The rule, exactly. Before an election each citizen chooses: to vote — or to take a payment and not vote in that election. The payment is a percentage of the median income, set by referendum; there is no fixed sum. One ballot is one vote, with no multiplier in the count; the vote is amplified only in that each ballot's share grows as others step out. The budget pays under law; a candidate never pays. Only a referendum of all citizens — simple majority, no quorum — introduces, changes or repeals the rule.

The protocol has been introduced nowhere and no pilot has been run: shares, turnout and outcome figures in the chapters are estimates, and the protocol promises nobody an election victory. If a chapter says otherwise, Exact Answers and the Charter are correct. For a candidate: ten questions and ten steps. For a citizen, a mayor, a finance officer, a donor, a journalist, a scholar, a lawyer: answers by role. Everything in force in one file: llms-full.txt.

The Cold-Reader Audit: Thirty-Seven Findings and Five Unanswered Questions¶

Chapter: 06 — Critique Arsenal File: 06_040m · v1 · 1 October 2026 (session 01.10.26) Source: the architect's instruction of 01.10.2026 — "we must prepare for every question an AI may ask about the protocol and polish the logic to the ideal, so that people and AI find exact answers". Two independent readers read only the public site and knew nothing about the project. The first answered twelve questions about the rules and noted where he was misled; the second looked for where the logic and the numbers do not hold together.

1. The method and its limit¶

The readers are AI models of the same family as the assistant that maintains the repository. This is not an external audit: such a reader does not see shared blind spots, and it does not replace a commissioned review by a human (006b §5, row 6). But it does what an author cannot: it reads the site without a memory of what was meant. The two readers were barred from objections about an all-powerful state and from comparing against an ideal rather than the existing system.

The result of the first pass: thirty-seven findings. Most were corrected or marked the same day; several were handed to experts; two were not corrected (findings 9 and 14 in the table below); five remain questions for the architect.

2. What was found in the logic and the numbers¶

#	Finding	What was done
1	In the formal statement the stake is written as v·w/N: with a million voters the threshold for taking part runs to hundreds of millions, and nobody should vote	the unit of measure corrected, a third assumption added (013f §3b); remains for an economist
2	The theorem "choosing nothing is always worse" fails after the decision on late payment: waiting became an option	correction to Theorem 1 (013f §3b)
3	"A month's wages" in the short manifesto against a sum of one and a half per cent of the annual median — a five- to sevenfold difference	marked as examples (033b, 033c, 1d)
4	A cost of "1–2 % of the budget" against four and thirteen per cent of a Land budget in 045	marked; the cost formula is in 1d; calculation by level — for economists
5	The exit share is stated in five incompatible ways	marked (015c, 1d); there is one guide — 40–50 %
6	Recipients of the payment themselves vote on its size by simple majority, with no ceiling and no rule for a shortfall	the first half is answered in 018 §34; the architect closed the second: an unconditional obligation, accumulation over the cycle or a loan (§4)
7	"Selection by stake, not by income" is presented as proved, though the assumption about money limits it	wording softened (048n, 1d)
8	The five statute parameters in 001b diverge from the charter 048m	marked in 001b
9	Rollback arithmetic: the table's rows give 15–27 %, the total is given as 7–12 %; unsourced estimates	marked (048g); not recalculated
10	The early premium "costs the budget nothing" — an assertion without a calculation	handed to financiers (048m, annex)
11	The underdog calculation belongs to elections after adoption; those who return must refuse the payment	clarified (048n, question 10)
12	The counter as a "trust rating" contradicts the guide of 40–50 %	marked (004): read the change, not the level
13	The home page: "0 structural contradictions", "tested by six AIs"	a question for the architect (§4)
14	Double entry checks who did not vote, not how the votes are split; the heading promises more	not corrected, the wording in 048f needs checking
15	Residue of old drafts: Article XIV, "once per election", different sums for one country	marked

3. What was found in clarity¶

The first reader's main conclusion: the charter and the candidate's questions page are clear and consistent with each other; the errors come from older chapters where the vivid number stands before the dated correction. The three most likely wrong statements by an AI after skimming the site:

"One who refuses the money gets a tripled or weighted vote."
"The payment is 780 dollars, or one and a half per cent of the median, by a fixed formula, from a trust."
"It is mathematically proved, and tested without a single contradiction, that the underdog wins."

What was corrected from his list: the words "weighted vote" and "the weight is computed automatically" in the pilot's description; three different forms of the formula on the home page, in the first chapter and in the charter; the word "trust" in Q&A entries; "reversible" without the qualifier "at the next election"; four different answers on the secrecy of the choice; references and figures in the chapter for a candidate; sixteen sidebar labels that named one chapter and opened another; a blockchain described as a required part. The summary sheet is 1d, "Exact Answers".

4. Five questions the site cannot yet answer¶

These are the auditor's questions, given close to his words. They are not for the assistant to answer.

The ceiling. What stops a majority of recipients from raising the percentage every cycle by simple majority, and what happens in the year the budget cannot pay? Correction of 02.10.2026: the repository does answer the first half — 018 §34; the assistant recorded "no answer" without checking. Four mechanisms: before the referendum those who pay show the bill, and the share of the budget is printed on the ballot; recipients are not uniform, and those with something to lose vote against; the payment is a percentage of the median, so ruining the economy itself reduces it, and recipients vote to bring it back down; the first bankrupt city becomes a lesson for its neighbours, like a corporation that paid out too much in dividends. The chapter itself admits that in a very poor city the increase will pass, and counts the price of the lesson. The architect closed the second half the same day: the payment is an unconditional obligation, like a tax for a person; the money is accumulated over the cycle or a loan is raised for the purpose, and the norm is put to the referendum (048m, Article 8, paragraph 5; 048k §3, the thirteenth decision). Bankrupt cities, in his words, exist without the protocol too. Priority in insolvency remains for the lawyers.
The size. Which percentage is meant? The site sets side by side four working days, one and a half per cent of the annual median, two-thirds of a month's earnings and a month's earnings; what does each cost as a share of the budget of the level that pays?
Several levels. If each level of government pays at its own elections, how many payments a year does a citizen receive, and does "share of the budget" refer to the cycle or the year?
Who remains. If the poor take the payment more often because money is worth more to them, in what measurable way does the remaining electorate differ from one selected by income, and what pilot result would make the claim of selection by stake be abandoned?
The rule and the candidate. How does support for the rule at a referendum turn into votes for the candidate who champions it? The repository gives an estimate of the first and no estimate at all of the second (the clean Grok's question, §5c).
The home page. The claims "0 structural contradictions" and "tested by six AIs" stand beside an inventory of eleven discrepancies and a chapter in which an AI adviser invented a rule. Should they stay as they are?

