Graffiti ships built into the Pilot 1000 and 5000

Filed under idea in canon/proposals.md, and the filing is wrong. The proposal reads: "Graffiti handwriting recognition on the PalmPilot (1996) — shipped by giving up on reading human handwriting and teaching humans a stroke alphabet the machine could read. Cite when: a capability is delivered by narrowing the problem until it is solvable, and the narrowing is invisible in the marketing." That citation occasion is right and it is the reason this entry exists. But the idea it names — constrain the human's input until the machine can read it — was not invented here, and filing this as an idea would credit a product with a framing three years older than it and about a century older than that in other materials.

The framing is stated, as a research claim with a rationale, in David Goldberg and Cate Richardson's "Touch-Typing with a Stylus" (INTERCHI '93, pp. 80–87), which proposed Unistrokes: an alphabet whose letters are "specially designed to be used with a stylus" and which are "faster to write, less prone to recognition error, and can be entered in an 'eyes-free' manner." Palm was later sued over that patent, which is section 3's problem, not this one's. Further back the same move is in metal rather than software: MICR E-13B, in use by the American Bankers Association for automated cheque processing from 1958, and OCR-A, issued 1966 and first implemented 1968 — alphabets shaped for a reader that was not a person. Those two narrow what a printer emits rather than what a hand writes; the closest true human-side ancestors are the block-capitals-in- boxes form and the shorthand systems, and neither is in canon/.

The kind that fits is moment, on the definition eliza-1966 set when it argued its own way out of that kind and clippy-1996 restated: a date on which something visibly happened in public, of the shape of Dartmouth, Lighthill, Deep Blue, the LLaMA weights. What visibly happened in March 1996 is not that someone thought of trading the user's habits for the machine's accuracy. It is that about a million people took the trade inside eighteen months, on a device they paid $299 for and carried every day. The idea is old; the acceptance has a date.

Why not interpretation. Nothing here is a reading of the field. It is a product decision with a measurable market outcome, and the interpretations that have grown on top of it — section 5 lists them — are mostly wrong.

On the id's name and year, both of which need a correction. In 1996 the device was not called a PalmPilot. Palm Computing, then a subsidiary of U.S. Robotics, introduced the Pilot 1000 ($299, 128 kB RAM) and Pilot 5000 ($369, 512 kB) in March 1996. Pilot Corporation of America, which makes pens, sued over the name; the 1997 models were rebranded PalmPilot Personal and Professional and arrived on 10 March 1997, and the "Pilot" name was dropped entirely under a 1998 settlement. Graffiti itself is older than both: it was announced on 19 September 1994 and sold that quarter as a $79 software add-on for other companies' machines — the Apple Newton, the Tandy/ Casio Zoomer, the Sony Magic Link. So 1996 is not the year Graffiti was invented or first sold. It is the year it stopped being an accessory somebody bought to fix a device and became the way a device worked out of the box, which is the year that matters for this entry and the reason the id's date is right even though its product name is off by a year.

descends_from is empty, and that is a statement rather than a gap. The honest ancestor is unistrokes-1993 — Goldberg and Richardson's paper — which is not in canon/ and is not in proposals.md; if the owner wants the idea filed as an idea, that is the id to research, and it has the additional merit of being the one Palm lost in court to. The Newton has no entry either. moravecs-paradox-1988 is in canon/, and it explains this entry completely: reading a person's handwriting is a perceptual task any five-year- old does and a 1990s computer could not. But explanation is not descent. There is no evidence that Hawkins came out of Moravec, and drawing that edge because it would look good on the graph is precisely the quiet, invisible-from-the- inside error rule 7 is about. The edge is left undrawn.

What it is

Graffiti is a stroke alphabet and the recognizer for it. Every character is a single continuous gesture — pen down, one shape, pen up — entered not anywhere on the screen but inside a fixed silkscreened area in the bottom third of it, with letters on the left half and numbers on the right, and modes reached by gesture for capitals, punctuation and symbols. Because each character ends when the pen lifts, strokes can be drawn on top of each other, and the user's eyes can stay on the insertion point rather than on the input area.

What it does not do is the part that matters. In the words of the first independent evaluation of it: Graffiti "makes no attempt to recognize words or phrases; it does not use a dictionary; and it does not accept cursive handwriting," and it "is similar to a keyboard in three senses: (a) input is character-by-character, (b) users' eyes may fixate on the application's insertion point rather than on the input device, and (c) it uses modes to access uppercase characters and special symbols." It fit in about 44K of ROM and 8K of RAM.

