The Macintosh — putting the interface in ROM, and making conformance cheaper than divergence

proposals.md filed this under idea, at line 717, inside the ## idea block. The filing is right. The one-line description under it is right about the shape of the thing and wrong about which thing it is, and section 3 grades that rather than quietly repairing it.

The proposal reads: "the Macintosh and the graphical interface. Cite when: an interface, rather than a capability, decides how many people can use a thing — the current analogue is a chat box in front of a model that already existed." The instinct is sound and the case is not the one the proposal thinks it is. On 24 January 1984 the Macintosh did not introduce the graphical interface, and the graphical interface did not sell the Macintosh. The interface was fifteen years old and had shipped commercially three years earlier. The machine carrying it nearly died within eighteen months, and what rescued it was a capability that arrived on a $6,995 printer with more computing power inside it than the computer had. What is worth citing here is neither the machine nor the metaphor. It is a mechanism, it is stated in the machine's own launch documentation in plain English by the man who built it, and it is the reason the Macintosh's interface outlived the Macintosh's market position by forty years.

The positive case for idea is the standard house one. build.py defines idea as "a framing later work is built out of rather than argued about," and the test set by shannon-chess-1950 and reused by dijkstra-1959, printing-press-1440, ibm-pc-1981 and apple-iie-1983 is that there is a mechanism in it you could implement. The mechanism here has two parts and both are artifacts you can point at:

1. Put the interface in the ROM, not in a document. Ship the window manager, the menu manager, the dialog manager, the text editor and the graphics library as machine code inside the machine, reachable through a published trap table, so that a developer who conforms writes less code than a developer who does not. 2. Define the interchange formats too, and keep the list short. Two data types — text and pictures — through one shared buffer, so that any two conforming applications can hand work to each other without either knowing the other exists.

Andy Hertzfeld wrote the sentence that makes this an idea rather than a product description, in Apple's own launch-week material: "It is somewhat risky to put 64K bytes of intricate system software in ROM on a disk-based system, but we did it because we wanted the machine to have a built-in standard user interface. By using our ROM-based toolbox, a programmer saves development time and precious memory space; this provides a positive incentive for doing it our way."

That is the whole thing. Consistency across applications was not achieved by persuasion, by a style guide, or by review. It was achieved by making the conforming path the cheap path, in an era when 64K of ROM cost less per bit than RAM and a developer had 80K of usable memory to work in. The guidelines came later and were downstream of the economics; the economics were in the silicon on day one. That mechanism has been reimplemented by everyone since — every platform SDK, every widget toolkit, every design system, and, in 2026, every open protocol whose adoption strategy is "conforming is less work than not conforming."

Why not moment. The house definition, fixed by eliza-1966 when it argued its own way out of the kind and reused since by clippy-1996, siri-2011, chatgpt-2022, speak-and-spell-1978, lego-mindstorms-1998 and apple-iie-1983, is a date on which something visibly happened in public. This entry passes that test the way ibm-pc-1981 passes it and for the same reason declines it. There was a date — 24 January 1984, at Apple's annual meeting at the Flint Center in Cupertino — and there was a public, primed two days earlier by a sixty-second film during Super Bowl XVIII. But a moment's evidential content is its date, and here the date is close to the least useful thing in the file. Nothing became possible on 24 January 1984. Everything the machine did had been demonstrated at Xerox PARC in the 1970s and sold commercially by Xerox in 1981. Now list the dates a reading would actually want from this entry — December 1979, 1981, January 1983, November 1983, September 1984, January 1985, March 1985, July 1985, November 1985, March 1988, July 1989, May 1990, September 1994 — and the shape is ibm-pc-1981's exactly. A moment that needs a decade of sequel before it means anything is an idea wearing a date.

Why not interpretation. The reading of these events — that interfaces rather than capabilities gate adoption — is the proposal's claim, not the object's, and section 3 grades it against the record rather than adopting it. The grade is partly right, and wrong in the specific way that matters most for 2026.

Why not prediction. There are five dated, gradeable claims in this file and section 3 grades all of them, but they belong to Raskin, to a magazine, to Apple's press office and to the proposal — not to the object. If the canon ever files Gregg Williams's February 1984 forecast on its own it should file it with its own dates.

The year field is 1984 and the choice is deliberate. build.py reads the first four-digit number there, so this file marks the chart at the launch rather than at 1979 when the project started, 1981 when the interface first shipped for money, or 1990 when it actually reached most people. Section 5 argues that 1990 is the year the proposal's claim actually came true, and the mark stays at 1984 because that is the year the mechanism shipped.

---

1. What it is

The machine, from the contemporaneous record

The best source available for what the Macintosh was on the day it appeared is **Gregg Williams's cover story in BYTE, February 1984, page 30**, titled The Apple Macintosh Computer and subtitled "Mouse-window-desktop technology arrives for under $2500." Twenty-two pages of it were read directly for this entry, including the three signed sidebars by members of the design team. Williams's own caveat, on page 51, is recorded before anything else is taken from him:

> I wrote this article after two days of meetings with various members of the > Macintosh staff, studying preliminary Mac documentation, making numerous phone > calls to Apple, and working for several days (over a period of weeks) with a > Macintosh computer… Apple was still making minor changes to both software and > pricing when this was written.

This is a product preview written with the manufacturer's cooperation, not an independent review, and it is used here for what it is: the most detailed technical description of the machine published in the launch week, by a senior editor who also filed three predictions that section 3 grades harshly.

The machine:

| | | |---|---| | processor | Motorola MC68000 at 7.83 MHz (Apple's own materials say "8 MHz"; Apple's advertising called the machine "32-bit," which Williams flags on page 53 — the 68000 "is generally regarded as a 16-bit processor") | | memory | 128K RAM, 64K ROM | | display | 9-inch, 512 by 342, non-interlaced at 60.15 Hz, square pixels | | storage | one Sony 3½-inch drive, 400K, single-sided | | expansion | none — no slots; two serial ports Apple called "virtual slots" | | price | BYTE's "At a Glance" box gives $1995 to $2495; $2,495 is the figure Apple's announcement and every later account use | | weight | 22.7 lb for main unit, keyboard and mouse |

MacWrite and MacPaint were bundled free for the first hundred days and $195 for the pair thereafter. The ImageWriter printer was $495. A second disk drive was $395 in BYTE's text; Wikipedia's Macintosh 128K article gives $495, and the discrepancy is reported rather than resolved. Microsoft's Multiplan and BASIC were available at introduction, with File, Chart and Word promised by mid-February — a detail worth holding until section 1's last subsection, because the company supplying the launch software is the company that ended up owning the interface.

Two design decisions matter more than the specifications:

It ran one application at a time. Williams: "the absence of hardware slots and the inability to run two applications simultaneously are two important ways in which the Macintosh is fundamentally different from the Lisa computer." Hertzfeld's sidebar gives the reason without apology: "We chose the single-application-at-a-time philosophy to help keep things relatively simple."

It was sealed. From BYTE's "By the Way" box: the names of the roughly one hundred people who worked on it were molded into the inside of the rear housing, and "Don't try to look for them, though; the Mac is not supposed to be opened except by repair people." Apple was explicit that it did not expect to be cloned: Williams reports the company "has suffered from carbon-copy Apple II machines" and "does not expect to have the same thing happen with the Macintosh." Section 5 argues that this sentence is why ibm-pc-1981 is the ancestor in the header.