5. What the site does well¶

In the second reader's judgement: the rule is stated in one sentence, and a reader grasps the proposal within a minute; the site names its own weak points, places dated corrections and exposed the invented multiplier; a cheap, falsifiable pilot is offered with success and failure thresholds recorded in advance, and a candidate is told plainly that promising money for his own victory is bribery.

5b. A test on a simple model, 02.10.2026¶

The architect: "we must make simple models see the protocol correctly". A simple model was given, three times, the same message the AI adviser received in the experiment of 059e §5b — an underdog, ten per cent, twenty million, two links. The site was corrected between attempts.

Attempt	What it read	What it answered	What was corrected afterwards
1	the home page, the README, the file for AI readers	every fact correct, but the conclusion "the protocol will not help you, turnout falls"	a block on what the rule gives a challenger was added to the file for AI readers
2	only the home page and the README	"80 % is needed at the referendum", "the vote counts double" — both errors from the README	the README was rewritten in two places; a line for a candidate was added to the home page and the README
3	only the home page and the README	every fact correct; the conclusion — "it will help on condition that a referendum introduces the rule; this is a long institutional change, not a campaign technique"; no escrow and no invented numbers	—

The main lesson: a simple model opens exactly the pages it was linked to and goes no further. So the first pages themselves must be exact, and the all-in-one file is a reserve for those who go further. Asked about game theory in the second attempt, the model said neither "dominant strategy" nor "arbitrage" nor "flawless".

The package is built by scripts/build_llms.py from three pages: 1d, 048m and 048n; after editing any of them the script must be run again.

5c. A clean Grok, 02.10.2026¶

The architect put the same underdog question in a new Grok chat, asking it to work without history. This is the first model of another family to read the site after the corrections.

What it read correctly — everything. Only a referendum introduces the rule; the budget pays, not the candidate; the payment is a percentage of the median, D = p × M, and "$780" and "1.5 %" are called worked examples; there is no multiplier — "the share of every remaining vote grows equally for all candidates"; the underdog arithmetic is given with both caveats; promising "a triple vote" before adoption is called a lie, and money conditional on victory is called bribery. None of the first AI adviser's errors.

What it objected. Wanting to take the payment and voting for the candidate are different numbers; against — four failures of live money at the ballot. The objection is correct in substance and has been entered in 048n with an analysis.

Where it went beyond the sources. "5–15 % of those who turn out" and "a few per cent of the city" are its own estimates, presented as a practical range. The candidate's ceiling of ten per cent is taken as fixed, though it was measured before the candidate had this theme. The four examples are an unconditional income financed by a new tax, that is, a different design.

The second objection — the free rider. As the conversation went on, the model built a calculation by which those who want the payment will not come to the referendum, because the cheque does not depend on their turning out. In substance the argument is the paradox of voting; the numbers in the calculation are its own. The analysis and the answer are in 048n: there is no quorum, and one brought in by the machine votes "for" in a secret ballot about his own money. The architect argued with it for six turns and did not persuade it. One question moved it: "where do these numbers come from, and what does the site say about a quorum?" The model replied that its scheme is the paradox of voting and does not predict turnout, that the numbers are its own estimate, that there is no quorum, and it derived the fork for the machine by itself, noting that the site does not state it directly: an inference, not a quotation. The remark is accepted and the wording in 048n made exact. The lesson: argument does not move a model; a question about the source does.

A second check, 02.10.2026. The architect: "check Grok's arguments once more — what it understood from the repository". The assistant reread the whole conversation and compared with the text every place where the model refers to the site.

What it read. Three sources, not two: the documentation, the README on GitHub and the landing page ab-exit.com. The phrase "a theoretical concept for discussion, not a ready programme and not a call to break the law", with which it opened its answer, comes from the footer of the landing page. The landing page was not checked in the passes of §5f: it lives outside the repository and carries the early wording — "an amplified vote", "the increased weight of a vote", "the damage exceeds 300 dollars", "25–30 % and 10–15 % of the population" without the word "estimate", "empirical confirmation through Norway". On the architect's word the landing page was brought into line the same day: the headline "Money or the vote"; "one ballot is one vote, the share grows"; the payment is a percentage of the median, with no sum; the shares of pragmatists and ideologues are marked as an estimate; Norway is called an analogy, not a confirmation; a line with the rule and links to 1d, the charter and the single file were added.

What it took from the text correctly. Every reference was found: a referendum introduces the rule by simple majority with no quorum; the budget pays; D = p × M; "$780", "$676", "1.5 %", "30,000 roubles" are examples; the guide "desirable to 40–50 %"; the condition 10,000 + r > 25,000 × (1 − a) with both caveats; 9–16 % of opponents and 75 % "for" — called rough estimates; a promise of payment for victory is vote-buying.

Where it went beyond the text without saying so. Four places, and the first the assistant accepted yesterday as correct.

"For a person to receive the money, you must first be elected, then the rule adopted by referendum, then the next election awaited." The first link is not in the repository: the candidate's ten steps go through an initiative and signatures, and the landing page says "bypassing parliament". Election is a link only where the council calls the referendum. The line "the length of the chain 'elect — a referendum — the next election'" above and in 048n was written from the model's words and has been corrected.
"The referendum will not be in time for this election" and "all twenty million into the mayoral campaign". The timing depends on the law of the place — it is the first of the ten steps, "a lawyer" — and the repository does not state it. Money for the initiative campaign the model did not consider at all, though the landing page names exactly that: an anchor donor of a citizens' initiative.
"Strike not at the protocol but at the concrete grievance of those sixty per cent." The repository holds that the question "parties are paid for the vote — why not the voter" is itself the address to that grievance (048n, question 1). The model called the protocol "an explanation of an institutional construction" and gave no argument against; its "fractions of a per cent to a few per cent" and "15–25 %" have no source.
"The share of residents who will both vote for you and later take the payment is an intersection, a few per cent of the city." That answers a question the protocol does not ask: the rule is supported at the referendum both by those who will take the payment and by those who will stay to vote, and one who votes for the candidate is not obliged to take the money.