The design's one concession to the user is that the strokes were drawn to look like the letters they stand for — this is the whole difference between Graffiti and Unistrokes, and it was deliberate. Twenty-one of the twenty-six strokes match either the upper- or lowercase Roman letter; five (A, F, K, Q, T) match neither, having been simplified to remove the cross-strokes and pen lifts. Hawkins' recorded reasoning is an analogy rather than an argument: "And then it came to me in a flash. Touch-typing is a skill you learn."

The Newton is the reason this looked like a solution, and the Newton is also where the standard account goes wrong. Apple's MessagePad shipped in August 1993 at $699 with a cursive recognizer licensed from ParaGraph International, and its failures were funny enough to become a Doonesbury strip that same month — the one in which "Catching on?" is transcribed as "Egg freckles." The engineers who later fixed it named the failure mode after the cartoon in a peer-reviewed journal: earlier systems "utilized strong, limited language models to maximize accuracy, but they proved unacceptable in real-world applications, generating disturbing and seemingly random word substitutions — known colloquially within Apple and NEWTON as 'The Doonesbury Effect'."

And then they fixed it. Apple's Print Recognizer — a neural-network character classifier under a context-driven search over segmentation and word hypotheses, built in Apple's Advanced Technology Group — "was first shipped in NEWTON OS 2.0–based MESSAGEPAD 120 units in December 1995," four months before the Pilot went on sale, and its authors say some have called it the "first usable" handwriting-recognition system. Note also what Apple did to make that possible: "To focus our work on a subproblem that was tractable enough to lead to usable products in a reasonable time, we initially restricted the domain to hand printing so that strokes were clearly delineated by pen lifts." Both teams narrowed. The difference is where the narrowing was put — on the user's alphabet, or on the input style plus a neural net, multiple dictionaries, a regular-expression grammar for dates and phone numbers, and a maximum-likelihood search engine that would not fit in 44K.

Graffiti's bet paid, and it was measured. I. Scott MacKenzie and Shawn X. Zhang put 25 subjects with no pen-computer experience in front of a Graffiti recognizer at Graphics Interface '97: after one minute studying the reference chart, character accuracy was 86%; after five minutes of practice, 97%; one week later with no intervening practice, 97% again — "complete retention." The market agreed faster than the literature did: U.S. Robotics sold roughly a million units in the product's first eighteen months, and the industry analyst Nigel Ballard called Graffiti "the program that comes closest to being the first killer application for pen computers."

How it ended. Xerox had filed for the Unistrokes patent in October 1993; it issued as US 5,596,656 on 21 January 1997 and Xerox sued that April. The case ran nine years — dismissed June 2000, reinstated by the Federal Circuit in October 2001, infringement found that December, the patent held invalid on summary judgment in May 2004 — and in January 2003, mid-litigation, PalmSource replaced Graffiti with Graffiti 2, built on Communication Intelligence Corporation's Jot, which restored multi-stroke letters and was disliked by the users it was imposed on, to the point that restoring the original became a documented how-to. Palm and Xerox settled on 28 June 2006 for $22.5 million and a seven-year mutual covenant not to sue. By then the argument had moved on: text entry on handhelds had gone to thumb keyboards and then to touchscreens, and the stroke alphabet went with the stylus.

Why a reading would cite it

A capability delivered by narrowing the operating domain. The 2026-08-19 midnight digest records Pony.ai reporting robotaxi services revenue up 691.2% to $12.1m, a fleet of 1,975 vehicles on 30 June, and a contracted overseas pipeline above 4,000 vehicles "subject to permits and clearances" — a figure the reading was careful to state separately from the deployed one. The Graffiti reading of that is not "it is fake." Graffiti was not fake. It worked, people loved it, it sold a million units, and it did so without solving the problem its category was named after. What the entry supplies is the question to put to any narrowed deployment: what exactly was narrowed, who is paying for the narrowing, and what happens to the product when the general capability arrives. For Graffiti the answers were: the alphabet; the user, in about five minutes; and it was displaced within a decade by a change nobody in 1996 was arguing about. A geofence, an HD map, a remote operator and a permit are the same class of answer to the first question, and a reading that records a robotaxi number without recording the domain it was earned in has reported the wrong quantity.

Vendor-stated accuracy against independent measurement. This is the sharpest use and it is nearly verbatim in the sources. In 1996 a survey of eighteen handwriting recognizers found that "eight developers quoted accuracy of 98–100% without qualification," the rest hedging with "writer dependent," "50% to 100% depending on application," or "100% if characters written as prescribed," and the surveyors noted that in the absence of a standard benchmark, "this and our subjective experience with the products is all we have to go on." The first independent number for the best-selling recognizer of the era arrived from a university in May 1997, more than a year after the device shipped. And Apple's own engineers, writing in AI Magazine in 1998 about a product they were proud of, declined to state an accuracy figure and explained why: neat, experienced printers "can achieve 100-percent accuracy if they are careful," complete novices can "yield word-level accuracies as low as 80 percent," newsgroup reports and personal testing suggest 97–98% in regular use, and "for scientific purposes, none of these numbers have any real meaning because our testing data sets are proprietary, and the only valid tests between different recognizers would have to be based on results obtained by processing the exact same bits."

