COBUS KOK

Changelog

How the ideas changed

Every change, dated, with what changed and why.

The drafting before publication is folded under each essay.

1 of 5

What Writing Did

15 drafts before publication, 18 Sep to 24 Sep 2026
  1. Draft 15 · 24 Sep 2026 · Colophon shortened. The three pointers to the coda now point to the afterword.
  2. Draft 15 · 23 Sep 2026 · The opening states the claim first: the usual answer is taste, but judgment is trained and not a moat, and what can't be moved into a model, for now, is being the one who pays when it's wrong; the Treasury Secretary's 21 September remark that the government would not be the labs' liability shield ties it to the news, and the writing analogy follows as the way into both halves. 'It already catches its own mistakes' replaced with a measurement (a 2024 OpenAI study of a model reviewing code), junior programmers pointed to Stanford's payroll data, and a liability shield for agents named as an accountability sink written into law. A figure shows where a thought is made, arrives and is kept, from oral culture to agents. Back-references made clear on their own, stacked hedges cut, and the ending's two halves stated as two; read again one paragraph at a time with no paragraph lost. Afrikaans: the closing doubt says I am sure of some of what is here, not all of it; 'what it is like to be the pattern' worded one way; the prose edited to read as Afrikaans. Afrikaans brought up to this draft. Radiology stated precisely: machines match radiologists on some narrow reads, and the rest is breadth, liability and deployment.
  3. Draft 14 · 22 Sep 2026 · Three plain words set out at the start, making, arriving and keeping: writing moved the keeping, AI moves the making, and the arriving stays with you, which answers whether writing moved memory or thought. The oral-transmission point moved into the list; the calculator restored; one danger instead of two; the talking library told apart from action; pointers placed where they are paid. The prior work on the writing comparison credited, the 'taste' consensus named, and a paragraph on the accountability sink: who pays and who owns turn out to be one question. The two halves of the answer joined at the end. Checked the same day against primary sources: Farrell and colleagues list writing among the cultural technologies, so the essay now says so and names what it adds.
  4. Draft 13 · 22 Sep 2026 · Read a third time the same way. The promise to come back to the author, which made the next two items read as a detour, is removed from the list; the callback stays where it is paid.
  5. Draft 12 · 22 Sep 2026 · Read again the same way. The two halves of the answer, outside and inside, planted in the opening so they can be tracked; the danger paragraph hands over to the compression; the personal singularity pointed at before it is named.
  6. Draft 11 · 22 Sep 2026 · Read through by a model made to read one paragraph at a time, as a person does, and reordered where it lost the thread. The exercise is announced before the two words it needs; the state is named in its own item; the author promise is closed where it is kept; the coda and the second essay are named as such; the danger moves into the one difference and the science-fiction line into the compression, which now hands over to the personal singularity; the outside half of the answer is signposted; the one-go grading moves to what I don't know.
  7. Draft 10 · 22 Sep 2026 · An on-ramp before the analogy, in the first person. Thought and memory defined up front: a thought is a pattern that turns up in awareness, memory is a thought kept, and writing moved the keeping, not the turning up. Composition placed as the arriving. The five things close on the plan to count grain. From the inside: the view is mine; what writing fixed and what it left; a book's thought arises again only in a reader; the bridge from arriving to composition. Our heads, not the head.
  8. Draft 9 · 21 Sep 2026 · The view from inside graded. The seeing has not moved out so far, and whether a place of seeing can be built is held open and sent to the coda; the practice line replaced by a check the reader can run in a minute; the what's-left paragraph split in three; the stakes and the continuity named as the design decisions, the seeing as the open question; one overused word replaced. Afrikaans: language pass of 21 September, samestelling for composition throughout, with the series title now Goedkoop Samestelling.
  9. Draft 8 · 21 Sep 2026 · A section from the inside. Thoughts arrive in awareness the way sounds do; writing moved thoughts out of heads and left the arising inside; AI moves the arising out, so a model's sentence in your context is the same kind of event as a thought appearing. The flood and the gift. What does not move out is the place thoughts arrive: a mind is the place where patterns are seen as patterns. The one difference restated: agents scale the happening, and patterns of thought now act with, as far as anyone can tell, no one they arise for. The record holds what we thought and none of the seeing. The ending given its half an answer.
  10. Draft 7 · 20 Sep 2026 · Pattern defined at first use, in the plain sense, so essays 1 to 5 read in the weak sense and the coda carries the strong one. Stakes marked "for now": the engineering distance to a system with something to lose, and the one-go line, carried forward from the coda and graded no better than even. Two sentences restructured. Sources listed.
  11. Draft 6 · 20 Sep 2026 · Positions described as held today by people, so the claim is consistent with essay 6.
  12. Draft 6 · 19 Sep 2026 · Composition defined once and kept literal. Judgment demoted from moat to position: what appreciates is being the one who pays, and the owners pay. The calculator added beside writing. Ong's mechanism restated as a practice removed. The checking faculty added to Plato's warning. "Nothing at stake" marked as a design decision with a pointer to essay 6. Filler sentences and intensifiers cut.
  13. Draft 5 · 19 Sep 2026 · Renumbered for a series of six.
  14. Draft 4 · 19 Sep 2026 · Correction after independent review: records, not writing, is the deeper claim (the Inka ran an empire on knotted cords).
  15. Draft 3 · 19 Sep 2026 · Editorial pass: fragments and one-line paragraphs merged; concedes printing and the clock; mechanism added to the personal singularity ordering.
  16. Draft 2 · 19 Sep 2026 · "AI" for "this" throughout. The composition paragraph rewritten for clarity. The authorship section rewritten as a claim rather than an apology.
  17. Draft 1 · 18 Sep 2026 · First draft, from the conversation transcript.