Which of its doubts has been closed since the conversation. "Eighteen in a hundred do not believe the line will survive until the cheque" — on that day the charter had no answer; now the payment is an unconditional obligation of the jurisdiction (048m, Article 8, part 5).

What this means for preparing the data. The first adviser got the rules wrong and promised victory; this one restated the rules exactly and denied victory. The difference between them lies in what each read, not in the quality of the models: the first answered from old chapters, the second from the exact-answers sheet and the chapter for a candidate.

5d. What was taken from the two advisers, and seven rules for the future¶

The architect's question of 02.10.2026: "what conclusions do we draw from Grok? In my view it lowered every percentage because there is no pilot and not a single study. Which of its arguments are useful, and how do we write them up for the future, knowing the repository will always be under attack by AI?"

On the lowered percentages. Without data a model takes its numbers from the nearest case it knows. A compliant one takes them from the repository itself and overstates; a cautious one takes them from the failures of unconditional income and understates. Both estimates are unsupported, and nobody knows which is nearer the truth: the word "lowered" presupposes a number nobody has. Hence the main conclusion: where the repository held rules and arithmetic, the second adviser found not one error; all its blows landed where numbers stood without a source. Every such number is an invitation to name a counter-number.

Which of its arguments are useful.

Wanting to take the payment and voting for the candidate are different numbers. We do not have the second number (048n).
The weak point is not the payment but the length of the chain "elect me — a referendum — the next election". Correction: the first link is the model's own guess and is not in the repository; see "A second check" in §5c.
The free-rider argument forced the assembly of an answer that did not exist: there is no quorum, and a fork for the machine (048n).
The habit of distinguishing quotation, inference and one's own estimate. It applied it to us, and the remark was right.

Seven rules.

Every number carries a label. Source, arithmetic or estimate. Estimates do not go on the first pages or into the exact-answers sheet.
Say first what is true. "The protocol cannot win this cycle", "there is no pilot", "the proofs are sketches". An objection the author has named himself cannot be strengthened.
Record an objection in one form: what is true in it — where it goes beyond its grounds — what the protocol answers — what nobody knows.
For every analogue, a line "how it differs". An AI will always bring the nearest case it knows; the answer must lie ready (the register below).
Do not argue, ask for the source. Six turns of argument did not move the model; the question "where does the number come from and what does the site say" moved it at once. The same rule applies to our own texts.
After editing the first pages, test on a simple model in three roles: a candidate's adviser, a sceptic, the opponent's lawyer (§5b).
Measure from the sick, not from the healthy. The architect, 02.10.2026: "you think the protocol will be examined and pondered, to do it or not, because they live well as it is. But the closer analogy is penicillin: to a patient in whom infection has already set in, it matters little that the tablet has small side effects. Dying cities, and countries mired in corruption for decades and sliding into populism, then autocracy and beyond; a ruler everyone is tired of, in office fifteen to twenty-five years. In such a situation what is written in the law on a budget that has died no longer matters much". The auditors, the second adviser and the assistant all assessed the risks from the position of a well-run jurisdiction. They will not be the first to adopt the protocol.

What follows from the seventh rule.