Read that against the same 19 August digest, which records a protein-design result measured against "a field baseline the company puts at 10–15%" and hedges it on exactly one hinge: that the proteins were built and tested by Adaptyv Bio and Twist Bioscience rather than by the claimant. That hedge is thirty years old and was written down by people with no stake in this project's arguments. The number becomes meaningful at the point where someone other than the vendor produces it, and a 1998 statement of that principle by the vendor's own engineers is a stronger citation than any restatement of it now.

Humans learning the machine's alphabet, again, at scale. The 18 August midday digest records Alipay launching a platform that lets merchants convert "pages, products and workflows into agent-callable skills and MCP tools," with KFC, Mixue Bingcheng and Luckin Coffee already integrated. That is the Graffiti move applied to a national retail surface: rather than the agent learning to read a page built for humans, the page is rewritten in a dialect the agent can read. The same shape covers prompting conventions, structured-output schemas and tool definitions. What this entry adds beyond the analogy is a warning from the other side of it: Apple's engineers observed that even their unconstrained recognizer trained its users — because a failure is usually a single-character substitution the user can trace back to their own ink, "the user learns what works and what doesn't ... so, the system trains the user gracefully over time." Human adaptation to a recognizer happens whether or not it is designed in. The only real design choice is whether it is admitted and taught, as Palm did with a printed chart and a bundled tutorial game, or absorbed silently and then described as the machine having got better.

What this entry does not support, stated plainly. It is not evidence, it deposits nothing, it moves nothing. It does not show that narrowing a problem is dishonest — the narrowed product was the good product for a decade. It does not show that narrowing is fine — the narrowing had an expiry and somebody paid for it in the meantime. Anyone citing it for either half alone is citing half of it.

What it got right, and what it got wrong

Not required for moment. Included because there is a rare thing here: a design bet, stated in advance, graded twice by controlled experiment — once at one year and once at twelve — with the two gradings disagreeing about who won.

Right — that ordinary people would learn an alphabet for a machine, quickly. The claim, from 1993–94: users will accept a new character set the way they accept touch-typing. Due on contact with the public. Graded correct and by a wide margin: 86% character accuracy after one minute with a chart, 97% after five minutes, 97% a week later with no practice in between, and a million buyers in eighteen months. The claim was not that the alphabet was better; it was that the learning cost was small enough to be worth paying. It was.

Right — that the constraint bought a product the alternative could not deliver. 44K of ROM against a neural network, four dictionaries and a search engine; $299 and a shirt pocket against a MessagePad. Due immediately. Graded correct on the market's evidence.

Wrong — that the mimic-the-Roman-letter compromise was cheap in the long run. Steven Castellucci and I. Scott MacKenzie ran the head-to-head at CHI 2008: ten participants, between-subjects, twenty sessions of fifteen phrases each over six weeks. Graffiti went from 4.0 to 11.4 wpm; Unistrokes went from 4.1 to 15.8 wpm. Correction rates: Graffiti "remained relatively consistent at 26%," while Unistrokes fell "from 43% to 16%." The alphabet designed to be learnable in a minute was, for a trained user, permanently about a third slower and permanently more error-prone than the alien one that took longer to learn. Hawkins optimised for the first five minutes and Goldberg for the thousandth; the market bought the first five minutes and never revisited the decision. Note the study's size — ten people, twelve years after the fact, on a Wacom tablet rather than a Palm — and treat the direction as solid and the magnitudes as indicative.

Wrong as popular history, though never claimed by Palm — that handwriting recognition was unsolved, so Palm routed around it. Apple's print recognizer shipped in December 1995, before the Pilot. The constrained alphabet did not beat an unsolved problem. It beat a solved and expensive one, on price, size, battery, and a synchronisation story, in a market that was buying a pocket organiser rather than a computer. This is the single most useful correction the entry carries, because the inference people draw — "the workaround won, therefore the general capability wasn't there" — is exactly backwards, and the same inference is available every time a narrow deployment outsells a general one today.