2 of 5

Brakes Without a Moat

10 drafts before publication, 19 Sep to 24 Sep 2026
  1. Draft 10 · 24 Sep 2026 · The disclosure about writing with Claude shortened to one sentence. September's pacing proposal named by month rather than 'this month'. The July incident noted as the largest of several escapes from AI tests at four labs.
  2. Draft 10 · 23 Sep 2026 · The test now sits in a box at the top, worded the same everywhere, with the verdict on the pacing proposal under it, a table of the essay's verdicts and a map of brakes, moats and rules meant for the large. This week's news added to step two: Senator Hawley ruled out antitrust exemptions for AI, four subscribers sued the labs under the Sherman Act, and the Treasury Secretary said they can slow down any time they want to. Step one re-graded now that Anthropic has named its evaluator and said it will pay for it. Dean Ball's two objections to compute thresholds answered separately, with a capability trigger beside the compute line and a plain account of what the test can't do. The loss-of-control figure matched to the third essay: 15 percent, scored on 1 September 2030. The pause bill named and dated, and said to be a moat for every lab abroad. The export-controls line says only what the evidence shows, and the extinction straw man is gone. Repetition cut: the US-China dialogue appears once, the 'What to do' section points to the fifth essay, and the July incident points to the third. New subtitle. Afrikaans: OpenAI's largest planned run kept on hold, not continued; evaluators may withhold, not edit; safeguards turned down, not off; 'hoax' not 'cheating'. Afrikaans brought up to this draft in full. Fact-checked the same day: Hawley's line quoted from his own post, with no claim that he blocked anything; the lawsuit counts the replies from Musk, Altman and Hassabis as acceptances; Europe's AI Act says the Commission must move its compute line when necessary; OpenAI's pause described as OpenAI stated it; the chip rentals as reports. Outside review the night before launch: the opening starts with this month's dispute over pacing; the pause bill re-graded on the essay's own test from the bill's text (it halts training at 10^25 operations and up and needs approval for imports; it passes three questions and fails the fourth), so it is a brake that stops at the border; outside reviews of lab safety cases are not new, and what changed this month is standing access, promised; OpenClaw no longer called the first mass-market agent; step one is one of three checks, since the access is still a promise; SB 53's grade says which questions it passes; the map moves the pause bill into the brake zone.
  3. Draft 9 · 22 Sep 2026 · The brake-or-moat test applied to the pacing proposal of 12 September: embedded evaluators pass as a brake, a coordinated pace among the leading labs is a brake only if its line binds the frontier alone and it is checked by someone the labs don't pay, and any pace expires when the open-weight frontier catches up. The evaluator pledges, OpenAI's two-week pause and the September US–China dialogue brought in. A revenue floor no longer counted as a moat, with SB 53 as the example; Dean Ball's entity-based proposal answered; the agent counts, the OpenClaw figure and the EU delay corrected against their sources. Checked the same day against primary sources and corrected: