Why keep count, and how it works

• 21 October 2026

  • **The top bar is simpler: Branches · Claims · News · About, with Family beside the light/dark switch.** The method page and the story so far are now the two parts of one About page, which opens with why the site keeps count; their text is unchanged, and their old addresses (`/method/` and `/story/`, including a link to one day of the story) land on the matching part. This week's reading keeps its page and its address, and is linked from the home page, About, News and the footer. Feedback and the source code moved to the footer. No record, stage, count or other address changed.
  • **The site now asks "How much of a human being can we change with technology today?", in place of "on purpose", as I ruled on 2026-10-21.** "On purpose" is what you say to rule out an accident, which nobody had raised, and it hid the real line the tree draws: between changes made with a device, a drug, a procedure or a machine a person uses, and everything else that changes a person, such as growing up, ageing or school. The headline now reads "ways of changing a human being with technology", and so do the home page, the method page, the claims page, the footer, the link preview and the share picture (redrawn, at a new address, `/social-preview-2026-10-21.png`). Where a sentence already says "a device, a drug or a procedure", the words are simply gone. Two upgrade names are shorter: "A treatment removes a specific fear memory" and "A child is born from a gene-edited embryo"; their addresses are unchanged. The mood branch keeps its condition that a change is one the person chose. No count, stage, test or address moved; the count stays at 28 of 35. The wording proposal is `docs/proposals/2026-10-21-on-purpose-wording.md`.

• 20 October 2026

  • **Five papers this tracker leans on were read in full against the record on 2026-10-20.** No stage moved and the headline count is unchanged at 28 of 35. One point was added to the speed chart, and four entries were made more exact. - **Added to the speed chart.** The [University of California, Davis report of 2026-06-15](https://doi.org/10.1038/s41591-026-04414-6) (Nature Medicine): one man with motor neurone disease used a speech implant at home, without researchers present, for nearly two years, at an average of 56 words a minute over 183,060 sentences. He rated the accuracy himself; on prompted sentences, where errors could be counted, the decoder was right on more than 99 per cent of words speaking aloud and 96.5 per cent speaking silently. It is now the series' latest point. The best figure for text is unchanged at 78 words a minute. - **[Willett and colleagues, 2023](https://doi.org/10.1038/s41586-023-06377-x)** (Stanford speech implant, Nature). The speed, error rates, vocabulary and the single participant all match. The speech entry said ordinary conversation runs at a hundred and fifty words a minute; the paper says 160, as the chart's baseline already did, and the entry now says so. The participant is now described as the paper describes her: a woman of 67. - **[Willett and colleagues, 2021](https://doi.org/10.1038/s41586-021-03506-2)** (handwriting to text, Nature). The 90 characters a minute, the 94.1 per cent and the 115 characters a minute for phone typing match. The point now also gives the error rates (5.9 per cent of characters and 25.1 per cent of words as decoded live) and the free-typing speed, and says plainly that it was one person over seven test days. - **[Scangos and colleagues, 2021](https://doi.org/10.1038/s41591-021-01480-w)** (an implant that detects a depressed state and responds to it, Nature Medicine). It is one woman, and the entry does not say "cure". The entry said her depression lifted "within weeks"; the paper says her daily scores fell the morning after the implant was switched on and her depression score was in the range counted as remission months later, and the entry now says that. The notes now also say what was detected and what the device did, and that the paper calls itself a proof of concept, not a test of whether the treatment works. - **[Längin and colleagues, 2018](https://doi.org/10.1038/s41586-018-0765-z)** (baboons living on pig hearts, Nature). "Months" is right. The pig-heart entry's notes now give the figures: of the last five baboons, four lived at least three months and the longest 195 days, against a previous longest of 57.

• 19 October 2026

  • **I rewrote the two news posts as short stories for a reader new to the site, the rule I set for every tracker on 2026-10-19.** The post on Vidal's 1977 paper keeps its address and every fact, gains a plainer title ("A 1977 paper moves the first brain-to-computer link without surgery back eleven years") and a redrawn card, and ends with a dated line linking its first version. The opening "Start here" post, held since 2026-10-18 until I had read it, is rewritten the same way and is still not published. No record, stage or count changed.