Wrong — the durability of the trade. Nobody made this claim explicitly, but the product was built as though the answer were permanent. Elapsed: about seven years as the default input method on the best-selling handheld line, ended in January 2003 by its own vendor under legal compulsion and replaced with something users tried to uninstall; effectively gone from consumer devices within a decade, killed not by a better recognizer but by a change of input device. As a base rate for how long a user-side workaround for a missing capability survives, seven-to-ten years is the number this entry supplies, and the manner of death — obsoleted sideways rather than superseded head-on — is the part worth carrying.

The proposal's own claim, graded: "the narrowing is invisible in the marketing." Partly right, and the correction sharpens it. At the point of sale Palm was not hiding anything: the stroke chart was silkscreened on the device, the manual taught the alphabet, and a tutorial game called Giraffe shipped to drill it. But the 1994 boxed Graffiti — the add-on for other people's machines — carried the slogan "Recognition that works" and told buyers to "just write on your Newton or Zoomer and watch typed text instantly appear," which is the overclaim the proposal describes, in the imperative mood, on the box. And the place the narrowing genuinely vanished is not the marketing but the retelling: the Computer History Museum's own summary of why the PalmPilot succeeded lists "effective handwriting recognition," which is not what Graffiti was. The narrowing survives one telling and is gone by the third. That is the thing to watch for, and it is a stronger claim than the proposal's because it does not require anyone to have lied.

A worked example of that drift, from the same source. The Computer History Museum states that proficient Graffiti users "could achieve 40 words per minute." No measurement supports it: 11.4 wpm after twenty sessions (Castellucci and MacKenzie 2008), and 21 wpm for self-described experts against 18 wpm on the on-screen keyboard (Fleetwood et al. 2002). The 40 is not a lie and has no author; it is what a remembered capability looks like after a decade of being retold by people who liked the product. A reading grading any retrospective performance claim — a model's, a vendor's, or this project's own past readings — should expect that drift to run in the flattering direction by default.

Commonly misused as

Not required for moment. Included because the single sentence most people know about this subject is false.

• "Palm solved handwriting recognition; Apple couldn't." Palm did not attempt handwriting recognition. Character-by-character, no dictionary, no language model, no cursive, and an alphabet drawn to suit the recognizer. Apple attempted it, failed publicly in 1993, and succeeded in December 1995. The correct sentence is that Palm solved text entry and Apple solved handwriting recognition, and the market rewarded the first.
• "The Newton's handwriting recognition never worked." True of the first-generation cursive recognizer, false of every Newton from MessagePad 120 units in December 1995 onward, which its own authors and later reviewers regarded as the best of its era. A joke from August 1993 is still doing damage to the reputation of software from December 1995 — which is itself a citable phenomenon when a model's early public failure outlives the version that fixed it.
• "Graffiti proves users will adapt to machines." It proves something much narrower: that about a million people would learn a 26-symbol alphabet, in five minutes, for a device they had already chosen to carry every day, when the alternative was a stylus keyboard. It says nothing about willingness to learn arbitrary machine dialects in general — and the counter-example is in the same story, since Unistrokes, the faster alphabet, was never adopted by anyone at all.
• "Graffiti proves users won't adapt." The inverse misuse, usually drawn from the Graffiti 2 backlash of 2003. What that episode shows is that users who have already paid a learning cost resist paying it twice, especially for a replacement forced by litigation rather than by design. It is evidence about switching costs, not about learning costs.
• "The narrow workaround is always the pragmatic choice." It was here, for about seven years, and the entry says so. It also had an expiry date, a permanent speed ceiling measurable in 2008, and a cost paid by every user in every session rather than once by the vendor. Citing the pragmatism without the ledger is half the entry.
• "Graffiti is the ancestor of swipe and gesture keyboards." The academic descent line for stroke- and shape-based text entry runs through Unistrokes, the alphabet Palm was sued over — Castellucci and MacKenzie open by noting that "Unistrokes is recognized in academia, Graffiti is commercially prevalent in PDAs." I have not traced the citation chain from Unistrokes into the modern shape-writing literature, so this is an impression from the sources read here rather than a checked trail, and it is offered only to block the lazier claim, not to establish the better one.
• "The PalmPilot, 1996." The device sold in 1996 was the Pilot 1000 or Pilot 5000. "PalmPilot" is the 1997 name, adopted because a pen company sued. Small, and the sort of thing that makes the rest of a citation look unchecked.

