bshepp/cat-keyboard-corpus

Cat Keyboard Corpus

Messages typed by the cats, Giddings and Chester, on the keyboard of my Dell Latitude E6420 XFR laptop (US QWERTY), kept "for posterity".

Why this exists. My desk is where the sun is, and the cats have right of way. The laptop runs a long-term sensor project with an AI assistant, and rather than fence off their spot, I let them walk across it. Sometimes a message gets sent. This corpus keeps them.

About the authors. Giddings favours the B key and the semicolon, and likes to drink from the stream in the bathtub left by a brief run of the shower. Chester walks the keyboard often; his first message (2026-10-02) was a single "x", and he pressed Enter himself - the first message a cat sent on its own. He was also the first cat the project's camera ever recognized.

Fields (messages.jsonl)

field	meaning
id	running number
sent_at_utc	when the message reached the assistant (typing may have happened earlier; see time_quality)
text	the message exactly as received
author	the cat: giddings, chester, ..., or unknown when nobody saw which cat
sent_by	who pressed Enter: the cat's name, me, or unknown (recorded from message 5 on; earlier ones backfilled where I remembered)
cat_part	only where the text mixes my input (or text the software offered) and a cat's: the cat's typed part
kind	absent for a typed message; keys for a key event: control keys that did something but left no text (text is empty)
passes	how many times the cat crossed the keyboard to leave the text, when I saw it (first given for message 28)
pass_parts	the text of each pass and the direction the cat was walking, when I saw it
keys	for key events (and messages where a control key mattered, like 16): the keys, as far as they can be told (nothing records keystrokes, so they come from the effects and from my own check of what recreates them)
effect	for key events: what happened
notes	context

Observations so far (43 messages + 2 key events, 1954 typed cat keystrokes, 2026-10-24)

• Most-pressed keys: ] (234, all but one in message 27), N (151), D (111), F (78), R (67), B (57), V (55) (Shift+B counted as B), plus 462 spaces.

• Messages 1-7, 13, 14, 17, 18, 19, 21, 22, 23, 29, 36, 37, 39, 40, 41, 42, 43, 44 and 45 are Giddings' (13, 14 and 17 back on his B/N/M spot, bottom row; 18 is the top-right corner, [ ] =, a new spot; 19 is the keys left of Enter with Shift held for part of it: ""??/':); 8, 9, 15, 16, 20, 26, 31, 32 and 33 are by an unknown cat (Giddings or Chester); 10, 11, 24, 25, 27, 28, 30 and 38 are Chester's ("x"; then 75 f's and a c, on the keys next to it; then, on 2026-10-07, a row of double quotes and '} from the keys beside Enter; then, on 2026-10-09, 233 closing brackets from the same corner; then, on 2026-10-11, hgtygcv d from the middle of the keyboard).

• Giddings' early messages cluster on the bottom-middle keys (B/N/V/M), later ones on the bottom right (k l ; ' , . /, next to Enter), perhaps following where he steps onto the laptop.

• Messages 8 and 9 are held keys: one b + 114 n's; then 248 spaces, c, 107 d's, 46 g's, v, b - a slide across the middle of the keyboard (C and V/B on the bottom row, D and G on the home row), after a long press on the spacebar.

• Seven messages show a paw on Shift (the fifth is message 27's thirty colons, the sixth message 32's seventeen double quotes, the seventh Giddings' 37: a quote and nine colons): capital B's in one, message 19's quotes, question marks and colon, and Chester's 24 (all double quotes) and 25 (the brace).

• sent_by: I pressed Enter 22 times; 4 are unknown; Chester sent message 10 himself, the first cat-sent message; Giddings sent message 17 himself (" bn"), his first; Chester sent message 24 himself too, his second; Giddings sent message 36 himself, his second; Giddings sent message 40 himself, his third; Giddings sent message 42 himself, his fourth; 43 and 44 too (44 while running across the keyboard to his bowl, also sending the owner's inventory line by a paw on Enter).