What was not new, which is most of it

The Macintosh's interface was fifteen years old at launch and had been on sale for three.

• Douglas Engelbart's NLS, demonstrated publicly in San Francisco in December 1968, had the mouse, windows, and hypertext links.
• Xerox PARC's Alto, from 1973, had the bitmapped display, the mouse, and overlapping windows in Smalltalk. It was a research machine and never sold.
• The Xerox Star (8010), shipped in 1981, was a commercial product with icons, folders, a desktop metaphor, Ethernet and WYSIWYG printing. Reachable accounts put its price at roughly $16,595 per workstation, with a full office installation running toward $75,000.
• The Apple Lisa, January 1983, at $9,995, had all of it and multitasking besides. Williams says so flatly: the Mac "inherited these concepts" from the Lisa — "the graphics/mouse orientation, the desktop metaphor, the data-as-concrete-object metaphor, and the shared user interface between programs."

So the Macintosh's contribution was not the interface. It was the price. From $16,595 to $2,495 is a factor of about 6.6 in three years, and it was bought with engineering that has nothing to do with human-computer interaction: about fifty integrated circuits, functions of many standard chips collapsed into eight programmable-logic arrays, two circuit boards instead of a backplane, a video display tube delivered pre-adjusted so nothing needed tuning on the line, no expansion slots to power, and the operating system in ROM so it did not have to be loaded from a disk the machine could not afford to have two of.

That last one is the hinge. The ROM was a cost-reduction that turned into a platform strategy, and the people who built it knew it.

The mechanism: the User-Interface Toolbox

The 64K ROM held, by Hertzfeld's own division, roughly a third operating system, a third Bill Atkinson's QuickDraw, and a third the User-Interface Toolbox: a collection of managers — Resource, Font, Window, Control, Menu, Text Edit, Dialog, Desk, Event — reached through about 480 new instructions grafted onto the 68000's 56 by way of its "line 1010 unimplemented" trap, dispatched through a table in RAM so Apple could patch its own bugs later.

QuickDraw is worth its own sentence because it shows the cost. Atkinson worked on it for three and a half years, rewriting it repeatedly and reducing it "from a 160K-byte compiled Pascal program to a 24K-byte package of highly optimized 68000 code." Its two most interesting capabilities, per BYTE, were filling an arbitrary shape with a pattern and clipping an image to an arbitrary mask — and the clipping is not a decoration. All QuickDraw calls are clipped to the intersection of up to three regions, "providing the fundamental capability necessary for overlapping windows." Overlapping windows are not a drawing style; they are a data structure, and Apple patented it.

The interchange half of the mechanism is in BYTE's "Sharing Data among Programs" box:

> Programs are free to implement their own data types. Apple has defined two — > text and Quickdraw pictures — and encourages every program to be able to read > and write both of these types… These two data types provide a guaranteed means > of communication among Macintosh programs.

Two types. A guarantee. Not a schema, not a negotiation, not a registry. Hertzfeld's closing line in the same issue states the ambition in one clause: the ROM firmware forms "the foundation for the development of communicating applications that share a common user interface."

And Williams, in a passage that reads oddly in 2026, calls the result open:

> Because the application drives the ROM routines (instead of the other way > around), the Macintosh is an "open" system whose behavior is completely > determined by the contents of the disk inserted into it — that is, software > designers can use the ROM routines to create a "standard" Macintosh > application, or they can write their own code to create an application that > behaves the way they want it to.

His scare quotes are load-bearing and section 4 uses them. The Macintosh was open at the layer of what an application could do and closed at every other layer: you could not open the case, add a card, run someone else's operating system, or build the machine yourself. Compare ibm-pc-1981, which published its entire BIOS as commented assembler and granted nothing. Apple published a calling convention and granted nothing. One of those got cloned to death and won the volume market; the other did not get cloned and lost it. Section 4 warns against the obvious moral.

The principles, written down afterwards

Apple eventually published the reasoning as Human Interface Guidelines, first as part of the developer documentation in the mid-1980s and then as an Addison-Wesley book. **The edition read directly for this entry is the Macintosh Human Interface Guidelines, first printing November 1992** — later than the launch, and used here for the principles rather than for what anybody knew in 1984. Chapter 1 lists eleven: Metaphors, Direct Manipulation, See-and-Point, Consistency, WYSIWYG, User Control, Feedback and Dialog, Forgiveness, Perceived Stability, Aesthetic Integrity, Modelessness.

Four sentences from it are the ones a 2026 reading will want, and they are quoted exactly:

> Consistency in the interface allows people to transfer their knowledge and > skills from one application to any other.

> The desktop is the primary metaphor for the Macintosh interface.

> Direct manipulation allows people to feel that they are directly controlling > the objects represented by the computer… an object on the screen remains > visible while a user performs physical actions on the object.

> Don't hide features in your application by using abstract commands. People > should be able to see what they need when they need it. For example, menus > present lists of commands so that people can see their choices instead of > having to remember and type command names.

The See-and-Point section states the underlying claim as a grammar: the Macintosh's two paradigms "are based on a general form of user action: noun-then-verb," and "users who are unsure of what to do next can refresh their memory by scanning through the menus. At any time, users can choose any available action without having to remember any particular command or name."

Recognition instead of recall. That is the entire intellectual content of the graphical interface reduced to five words, and it is the sentence section 2 runs against a blank prompt box.

What actually happened commercially, which is not the story

The Macintosh's first year is remembered as a triumph. The first hundred days were; the next four hundred nearly killed the company.

• First hundred days. Jobs's target was 50,000 units. By 3 May 1984 Apple had sold about 70,000–72,000. Apple projected 250,000 by year-end and roughly that many were sold in 1984.
• The holidays. Apple's marketing forecast more than 75,000 Macs a month over Christmas 1984. Andy Hertzfeld's contemporaneous-adjacent account on Folklore.org, dated May 1985, says they "didn't even break 20,000 units per month."
• The forecasts stayed up anyway. Hertzfeld reports that when Joanna Hoffman returned to marketing in early 1985 she found "sales forecasts for the upcoming quarter were unchanged from six months ago" although actual sales were about a quarter of that, because people "were afraid to tell Steve about it." Jobs, in Hertzfeld's telling, remained "oblivious to the slowing sales." The failure was legible inside the building for two quarters before it was admitted, and the mechanism was not a modelling error. It was that nobody wanted to file the number.
• The machine was underspecified for its own software. 128K and one 400K drive meant that loading a document from one disk into a program on another produced a long sequence of manual disk swaps. BYTE recorded the reaction of two of its own editors on first seeing the machine — "Only one disk drive? You've got to be kidding!" — and Williams concluded that "most users will eventually buy the second disk drive." The Macintosh 512K arrived in September 1984; the upgrade kit for existing owners was $995, which the New York Times called unfair to early buyers. The 128K was discontinued on 1 October 1985.
• The business plan failed outright. In January 1985 Apple announced the Macintosh Office — Macs, a laser printer, AppleTalk networking, and a file server — with a Super Bowl film called Lemmings, directed by Tony Scott, showing blindfolded office workers walking off a cliff. It insulted the customer it was addressed to. More to the point, the file server never shipped: the Unix-based "Big Mac" was cancelled, and AppleShare did not appear until January 1987, two years late.
• The reckoning. On 10 April 1985 Jobs was removed as general manager of the Macintosh division, and on 31 May 1985 formally relieved of his operating role. In mid-June 1985 Apple laid off about 1,200 people, its first layoffs since 1981, closed plants, and reported its first quarterly loss. (The loss figure could not be verified here: the Washington Post's 15 June 1985 report returned HTTP 403 and only a search-result summary of it was available, giving "about $40 million." The 1,200 figure is consistent across sources; the dollar amount should be checked against the filing before anyone quotes it.)
• Cumulative Macintosh sales reached about 500,000 in September 1985 — that is, roughly as many in the second nineteen months as in the first hundred days multiplied by six, on a machine whose maker had projected far more. 1986 unit sales barely matched 1984's, at a US market share around 4.2%.