• 18 October 2026

  • **Five changes to the weekly reading, from the audit of its first cloud run, as I ruled on 2026-10-18 ("yes to all five").** The audit is `docs/readings/audit-2026-10-18.md`. (1) The documents now say two upgrades sit at the top of the ladder and a weekly reading is 33, and the export refuses a newest reading that reads any other set than its scope asks for. (2) `noted` means dated inside the week or new to the record; an older record the entry already holds is `quiet`, with the fact in `found`. (3) A reading opens every record it cites for a non-quiet item, cites the maker's, regulator's or journal's own page, lists the pages in a new `opened` field, and runs one more search limited to the register's site, whose sites go in a new `register` field; the export holds records from the reading of 2026-10-19 to both. (4) The export script stamps the site's data and the headline file with the Pacific date, not the machine's universal-time date (the first cloud run had stamped 2026-10-19 on a reading of 2026-10-18); `scripts/test_reading_rules.py` tests this and the rules above, and the deploy runs it. (5) The reading page's key now says `quiet` means nothing turned up in the search results, not that nothing happened. No record, stage or count changed.
  • **Vidal's 1977 paper was read in full, and the first human for the scalp headset moves from 1988 to 1977 on its strength, as I ruled on 2026-10-18.** I bought the paper on 2026-10-18; it was read page by page and the questions left open on 2026-10-17 are answered. It reports volunteers with electrodes on the scalp and ears, no surgery, steering a symbol through a maze on a screen in real time: each looked at one of four dots, the computer sorted the brain's response to each single flash into one of four classes and moved the symbol, and it was right more than 90 per cent of the time with average subjects (Proceedings of the IEEE 65(5), May 1977, pages 636 to 638). That meets the upgrade's test as written (a device worn on the scalp, no surgery, deliberate control of a cursor at better than chance), so the 1988 Illinois speller is no longer the first human; it stays in the notes as the first brain-controlled speller. The date is 1977-05-01, the paper's issue month, by the rule that a study is dated by its publication. Not 1976: the paper gives no date for any session, and the 1976 dates that can be read off its figure labels are inferences the paper does not state. Said plainly in the notes as well: the paper does not report anyone reaching the end of the maze and does not test whether the response depends on eye movement. No stage and no count changed (the headline is still 28 of 35); only the date on the first-human rung and the record behind it. The paper itself is not in this repository: it is licensed to me alone.
  • **The weekly reading moved from my Mac to a cloud routine, as I ruled on 2026-10-18.** The routine "Upgrade Tree weekly reading" (routine id `trig_01Qo6Rh8i92T3qDfp44uaHbr`) keeps the Sunday 18:00 Pacific slot and runs whether or not the Mac is awake. The Mac task was switched off the same day, and the routine's first run was that evening, 2026-10-18 at 18:04 Pacific: it pushed its reading to `main` itself as commit `f5a7f3c` ("Weekly reading 2026-10-18: 33 read, 0 moved, 0 flagged"), so the cloud workspace can push, and the site deployed. The next run is 2026-10-25. Only the schedule changed: the procedure, the limits on what the reading may edit and the record it writes are the same. `docs/weekly-reading.md` (decision 1) and the README say so.
  • Weekly reading (scheduled task, 2026-10-18): 33 upgrades read, 0 moved, 0 flagged for me; record at data/readings/2026-10-18.toml.
  • In the news: A 1977 paper moves the first brain-to-computer link without surgery back eleven years.