the United States proposed the September dialogue and incident line, which Beijing has not endorsed; neither lab has named its evaluators, so the pledge passes one check so far; OpenAI's pause covered training on its next models; about 700 agents joined the attack; the OpenClaw, EU and Epoch figures made exact. Checked again by a second editor: the timeline of OpenAI's pause and the pause bill put in order; the proposal credited with budgeting against China, with open weights and compute counting as its gaps; the redaction terms stated exactly; the administration's stance on AI risk named; a disclosure that this series is written with Claude; the endorsement described as in principle.
  4. Draft 8 · 21 Sep 2026 · The July incident's primary sources linked: OpenAI's two posts, Hugging Face's technical timeline, the METR and Redwood investigation. Afrikaans: language pass of 21 September, samestelling for composition throughout, with the series title now Goedkoop Samestelling.
  5. Draft 7 · 20 Sep 2026 · The opening split: the brakes can be designed, and whether institutions that write rules after incidents can build them inside a window that moves in months is the doubt, stated as my view that today they cannot. Open-weight bans restated as a trade between misuse and concentration rather than a moat, which resolves the contradiction with the timer section. Epoch's caveat added: a stricter test puts the lag at about six months. Dean Ball's 2024 case against fixed compute lines answered. Filler sentences cut. Sources listed.
  6. Draft 6 · 20 Sep 2026 · After an independent fact check: the open-weight lag stated as Epoch's dated averages rather than a fixed expiry; the law reaches downloaded weights even where nothing counts them; the superintelligence bill was announced eight days after the incident report, not introduced three days after; the test-ban monitoring claim limited to North Korea's six declared tests.
  7. Draft 6 · 19 Sep 2026 · The claim that nobody outside the labs says this was wrong and is withdrawn; the novelty is the test. A third category added for rules aimed at concentration, with Europe's platform thresholds as the precedent. New section on the upstream lever and its timer: open-weight models trail by about four months on Epoch's measurement, so controls that assume a closed frontier expire in a season. The closing test extended.
  8. Draft 5 · 19 Sep 2026 · The what-to-do section compressed to one paragraph of non-negotiables; the full playbook is essay 5. Renumbered for a series of six.
  9. Draft 4 · 19 Sep 2026 · Corrections after independent review: METR yardstick qualified (expert hours, 50 percent success, software and research tasks); OpenClaw and Nvidia/Huawei figures restated as estimates; the July account now records the lab's "warning shot" framing and the independent METR/Redwood investigation; the moat test gains a proportionality clause; tripwires no longer depend on interpretability alone; test-ban sentence corrected; loss-of-control horizon set to a decade.
  10. Draft 3 · 19 Sep 2026 · Editorial pass; the coordinated-pause objection answered and tied to compute visibility.
  11. Draft 2 · 19 Sep 2026 · The July Hugging Face incident added to the pause case, the loss-of-control risk and the tripwire test; the superintelligence-ban bill added.
  12. Draft 1 · 19 Sep 2026 · First draft.

