 6025005**. Both refer to the August 1981 manual. The discrepancy was not resolved here and neither number should be quoted as authoritative without checking the title page.
• PCjs Machines, "IBM PC (Model 5150) BIOS Sources." Read via fetch. Independently corroborates the 5,940-line count for the August 1981 listing and gives 6,207 lines for the April 1984 revision. Also the source for the copyright string DB '5700671 COPR. IBM 1981' appearing in a later BIOS revision — a detail that was not visible on the Appendix A pages read here, and is reported on PCjs's authority rather than confirmed.
• Wikipedia, "IBM Personal Computer." Read via fetch. Source for the announcement date and $1,565 introductory price; Lowe and Estridge; the independent working group, the thirty-day prototype and one-year design mandate and the 150-person cap; the off-the-shelf-parts decision and the dealer-repairability finding; the immediate publication of the bus specification and memory map; the 220,000-over-three-years business plan against sales exceeding it "by as much as 800%" and 40,000 units a month; 750,000 machines in 1983; $4 billion 1984 PC revenue against Apple's; the 1985 Fortune 56%/16% figures; and Columbia Data Products, June 1982.
• Wikipedia, "IBM PC compatible." Read via fetch. Source for the Franklin-to-clean-room logic; the Compaq Portable announcement of November 1982; Phoenix BIOS in 1984 and the AMI/Award followers; the Eagle, Corona and Handwell infringement actions; IBM at "about 25% of sales of personal computers between $1,000 and $10,000" in 1983; and the Gartner Group 1988 estimate of 1.5 clones per IBM PC.
• Wikipedia, "Acquisition of the IBM PC business by Lenovo." Read via fetch. Source for 7 December 2004 announced, 3 May 2005 completed, $1.25 billion cash and stock plus $0.5 billion assumed liabilities.
• Christian Science Monitor, "Low-Cost Strategy Makes Compaq the Leader of the Pack" (7 October 1994). Read via fetch. A contemporaneous source, which is why it is preferred here over the retellings: IDC's first-half-1994 figures of Compaq 10.4% against IBM 8.5%, and Compaq's September 1993 statement of intent to displace IBM by 1996.
• Jonathan Scott, "The Rise and Fall of the IBM PC" (retrotechreads, Substack). Read via fetch. A secondary blog account, used only for the Micro Channel licensing terms — 5% forward plus a retroactive 5% on prior ISA machines, and an estimate of about $150 added per system — and for Bill Lowe's 1987 remark about engineers and lawyers. Its market-share series (63% in 1984 falling to 38% in 1987) is reported in this file as unverified because no second source for it was found; the Wikipedia and IDC figures above are preferred wherever they overlap.
• **Search results on Apple Computer, Inc. v. Franklin Computer Corp., 714 F.2d 1240 (3d Cir. 1983). Opinion not read in full**; the citation, the 30 August 1983 decision date, the holding that object code in ROM including an operating system is copyrightable, and its status as the first US appellate ruling to that effect are taken from search summaries of the Justia and Wikipedia pages rather than from the text of the opinion.
• Wikipedia, "Extended Industry Standard Architecture" and "Micro Channel Developers Association," and All About Circuits, "Tales from 80s Tech." Search summaries only, not fetched. Source for the September 1988 EISA announcement, the nine-company membership list, the 33.5% combined 1987 share, and the "Revolt of the Clones" label.
• Compaq first-year figures — 53,000 units and more than $111 million, "the most by any first-year company in U.S. business history." Search summaries only, appearing consistently across the Texas State Historical Association handbook entry, TechSpot and Britannica.
• **US Department of Justice, United States v. Microsoft, consent decree entered 15 July 1994. Search summaries only**, from the Washington Post contemporaneous reporting and the Stanford CS201 archive. Source for the per-processor licensing practice, its introduction in 1988, the discounts of up to 40%, and the decree's prohibition.
• Microsoft–IBM contract of 6 November 1980 — the non-exclusive licence permitting Microsoft to sell to other manufacturers, Microsoft's own non-exclusive 86-DOS licence from Seattle Computer Products, and the flat fee of roughly $250,000 in place of a perpetual royalty. Search summaries only, from secondary sources of varying quality. The non-exclusivity is not in dispute anywhere; the $250,000 figure should be treated as the commonly repeated number rather than a verified one, as the contract itself was not read.
• Philip Don Estridge's death, 2 August 1985 aboard Delta Air Lines Flight 191 at Dallas–Fort Worth, aged 48, with his wife Mary Ann; 135 dead in a microburst-induced crash. Search summaries only, from the Wikipedia pages for Estridge and for the flight.
• IBM's own history page for the Personal Computer (ibm.com/history/personal-computer) returned HTTP 403 on fetch and could not be read. IBM's own account of its own decision is therefore not represented in this file, which is a real gap: everything above about IBM's intent is inferred from the artifact and from third-party accounts. A later session that gets the page should revisit section 3's first claim, which is the one that would most benefit from IBM saying in its own words what it thought the BIOS copyright was for.