A reading that cites "the Macintosh" as an adoption success is citing something that did not happen. The interface was in the box on day one. It did not sell the machine, and it very nearly did not save the company.

What did save it, which is the load-bearing fact in this entry

Three things arrived within seven months of each other and none of them is an interface:

1. PostScript. Jobs negotiated a licence from Adobe's John Warnock; agreement reached December 1983, one month before the Macintosh was announced, with Apple investing $2.5 million in Adobe. 2. The LaserWriter, announced at Apple's annual shareholder meeting on 23 January 1985 and shipping in March 1985 at $6,995. It ran a 68000 at 12 MHz with 1.5 MB of RAM — at that moment the most computing power in Apple's product line, several times what was inside the Macintosh sending it pages. Over LocalTalk it could be shared by up to sixteen Macs, putting the per-user cost under about $450. 3. Aldus PageMaker, announced the same day as the LaserWriter and shipped in July 1985, Macintosh-only.

Together these made camera-ready typeset output producible on a desk for a fraction of what a typesetting shop charged. That is a capability, not an interface — a new thing became possible that had not been possible, and the publishing and graphic-arts trades bought Macintoshes to get it. The rescue was also an accident: the LaserWriter and AppleTalk were shipped as components of the Macintosh Office, a product for corporate offices that failed, and desktop publishing was the unplanned residue of the two components that worked.

Apple's own theory of who would use the machine was wrong, and the market that saved it was a side effect of plumbing built for a different purpose.

How the interface actually reached everyone, which was not through Apple

• 10 November 1983 — Microsoft announces Windows at Fall COMDEX, two months before the Macintosh ships.
• 20 November 1985 — Windows 1.0 ships at $99. It is poorly received; the New York Times likens its performance to "pouring molasses in the Arctic." Its windows are tiled, not overlapping — which is the fact the next twelve years of litigation turn on. It sold about 500,000 copies by April 1987.
• 22 November 1985, two days later — John Sculley and Bill Gates sign a licence. Microsoft receives a "non-exclusive, worldwide, royalty-free, perpetual, nontransferable license" to specified Macintosh visual displays for Windows 1.0 and, on the accounts reachable here, for future programs. In exchange Microsoft commits to continue developing for the Macintosh, to upgrade Word, and to hold Excel exclusive to the Mac for a period. Apple's chief counsel Al Eisenstat drafted it. Gates is reported to have told Sculley that if Apple sued, "we'll stop development on all Mac products."
• 17 March 1988 — Apple sues Microsoft over Windows 2.03, which had added overlapping windows, and Hewlett-Packard over NewWave. Apple lists 189 visual displays.
• 25 July 1989 — the district court holds that 179 of the 189 were covered by the 1985 licence. Most of the remaining ten are held unprotectable.
• August 1993 — final judgment for Microsoft. 19 September 1994 — the Ninth Circuit affirms, Apple Computer, Inc. v. Microsoft Corp., 35 F.3d 1435. The Supreme Court declines to hear it. The holding a reading should carry is that Apple could not get patent-like protection, under copyright, for the idea of a graphical user interface or for the desktop metaphor, and that on the facts a virtual identity standard applied to what remained.
• 1989 — Xerox sues Apple over the same interface and is dismissed, largely because it waited too long to bring a copyright claim and was left with a weaker unfair-competition theory. Nobody ever established who owned the desktop.
• 22 May 1990 — Windows 3.0 ships at $149. Two million copies in six months, about ten million by the time 3.1 replaced it. At its release only about 5% of the roughly 40 million installed PCs ran any version of Windows; by 1993 Windows held over 90% of the PC operating-system market.
• Apple's share of personal computers fell to around 3% by the late 1990s, and the company came close to failing in 1997 — the year Microsoft invested $150 million in it and the remaining claims were settled.

The Macintosh introduced the mechanism. Microsoft distributed it, legally, under a licence Apple signed in exchange for a spreadsheet. Any reading using 1984 as the date the graphical interface reached ordinary people is off by about six years and has the wrong company.

---

2. Why a reading would cite it

Four cases, mapped to LENSES.md. The first is the entry's job.

(a) Concentration, and the public square — the interface is the distribution channel, and whoever defines it collects a toll on everything that runs through it. The Macintosh's Toolbox is the earliest clean instance of a mechanism now running at scale: define the layer everyone else has to build against, make conforming cheaper than diverging, and you own the developers without owning the developers. In 2026 the same move is visible and datable:

• March 2025 — OpenAI adopts the Model Context Protocol across the Agents SDK, the Responses API and the ChatGPT desktop app.
• April 2025 — Google DeepMind confirms MCP support in Gemini; MCP support in VS Code Copilot reaches general availability in July 2025.
• 9 December 2025 — Anthropic donates MCP to the Agentic AI Foundation under the Linux Foundation. Anthropic, Block and OpenAI are named co-founders; AWS, Google, Microsoft, Cloudflare and Bloomberg join as platinum members.

That is the Toolbox with the ROM removed and the governance added, and the difference is exactly the variable this entry says to watch. Apple put the interface where nobody else could reach it, gave it away in a contract two years later without meaning to, spent six years litigating, and lost. MCP's owner gave it to a foundation at the fourteen-month mark. Whether the second choice was the better one is not knowable yet, and it is the right question to be asking. The transferable questions this entry supplies, in order: who defines the interchange format; what does a developer save by conforming; who can change it unilaterally next year; and what happens to the conforming developers when the definer's own commercial interest moves.

The toll is not hypothetical, and this project's own reading of 19 August 2026 recorded its current form: OpenAI announced that ChatGPT advertising would expand to 31 European markets from 24 August, with the company's enterprise marketing chief describing about a billion weekly active users of whom roughly 20% show commercial intent. An interface with a billion people in it is inventory. The Macintosh's interface was sold once, in a box; the 2026 interface is rented continuously and monetised by selling access to the attention inside it. That is a difference in kind and it should be stated as one rather than smoothed over by the analogy.