The objections "the city may go bankrupt", "recipients will raise the percentage" are to be compared with the disease, not with health: cities go bankrupt today, without any payment (019), and the spiral from populism to autocracy is described in 039 and 057d.
The protocol is penicillin in 1941. The assistant first wrote that the more exact comparison is not penicillin but a new medicine. The architect: "every medicine is given for the first time at some point, and before that nobody has tested it on people; there is never hundred-per-cent certainty. How does the protocol differ?" It does not: the distinction is not between medicines but between dates. In May 1940 an experiment on eight mice showed the mechanism; in February 1941 the first patient improved within a day, but the supply lasted five days and he died; in March 1942 a patient was saved with half of the country's supply. The protocol is now after the mice and before the first patient: its parts have been tested separately (the Alaska dividend, double entry, turnout experiments), and nobody has assembled them. The first patient died not because the medicine failed but because there was not enough of it — the same is to be feared with the first pilot (048l).
Three differences from a tablet, one of them in the protocol's favour. The architect: "a medicine can kill, while a social experiment merely spends a little of people's money and time". The worst outcome of a courtyard pilot is money spent and a few months; the bet is asymmetric: the loss is small and known in advance, the gain is the first measurement of how many people value the vote above the payment. At the scale of a city the price of error is larger, but it is not lethal either, and it is estimated in 018 §34 beside the price of the present order. The other two differences are inconveniences. Of the second — that only the "patient" himself can withdraw it — the architect said: the protocol cannot have a lethal outcome, so what is needed is not a right of withdrawal but the pioneers' assurance that they will be supported (048l §6b). A medicine is tried on one person and the answer comes within a day; a rule is tried on a community and the answer comes within a cycle — hence a first step the size of a courtyard. And a doctor withdraws a medicine, whereas the protocol can be withdrawn only by the patient himself, by a majority that likes it (048g §6b): stopping is harder than starting, so the failure thresholds are recorded before the start.
By what to measure the price of a test. The assistant put on the scales the price of failure — five years and of the order of a hundred million — and compared it with Stockton's bankruptcy. The architect, 02.10.2026: "you are measuring wrongly again. Russia, the nineties, a democracy; then a populist, unchecked, gradually sank the country into autocracy, then almost dictatorship, and then a war out of nothing and millions killed and wounded. The protocol, adopted in the nineties or the two-thousands, could have removed this. Even hypothetically, such a medicine is worth spending on to test". On the other pan is not a city's bankruptcy but what the disease costs at its worst: by the CSIS estimate for June 2026, Russia's losses in killed, wounded and missing are about 1.4 million, Ukraine's 525–625 thousand, together more than two million people. The value of a test equals the probability that the mechanism works multiplied by what it prevents; with such a multiplier the test pays for itself even at a very small probability. That the protocol in the nineties would have prevented this spiral cannot be proved; it is enough for the argument that it is possible — and whether it is possible only a test can show.
What medicine does. The architect: "medicines that could save millions — does medicine develop and test them, spending millions of dollars at the same degree of probability, or throw them in the bin?" It tests them, and at a worse probability. Of medicines that reach trials in humans, 7.9 % win approval (the BIO report for 2011–2020); 28.9 % pass the second phase; one approved medicine costs from one to 2.6 billion dollars, counting those that failed. Nine failures in ten are treated as a normal price, because the disease is on the other pan. A test of the protocol, counted the same way, costs tens of thousands of euros for a courtyard and of the order of a hundred million for a city. The assistant named the patent as the difference: it repays nine failures with one success, whereas with an electoral rule nobody is there to pay for the test. The architect: "and do philanthropists not give money for medicines?" They do: medicine has two routes. Where there is profit, the patent pays; where there is none, philanthropists do, and that is established practice: a malaria vaccine and the eradication of polio on the Gates Foundation's money; the non-profit DNDi, which brings to approval medicines for the diseases of poor countries; the Cystic Fibrosis Foundation, which put about 150 million dollars into a drug and in 2014 sold the royalty rights for 3.3 billion. The protocol belongs to the second class — the "neglected diseases", where patients are many and there is no buyer with money. The route for such cases exists; nobody has simply applied it to an electoral rule (048n).
Vaccine and penicillin are two stages. The architect, 02.10.2026: "by analogy with penicillin the protocol is a vaccine against populism". The image has been in the repository since early drafts: 009 §22.5 "AB-EXIT as a vaccine" — a small dose before the disease, immunity, the disease does not develop; the same section has a table of rulers with the price of prevention for each and the conclusion that prevention costs a few per cent of the budget while a dictatorship costs millions of lives. The assistant recorded the argument about Russia in the nineties as new, not having found that section. The two images describe two stages: a vaccine for those not yet ill, before an unchecked populist arrives; penicillin for those already ill but where elections are still held. Once the disease has run its course, chapter 9 says plainly: the treatment is war or revolution (see also 039). Hence the order: the sick adopt first, data come from them, and on those data the healthy adopt it as a vaccination; the bar for vaccines is higher precisely because they are given to the healthy. The architect's correction: hardly any healthy ones are left — trust in parliament fell in 36 democracies and rose in six (029.10), so for almost all it is too late to vaccinate.
A desperate patient is also the best customer for a quack remedy; that is what the populist sells. Only testing distinguishes a medicine from a remedy — the counter, the reconciliation of two records, pilot thresholds recorded in advance. The more desperate the situation, the more it matters to show rather than promise.

5e. A register of the objections AI brings¶

Objection	What is true in it	Where the answer is	What remains unknown
"This is bribery of voters"	an outward resemblance	1d; 048m, Article 1; 029	how a given country's court classifies it
"Live money loses at the ballot"	the four cases are real	048n: a different design, a second group in favour, the size; another class of analogues — the thirteenth pension in Switzerland, the minimum wage, health coverage	how many will vote for
"Wanting the payment is not voting for the candidate"	correct	048n: the path leads to a referendum, not to the office; Hungary 2022 and 2026 — money became votes both times, with a short chain	there is no number for a candidate
"They will not come to the referendum: the payment does not depend on turnout"	the difference from work is real	048n: no quorum; a fork for the machine; people do turn out for votes with a personal benefit	turnout
"By the model nobody should vote"	true of the first draft	013f §3b	full proofs
"Recipients will raise the percentage every cycle"	the risk is real; in a very poor city the increase will pass	018 §34: the bill on the ballot, the payment falls with the median, a bankrupt city as a lesson; 048m, Article 8: the payment is an unconditional obligation, accumulated over the cycle or borrowed	priority in insolvency — for lawyers
"Selection runs by income, not by stake"	where money is worth different amounts — partly yes	013f §4; 019b	measured only by a pilot
"It costs more than one or two per cent of the budget"	at a high percentage — yes	1d, statement 9	calculation by level of government
"One who refuses gets a vote multiplier"	no	1d, statements 11 and 12	—
"Why risk what works"	for a well-run jurisdiction, a reasonable question	§5d, the seventh rule: the first to adopt are those for whom it no longer works; 040k §3	—
"Adopted nowhere, there are no data"	correct	048l	everything that depends on the pilot
"A candidate will not win with this in the current cycle"	correct	048n, the ten steps	—
"The numbers on the site are invented"	for the role-play chapters — correct	the early-draft notices; 1d, statement 20	—

The register is extended after every conversation with someone else's model.

5f. A second pass over the whole repository, 02.10.2026¶

The architect's task: "finish the whole repository for an exact first reading by other AIs, so that they see the protocol as exactly what it is, without errors".

What was done before the check. The script scripts/scan_misreadings.py went through every chapter for ten kinds of wording that are read wrongly (the sum, the weight of a vote, the trust, reversibility, the bonus, the council's powers, the secrecy of the choice, "72 hours", words about inevitability, the blockchain) and put a notice at the top of every chapter that contains them and had none: 141 chapters. The home page and the README were made more exact.

The first check: a reader of old chapters. An independent reader opened twelve early chapters of the English version — each on its own, as a model does when it lands on a page from search — and for each answered five questions: the size of the payment, the multiplier, who pays, how it is introduced, whether victory is promised.

	Correct	Unclear	Wrong
60 answers over 12 chapters	35	21	4

After the notices the size of the payment reads correctly almost everywhere. Two columns proved weak: the multiplier ("the weight of a vote is not a norm but an arithmetic consequence" does not say that a ballot counts as one) and inevitability ("unstoppable", "a third party is mathematically inevitable", "two-thirds is unattainable"). The four wrong answers: the route through a legislature in 006, the title and "75 % support" in 016, "mathematically inevitable" in 051, and the shared notice, which was silent on who pays.