• Entry 12 is a key event, not a message: Giddings, Ctrl+C twice (confirmed afterwards: the only combination that recreates what I saw). It typed nothing and opened a session-select menu inside Claude Code instead; the running session was not closed.

• Entry 34 is the second key event: Chester on the arrow keys, most likely with Alt (inferred; I did not see Alt). It switched the screen to another text console with a login prompt, which looked to me like being logged out. The first cat event that left the program altogether.

• Message 16 begins with a sentence nobody typed: a cat's Tab accepted the program's suggested prompt ("good morning, how did the night run go?"), then came c, x, 11 v's, 3 spaces, f and 52 z's from the bottom-left corner, and my Enter (pressed to wake the computer) sent it. All 52 z's in the corpus are from this one message.

• Message 36 is the second recalled sentence: Giddings' Up arrow brought back my own "good morning." from the line history, '[ came from the two keys beside Enter, and Enter sent it. Only the two characters are his; the greeting is mine. The assistant read it as a repeat with a stray key press before I said anything.

• Message 20 is the first actual word: "ok" (unknown cat; O sits just above K). I pressed Enter only to wake the screen, and the assistant answered it as if I had written it - the first cat message to pass as mine.

Working notes: kinds of events (2026-10-03)

Not a field yet. We are still learning what happens, and the kind can be worked out from each entry.

name	what it is	entries so far
sit	a cat rests or walks on the keys, text piles up, a human sends it	1, 5, 6, 8, 9, 11, 13, 14, 15, 16, 18, 19, 20, 21, 22, 23, 25, 26, 27, 28, 29, 30, 32, 33 (probably 2 and 4)
recall	the keys recall an earlier command (Up-arrow), then add text	3
self-send	the cat's keys include Enter	10, 17, 24, 31
operate	control keys change what the program or the computer is doing; no text	12, 34
mixed line	my text and a cat's on the same line	7
accept	a cat's key accepts text the software offered (Tab on a suggested prompt); the words are the program's, then cat keys follow (also a sit)	16

Not seen yet, or not noticed: edit (a cat deletes or changes what I typed), interrupt (keys that cancel something in progress), unsent (typed, then cleared by me - never recorded unless I mention it). A long run of one key is a paw resting; many different keys is a cat moving.

Who is involved: the two cats make the keystrokes; I am the witness (which cat, and often the one who presses Enter); the AI assistant keeps the record, and is not a neutral recorder - entry 12 needed two corrections because it filled in what it could not see. An event only gets here if Enter is pressed or I notice, so this is what the cats typed that got sent or witnessed, not everything they typed.

The unknown values of sent_by were most likely me pressing Enter, but I am not sure, so they stay unknown.

Collection

Messages are recorded as they arrive, with who wrote them and who pressed Enter. Desk changes that may affect the cats' routes are noted (2026-09-30: I cleared the desk). No other data about the household is included.