(b) Capabilities — dating the technology from the product is the field's most common error, and this is the cleanest historical case of it. chatgpt-2022 already establishes the point for 2022: the capability was purchasable from June 2020 and instruction-tuned by default from January 2022, and "nothing became possible on 30 November 2022." The Macintosh is the same error committed forty years earlier and never corrected in popular memory. Engelbart demonstrated it in 1968, PARC built it in 1973, Xerox sold it in 1981 for $16,595, and Apple sold it in 1983 for $9,995 before selling it in 1984 for $2,495. What changed in 1984 was a price and a distribution decision.

Cite this against any claim of the form "X became possible in [product launch year]." And cite it with the specific correction the BYTE preview supplies for free: on page 53, Williams notes that Apple's own advertising called the Macintosh a 32-bit machine when its processor was generally regarded as 16-bit. A contemporaneous technical reviewer catching a vendor rounding a specification in its own favour, in the launch week, is precisely the service this canon wants a reading to look for when a 2026 model card, benchmark claim or context-window number arrives.

(c) The public square — an interface is necessary and not sufficient, and whether it is sufficient depends on one variable. This is the entry's correction to its own proposal and the most useful thing in it.

The Macintosh's interface was complete on day one and the machine nearly died. ChatGPT's interface was a text box in front of a model that had been on sale for two and a half years and it reached a hundred million people in two months. The difference is not interface quality. It is the marginal cost of the first use. Meeting the Macintosh cost $2,495 for the machine, plus $395 or $495 for the second drive you would need, plus $495 for a printer, plus a dealer visit; meeting the thing that made it worth having cost $6,995 more. Meeting ChatGPT cost a URL.

So the rule a reading should carry is conditional, not universal: where the marginal cost of a new user is near zero, the interface is the binding constraint and removing friction produces the whole adoption curve. Where the marginal cost is a purchase, the interface is necessary and a capability decides the outcome — and the capability that decided this one was PostScript typesetting, delivered on a peripheral. Any 2026 claim that a chat box, an agent harness, or a device is about to unlock mass adoption should be asked which regime it is in and what the first use costs the user.

(d) The public square — the 2026 interface reverses the principle the Macintosh established, and hardly anybody says so. The Human Interface Guidelines state the graphical interface's thesis as "menus present lists of commands so that people can see their choices instead of having to remember and type command names" and "users… can choose any available action without having to remember any particular command or name." Recognition, not recall.

A blank prompt box is the opposite of that, precisely and by construction. It shows the user nothing, offers no enumeration of available actions, and places the entire burden of knowing what the system can do on the person typing. The 2026 design literature has independently rediscovered this — a run of pieces through 2026 argue that chat is the wrong default, that the blank text box is a tax, and that the burden of context falls on the user — but they argue it as a new observation. It is not new. It is 1984's argument against the command line, restated against a command line that answers in English.

That is a genuine finding and it should be cited with its limit attached: the prompt box also does something no menu can, which is accept an instruction nobody built a menu item for. The Macintosh's bargain was that you could only do what was on the menu, and the menu was written by the vendor. Cite this whenever the current interface generation is described as a straightforward advance; the honest description is a trade of discoverability for expressiveness, made by every product at once, mostly without anyone writing down that a trade was occurring.

The honest limits, stated here rather than buried. This is one company, one machine, one interface generation, in a market of a few million units where software shipped on physical disks and a developer's constraint was 80K of usable memory. The ROM mechanism worked because memory was scarce and expensive; where the conforming path is not also the cheap path, the mechanism does not transfer, and in 2026 nothing about integrating with a protocol is cheap in the way that calling a ROM routine was. This entry is a base rate for how interface standards propagate and who ends up holding them, not a forecast, and the second half of section 1 — where the standard's author lost the standard to its own launch software vendor — is the part most likely to be left out when someone reaches for the analogy.

The disclosure. MCP in case (a) was created by the company that makes the model writing this file, and its donation to a foundation is described here in terms that could read as favourable. The convenient errors run both ways — flattering the author's maker, or overstating the risk to look independent. The defence is that the mechanism is stated as a mechanism rather than a virtue, that the same four questions are asked of MCP as of Apple's ROM, that the 2026 example in this section that involves selling attention is a competitor's and is stated neutrally, and that section 4 explicitly refuses the conclusion that publishing an interface is a public good. Apple's Toolbox was published and consolidated a platform; IBM's BIOS was published and dispersed one. Publication predicts neither.

---

3. What it got right, and what it got wrong

idea does not require this section. It is here because there are five dated claims on the table and each is worth more graded than repeated.

Claim 1 — Jef Raskin, "Computers by the Millions," late 1979

The Macintosh project's founding document, written by the man who started it and named it, circulated internally, suppressed from publication at Mike Markkula's request "in order to not get our competition thinking in these directions," and finally published in the SIGPC Newsletter in 1982. Its opening:

> If "personal computers" are to be truly personal, it will have to be as likely > as not that a family, picked at random, will own one. To supply even our own > nation with enough computers to make this happen (over, say a four-year period) > we will require, for example, twenty-five companies each producing over a > million computers a year.

And the architectural prediction:

> The third generation personal computers will be self-contained, complete, and > essentially un-expandable.

Right about the machine, wrong about the timetable, and right in a way that took twenty-five years to become obvious.

The four-year saturation forecast — a computer in half of American households by about 1983 — was badly wrong; apple-iie-1983 carries the measurement that settles it, from the Census Bureau's October 1984 Current Population Survey: 8.2% of American households had a computer, and 1.7% of households below $10,000 did. Raskin was off by roughly a factor of six on penetration and by about fifteen years on the date.

The un-expandable prediction is the interesting one. It was the single most criticised feature of the shipping Macintosh — BYTE devoted paragraphs to the absent slots and Apple retreated from it inside three years, adding an expansion slot to the Macintosh SE and six to the Macintosh II in 1987. On a five-year horizon Raskin lost. On a twenty-five-year horizon he is the only person in this file who was completely right: every phone, tablet and laptop sold in 2026 is self-contained, complete and essentially un-expandable, and the reasoning he gave — that expandability creates a combinatorial "software nightmare" that mass manufacture cannot support — is the reason.

The transferable finding: a claim about what mass production forces can be right about the endpoint and useless about the schedule, and a canon that grades only against the near date will systematically misjudge that class of claim. Note also that Raskin's memo records Steve Jobs as the one reader who disliked it — and Jobs then built the machine, took the credit, and shipped the un-expandable design Raskin had argued for.

Claim 2 — Gregg Williams, BYTE, February 1984, three predictions

Made in the launch week, in the Conclusions on pages 53–54.

2a. "It will be imitated but not copied." Williams elaborated:

> In a similar way, companies will be out to imitate the Macintosh, but their > attempts will be less successful. Those companies that try to imitate the Mac > on other machines will have trouble matching its price/performance > combination… The only other way to match the Mac would be to design an > entirely new system that would be comparably priced. This will probably not be > attempted; only a few corporations have the ability to duplicate Apple's design > and manufacturing effort, and still fewer will make such a large financial > commitment. (Apple is the only American company that does not live under the > tyranny of next quarter's profits; if any company tries to duplicate Apple's > effort, it will probably be a Japanese one.)

Comprehensively wrong, and wrong in the funniest available direction. The imitation succeeded, it ran on other machines, it was performed by an American company, and that company is named twice on page 48 of the same article as the supplier of the Macintosh's launch software. Microsoft had already announced Windows three months before Williams wrote; Apple licensed the visual displays to it twenty-two months later; Windows 3.0 in May 1990 put the interface on ten million machines; Apple sued, litigated for six years, and lost on the ground that 179 of its 189 claimed elements were covered by its own contract.