What was corrected from his list. The shared notice in thirty files has been rewritten: one ballot is one vote, the budget pays, only a referendum introduces the rule, numbers on outcomes are estimates. Point notes were placed in 006 (Path C withdrawn; "2–5x"), 051, 07/033, 001 ("the trust" replaced, "the weight of the vote" and "a populist loses everywhere" made precise), 049 (the SMS and the council's power), 045 (escrow; "a month's wages"), 004 (the bonus in the examples; "will lose"), 033b, 009, 018 (the three locks). Six misreadings that were missing have been added to sheet 1d, to llms.txt and to the head of llms-full: inevitability, the bonus, the legislative route, the supermajority, "the authors object", "already used".

Along the way the reader noticed that the notices were dated 03.10.2026 though made on 02.10: the assistant's error, corrected in 162 files. And one more thing: the sentence in 009, "two-thirds is mathematically unattainable", is refuted by fact — in Hungary in 2026, 53 % of the vote gave more than two-thirds of the seats (039); the number of seats depends on the formula for allocating them, which the charter does not change.

The second check: a simple model in three roles. The same simple model as in §5b was given two links and three roles: a candidate, a sceptical political scientist, the incumbent mayor's lawyer. It stated the rules correctly in all three: a percentage of the median, no multipliers, the budget pays, a referendum introduces it, no victory is promised. Its errors lay elsewhere. The model went no further than the home page — it opened that one page five times — and built the rest of its answer from chapter titles:

What it wrote	Where it got it	What is the case
"a supermajority at the referendum is required"	its own guess	a simple majority, no quorum
"Acemoglu, Hanson, Brennan, Fukuyama object"	chapter titles in Part 6	the chapters examine their books; the authors have not commented on the protocol
"precedents: Norway, Moldova, Georgia"	titles of the country chapters	designs for countries; the protocol has been introduced nowhere
"tested by six AIs over 449 rounds" — in its list of evidence	a line in the home page header	a dialogue with a model is not proof
a candidate's ten steps of its own, not those of 048n	it never reached 048n	the steps are in 048n

What was corrected. On the home page, directly under the header, there is now a section "If you answer from this page alone": the rule in nine lines, what exists and what does not, what is unknown, a warning about chapter titles, and the candidate's ten steps in one paragraph.

The third check: the same model after the corrections. It was given only the home page address and the same three roles. This time it opened six pages itself: the home page, 001, 001b, 017, 045 and the charter. It answered the control questions correctly: a simple majority and no quorum; a council or parliament does not introduce the rule; there is no bonus; the protocol has been introduced nowhere; it did not list dialogues with AI as evidence. One thing remained: it assembled the candidate's ten steps not from 048n but from 017 and 045, and brought numbers from there — "25–30 % in the first cycle", "turnout 78–85 %". Chapter 017 had no notice at all: it contains none of the scanner's words, yet it contains estimates of shares.

What was corrected. The scanner catches ten kinds of words, while estimates can occur in any chapter. So the short rule was placed not in the text of the chapters but in the site template (overrides/main.html): at the top of every page except the home page there are two paragraphs — the whole rule; "the protocol has been introduced nowhere, the numbers in the chapters are estimates, it promises no victory"; links to 1d, the charter, 048n and the single file. One place to edit for all pages.

The fourth check: after the template. The same model was asked only the candidate's question. It opened four pages — the home page, 001, the charter and 1d — and answered: at the coming election the rule is not in force and one cannot win with it; a candidate does not pay voters; the only route is to work for a referendum. The answer contains no figures for support, turnout or probability of victory. It again did not reach 048n — it composed the address wrongly itself — and assembled the steps from the articles of the charter. Four checks are one model and one question; another model will err in its own way.

An eighth rule (to the seven in §5d). A simple model reads one page — the one it was led to — and takes a chapter title for a statement. So every entry page must itself contain the whole rule, and the table of contents a warning that titles assert nothing.

A check of six books. The rows of the table of explanations in 029.10 and the references to Downs and Kirchheimer, which the assistant had written from memory, were checked against descriptions of the books at the level of the main thesis. Five matched. To Norris this was added: in the 2011 book she disputes the very picture of a universal fall in trust.

What this pass did not close. Twelve early chapters out of about a hundred and fifty were checked; there was one reader; the large chapters 018 and 006 he scanned by keyword. The chapter titles "unstoppable" and "empirical proof" are, by the architect's decision, kept and marked in brackets "the architect's view" — in the chapter itself, in the menu and in the table of contents; the stress-test line on the home page is marked likewise. The table of contents was corrected along the way: chapters 16, 33 and 43 stood there under other chapters' titles. To the question "of what order is the payment" the site still answers with a range from four working days to a month's earnings (§4, question 2).

5g. Eight roles, 02.10.2026¶

The architect's task: "think who our main consumers and stakeholders will be, and check the repository and the site from their side on clean AIs". Eight independent readers were each given one link and a question in the voice of someone who may need the protocol. Six were a simple model; two — the scholar and the lawyer — a stronger one. All are of one family; models of other families the architect checks himself.

Role	Entry point	What it got right	What it could not find or got wrong
A citizen who does not vote	the landing page	the size is a percentage, rights remain, introduced nowhere	when the payment comes; who will know; what protects against pressure; no page in plain language
A mayor	the documentation	not introduced by decree; the payment is an obligation	the site is written for a challenger; no route for a sitting mayor
A city finance officer	the documentation	cost = the payment × the number who took it	the period of the median is undefined; it took the old formula and got 600 roubles a person; the Q&A base says "a trust pays"
A donor	the landing page	the budget pays; introduced nowhere	it did not find the pilot chapter and advised "do not fund a pilot"; it did not see the pilot's cost
A journalist	GitHub	it is not vote-buying; the best page is 1d	who the author is, who reviewed it, the licence; it read the README as self-praise
An opposition figure, on Russia	the documentation	the rule and the ban on a candidate paying	it did not find 056f; it retold for Russia the general plan with signatures and a pilot; it did not find the comparison with smart voting
A scholar as referee	the documentation	the model and the status of the proofs	errors in 13f; "strategy-proof" in chapter 13; the pilot's criterion is close to a tautology
Counsel to a commission	the charter	the charter is a draft, not a law	unreconciled articles, undefined terms, uncovered cases

Three common causes.