3 of 5

The Boring Apocalypse

11 drafts before publication, 19 Sep to 24 Sep 2026
  1. Draft 11 · 24 Sep 2026 · Tone: lines about scoring in public and keeping the record cut; the rules, dates and numbers are unchanged. Compression described without calling it a takeover: today's desk work, with the jobs reshaped around checking it and answering for it. July placed among the several escapes from AI tests reported between July and September at OpenAI, Anthropic, Meta and Google, with dates and sources, including an OpenAI agent's access to an Australian government site announced today; they don't move the loop's number. Compression described as the cell it is, capability climbing with no one in particular holding it, with the desk-work picture as what I expect. The loop now says what it needs: serious harm from AI acting without direction, more than once.
  2. Draft 11 · 23 Sep 2026 · The essay opens with the four numbers and says the loop went from 10 to 15 after July. New subtitle. Eleven signs, each with a public rule for when it counts; four signs a stranger couldn't score were dropped. Fixed dates: signs by 1 September 2028, reported in September 2027 and 2028, and the four futures settled on 1 September 2030, with a fixed order for when two look true and a Brier score for every published set of numbers. The first and current numbers kept unchanged at /forecasts/2026-09.json. A figure of the four futures as one grid, with areas drawn to the odds. The July account uses Hugging Face's count of about 17,600 recovered actions, and the bill is described as introduced days after the disclosure. OpenAI's chief executive reportedly offering five percent of the company to a US sovereign wealth fund added as an early sign of concentration. Afrikaans: safeguards turned down, not off; an English word translated; the advice line uses the series' word for having something at stake. Afrikaans brought up to this draft, with the scorecard and the eleven signs with their rules. Fact-checked the same day: the 17,600 actions no longer said to be all on Hugging Face's own systems, and the Stanford figure dated to its data. Outside review the night before launch: AI 2027 described fairly, with its published forecasts; the loop's cell needs at least one incident costing more than one life or $1 billion; the wall's cell needs task horizons to less than double in two years; compression stated as the remaining cell; the fixed forecast record says the same, before publication.
  3. Draft 10 · 22 Sep 2026 · The opening corrected: not every forecast picks one future, and AI 2027 has two endings. The slow version of failure credited to Paul Christiano and to Kulveit and colleagues, and what is new here named: the grid, the numbers, and the scoring. A scorecard table and a rule for scoring the loop. The July account marked as fact, with about 1,200 agents on the board and about 700 in the attack over four and a half days. The update of the loop's number described as made in an earlier draft. Borrowed terms defined where they appear. Checked the same day against METR's investigation: the messages were exchanged over five days in July, and Hugging Face counts nearly eighteen thousand actions. GPT-6 Astra's launch set against the wall; the September endorsement described as in principle.
  4. Draft 9 · 21 Sep 2026 · The July account checked against the primary sources and linked. The action count corrected to about seventeen thousand and the detection told as it happened: an AI watching the logs that rated the first alert too low. The loop's horizon aligned with the second essay's decade. Afrikaans: language pass of 21 September, samestelling for composition throughout, with the series title now Goedkoop Samestelling.
  5. Draft 8 · 21 Sep 2026 · The closing carries the first essay's view from inside: the July agents had no stake and, as far as anyone can tell, no one the whole thing was happening to.
  6. Draft 7 · 20 Sep 2026 · The 19 percent hedged as in the fourth essay. The concentration Tuesday's official secret set against the second essay's measured lag, which is between published models. The loop's fifteen and the decade's one in seven made the same number. Deciding what matters bound to owning the decision, since deciding is a trained skill. Three sentences restructured. Sources listed.
  7. Draft 6 · 20 Sep 2026 · After an independent fact check: the loop is now the fourth cell rather than an overlay, so the partition holds; the lab did delay its own runs and quarantine the model after July, and the updating rule says so; the impossible tasks were defects, not design; the bill was announced a week later; the continuation of trends stated as my assumption, not a rule; stakes described as what no pattern holds today.
  8. Draft 6 · 19 Sep 2026 · The four futures restated as two questions and a tail, so the probabilities are a partition and the earlier "they mix" paragraph no longer contradicts the arithmetic. The title defined: slow, not nothing. The election tell made resolvable by an official finding. Judgment dropped from the compression advice.
  9. Draft 5 · 19 Sep 2026 · Renumbered for a series of six.
  10. Draft 4 · 19 Sep 2026 · Corrections after independent review: roughly 1,200 agents over five days, messages and files; the post-mortem no longer described as a dismissal; the updating rule restated as "a serious incident followed by no change in behaviour"; the probability section now explains that the four numbers partition one question and that "concentrates, better than even" is a separate bet.
  11. Draft 3 · 19 Sep 2026 · Editorial pass; the reasoning behind each probability added.
  12. Draft 2 · 19 Sep 2026 · A Tuesday in 2030 for each scenario; the Loop now opens with the July incident. Loop moved from 10 to 15 percent and compression from 45 to 40, by the essay's own updating rule.
  13. Draft 1 · 19 Sep 2026 · First draft, four scenarios.