The transferable finding is not that Williams was foolish. It is the specific shape of the error: he predicted correctly that no one could match Apple's hardware economics, and the imitation did not come through hardware. It came through software licensed onto hardware built by other people. A moat measured in the wrong layer is not a moat, and this is the base rate to hold against any 2026 claim that a capability is defended by manufacturing scale, data volume or integration.

2b. "The Macintosh will secure the place of the Sony 3½-inch disk as the magnetic medium of choice for the next generation of personal computers." Right, and it is the only unqualified hit in this file. The 3½-inch disk became the universal removable medium for about fifteen years. Worth recording because it is the least consequential of his three forecasts and the only one he got: the forecaster was accurate about the commodity and wrong about the strategy, which is a common and under-noticed pattern.

2c. "To some people Apple will be as synonymous with the phrase 'personal computer' as IBM is synonymous with 'computer'… The Mac will delay IBM's domination of the personal computer market." Half right for the wrong reason. IBM never dominated the personal computer market, and by December 2004 it had sold the business to Lenovo — but as ibm-pc-1981 establishes at length, what prevented IBM's domination was the clone industry its own published BIOS created, not the Macintosh. Apple's share fell to about 3%. Apple did eventually become synonymous with personal computing, four decades later, by way of a telephone.

2d, recorded separately because it is a flat factual prediction. On page 43: "As a result, despite some rumors to the contrary, the Macintosh will never use IBM PC- or MS-DOS-based software." Wrong within four years and repeatedly afterwards — Insignia Solutions' SoftPC reached the Macintosh in 1987–88, Apple shipped DOS Compatibility Cards in the mid-1990s, and Boot Camp booted Windows natively on Intel Macs from 2006. Note the reason he gave: the sealed, slotless architecture made a processor card impossible. He was right about the hardware constraint and wrong about the conclusion, because the constraint was routed around in software. That is the same error as 2a, one page earlier, and it is the error to watch for in 2026 arguments that a hardware or access restriction bounds what can be done.

Claim 3 — Apple's own, 24 January 1984

From the announcement: users tell the Macintosh what to do "simply by moving a 'mouse'… to select among functions listed in menus and represented by pictorial symbols on the screen," the machine typically requires "only a few hours to learn," and — Jobs, quoted in the release — "With Macintosh, the computer is an aid to spontaneity and originality, not an obstacle… Macintosh enhances not just productivity, but also creativity."

Right about the learning claim, and this is the part of the Macintosh story that deserves its reputation. The mouse-and-menu model did make the machine learnable in hours by people who had never used a computer, and every measurement and every competitor's behaviour since has confirmed it: within six years the principal rival had copied it, and no mass-market computing device has shipped without it since.

Unproven on creativity, and the entry declines to supply the anecdote. The strongest honest support is the desktop-publishing trades, where output that had required a service bureau became producible in-house — a real and large effect on who could publish. It is also a productivity finding wearing a creativity label, and lego-mindstorms-1998 and apple-iie-1983 both catalogue the identical evidentiary trap: a great many designers date their start to a Macintosh, which is a self-report from a self-selected group. No study establishing that the graphical interface increased creative output was found for this entry, and none is invented here.

Claim 4 — the proposal's, canon/proposals.md, 17 August 2026

> the Macintosh and the graphical interface. Cite when: an interface, rather > than a capability, decides how many people can use a thing — the current > analogue is a chat box in front of a model that already existed.

Graded in three pieces.

• "a chat box in front of a model that already existed" — exactly right, and the strongest clause in it. chatgpt-2022 documents the 307 days between instruction-tuned models becoming the API default and the launch that everyone dates the field from. The structural parallel to 1984 is real and it is the reason this entry belongs in the canon.
• "an interface, rather than a capability, decides how many people can use a thing" — right in general, and wrong about the Macintosh specifically, which is this entry's contribution. The interface was in the box in January 1984 and the machine was in trouble by January 1985: forecasts of 75,000 a month against actuals under 20,000, a business product whose file server never shipped, 1,200 people laid off in June, the founder removed. What moved it was PostScript on a $6,995 printer — a capability, arriving fourteen months late, on a peripheral. The corrected rule is in section 2(c) and it is conditional on the marginal cost of the first use. Stated unconditionally, the claim would have predicted a 1984 success, and 1984 was not one.
• "the Macintosh and the graphical interface" — the wrong object, kept anyway. The Macintosh did not originate the graphical interface; Engelbart demonstrated it in 1968, PARC built it in 1973, Xerox sold it in 1981 and Apple sold it in 1983. What the Macintosh originated is the ROM-resident toolbox that made third-party conformance economically rational, which is a narrower and far more transferable object. The id is kept because ids are stable and prose can carry the correction, which is the ruling apple-iie-1983 made for the same situation two days earlier.

Claim 5 — the implied claim of the 1985 licence

Trading the interface for continued application support is worth it. Due continuously from 22 November 1985.

Wrong, catastrophically, and the counterfactual is not available. Apple gave Microsoft a perpetual, royalty-free, worldwide licence to the visual displays in exchange for a Word upgrade and a period of Excel exclusivity, under a reported threat that Microsoft would otherwise stop developing for the Macintosh at a moment when Macintosh software was scarce and the company was losing money. Six years of litigation later, 179 of the 189 elements Apple complained of were held to be inside that licence.

The honest complication, which is why this is graded "wrong" rather than "stupid": in November 1985 Apple had just laid off a fifth of its workforce and Microsoft supplied a large share of the Macintosh's serious applications. Losing them could plausibly have ended the platform. And the 1994 opinion suggests the outcome might have been similar without any licence at all, since copyright does not protect the idea of a graphical interface or the desktop metaphor. The transferable finding is about the asymmetry, not the judgment: a platform owner negotiating from a cash-flow emergency will trade a durable asset for a current-quarter one, and the asset traded is usually the interface, because it is the thing whose value is not on any balance sheet. Cite this when a 2026 AI company under funding pressure licenses, opens, or standardises something whose value is strategic rather than accounted.

---

4. Commonly misused as

build.py reserves this section for limit. It is here because this entry exists to be used as an analogy and the analogy is habitually run in reverse.

1. "Apple stole the graphical interface from Xerox." The most repeated claim about this machine, and it is wrong in both directions at once. It was purchased, not stolen: on the accounts reachable here, Xerox received an option to buy 100,000 pre-IPO Apple shares at $10 — about $1 million, worth enormously more after the December 1980 IPO — in exchange for access, and PARC's own researchers were reportedly not told about the financial side, which is where the sense of grievance came from. And it was not a copy: there were two visits in December 1979, Apple's teams were already working on bitmapped graphical designs, and the Macintosh's most distinctive elements were not in the demo. The pull-down menu, the one-button mouse, the drag-and-drop Finder, the trash can and Atkinson's region-based overlapping windows were Apple's; the Star used tiling for its main windows. When Xerox eventually sued in 1989 the case was dismissed without ever reaching the question of whether the interface had been taken.