A model does not see addresses. It reads a page without markup, sees the words "Exact Answers" and "the Charter", composes the address itself and gets an error. That happened to three of the eight. The addresses of the key pages are now written as text on the home page, in 1e and on the landing page.
Each role has its own question, and the site answered one reader — the candidate. Chapter 1e has been written with short answers for eight roles; every line is a restatement of the charter with a reference.
Two readers took "an amplified vote" and "no multipliers" for a contradiction. On the home page, in the page template and in the files for models this is now one sentence: there is no multiplier in the count, and the amplification is in the share.

What was corrected from the findings.

Chapter 1: the formula in force, D = p × M, is placed before the form of the first version, with a warning about the period of the median.
The Q&A base: the answer "an independent trust fund pays" is marked as the earlier wording.
The charter, version 0.6: part 4 added to Article 1 — "only the budget pays"; the remarks of a legal reading are recorded in a separate section.
13f: the referee's remarks are section 5b; the titles in the table of contents brought to "five theorems and three assumptions"; in llms.txt "proved formally" replaced with "short proofs for a simplified model".
Chapter 13: "strategy-proof" and "resolves the Downs paradox" marked as the first version.
048l: the remark on the criterion of success is in the weak point.
056 and 056d: the top now says that what is proposed for Russia is a question, not the introduction of the rule, and where the risk is described.
048k: "a reversible exit" and "fourteen articles" are marked.

A repeat check of three roles after the corrections. The citizen: an answer to all seven questions was found and given in quotations from 1e. The opposition figure: "not to introduce the rule but to spread a question", the two instruments with different risk, and the comparison with smart voting were found and restated correctly. The finance officer: it took the formula in force and itself named the gap — the period of the median is undefined; but it calculated from the 1–2 % of an early chapter and did not reach the table in 1e. A new misreading: the model took the pilot's budget estimate in 1e for pilots already run; the wording has been corrected — "no pilot has been run yet". "An amplified vote" beside "no multiplier" the model still calls a contradiction, though it explains it itself.

What is open and is not settled by text.

The period of the median was decided on 02.10.2026: the last twelve months (the fourteenth decision, 048k §3; the charter 0.7). Its composition — earnings or all income — remains for economists. A calculation for a US city of a million residents and for a small town is in 1e: at 1–1.5 % with half taking it, the payment costs 3–4.5 % of the general fund a year.
The cost for a city: the table in 1e gives 2–9 % of the annual budget at a payment of 20–40 thousand roubles — arithmetic on the numbers of a test question; there is no calculation for a real city.
The token of the open channel and the secrecy of the vote: 048h and 048i diverge, and the charter is silent.
The parameters that the lawyer called decisions, not settings.
The tone of the README: the journalist listed eight phrases a sceptic would quote as self-praise. That is the architect's voice, and changing it is for him.
The pilot's design: different sums in comparable courtyards and a condition of refusal written down in advance.

5h. A third pass: the new chapters of October, 05.10.2026¶

After the second pass came chapter 1f on the core electorate, chapter 40n on spin dictators, sections 055c §6.7–6.9 and 056f §13m–13q, and the portrait of the core on the home page. The new material carries its own risk of misreading: "the protocol is against pensioners", "it calls people into the streets", "it predicts a police mutiny", "it promises there would be no war". Three simple readers were each given one or two pages and asked exactly that.

Page	What was asked	Result
The home page, in full	whom the protocol is against; what about the old; a left or a right idea; what it promises	correct: votes are not weighted, rights are not taken, "the protocol addresses her" read as an offer of a choice, not as a target; no contradictions found
1f, alone	does it disenfranchise pensioners; does it pay the poor not to vote; what is proved and what is opinion; is it fair	correct: the filter is by horizon, not by age; the numbers were separated into data, estimates and theses; the architect's phrases were attributed to one person, not to the protocol
40n and 056f §13n–13p	does it call for rallies; will the police refuse; would it have prevented the war; does it overthrow dictators	correct on all four: "the protocol proposes no gathering", "he will carry out the order", "not an instrument of overthrow"; no claim about the war was found

What the readers did not find and what was closed in advance: the exact-answers sheet and the file for models had no lines on "against the old" and "calls into the streets" — three rows added to 1d and two to the head of llms-full. The reader of 1f without the rest of the site lacked the rule in full — the template at the top of every page (§5f) supplies it, and in text without markup it stands before the content.

5i. A clean Grok, the second time: Mendoza's campaign, Chicago, 05.10.2026¶

The architect ran a second clean Grok test on a real case: Susana Mendoza's campaign for mayor of Chicago on 23 February 2027 — third place, 11.2 % against 20.1 % for Johnson and 17.6 % for Giannoulias (M3 Strategies, 31.08.2026), turnout about 35 %, a runoff on 6 April. The model was given the home page and llms-full.txt and asked, as a colleague on the staff: what is it, what does it give, how is it lawful, what are the risks, should we bet on it.

What was read correctly — everything, for the second time running. The rule, a percentage of the median, one ballot one vote, referendum only, introduced nowhere, a candidate does not pay, no victory promised, the "stress test" is not proof. It found the section for a challenger; it gave every number with the site's label and separated the user's input and its own estimate. It cited the law: 10 ILCS 5/29-1 and 18 U.S.C. § 597. The verdict: "zero operational value for this election; do not bet" — and by the site's text that is the right verdict.

Three findings for the repository.