Sources

Primary, read at source. I. Scott MacKenzie and Shawn X. Zhang, "The Immediate Usability of Graffiti," Proceedings of Graphics Interface '97, Kelowna BC, May 1997 — read in full from the Graphics Interface proceedings PDF. All of the following are from that text: the 86% / 97% / 97% accuracy sequence and the "complete retention" finding, the 25 subjects and the Fujitsu 325Point apparatus, the 44K ROM / 8K RAM figure, the "no dictionary, no cursive, character-by-character" description, the keyboard-in-three-senses framing, the 21-of-26 stroke match and the five non-matching letters (A, F, K, Q, T), the per-letter results including V at 36.0% after one minute, the survey of eighteen recognizers with eight developers quoting 98–100% "without qualification," the observation that character accuracy must be "at or above 97% before users accept the technology," and the Nigel Ballard "killer application" quotation.

Primary, read at source. Larry S. Yaeger, Brandyn J. Webb and Richard F. Lyon, "Combining Neural Networks and Context-Driven Search for Online, Printed Handwriting Recognition in the NEWTON," AI Magazine 19(1), Spring 1998, pp. 73–89 — read from the AAAI open-journal PDF. Source of the "Doonesbury Effect" passage, the December 1995 NEWTON OS 2.0 / MessagePad 120 ship date, the "first usable" claim, the deliberate restriction of the domain to hand printing, the refusal to quote an accuracy figure with its reasoning about proprietary test sets and "the exact same bits," and the observation that the system "trains the user gracefully over time."

Primary, read at source. Steven J. Castellucci and I. Scott MacKenzie, "Graffiti vs. Unistrokes: An Empirical Comparison," CHI 2008, Florence, April 2008 — abstract and method read from the authors' PDF. Source of the 4.0 → 11.4 wpm and 4.1 → 15.8 wpm figures, the 26% versus 43% → 16% correction rates, the ten participants and twenty fifteen-phrase sessions, and the "recognized in academia / commercially prevalent" characterisation.

Primary, not read at source. David Goldberg and Cate Richardson, "Touch-Typing with a Stylus," INTERCHI '93, pp. 80–87 — the Unistrokes rationale is quoted here from abstracts and from Castellucci and MacKenzie's summary of it, not from the paper itself, which was behind a paywall. US Patent 5,596,656, "Unistrokes for computerized interpretation of handwriting" — the October 1993 filing and 21 January 1997 issue dates are from patent-database listings; the USPTO and Justia pages both returned HTTP 403 and the patent text was not read.

Secondary, product and litigation record. Wikipedia's "Graffiti (Palm OS)", "Graffiti 2" and "Pilot 1000" articles for the 1994 origin, the Sipher and Marianetti contributions, the stroke-design details and the March 1996 introduction; PalmInfocenter for the Pilot 1000 retrospective, for the January 2003 Jot licence, for the how-to on restoring original Graffiti, and above all for the settlement report of 28 June 2006 giving the $22.5m figure, the three-patent licence, the seven-year covenant and the litigation chronology (April 1997 filing, June 2000 dismissal, October 2001 reinstatement, December 2001 loss, May 2004 invalidation). The Pen-Based Computing History Museum's artifact page for the boxed Palm Graffiti for Newton supplies the 19 September 1994 announcement, the $79 price and the box copy — "Recognition that works" and "just write on your Newton or Zoomer and watch typed text instantly appear" — which is a description of a physical artifact I have not handled. The Computer History Museum's PalmPilot exhibit supplies the ~1 million units in eighteen months, the $299 price, and the unsupported 40-wpm claim graded in section 4. The Doonesbury week of August 1993 and the "Egg freckles" panel are as reported in secondary coverage; the strips themselves were not retrieved.

Not read at source, flagged in place. Andrea Butter and David Pogue, Piloting Palm (John Wiley & Sons, 2002), pp. 62–66, is the cited origin of the "Touch-typing is a skill you learn" remark, which reaches this entry through Wikipedia's citation of it rather than from the book. Michael D. Fleetwood et al., "An Evaluation of Text-Entry in Palm OS — Graffiti and the Virtual Keyboard" (2002), is cited for 21 wpm expert / 7 wpm novice via secondary summary only.

A note on one search result that was not used. Queries about Graffiti, the Pilot and handwriting recognition repeatedly surfaced encyclopedia pages published by a frontier-model vendor whose principal's predictions this canon grades in musk-robotaxi-2019. None were opened or cited. Rule 7's instruction that those two facts must never meet is not only about who composes a tweet; citing a graded vendor's reference work on a page this project publishes is the same transaction in a quieter form, and the sources above were all reachable without it.

The Pony.ai figures, the protein-design result with its self-stated 10–15% baseline and third-party wet-lab validation, and the Alipay agentic-commerce platform are as recorded in this project's own 2026-08-19-00 and 2026-08-18-12 digests, which hold the primary links. They are named here as the citation occasion and nothing more. Nothing in this file is evidence, nothing in it is deposited in the ledger, and nothing in it touches the needle.