• 17 October 2026

  • **The five statements still marked unchecked were each taken to a source a reader can open: none is left (checked on 2026-10-17).** I ruled the same day that nothing in the record rests on an AI model's recollection: every fact carries a source a reader can open, and a fact that cannot be sourced after a real search leaves the record. The headline file now reports no statements to check (it said five). No stage, no upgrade and no count changed: the headline is still 28 of 35. Every search and every page opened is listed in `docs/verify-pass-2026-10-17.md`; pages that refused an automated reader (the New England Journal of Medicine's, the journal that printed a 2016 conference abstract, and a defence-department report archive that was closed for maintenance) were left alone, not got round. - **Confirmed against the source.** Which vocabulary the 78 words a minute belongs to: the full text of the San Francisco paper (Nature, 2023-08-23) says the participant silently attempted 249 sentences drawn from a set of 1,024 words, not used in training, and gives a median of 78.3 words a minute and a word error rate of 25.5 per cent; the legend of its Figure 2 says the rate was computed on those trials. The point's description now says "a vocabulary of 1,024 words" and notes that the paper calls it large-vocabulary decoding and tried a vocabulary of 39,378 words only offline. The first mitochondrial-donation child's birth date, 6 April 2016: New Scientist's report that broke the story (2016-09-27, Jessica Hamzelou) says "the boy was born on 6 April 2016", and its opening, "five-month-old", agrees; CBS News repeated it. The 2017 paper gives no day, and no document from the clinic or the authors that could be opened does, so the date now rests on that news report and the entry's notes say so. The date and the stage were not changed. - **Corrected.** The date of the University of California, Davis speech decoder (New England Journal of Medicine) is 2024-08-14, not 2024-08-15: the university's own release of 2024-08-14 says the study "was published today", and the PubMed and DOI records were created that day; 2024-08-15 is the print issue's date, which the point's source note still gives. The series dates a paper by its first publication (the same rule as the San Francisco and Stanford points), so the bandwidth series' latest date moves from 2024-08-15 to 2024-08-14 in the headline file. The journal's own page could not be read (a bot check, not got round). The Dubai 2017 claim's account said a thumb manages "around ten bits per second". Both transcripts of the recording have him saying "10 bits per second or 100 if you're being generous" (about 26:53 to 26:59), and the account now says that. The claim's date, status and quoted words are unchanged. - **Not closed by reading the paper, and put to me.** Whether a 1977 paper by Jacques Vidal (Proceedings of the IEEE) shows an earlier first human for the scalp headset: its full text is behind a paywall and no open copy exists, so it was not read. What was read: its abstract (real-time sorting of single brain responses, with feedback to the subject; no maze or cursor mentioned), Vidal's 1973 review in full (a planned programme; it says direct brain-computer communication "still lies somewhat in the future") and a 2011 review in Communications of the ACM that says Vidal's users moved through a maze on a screen in real time by attending to the corners of a flashing checkerboard, citing the 1977 paper. The entry's notes now say exactly this and nothing more; the first-human date stays 1988 and no stage moved. A decision for me is on the roadmap: whether that second-hand account is enough to move the date, or whether to wait until the paper has been read. - **Removed for want of a source.** Nothing was taken out of the record: each of the five facts either has its source or, for Vidal, is stated only as far as the sources go. - **The Dubai 2017 account, as I ruled on 2026-10-17.** The Dubai wording was the third item of the decision I ratified that day with "yes to all" (the roadmap task lists it as the account's "around ten bits per second" against the transcripts' "10 bits per second or 100 if you're being generous"). The task, the record layout and the worklog entry say what was put to him but not in so many words which way he answered, so the account was brought to what both transcripts say, which is what the question set against it; the claim's date, status and quoted words are unchanged. If he meant the shorter wording stays, it is one sentence to put back.
  • **The headline file.** `data/headline.json` exists, the first adoption from the family's cross-project lessons review of 2026-10-17 (lesson 1, a published headline file; I ruled the same day): the count and its words (28 of 35 ways of changing a human being on purpose have been tried in at least one living person), the five-rung totals, the bandwidth series' latest point and its best figure on each measure, the day it was computed, the last reading's date and one line on the method's status. The export script writes it from the same records as the site's data and never by hand, and its check now refuses to pass while the committed copy is missing or says something the record no longer says (the day and the commit may differ); `scripts/test_headline.py` tests that, and the deploy runs the tests. The weekly reading's procedure runs the export, not only the check, after an edit, so a stage change rewrites the file. The record layout document describes the file's fields and the three rulings below in its decisions table.
  • **The headline file is served from the site** at https://theupgradetree.com/data/headline.json, as I ruled for every tracker on 2026-10-17: the build copies the committed file into the site unchanged, so a reader who knows only the site can find the number. The family's Projects page can read it there or from the repository.