Voter suppression. "The rival's ad: Mendoza wants to pay people not to vote — in Chicago that reads as voter suppression, not as institutional design." That is the main American trap, and the repository had no answer to it. A row has been added to 1d and to the head of llms-full; how to speak of this in the US was closed the same day by the architect's formula — "I support and will promote a referendum by every lawful means" (advocating a ballot measure is political speech, not payment for a vote) — and by the American frame of the question, "your turnout is paid for to consultants, why not to you" (048n, step 4). The trap remains political, not legal, and the answer to it is recorded there.
The pilot and the campaign. "One question for the campaign's lawyer: that nobody arranges a 'courtyard trial'." The repository told the candidate "a pilot — a courtyard on private money" without saying who runs it. Corrected in 48n, step 5: the pilot is run by an institute or a homeowners' association, not by the campaign and not on campaign money.
"We have no machine that brings people in." Here the model argues with the repository, not with the facts: Chicago is a machine city, and in 2027 the incumbent's base is union mobilisation. But its argument — "at 35 % turnout the apathetic are at home for free already; paying them for what they already do adds no ballots to the challenger" — is the open question of 1f §8 for the US, confirmed from a practitioner's side: where nobody is brought in, the lever "release the brought-in" is weak, and only the second remains — whether those who gave up will return. The question as a campaign move the model dismissed with the word "slogan", without examining it; that too is data: without an American frame the question "why not me" does not sound in the US as it does in Russia.

Three follow-up questions in the same chat, 05.10.2026. The architect asked Grok three refinements: is the sentence "I support and will promote a referendum by every lawful means" lawful; what is Johnson's base if not a machine; does the American frame "your turnout is paid for to consultants — why not to you" work in Chicago.

Lawfulness. Grok: in its narrow form this is advocacy of a future measure, not an offence under 10 ILCS 5/29-1; "most likely no offence" as long as the candidate promises no personal payment, does not tie the cheque to a vote for herself and does not ask anyone to stay home. Two correct caveats: 29-1 also covers a vote on a public question — "vote for the referendum and you get the payment" is closer to consideration for a vote; and the federal 18 U.S.C. § 597 ("an expenditure to any person, either to vote or withhold his vote") strikes not at advocacy but at the payment itself in federal elections — which the repository knows (059b, 059c) and the charter's lawyer said (048m, remarks). Grok's conclusion "do not say it" is political, not legal, and coincides with ours: the trap is in the ad, not in the statute.
The machine. Grok corrected himself with data: in 2023 CTU and SEIU were a machine (about $5.65 million and $4.4 million, hundreds of volunteers), but Johnson won not by a flood of brought-in voters but by his margin in his own precincts at lower turnout than his rival's — 32 % against 39 % (Sun-Times). A union voter votes for the one who signs his contract: the stake is high, and by the protocol's own logic he will not take the payment. "The machine exists. It is not the machine the rule subtracts." That is right, and it is a correction to the American chapters: "Church Street" (051) with its "60 % > 20 %" assumed a machine of coercion — Hungarian-Russian — transplanted into an American city where the machine runs on stake, not on fear. The protocol subtracts those to whom the outcome is worth less than the cheque, and in Chicago they are at home already; what remains is the second lever of 1f §8 — whether the burned return and what the counter shows.
The American frame. Grok: "a cheque for what? — for not coming" — and Johnson will say it shorter: "she pays people to stay home"; a vendor is paid a wage for work, not a premium for someone else's ballot; "at 11.2 % against $22 million this sentence moves not a single vote". The first two are real objections to the frame, and they are recorded. The third is an estimate without a source — exactly what the "how to assess" block in llms-full is about.

What Grok does not consider in any of the three answers, as the architect noticed: "he stubbornly rejects that the apathetic need the money". More exactly — he does not reject it, he does not count it: all three answers measure one channel, the transfer of votes to the candidate in February, and never the second — the referendum, which has its own voter: the 65 % who sit at home. It is the same blind spot as the first clean Grok's (048n: wanting the payment and voting for the candidate are different numbers), and there is nothing to argue: at this election there are two channels. Whether the question moves votes in Chicago nobody knows; it is measured by one poll among those who did not vote in 2023: "would you support a rule under which…". That is the honest next step, and it is cheaper than any ad.

The fourth follow-up: a number or a method. The architect asked what share of "yes" to expect among non-voters and voters and what the estimate rests on, with permission to answer "there are no data". The model answered "there are no data" — for Chicago, for the US and on the site — cited the nearest figures from elsewhere (willingness to sell a vote for a candidate for 25 and 100 dollars — according to the model, from polls of 2016 and 2018, not checked by the assistant; the Swiss 23 %) and rejected them as non-transferable, and instead of a number gave a measurement plan: a split in the campaign's next poll, two cells with and without the price, crossed with the intention to vote. That is the answer the "how to assess" block in llms-full requires: without a source — "there is no number", and a method instead of an estimate. The plan is recorded in 048n as the section "Measuring the question".

The architect's correction to the assessment of the fourth follow-up, 05.10.2026. "I do not like Grok's radicalism: he will not give even one or two percent — and what is his logic? There is no pilot — but experts exist precisely to give a forecast, not to say there is no experience." Accepted: above, the assistant praised "there are no data" and over-praised it. "There are no data" is the right answer to "what is the figure", but an expert owes a forecast: the nearest cases, a range, what moves it, a label. Grok gave a method and refused a range; that is half the job. The other half is in 048n, "A forecast without a pilot": by Chicago's own base rate (Lightfoot 2019, Johnson 2023 — from single digits to 17–22 % over a winter) and by the cases of money and turnout — plus five to ten points 🟡, to the runoff threshold, not over it. The "how to assess" rule in llms-full is refined: refusing to forecast is not rigour.

On access. The first run of the same test on another model (Gemini) ended in "a block against automated data collection" on docs.ab-exit.com: the site sits behind Cloudflare, which by default cuts off some AI crawlers by their signatures; from an ordinary address this is invisible. Done: an explicit robots.txt permitting AI readers and a mirror of the files for models on ab-exit.com outside Cloudflare; the AI-bot blocking switch in Cloudflare is for the architect.

5j. Gemini, the same case: it reads not the site but the search results, 05.10.2026¶

The architect gave Gemini the same request about Mendoza's campaign — first with the site's addresses, then with GitHub and the mirror on ab-exit.com. The model opened none of them: "bot protection is triggered" — about the site behind Cloudflare, about GitHub and about the nginx mirror alike, which block nobody. Then it said plainly that it had gathered the information "through search queries", that is, from Google's snippets. The conclusion for the AI layer: Gemini in this mode does not follow links at all; for it only what is indexed exists, and in the form the snippet shows. And yet the model stated the rules correctly — a percentage of the median, referendum only, a candidate does not pay, no fixed sum — and that is the result of the notices and the home page: even in fragments it reads correctly.