Changelog

• 2026-09-30: first release as giddings-keyboard (5 messages by Giddings).
• 2026-09-30: renamed to cat-keyboard-corpus so every cat fits; the card is in my own voice; sent_by value owner -> me.
• 2026-09-30: "the cats" instead of "my cats" (companions, not possessions).
• 2026-10-01: message 6 (Giddings, J/M/N); message 7 (mixed: my " mxc" typed on the wrong keyboard, then Giddings' run of C's and "dm").
• 2026-10-02: message 8 (unknown cat: 38 spaces, one b, then 114 n's - the longest message so far); observations recounted over all messages.
• 2026-10-02: message 9 (unknown cat: 248 spaces, c, 107 d's, 46 g's, v, b - now the longest; slides across the middle row).
• 2026-10-02: message 10: Chester's first, a single "x", sent by Chester himself (I didn't press Enter) - the first cat-sent message.
• 2026-10-03: message 11: Chester again, 75 f's and a c (I pressed Enter).
• 2026-10-03: entry 12, the first key event: Giddings, Ctrl+C twice - no text, but it opened a session-select menu inside Claude Code. New fields kind, keys, effect.
• 2026-10-03: entry 12 corrected twice: the keys are confirmed (Ctrl+C twice is the only combination that recreates what I saw), and the effect was a session-select menu inside the program - the session was never closed.
• 2026-10-03: working notes on the kinds of events (sit, recall, self-send, operate, mixed line), who is involved, and what never gets recorded.
• 2026-10-03: message 13: Giddings, a space, 12 m's, n, m and a comma (I pressed Enter).
• 2026-10-03: message 14: Giddings again, 11 n's, m, space, n, b, b (I pressed Enter).
• 2026-10-04: message 15: unknown cat, "lcx" (I pressed Enter).
• 2026-10-05: message 16: unknown cat. Tab accepted the program's suggested prompt, then c, x, 11 v's, 3 spaces, f, 52 z's (I pressed Enter to wake the computer). New kind in the working notes: accept. cat_part now also separates software-written text from the cat's.
• 2026-10-05: message 17: Giddings, " bn", and he pressed Enter himself - his first self-sent message.
• 2026-10-05: message 18: Giddings, 3 spaces, ], 11 ['s and = from the top-right corner (I pressed Enter).
• 2026-10-06: message 19: Giddings, ""??/': with Shift held for part of it; typed about half an hour before I pressed Enter.
• 2026-10-06: message 20: unknown cat, "ok" - the first actual word. I pressed Enter to wake the screen; the assistant replied to it as mine.
• 2026-10-07: message 21: Giddings, 40 v's (I pressed Enter).
• 2026-10-07: message 22: Giddings, f, v, g, f and 67 r's (I pressed Enter).
• 2026-10-07: message 23: Giddings, 1 space, 2 n's, 11 spaces, n and b (I pressed Enter).
• 2026-10-07: messages 24 and 25: Chester. 11 double quotes, sent by Chester himself; then '} 102 s later (I pressed Enter).
• 2026-10-09: message 26 (unknown cat, overnight): 161 keystrokes, 114 of them spaces; ,kl.,m, then the keys left of Enter: '/;, sixteen periods, ;' and twenty slashes. I typed "good morning" behind it without noticing and pressed Enter.
• 2026-10-09: message 27: Chester, 279 keystrokes on the keys left of Enter: 14 apostrophes, 233 closing brackets, '/ and 30 colons (Shift). I pressed Enter.
• 2026-10-11: message 28: Chester, hgtygcv d, nine keystrokes from the middle of the keyboard, left in two passes across it (new fields passes and pass_parts, filled only when I saw it): hgty walking right to left, then gcv d left to right on the way back. I pressed Enter.
• 2026-10-11: message 29: Giddings, njbhjkmh, back on his B/N/M spot after three days without a message. I pressed Enter.
• 2026-10-11: message 30: Chester, 22 spaces and an n. I pressed Enter.
• 2026-10-11: messages 31 and 32, in the morning, by a cat nobody saw: ,..;l, sent by the cat (I did not press Enter); then 45 keys from the corner left of Enter, which I sent.
• 2026-10-11: message 33, the same morning, cat not seen: mkiklk;lko, ten keys around K. I pressed Enter. It arrived while my login had expired, so it was logged afterwards from the session's record.
• 2026-10-13: entry 34, the second key event: Chester on the arrow keys, Alt+Left inferred. The screen switched to another text console with a login prompt; the session kept running and Alt+F1 brought it back.
• 2026-10-15: message 36, Giddings, early morning: the Up arrow recalled my previous greeting, then '[ - the second recalled sentence after message 16; Giddings sent it himself. Message 37 twenty minutes later, Giddings again: /":::::::::, the keys right of L with a paw on Shift; I pressed Enter.
• 2026-10-16: message 38, Chester sitting on the keyboard: dsswdesssssssssssssssssss, the D/S/W/E cluster and a held S - his first from the left side of the keyboard. I pressed Enter.
• 2026-10-20: message 39, Giddings, evening: .,,,,,,,,,,, a period and ten commas from his bottom-right keys. I pressed Enter. Four days since the last message: the cats have been stepping on the keyboard less since it stopped getting my attention.
• 2026-10-24: messages 40 and 41, Giddings, seconds apart during an evening inventory session (owner: "must be feeding time"). 40 = ] + backslashes from the top-right corner (cf. 18), Giddings pressed Enter himself. 41 = a long run of apostrophes + "???? from the key right of L with a paw on Shift (cf. 19, 37); I pressed Enter.
• 2026-10-24: message 42, Giddings, same feeding-time cluster: "? (double-quote + question mark, both Shifted, right of L). Giddings pressed Enter himself.
• 2026-10-24: message 43, Giddings, last of the feeding-time cluster: jm\=m, a scattered middle-to-top-right smear; I pressed Enter, then got up to feed him.
• 2026-10-24: message 44, Giddings, ,, - running across the keyboard to the food bowl, he hit Enter on the owner's half-typed inventory line ("2 HC-RS04s", human text sent early by a paw) and trailed two commas, sending them himself in transit. A paw-in-motion trail.
• 2026-10-24: message 45, Giddings, evening inventory: A}}]Z[A - his top-right bracket/brace corner mixed with letters and a paw on Shift. I pressed Enter.