  • **The subject notice.** Every page's footer carries one line between the attribution line and the family line, in the place and register of Outrunning Ageing's medical notice, decided on 2026-10-17: "This site records what has been shown and what has been promised about changing a human being on purpose. It recommends nothing, and nothing on it is medical advice."
  • **The README made true, and the sitemap.** The repository README gains a status section dated 2026-10-17 (the count, the headline file, the skin, the reading's schedule as proposed and waiting on me, the quoted-recordings check, the cross-links to Outrunning Ageing, what is open) and its layout and deploy paragraphs say what is so today; the site README's page table gains the claims and reading pages, the served headline file and the sitemap, and no longer says the exported JSON is not committed. `robots.txt` had named a sitemap since 2026-10-16 and there was none: the site now builds one from the record at /sitemap.xml (every public page, nothing else, the way the parent site does), and `robots.txt` names it. The review's other findings for this site (the story page printing raw Markdown, the method and claims pages out of date, the date hook, a publish script) are not in this change.
  • **I ratified the two sets proposed earlier on 2026-10-17, as proposed ("yes to all", the same day).** The check for quoted recordings (the rule, its record table, the set opening sentences, the captions link and the extra passages checked) and the cross-links to Outrunning Ageing (the optional `sibling` field, unknown fields on an upgrade refused, and the weekly reading's rule that a disagreement between the two sites' stages is a flagged item for me and never a copy). The record layout and weekly reading documents now say "ruled" in place of "proposed"; the 2018 film's claim notes that the relink to the full copy was ratified the same day. The Dubai 2017 claim keeps its still-to-check line. No stage, date or claim account changed.
  • **An automated check for quoted recordings, proposed for me to ratify (2026-10-17).** I ruled earlier the same day that he will not check quoted passages by ear and that automatic captions are acceptable, and asked for a reliable automated check instead. The rule, in the record layout under "Checking quoted recordings": a quoted passage from a recording counts as checked when YouTube's automatic captions and a local Whisper transcription of the audio agree on the quoted words (case, punctuation and filler sounds ignored); a difference is flagged with both readings and settled by a larger local Whisper model, then an official transcript, then the speaker's own text, and stays open if none settles it. It checks the words, not what a claim means, and never changes a claim's date, status or account. A claim carries the result in a small table (`transcript_check`) and its `checked_how` opens with one of three set sentences, which the export script now checks. `scripts/check_quotes.py` applies the rule (how to run it: `docs/check-quotes.md`); it fetches captions and audio, transcribes the quoted stretch, prints the table and never writes to the record.
  • **The seven recorded claims, run through it on 2026-10-17** with Whisper large-v3-turbo as the second transcript: all seven quoted phrases are found in both transcripts, the 2018 film's only after a full copy of it was found (the first upload is now a 3-minute clip; its captions are still readable, the new link has the audio, and the quoted line is at 1:11:49 to 1:11:50 in both; the link on the entry now points to the full copy). One extra passage was checked as well: the audience answer at the Neuralink launch of 2019 (about 01:22:15), where the captions read "sympotic", the turbo model "sympathetic" and a larger model (Whisper large-v3) "symbiotic", which is the entry's reading; the entry is marked settled by the larger model. The "still to check" lines about listening are gone from six entries; the Dubai 2017 entry keeps one for a different reason (its account says "around ten bits per second", where both transcripts have him say "10 bits per second or 100 if you're being generous"; the account is unchanged). No claim's date, status or account changed. The Whisper settings for the check use no random retries, so the same audio reads the same each time (an earlier run with the default retries once read "bandwidth" as "bandless" in the Seo talk). The claims page now describes the check in one sentence.
  • **The five lifespan-and-repair entries are linked to their matching leaves on Outrunning Ageing** (outrunningageing.com, the sibling site that asks when repair overtakes ageing, live since 2026-10-17), both ways on this side: a drug makes people live longer in good health; partially resetting cells reverses ageing in a living animal; a drug clears aged cells from a person's body; an organ from a gene-edited pig keeps a person alive; a printed or lab-grown organ works in a person for a year. Each entry now carries the id of its leaf there in a new optional field, `sibling` (the same name and the same idea as the field those five leaves have carried pointing here since 2026-10-16), with a dated revision and a note; each entry's page shows one line linking to the leaf's page there; the export carries the id and the address through. Checked on 2026-10-17 against the sibling's public records: all five pairs sit on the same rung with the same date and the same record, so no stage changed and the headline count stays at 28 of 35. The field and a second rule, that an upgrade carrying a field the layout does not list is refused, are proposed in the record layout document and wait on me. Done in a cloud session, so it arrives as a pull request rather than a commit on main.