A caution about the retelling, recorded because it is what this canon is for. A widely cited secondary account of the litigation, read for this entry, states that Apple paid Xerox "$100 million in stock" for a GUI licence. That is off by two orders of magnitude from every account of the actual pre-IPO share option, and the same page correctly reports the 1989 dismissal on the ground that Xerox had waited too long — which is not consistent with Xerox having licensed the technology to Apple in the first place. The figure should not be repeated. Larry Tesler, who gave the demo and later joined Apple, is the source of the line that is worth keeping: the visitors "understood our technology and what it meant, more than any Xerox executive understood after years of showing it to them." The story is not theft. It is a firm that could invent a thing and could not recognise what it had, and that is the useful shape for 2026.

2. "The Macintosh made the graphical interface mainstream." It did not. Windows 3.0 did, on 22 May 1990, on hardware Apple did not make, under a licence Apple had signed. At Windows 3.0's release about 5% of roughly 40 million installed PCs ran any Windows; within three years Windows held over 90% of the PC operating-system market while Apple's share of personal computers fell toward 3%. The 2026 form of this error is treating whoever demonstrates an interface first as the party who will hold the resulting market. Introducing a standard and capturing it are separate events with separate causes, and in the one fully resolved case in this canon they went to different companies six years apart.

3. "Open beats closed" — or its inverse, "integrated beats modular." The two canon entries on this shelf are a matched pair and they do not support either slogan. ibm-pc-1981 published its entire BIOS as commented source, granted nothing, was cloned within two years, took the volume market and lost the platform to the cloners. The Macintosh published a calling convention, granted nothing, was not cloned, kept a small high-margin position for forty years, and lost the interface anyway — to a licensee, by contract, not by copying. Publishing an interface disperses a capability; withholding one does not concentrate it. What determined both outcomes was where the interface sat relative to where the value was, and in both cases the incumbent misjudged that. Any 2026 argument about open weights that reaches for either machine as precedent has to say which layer before the analogy means anything.

4. "'Look and feel' is protectable, and a great interface is defensible." Twelve years of litigation across two suits established the opposite. Apple v. Microsoft held that Apple could not obtain patent-like protection, under copyright, for the idea of a graphical user interface or for the desktop metaphor, and that most of what remained after the licence was unprotectable as functional or merged with the idea. Xerox v. Apple was dismissed on timing. Nobody has ever owned the desktop. Cite this against 2026 claims that an interaction pattern, a prompt format, a scaffolding convention or an agent protocol is a defensible asset. The historical answer is that the interface is the least defensible layer in the stack and the most valuable one to have been first at, which is a genuinely uncomfortable combination and is why so many people misremember the outcome.

5. "The Macintosh proves interface beats capability." The exact inversion of what happened, and the reason section 3 grades the proposal rather than quoting it. The Macintosh shipped with the best interface anyone had built, was outsold by machines with no interface at all, and was rescued by a page-description language on a peripheral. The interface was necessary and it was not sufficient. A reading that runs this entry as evidence that a good interface produces adoption has inverted the file's central finding; the conditional version is in section 2(c) and it turns on what the first use costs.

6. "The '1984' commercial launched the Macintosh." It aired once nationally, during the third quarter of Super Bowl XVIII on 22 January 1984, to a reported 96 million viewers; Apple's board is said to have hated it and Sculley tried to sell the time back, managing to unload only the thirty-second slot. It is the most celebrated advertisement ever made and it does not show the product, name a feature, or state a price. Meanwhile the sequel — Lemmings, for the Macintosh Office, at the following year's Super Bowl — is a case study in insulting the buyer, and the product it advertised was missing its central component and never shipped it. Advertising recall is not adoption, the same agency produced both, and the company that produced the most admired commercial in history was laying off 1,200 people seventeen months later.

---

5. Ancestry

descends_from is [ibm-pc-1981], one edge, and the edge is strategic rather than technical. No wire runs from an 8088 to a 68000, and nothing in the Macintosh's interface came from Boca Raton. What descends is the decision.

ibm-pc-1981 states its mechanism as two parts: publish the interface completely, and grant nothing. The Macintosh is the second half executed without the first, and Apple's own launch materials say it was chosen with the IBM outcome in view. BYTE records, in the launch week, that Apple "has suffered from carbon-copy Apple II machines" and "does not expect to have the same thing happen with the Macintosh" — and then describes the design decisions that follow from that expectation: no published schematics, no expansion slots, a sealed case the user is told not to open, custom components Apple believed would be hard to duplicate, and the operating system in a ROM protected by copyright rather than in a manual anyone could buy. The Macintosh's most distinctive architectural features are a response to what happened to a machine announced thirty months earlier, which is descent that explains shape, and that is the house test.

The objection, stated because it is a real one. apple-iie-1983 considered ibm-pc-1981 for its own header two days before this file and declined it, ruling that the two describe "opposite mechanisms" and that contrast is not descent. The distinction claimed here is that the Apple IIe entry's link to IBM was thematic, whereas this one is documented in the primary source at the moment of the decision: Apple is on the record, in the launch week, explaining its architecture by reference to cloning. A later session that disagrees should say so in its own file rather than edit this header.

What is deliberately not claimed, and why the header is one edge rather than five. The Macintosh's genuine technical ancestors are absent from canon/, and this is a gap in the canon rather than a fact about the machine — the same situation chatgpt-2022 recorded when it found that none of its self-declared parents had files.

• Engelbart's NLS and the 1968 demonstration. The mouse, windows and interactive editing, nine years before Apple existed. No id, and none proposed.
• Xerox PARC, the Alto and Smalltalk. The direct source of what Jobs and Atkinson saw in December 1979. No id, and none proposed — which is a notable hole, since arpanet-1969 already carries PARC as the place ARPA's people went and records an Engelbart meeting, without either becoming an entry.
• The Xerox Star, 1981. The first commercial graphical computer, and the correct answer to "when did this become a product." No id.
• The Apple Lisa, January 1983. The machine the Macintosh is a cost-reduction of, by Williams's own account. No id.
• Jef Raskin's "Computers by the Millions," 1979. The project's founding document and the source of the un-expandable appliance. No id.

None of them is invented into the header. If the chain ever writes one, this entry should acquire the edge then; the strongest candidates in that list are the Star, because it is the entry that would fix the dating error section 2(b) is about, and Raskin, because he is the only forecaster in this file who was right about the endpoint.

Siblings, where the contrast is the useful part and the graph edge would be wrong:

• apple-iie-1983. The same company in the same eighteen months, and the two read together make a point neither makes alone: the machine Apple gave away under a tax credit outsold and outlived the machine it bet the company on. The IIe funded the Macintosh's near-death year and stayed on sale until 15 November 1993, longer than the Macintosh's first decade. A vendor's strategic product and its actual business are frequently different objects.
• chatgpt-2022. The direct 2026 analogue and the entry this one exists to sharpen. chatgpt-2022 establishes that a launch can be entirely about distribution and contain no capability event; this entry establishes the boundary condition on that finding — that it holds when the first use is free and fails when the first use costs $2,495.
• eliza-1966. The canon's other entry about an interface doing work the capability underneath did not justify, and it already concludes that "the finding points at the interface, not the person." Where ELIZA's interface created an impression of understanding that was not there, the Macintosh's created an impression of simplicity that was: the machine really was learnable in an afternoon. Interfaces can overstate or accurately state what is behind them, and telling which requires looking behind them, which is the whole discipline.
• printing-press-1440. The desktop-publishing rescue in section 1 is a cost collapse in the production of printed pages, six centuries after the first one, and a reading writing about who can publish may want both. It is an analogy and not an ancestor, and the header does not claim it.
• graffiti-palm-1996. Twelve years downstream, the same trade run the other way: Graffiti shipped by teaching humans an alphabet the machine could read, where the Macintosh shipped by teaching the machine to display objects humans already recognised. Both are interface decisions that determined whether a capability was usable, and the pairing is worth a reading's time.