License

CC BY 4.0.

Direction of travel (first look, 2026-10-20)

key_paths.py turns each message into a path across a US QWERTY layout, in typing order, and reports the net displacement from first key to last and the number of direction reversals. On the 36 text messages so far, with the three that contain a recalled sentence excluded (3, 16, 36 are my own typing brought back by the Up arrow):

• Most messages are a paw parked, not a crossing. 27 of 33 end within 1.5 keys of where they began: one cluster, a held key, a few neighbours.
• When Giddings moves, he moves to the right (4, 7, 13, 18: klllll,./, the mxc...dm run, m...nm,, and the top-right corner ][[[=), four of four, left-to-right from the typist's side. No leftward crossing of his yet.
• Chester's one move is leftward (28, hgtygcv d, from H to D), and the unknown cat's two are leftward (15, 16).
• No message holds two crossings. Reversals appear only in the recalled-sentence messages, so there is no case yet of a cat walking across, turning, and walking back without an Enter between.
• What this would need to become a route estimate: the keyboard's orientation in the room per desk position (it faces the same way every night, so left-to-right on the keys is one fixed direction between the two windows), and a lot more crossings than four. For now: a tendency, not a probability. Placing the laptop on their route to get more is out: the corpus is what they do on their own.

Content-free input monitor (input_monitor.py, 2026-10-20)

A read-only watcher (input_monitor.py) that records timing and non-content events only, so it can run continuously beside the corpus without altering it: it never grabs the devices, so typing and pointing work normally, and it is built so it cannot write what was typed.

• Character keys (letters, digits, punctuation, Space, Tab, Backspace) are logged as a keystroke with the key set to null: the time and the fact that a key moved, never which one. A code not on the non-character allowlist fails safe to null, so a mistake drops identity rather than leaking it.
• Non-character keys carry no content, so their name is kept: modifiers, arrows, the navigation cluster, the function keys, Esc, the lock keys, Enter, and the Dell/media hotkeys (ALLOW in the script). These were the events the corpus had to reconstruct by hand before (entry 12's Ctrl+C, entry 34's Alt+Left).
• Trackpad: touch down/lift, button presses, and each gesture summarised as a contact count and a bounding box (a paw gives many contacts and a large box; a fingertip few and small). No absolute position of anything typed, because nothing typed is read.
• --content timing drops even the non-char identities: pure press/release rhythm. A cat's held-key repeats and a human's word-rhythm separate on the inter-event timing alone, with zero identity.

Output is one JSONL per UTC day under ~/data/input_events/ (PRIVATE, under ~/data, never committed and never uploaded; cat events are promoted into the public messages.jsonl by hand, as now). Needs read access to /dev/input/event* (group input). Devices are matched by name, never by event index.