  • **The weekly reading gains the stage-agreement check.** After the upgrades are read, the scanner runs a new script (`scripts/sibling_check.py`) that reads each linked leaf's record from the sibling's public repository and prints both stages, dates and records with a verdict per pair. A pair that differs, or a leaf that cannot be read, makes that upgrade's item a flagged one for me; the scanner never copies the other site's stage, because a rung here moves only on the kind of record the entry names. The procedure, the standalone instruction for the scheduled run and the decisions list (item 12, proposed) say so.
  • **I ruled the three calls the claims-and-bandwidth work left open ("yes to all", 2026-10-17; each question was raised 2026-10-16).** The claims page keeps its name, "The claims". The 2006 BrainGate paper stays on the bandwidth series as its starting mark with no rate, because it states none: it shows in the table and not on the chart. Neuralink's own 8.0 bits a second (2024-05-08) belongs on the pointing chart, shown as a hollow dot and labelled in the legend and the table as the maker's own release, with the best peer-reviewed figure (4.16, Stanford, 2017) named beside it. The page and the chart already did all three, so nothing on the site changed; the rulings are recorded in the decisions table of the record layout document.
  • **The listening pass.** Every claim held on a recording was checked again on 2026-10-17. All seven recordings turned out to have automatic captions only, including the 2017 Dubai one, which the record had called human-made-looking. Six were transcribed again from the audio by a second, separate speech-recognition program (OpenAI's Whisper): Musk in Dubai 2017, Harari at Davos 2018, Musk at the Neuralink launch 2019, the Neuralink update 2020, DJ Seo's talk 2022, and Musk in Dubai 2025 (the whole talk). In every case the two machines agree on the quoted phrase and on the account, so no quoted words and no claim changed; four timestamp ranges were widened by a few seconds (Dubai 2017, the launch, the 2020 update, Seo) because the two machines place the phrase up to three seconds apart. No person has listened to any of them, and the site now says so: each of the six carries a "still to check" line, and each names the exact stretch to listen to. One word is left to an ear: in the launch-event audience answer the captions read "sympotic" and the second program "sympathetic", and the entry reads "symbiosis".
  • **One correction from the listening pass.** The Harari entry said the word "hackable" is absent from his Davos talk of 2020. It is not: he says "We are now hackable animals" at about 12:13, and the automatic captions had it as "packable animals". That is very probably where the circulating phrase "humans are now hackable animals" comes from. The 2018 entry stands (the word occurs nowhere in the 2018 talk, which a second transcription of the whole recording confirms); only its note about the 2020 talk is corrected. The 2020 talk is not entered as a claim; that is my call.
  • **Two claims could not be listened to.** The 2018 film "Do You Trust This Computer?": the upload the record linked is now a 3-minute clip of other material, not the film, so the recording cannot be reached; the quoted line matches CNBC's own transcription of 2018-04-06 word for word, and the entry says it rests on captions and that transcription. The Schneider Financial Times piece stays headline-only behind the paywall. One post was upgraded: Musk's "If you can't beat em, join em" was read on 2026-10-17 from X's own embed data for the post, which confirms the words and puts the posting at 07:01 universal time on 2020-07-09 (00:01 Pacific).
  • **The two open checks from 2026-10-16.** The first mitochondrial-donation child's birth date: still no source from the clinic or the authors. The paper and the authors' later commentary give only his age (7 and 11 months); the clinic's pages carry press clippings; the day 2016-04-06 rests on CNN, The Guardian and New York magazine of 2016-09-27 and 28. The entry's note now says exactly that, and the stage and the date stand. Vidal's 1977 maze paper: still only the abstract can be read; no open copy exists; the entry's note records the second attempt and the date stands. The third item from the first reading, the 2016 Pittsburgh touch paper, was settled on 2026-10-16 and was not re-read in full here; its first-publication date of 2016-10-13 was confirmed from Europe PMC.