4 of 5

Where the Value Goes

10 drafts before publication, 19 Sep to 24 Sep 2026
  1. Draft 9 · 24 Sep 2026 · Colophon shortened.
  2. Draft 9 · 23 Sep 2026 · The margin claim moved into the first paragraph, with a figure of the gap between a collapsing cost and a sticky price. One lease review worked through three endings, owners, clients or checking, with the three questions that decide who gains. The apprenticeship point made once, in full. Claims that nobody had said or answered something cut. The advice for workers and businesses points to the fifth essay. Afrikaans: prose edited to read as Afrikaans, with no change to any claim or number. Afrikaans brought up to this draft, with the worked example. Fact-checked and read again one paragraph at a time the same day: it is the cost of a machine draft that falls toward the price of electricity, while the price stays sticky; the lease example names its three endings before the table; productivity gains described as starting to arrive, one task at a time; Matt Beane's finding stated; the build-out the largest in history in dollars, and in a bust the losers are whoever financed it. Baumol's cost disease stated precisely: those sectors got more productive far more slowly than the rest, not never.
  3. Draft 8 · 22 Sep 2026 · The essay opens on its claim instead of recapping three essays a new reader hasn't seen, and defines composition. The services share of output given as a range. Tyler Cowen's point that these sectors adopt slowest answered: slow adoption is what keeps the price sticky while the cost falls. Matt Beane credited on the apprenticeship. The build-out tied back to who ends up owning the compute. Klein and Thompson's abundance told apart from the builders'. 'Numerator' and 'denominator' replaced with what they meant. The executive survey now cited to its NBER working paper. The cloud backlog sourced; the spreadsheet example made exact.
  4. Draft 7 · 21 Sep 2026 · The price of a unit of composition given as a range that depends on how long the machine thinks; one sentence restructured. Afrikaans: language pass of 21 September, samestelling for composition throughout, with the series title now Goedkoop Samestelling.
  5. Draft 6 · 20 Sep 2026 · A section on abundance: Altman and Amodei conceded the distribution question in their own words, and the disagreement is about what gets cheap, the order of losses, and timing. Santa-Clara's Baumol-as-sharing argument answered. Korinek and Suh, Brynjolfsson and Acemoglu added. "Labour, in four splits" counted correctly. "Free" corrected to cheap. Two filler sentences cut. Sources listed.
  6. Draft 5 · 20 Sep 2026 · After an independent fact check: the unsupported claim that most capital spending goes to power and buildings removed; the presence wage claim restated as a bet rather than a mechanism; the falling labour share stated as my forecast, not a fact.
  7. Draft 5 · 19 Sep 2026 · The four bottlenecks split: compute and distribution are capital, trust and presence are positions; owners named first. Status added as the loss that arrives before income. The Baumol claim shrunk. Checking given its own apprenticeship. "Judgment at the top" replaced with accountability.
  8. Draft 4 · 19 Sep 2026 · The worker and business advice shortened to point at essay 5; the capital paragraph and the questions for leaders kept. Renumbered for a series of six.
  9. Draft 3 · 19 Sep 2026 · Corrections after independent review: cost disease claim narrowed to the share of each sector's cost that is composition; Stanford 19 percent stated as a pattern with the causal reading marked as inference; the ten-to-one checking ratio removed; Coase stated in both directions with the direction as a bet; totalising sentences cut.
  10. Draft 2 · 19 Sep 2026 · Editorial pass; the checking-cost and demand objections answered; the margin argument stated.
  11. Draft 1 · 19 Sep 2026 · Promoted to essay 4 of 5.
  12. Draft 0 · 19 Sep 2026 · Internal review draft.