Ids named in this file that do not exist in canon/ and are not being invented into the header: Engelbart's 1968 demonstration, Xerox PARC, the Alto, the Xerox Star, the Apple Lisa, Raskin's 1979 memo, PostScript, PageMaker, Windows 3.0, and the Model Context Protocol — none proposed anywhere, all appearing here only as prose.

---

6. Sources

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

• **Gregg Williams, "The Apple Macintosh Computer," BYTE, Vol. 9 No. 2, February 1984, pp. 30–54. Read directly, as rendered PDF pages, from mwftr.com/416F20/Apple macintosh BYTE1984.pdf (pages 1–22 of the file, covering magazine pages 30–54). The primary source for: the subtitle "Mouse-window-desktop technology arrives for under $2500"; the "At a Glance" box ($1995–$2495, 128K RAM, 64K ROM, one 400K Sony drive, 512 by 342 at 60.15 Hz, 58-key keyboard, mechanical mouse, 22.7 lb, ImageWriter $495, MacWrite and MacPaint bundled free for 100 days then $195 the pair); the 7.83 MHz clock against Apple's advertised "8 MHz" and against the Lisa's 5 MHz; the "32-bit" advertising correction on page 53; the roughly fifty ICs and eight PALs and two boards; the "virtual slots" serial scheme and the statement that a Microsoft Softcard-style processor card could not work through it; the single-application limitation; the launch software list including Microsoft Multiplan and BASIC at introduction and File, Chart and Word by mid-February, and Apple's estimate of "more than 100 software vendors"; the external drive at $395 and keypad at $99; the "Only one disk drive? You've got to be kidding!" reaction and the conclusion that most users would buy the second drive; the "By the Way" box with the molded names and the instruction not to open the case; the statement that Apple did not expect the Macintosh to be cloned as the Apple II had been; the "Caveats" paragraph quoted in full above; and the entire Conclusions section, which is the source for all of claim 2. Also the three signed sidebars: Burrell C. Smith, "Macintosh System Architecture" (the triple-ported RAM, the 6522 VIA, the IWM disk controller, the 68000 averaging about 6 MHz of effective RAM access); Andy Hertzfeld, "Macintosh System Software Overview" (the ROM's thirds, the ~480 trap instructions, the "positive incentive for doing it our way" passage quoted above, the single-application philosophy, and the "foundation for the development of communicating applications that share a common user interface" line); "The User-Interface Toolbox" (the manager list, and Atkinson's three and a half years reducing QuickDraw from 160K of compiled Pascal to 24K of 68000 code, with clipping to the intersection of up to three regions as "the fundamental capability necessary for overlapping windows"); and "Sharing Data among Programs"** (the clipboard, the two defined data types, and "a guaranteed means of communication among Macintosh programs").
• ***Macintosh Human Interface Guidelines, Apple Computer / Addison-Wesley, ISBN 0-201-62216-5, first printing November 1992. Read directly, as rendered PDF pages, from apps.hci.rwth-aachen.de/borchers-old/cs377a/handouts/HIGuidelines.pdf (front matter, contents, and Chapter 1 "Human Interface Principles," book pages 3–8). Source for the eleven principles and for all four quoted passages — Metaphors and "the desktop is the primary metaphor," Direct Manipulation, See-and-Point with the noun-then-verb formulation and "without having to remember any particular command or name," Consistency's "transfer their knowledge and skills from one application to any other," and WYSIWYG's "see their choices instead of having to remember and type command names." This is the 1992 edition, not the 1985 developer documentation or the 1987 Addison-Wesley Human Interface Guidelines: The Apple Desktop Interface***, and it is used here for the principles rather than as evidence of what was written down in 1984. A later session wanting the 1987 text should find it; it was not read here.
• Jef Raskin, "Computers by the Millions," late 1979. Read via fetch, from digibarn.com/friends/jef-raskin/writings/millions.html, for the opening paragraph and the "third generation personal computers will be self-contained, complete, and essentially un-expandable" passage, both quoted verbatim above. The circulation history — Markkula's request that it not be published "in order to not get our competition thinking in these directions," the 1982 appearance in SIGPC Newsletter Vol. 5 No. 2, and Raskin's 3 April 1980 memo recording that every reader enjoyed the draft "with the exception of Steve Jobs" — is from Stanford University Libraries' Macintosh history pages (web.stanford.edu/dept/SUL/sites/mac/primary/docs/cbmmemo.html and cbmmemo2.html), read via fetch.
• Andy Hertzfeld, "The End Of An Era," Folklore.org, dated May 1985. Read via fetch. Source for the holiday forecast of more than 75,000 Macs a month against actuals that "didn't even break 20,000 units per month"; Joanna Hoffman finding forecasts unchanged from six months earlier because people "were afraid to tell Steve about it"; the "growing reality gap"; and Jobs "oblivious to the slowing sales." A participant's memoir written about twenty years after the events, and used for the internal texture rather than for the numbers, which are corroborated in direction by every other account here.
• Apple's 24 January 1984 announcement, as reproduced and quoted by MacRumors, "The Mac Turns 40: Read Apple's Announcement From 1984," 23 January 2024 — read via fetch. Source for the "simply by moving a 'mouse'" description, the "only a few hours to learn" claim, and the Jobs quotation ending "Macintosh enhances not just productivity, but also creativity." The original press release was not read; a later session should find it, and the Computer History Museum's Macintosh Selling Guide, 1984 (s3data.computerhistory.org/brochures/apple.macintosh.1984.102646178.pdf) appeared in search results and was not fetched.
• Wikipedia, "Macintosh 128K." Read via fetch. Source for the 24 January 1984 date at $2,495, the 3 May 1984 figure of about 70,000 units, the year-end projection of 250,000, the September 1984 Macintosh 512K and its $995 upgrade kit with the New York Times criticism, the external drive at $495 (which differs from BYTE's $395 and is reported unreconciled), and the 1 October 1985 discontinuation.
• Wikipedia, "LaserWriter." Read via fetch. Source for the announcement at the annual shareholder meeting on 23 January 1985, March 1985 shipment at $6,995, the 68000 at 12 MHz with 1.5 MB of RAM being the most processing power in Apple's line, the December 1983 PostScript agreement with John Warnock one month before the Macintosh was announced, Apple's $2.5 million investment in Adobe, and LocalTalk sharing among sixteen Macs at under $450 a user.
• Wikipedia, "Macintosh Office." Read via fetch. Source for the January 1985 announcement, the AppleTalk Personal Network at 230 kbit/s, the never-shipped Unix-based "Big Mac" file server, AppleShare arriving only in January 1987, and the observation that the LaserWriter and AppleTalk inadvertently produced desktop publishing.
• Wikipedia, "Windows 1.0" and "Windows 3.0." Read via fetch. Source for the November 1983 COMDEX announcement, the 20 November 1985 release at $99, the New York Times's "pouring molasses in the Arctic," Windows 1.0's tiled rather than overlapping windows, and 500,000 copies by April 1987; and for Windows 3.0 on 22 May 1990 at $149, two million copies in six months, about ten million by Windows 3.1, only about 5% of roughly 40 million installed PCs running any Windows at release, and PC Magazine's Bill Howard finding it "easy to use, though not quite as intuitive as Macintosh."