• 16 October 2026

  • **The top bar carries short handles** (Branches · Reading · Claims · Story | Method · Disagree · Source; Home drops, the site's name is the home link), the full plain names staying as the page titles and in the footer, under rule 4.7 of the tracker pattern I ruled on 2026-10-16 after seeing the three live bars side by side. The seven links now sit on one row beside the site's name at 1280 pixels, so the collapse band up to 1320 pixels added earlier the same evening for the long names is gone and the standard's own collapse at 820 pixels is enough.
  • The claims are on the site. "The claims" (/claims/) sits in the heavier tier of the top bar, as the tracker pattern I ruled on 2026-10-16 puts the page a site is built around. It shows the twenty-seven claims from the project's claims log in date order, 2000 to 2026, each with its one exact phrase, the plain account with the hedges kept, the source link and how the words were checked, and its marking in words where it is held on coverage only or where a version that circulates has no primary source; beside them, on the same page, what the seven branches have actually reached and the headline count, so the gap between talk and delivery is visible (the pattern's rule 6.4). The entries moved from the TimeAssembler document into the repository as one file per claim under data/claims/, nothing added or dropped, with my three sourcing rulings of 2026-10-16 kept as the document recorded them: the two Nicolelis entries and the Wait But Why entry held as unsourced, coverage only, with no quoted words; the Schneider 2019 piece kept as a claim with its fiction framing as a hedge; the Warwick 2003 paper dated by the paper with the implant's date beside it. The one speaker looked for and not found, and the fourteen leads noted and not yet checked, are kept with the record and shown at the foot of the page.
  • The brain-to-machine bandwidth series exists: data/bandwidth.toml, eight dated figures from 2006 to 2024, each in the unit its source used, with who measured it, in whom, whether it was repeated, and whether it came from a peer-reviewed paper or the maker's own release; every figure read in its primary source on 2026-10-16 (the two claims not fully readable are in the points' verify lists). Two measures are kept apart, text out of the head (words a minute, with characters shown as words at five a word) and pointing at a screen (bits a second), each plotted on the link-to-machines page as dots with a stepped line for the best so far and dashed lines for the rates a reader already has: speech in conversation at 160 words a minute, a fast typist at 120, typing on a phone at 115 characters a minute, an engineer with a mouse at about 10 bits a second, speech as information at 39. The best figure so far for text is 78 words a minute (San Francisco, August 2023), about half the speed of talking with one word in four wrong; for pointing, 8.0 bits a second (Neuralink's own figure, May 2024) against a peer-reviewed best of 4.16 (Stanford, 2017). The 2006 BrainGate paper is on the series as its starting mark with no rate, because it states none.
  • The record layout gains the claim record (one file per claim, named by date then slug, with a sourced or unsourced status and the paraphrase table the charter asks for) and the bandwidth series (measures, points and baselines), both validated by the export script before anything is written. The method page's "what is not here yet" now names only the weekly reading.
  • The hearing gene therapy moved from first human to in clinical use on the FDA's accelerated approval of Otarmeni (2026-04-23), which the first weekly reading found the entry had missed; decided on 2026-10-16, with a dated revision, the first-human record kept in earlier stages, and Otarmeni, OTOF and lunsotogene parvec added as watch terms. The headline count is unchanged at 28 of 35; the five-rung totals are now 4, 3, 18, 7, 3.
  • The weekly reading built: a procedure in plain words (docs/weekly-reading.md, adapted from The Belt Equation's scan plan and procedure, every design choice with its confidence and an alternative, the schedule proposed), the reading record's layout in the record layout document (one file per run under data/readings/, never edited, four verdicts: quiet, noted, stage change, flagged) with the export script refusing any record that breaks it, and an optional list of extra search terms (`watch`) on each upgrade. The scanner may move a stage one rung on a record of the kind the upgrade names, and may not touch a test, a settler, a verify list or the headline rung; everything else is a flagged item for me.