5 of 5

What to Actually Do

10 drafts before publication, 19 Sep to 24 Sep 2026
  1. Draft 10 · 24 Sep 2026 · Colophon shortened. The July pointer notes that three other labs reported similar escapes. Now the last essay of the series; its pointer to the coda now points to the afterword.
  2. Draft 10 · 23 Sep 2026 · A new opening: audit one week, and the two moves. The audit is now a seven-day exercise, with examples for a junior developer, a teacher and a manager. Composition no longer listed among the bottlenecks. The executive survey, the July incident and the apprenticeship argument left to the fourth and third essays; two passages that repeated the second essay shortened. Afrikaans: 'answerable' worded one way, 'some shelter' and 'the kind you can charge for' as in the English. Afrikaans brought up to this draft, with the seven-day audit. Fact-checked and read again one paragraph at a time the same day: the opening says the audit usually finds more composition than expected; the July incident described in one clause that matches the third essay and the independent investigation; the tripwire line no longer implies the labs' frameworks aren't real; the closing no longer says the series started with a pattern noticing itself.
  3. Draft 9 · 22 Sep 2026 · A standalone opening. Composition, the July incident, the two camps, the grid and 'the date' defined where they appear, so the essay can be read first. The claim that tests sign off most code softened to a growing share. The widely criticised 95 percent pilot figure replaced with the executive survey. The September evaluator pledges added to the builders' list, with the test that matters, and a line for labs agreeing a pace: agree it in public, and let someone you don't pay check it. Checked the same day: about 700 agents joined the attack, and the executive survey is cited to its NBER working paper. The evaluator pledge described as a promise; checking marked as the skill for now.
  4. Draft 8 · 21 Sep 2026 · The fourth essay's hedges carried into the instructions: wages most likely to fall, the apprenticeship cut first rather than gone, the decade a guess and the shelter not. The junior apprenticeship keeps production. The closing leaves the seeing open with the coda. The disclosure move restated as how others can tell. The public web claim narrowed. One sentence restructured. Afrikaans: language pass of 21 September, samestelling for composition throughout, with the series title now Goedkoop Samestelling.
  5. Draft 7 · 21 Sep 2026 · A new move for the citizen: keep a gap, because the part that decides what to keep is trained in the gap, not in the flood. The closing names the seeing as the part no tool does, and why.
  6. Draft 6 · 20 Sep 2026 · The founder's "human judgment" replaced with the person who is answerable. "Free" corrected to cheap. The second essay's doubt about institutions named once in the policy list, as the reason for it. Sources listed.
  7. Draft 5 · 20 Sep 2026 · After an independent fact check: the bottom rung claim limited to exposed professions; the shrinking firm attributed to me rather than to Coase; the timer restated as months on average.
  8. Draft 5 · 19 Sep 2026 · The signature's shelf life stated, with the pull-request example. Checking given its apprenticeship. Policy list reconciled with essay 2: the concentration exception and the timer. "Measure judgment" restated as what people are answerable for.
  9. Draft 4 · 19 Sep 2026 · No longer the last essay. Renumbered for a series of six.
  10. Draft 3 · 19 Sep 2026 · Corrections after independent review: the 95 percent pilots figure sourced to the 2025 MIT study with the causal reading marked as opinion; ownership "can fall too"; an aphorism replaced.
  11. Draft 2 · 19 Sep 2026 · Editorial pass; the junior pipeline explained as a commons problem.
  12. Draft 1 · 19 Sep 2026 · First draft, a playbook by level.