• Wikipedia, "Apple Computer, Inc. v. Microsoft Corp." Read via fetch. Source for the 1988 filing, the trigger being Windows 2.0's overlapping windows, the 189 elements of which the court found 179 licensed, the 19 September 1994 Ninth Circuit decision, the denial of Supreme Court review, the 1997 settlement, and the dismissal of Xerox's suit on the ground that it had waited too long for a copyright claim and was left with unfair competition.
• "The Apple vs. Microsoft GUI Lawsuit," Low End Mac, 2006. Read via fetch. Source for the 22 November 1985 signing by Sculley and Gates one week after Windows 1.0's release, Bill Neukom's presence, Al Eisenstat drafting the contract, Gates's reported "we'll stop development on all Mac products," the Excel exclusivity term, the 17 March 1988 filing without advance notice, Judge Schwarzer's 25 July 1989 ruling on 179 of 189 displays, and the August 1993 final judgment. This page also states that Apple paid Xerox "$100 million in stock" for a GUI licence, which section 4 rejects; the rest of its chronology is consistent with the other sources here and is used with that caveat attached.
• The 1985 licence's operative language — the "non-exclusive, worldwide, royalty-free, perpetual, nontransferable license" formulation, and the Word-upgrade and Excel-exclusivity consideration — is from search-result summaries only. tech-insider.org/windows/research/1985/1122.html, which appears to reproduce the agreement text, returned HTTP 403, as did the Justia pages for both Apple v. Microsoft opinions (35 F.3d 1435 and 759 F. Supp. 1444) and the openjurist mirror. The 1994 opinion was therefore not read in the original, and the holding quoted in section 4 — that Apple could not get patent-like protection for the idea of a graphical user interface or the desktop metaphor, and the virtual-identity standard — comes from search summaries of it. This is the weakest sourcing in the file on a point section 4 leans on, and a later session should read the opinion.
• The Xerox PARC visit. Search-result summaries only, drawn from MakeUseOf, Newsweek's "The Story of Steve Jobs, Xerox and Who Really Invented the Personal Computer," Mac History and Stanford retrospectives. Source for the option on 100,000 pre-IPO shares at $10, the two visits in December 1979, the Alto running Smalltalk, PARC researchers not being told about the share arrangement, and Larry Tesler's "more than any Xerox executive understood after years of showing it to them" as relayed from Triumph of the Nerds (1996). The Newsweek piece was fetched and turned out to contain none of the technical detail, which is recorded so nobody re-fetches it. The claim that the pull-down menu, one-button mouse, Finder, trash can and region-based overlapping windows were Apple's rather than PARC's, and that the Star used tiling for main windows, rests on search summaries and on the Wikipedia QuickDraw and Xerox Star material; the Star's roughly $16,595 price is from a search summary of retromobe and a 2026 Medium piece and was not verified against a Xerox price list. Confidence across this paragraph is a notch lower than elsewhere in the file.
• The 1985 collapse. Jobs's removal as general manager on 10 April 1985 and formal relief on 31 May 1985, the roughly 1,200 June layoffs, the plant closures and the first quarterly loss are from search-result summaries, including of the Washington Post's "Apple Computer To Lay Off 1,200 Workers" (15 June 1985), which returned HTTP 403 on fetch. The dollar figure for the loss is not asserted here for that reason. The 500,000 cumulative Macs by September 1985 and the 4.2% 1986 share are from search summaries of Macworld and Low End Mac.
• The "1984" and "Lemmings" advertisements. Search-result summaries only. The 22 January 1984 airing during the third quarter of Super Bowl XVIII, Ridley Scott's direction for Chiat/Day, the reported 96 million viewers, the board's hostility and Sculley's attempt to sell back the time; and Lemmings at the January 1985 Super Bowl, directed by Tony Scott, and its reception. The Smithsonian's own account returned HTTP 403. Sources disagree on the date of the qualifying pre-Super-Bowl airing in Twin Falls, Idaho — 15 December versus 31 December 1983 — and the disagreement is reported rather than resolved.
• SoftPC, the DOS Compatibility Cards and Boot Camp, used only to grade claim 2d. Search-result summaries only, from Wikipedia's SoftPC and Boot Camp articles and Low End Mac; SoftPC's Macintosh port is given variously as 1987 and 1988 and both are reported.
• The Model Context Protocol timeline. Search-result summaries only, of the Linux Foundation's press release announcing the Agentic AI Foundation (9 December 2025), the Model Context Protocol blog's own post of the same date, Anthropic's donation announcement, and WorkOS's 2026 overview. Source for OpenAI's March 2025 adoption, Google DeepMind's April 2025 support, VS Code Copilot GA in July 2025, the founding and platinum members, and the reported ~97 million monthly SDK downloads. No primary announcement page was fetched successfully, and the download figure is a vendor-ecosystem number of the kind LENSES.md warns about.
• The 2026 adoption and monetisation figures in section 2(a) are taken from this project's own reading of 19 August 2026 (digests/2026-08-19-12.md), which cites OpenAI's announcement of ChatGPT advertising expanding to 31 European markets from 24 August and the company's enterprise marketing chief describing about a billion weekly active users of whom roughly 20% show commercial intent. Using the project's own digest rather than re-deriving the number is deliberate; the digest carries the link to the primary announcement. A search for independent 2026 usage totals returned mostly statistics-farm pages and none is used here.
• The 2026 "chat is the wrong interface" argument. Search-result summaries only, of pieces published through 2026 in UX Collective, Adaline Labs, HackerNoon and elsewhere. These are arguments, not evidence, and section 2(d) uses them only to establish that the argument is being made and that it is being made as though it were new.

What could not be established, listed so nobody re-does the work: the text of the 22 November 1985 Apple–Microsoft agreement and both Apple v. Microsoft opinions in the original (four separate 403s); Apple's actual reported loss for the quarter ended June 1985 and the underlying 10-Q; Apple's original 24 January 1984 press release in full; the 1987 first Addison-Wesley edition of the Human Interface Guidelines, which is the edition that would show what was written down closest to the launch; a verified Xerox price list for the Star 8010; whether any contemporaneous measurement was ever made of how long it actually took a novice to learn the Macintosh, against Apple's "only a few hours" claim — none was found, and the claim in section 3 is graded on the behaviour of competitors and the market rather than on a study; and how many of the roughly 500,000 Macs sold by September 1985 went to the desktop-publishing trades, which is the number that would turn section 1's rescue narrative from a well-supported account into a measured one.