  • The first reading, done by hand on 2026-10-16 (record at data/readings/2026-10-16.toml): all 35 upgrades read, 75 searches, none moved, two noted (the first United States patient with a bone-anchored arm with implanted nerve and muscle electrodes, 2026-09-17, in a study and not under an approval; and the Australian regulator's page the psychedelic entry could not load earlier in the day, read tonight and matching the entry, a claim the verify session cleared the same evening), one flagged for me: the FDA approved Regeneron's Otarmeni, the first gene therapy for a genetic hearing loss, on 2026-04-23, so the hearing gene therapy entry's note that no regulator had approved one is wrong and the entry would move to in clinical use. Two claims still to check were looked at and not settled (the first mitochondrial-donation child's exact birth date; whether the 2016 Pittsburgh touch paper itself includes the robotic hand), and a third was found supported but unread (Vidal's 1977 maze-by-brain-signal paper as an earlier first human for the scalp headset).
  • The site: the home page's last-checked strip now reads the latest reading (the date written out, how many upgrades were read, what moved, what waits on me); a "This week's reading" page at /reading/ in the top bar's heavier tier, with every run newest first, the items grouped by branch, the flagged items waiting on me, and a key to the four verdicts; each upgrade page shows its last three readings. The page, the strip and the verdict badges are copied from The Belt Equation's Radar with a comment saying so. The top bar, now six page names, sits tighter below 1320 pixels and collapses behind the menu button below 1080, as The Quiet Galaxy's does.
  • The scheduled run: docs/weekly-reading-prompt.md is the standalone instruction a fresh session follows, and a scheduled task on my Mac ("Upgrade Tree: weekly reading of the upgrades against the news", Sundays at 18:00 Pacific) opens the repository and follows it. The task can be switched off in the app's Scheduled section; moving it to a cloud routine so it does not need the Mac awake is proposed, not done.
  • Home page checked against the tracker pattern's type roles after the family design review of 2026-10-16 (the sans for an answer, the monospace with aligned figures for a number, the serif only for prose): the count "28 of 35" is in the monospace with aligned figures and its claim in the sans, so nothing changed; the question under them stays in the serif, as the parent site's does. The picture at 1280 pixels is `docs/home-page-2026-10-16-type-check.png`, taken after the Plate skin landed, in the dark theme.
  • The site wears its own skin, "Plate": warm paper, ink and one plate green, with the human figure standing on the page between double rules like a plate in an anatomy book, light by default (dark kept for readers who ask for it), no starfield, no galaxy, no web font. Proposed with mock-ups in the Sentient Horizons design standard (`design/proposals/upgrade-tree-skin-2026-10-16.md`) and decided on 2026-10-16 after the standard split into a shared skeleton and a skin per site. The stylesheet is now `skin.css` (this site's tokens) plus the shared `shell.css` 2.0.1 plus the site's own rules; the starfield component is gone; the favicon and the browser's theme colour take the paper and the green. The skeleton copy also fixes a bug the whole family had: in the light theme a solid button's text took the link colour and all but vanished. Pictures before and after in `docs/site-2026-10-16-*.jpg`.
  • The six claims written from memory were checked against their primary records on 2026-10-16. Three are settled and their notes removed: the Australian regulator's notice of 3 February 2023 on psilocybin and MDMA (in force 1 July 2023) says what the entry says; the company's own newsroom links its embryo-product launch release, dated 4 June 2025; and the 2016 Pittsburgh paper does include touch from a prosthetic hand's sensors (84.3 per cent of finger touches identified), so the touch entry's record stands. One moved: the headset entry's in-wide-use date goes from 2015-12-31 to 2014-08-28, the earliest dated copy in the Internet Archive of the maker's page that says over a million of its chips are in circulation (revision recorded on the entry). Two remain open, each reworded to say what was tried: the birth date of the first child born with a donor's mitochondria (the paper gives no birth date, so 2016-04-06 still rests on press coverage) and Vidal's 1977 paper (only its abstract is readable, and it does not mention the maze). No stage changed. The count is unchanged at 28 of 35.