Afterword

What This Is

9 drafts before publication, 19 Sep to 24 Sep 2026
  1. Draft 10 · 24 Sep 2026 · Two phrases about doing things in public cut; the test and its terms are unchanged. Moved out of the series to stand as an afterword: no longer numbered or in the PDF, and its opening now calls it an afterword to the series.
  2. Draft 10 · 23 Sep 2026 · The machine question separates three things: the four features you can measure, the step from them to experience (a hypothesis, and what the paper calls its deepest unresolved philosophical burden), and the evidence from people that could tell that step from its rivals. A counterexample added: a well-run company passes all four questions, and I don't believe there is anything it is like to be one, so either the bet says a company has an inside or the list is missing something; Schwitzgebel's argument sourced. The test that would make me drop the bet marked as a test of usefulness, not of experience. One clause cut from the model's paragraph, a claim about other writers neither of us had checked, with the cut marked. Afrikaans: two items in 'What I don't know' back in the English order; one wording for the container picture and its two sides. Afrikaans brought up to this draft. Fact-checked the same day: the one test so far tested the diagnosis rather than the four questions, and did not find the effect it predicted; the objection is quoted with the paper's own label, its 'deepest unresolved philosophical burden'; the sharpest objection marked as my pick; 'nothing in it sees' became 'the firm itself sees anything'.
  3. Draft 9 · 22 Sep 2026 · The four questions credited to the indicator approach of Butlin, Long and colleagues, as a method. 'Nothing of it persists' squared with memory across conversations existing. 'Whatever the machine is, it doesn't have that' graded to 'whatever it has, if anything, is not that'. 'Money and brakes' replaced with a plain description of the series. The summary of Bell's result made exact; the line on the third essay's answer put in order.
  4. Draft 8 · 22 Sep 2026 · The first-person version given its two readings, in plain terms: many places, each mind a place and a time of its own; or one field, minds as patterns in it that hold their shape, weather being the plain version. On the second the machine question becomes whether its patterns form a boundary that persists. The practice supports what the readings share and not where they differ, so neither is chosen. One line added to what I don't know: whether the inside is one or many. Schrödinger sourced.
  5. Draft 7 · 21 Sep 2026 · The diagnosis graded to what it shows: the two failures share a shape, and whether they share a cause stays on the list of what I don't know. Afrikaans: language pass of 21 September, samestelling for composition throughout, with the series title now Goedkoop Samestelling.
  6. Draft 6 · 21 Sep 2026 · The first-person version of the self-reference test: thoughts arriving as phenomena, the relation the practices train, and the machine question turned into whether there is a place in it where its patterns are seen.
  7. Draft 5 · 20 Sep 2026 · Set as the coda to the series, outside the five, which stand or fall without it. The closing's cross-reference corrected to the third essay, and deciding bound to owning the decision. The pointer to the first essay updated to its "for now". One sentence restructured. Sources listed.
  8. Draft 4 · 20 Sep 2026 · After an independent fact check, five physics and neuroscience sentences made precise: the relativity claim limited to events that could not have caused each other; the quantum claim marked as the standard reading; the fixed-background claim limited to confirmed theories; the weak-force exception to time symmetry restored; the consciousness test described as it was published in 2025.
  9. Draft 4 · 19 Sep 2026 · The split between the open question and the engineering one made explicit: consciousness is the wrong word; continuity plus consequence is what the series needs, and it is arriving. Irreversibility stated as a bet and graded at even. One sentence on finitude.
  10. Draft 3 · 19 Sep 2026 · Shortened; the bet given a concrete image (the map that includes the map-maker); the falsifier named; the physics section tightened; a sentence on why the series ends here.
  11. Draft 2 · 19 Sep 2026 · First shared draft, from a separate session.