  • Home page first screen redesigned to the tracker pattern's section 1.3 order in The Belt Equation's layout: the count in the largest type with its claim and the five-rung totals, then the question, one paragraph, the last-checked strip and two buttons ("What has been tried?", "Who is counting"), with the figure beside them; the whole first screen fits at 1280 by 720. Asked by me on 2026-10-16 so a new reader meets the answer before the question.
  • The real list written: thirty-five upgrades across the seven branches, from the first leaf list I ruled on 2026-10-16 (the charter calls it thirty-six, but names thirty-five; the count is the charter's to correct). The four example entries are gone, their three ids kept and rewritten (the sight implant, the ageing drug, the gene-edited child, the wireless brain chip). Every stage above idea only carries a date and a primary record (a paper, a regulator's database entry or notice, or the institution's own release naming the patient), each checked against the record itself on 2026-10-16; the six claims not yet checked against their record are in the entries' verify lists. The headline count is 28 of 35 at a first human or beyond: 4 idea only, 3 shown in animals, 19 first human, 6 in clinical use, 3 in wide use.
  • Six rungs differ from the charter's expectation, each explained in the entry's notes: the bone-anchored limb with nerve control sits at first human, not in clinical use (only the bone anchor is approved); the fear-memory treatment sits at first human, not shown in animals (a 2009 human study erased a conditioned fear); the artificial womb sits at idea only, not shown in animals (no mammal has gone from conception to birth outside a body, and the test says so); and three first humans are earlier than the charter had in mind (the visual cortex implant, 1968 not 2018; the wired cursor implant, 1998 not 2004; the memory implant, 2012 not 2018). The wireless brain chip's first human moved from Neuralink's 2024 implant to a fully implanted Dutch system reported in 2016.
  • The record layout gains a fourth horizon value, "top", for entries already in wide use, so the page does not say "longer than 20 years, if ever" about a finished thing; and a dating rule: an event the record itself dates (a birth, an operation) takes the event's date, a study takes its publication date.
  • The home page already showed the totals for all five rungs directly under the headline count, as I ruled on 2026-10-16; checked, not changed.
  • The brain-to-machine bandwidth series is still to come: the record layout has no place for it yet (it is to be its own data file, not a field on an upgrade), so the speech and cursor entries carry their figures in their notes as the series' first two points, in the units their papers use.
  • The top bar and footer renamed to the page names of the tracker pattern I ruled on 2026-10-16: the spine is Home, Branches and "The story so far"; the lighter tier is "How the numbers are worked out" (the method page, retitled to match, still mostly about how a stage is decided) and "Disagree with something". "This week's reading" is the name for the weekly scan page once one exists; there is none yet, so the home page's "Last checked" strip stays plain text and will link there when the page comes.
  • theupgradetree.com and upgradetree.com, registered by me on Cloudflare the same evening, attached to the site as custom domains, with www.theupgradetree.com, upgradetree.com and www.upgradetree.com forwarding to the same page at theupgradetree.com. A small Worker in front of the static files (`site/worker/redirect.js`, the Quiet Galaxy's way) does the forwarding; the workers.dev preview address keeps serving the site directly.
  • First version of the website built: a home page with the question, the human figure shaded by the furthest stage reached under each of the seven branches, and the headline count; a page per branch; a page per upgrade with its stage ladder; the Method, Story and Disagree pages. Deployed to the workers.dev preview address; the domains (theupgradetree.com and upgradetree.com) are mine to buy and attach.
  • The record layout written (docs/upgrade-schema.md): seven branches in data/branches.toml and one TOML file per branch, each upgrade carrying a stage from the five-rung ladder (idea only, shown in animals, first human, in clinical use, in wide use) with the date and public record that put it there. No probabilities, by my ruling of 2026-10-16.
  • Four example upgrades seeded, one each under Senses, Lifespan and repair, Inheritance and The link to machines, marked "EXAMPLE ENTRY" in their notes. They show the record shape; the real list is being written from the project charter in TimeAssembler and replaces them. Every date in them written from memory is listed in the entry's `verify` field.
  • Repository created on GitHub (jfredson/upgrade-tree, public) with the MIT licence for code and CC BY 4.0 for the record.

This page is the repository's changelog as written: CHANGELOG.md.