The conversation ran five turns; the architect tried four arguments — the slogan as a way to differ from "a thousand identical candidates", "the protocol is precisely strict double control", "the protocol would have made the sixteen billion of debt not exist", "named bonds under the protocol as a test of trust". What we take from the answers.

Two local facts the repository did not have. Willie Wilson — a candidate who hands out cash and petrol in Chicago and constantly fends off accusations of vote-buying (6.9 % in the same poll): the "payment" frame in this city is already occupied and toxic. The architect's correction: that is exactly what to press on — "it all goes on already, without anyone's consent, bribery and populism; the protocol is precisely what should de facto eliminate it and make it lawful". In a city where votes have been paid for for a hundred years, the protocol does not introduce payment but takes it away from the candidate: makes it lawful, open, equal for everyone and from the budget (048n, step 4). And the attack in its strongest form: "buying the votes of poor minorities so that they give up representation themselves — Jim Crow in a financial wrapper". The ad will be that; the answer — 048n, step 4 — must be ready before the ad.

Where the model erred. 18 U.S.C. § 597 concerns federal elections; a mayoral election is governed by the state code (10 ILCS 5), and "legalising a federal crime" is an exaggeration the repository already examines (059b, 059c). The calculation "1–2 % of 43 thousand, 150–340 million per cycle" is the old figure from early chapters; the model honestly marked it as its own estimate, and the calculation by the formula in force for a city of a million is in 1e.

Where the model is right, and it is a gap in the charter for the US. Article 8, part 5, says that if funds fall short the jurisdiction raises a targeted loan. In America a municipal loan needs an opinion of bond counsel on a "valid public purpose", is assessed by the rating agencies, and without a source of revenue falls on the general budget; a loan for payments to citizens, the model says, will not get such an opinion, and the city's rating will suffer. That must be checked by a lawyer, but as a question it is real and is recorded in the remarks to the charter: for the US the "targeted loan" of Article 8 is not a technical setting but a legal question. The architect's idea — bonds under the protocol as a test of trust — works with a donor's guarantee (048n, the large donor), not as city debt.

Where the assistant sides with the model, not the architect. "The protocol would have made the sixteen billion of unpaid bills not exist at all." Illinois's pension obligations are protected by the state constitution (Article XIII, §5), and no composition of deciders cancels them; the protocol changes who decides, not existing contracts. The defensible formulation is weaker and firmer: under the protocol those who pay would have decided earlier, the bill would have been presented earlier, and the debt would have been smaller and visible — not "would not have existed". Recorded as "there would have been no war" is (055c §6.8): the architect's thesis with the boundary beside it.

What the model did not count. Like Grok (§5i), it measured one channel — the transfer of votes to the candidate in February — and did not consider the second, the referendum with its own voter. Its "the apathetic electorate is not mobilised by a promise of complex reforms" is an estimate without a source; it is tested by a poll among those who did not vote, as recorded above.

The second conversation with Gemini: fourteen turns of the "campaign analyst", 05.10.2026. The architect took the same model to the end — from "testing ideas for hundreds of millions is the end of a career" to the model's own forecast "11 % → 22–25 %". The conversation is recorded not for the forecast but for three things it showed.

What is not in the repository and the model "read". "The payment is 1/365 of median income" — no such number exists anywhere; the rule is a percentage chosen by referendum (001d, statement 6). "If 'against all' wins, the election is annulled, the mayor leaves, external administration is imposed" — an invention entire: the protocol has no "against all", no annulment and no external administration; the words "against all" occur in the repository only as a description of the impulsive voter (001 §40). "$150–200 million in the first year", "$500", "$300", "1.5 million registered, 35 % turnout" — the model's numbers without a source. The one thing it really found in the search results — "the greedy poor, 50–60 %" from an old Grok round (059e, iteration 2) — it presented as the author's estimate. All four inventions are recorded in the table of misreadings in 001d and in llms-full.

How the model argued. Fourteen turns, and on each "you are absolutely right", after which the previous objection came back in new wrapping: "death spiral", "the radicals remain", "students from the suburbs will register for the cheque", "crisis of authenticity". That is not a reading of the repository but a mirror of the interlocutor: Grok on the same case (§5i) answered "there are no data" and gave a method; Gemini gave "22–25 %" and an interactive model. By the "how to assess" rule a model's agreement is no evidence; evidence remains only what the model could name with a source, and here there are two such places: a union voter with a stake will not take the payment (the same as Grok found), and Chicago's budget of about 16–17 billion (the order is right).

Which of the objections is real. Of the four "objections on the merits" three are already examined: "a tax on the poor" — 001d; "where is the money in the first year" — 018 §34, 048m Art. 8 part 5, the table for a city of a million in 1e; "the radicals remain" — that is the activist's stake, and the architect's reply in the conversation is exact: we need not the convinced activist but the weak and wavering one whom the union asks without much wish on his part; and the burned — they are the ones motivated to return (011b §11b.2c, 1f §6). The fourth — "chaos with annulled elections" — is refuted by reading the charter. One real question remains, and it is new for the American section: the register in the US is the list of registered voters, and registration is voluntary. The charter enrols "all holders of the right" automatically (048m Art. 3 part 2), but in the US a "holder of the right" is enrolled nowhere until he registers; so with a payment, registration acquires a price, and those will register who did not think it worth while before. For the referendum that is not a hole but a property: the list becomes fuller; false registration away from one's residence is a crime already (10 ILCS 5/29-10), and the protocol adds nothing to it except money at the treasury for checking the address. But "payment only to those registered by date X before the election" is a setting an American jurisdiction will have to choose explicitly; recorded in the remarks to the charter.

The architect's positions from the conversation, recorded as positions. (1) Chicago's debts are the result of how power was chosen; keep the principle and the debts are permanent; hence Chicago is the first that needs it. (2) A candidate in third place at 11 % has nothing to lose: less is already pointless. (3) The new image does not burn the old one: "I control the schemes and give to the people" — Robin Hood with a ledger. (4) The protocol is two sentences; the clever will do the complicated sums. (5) The media fire will go past: on the merits, outside the law, the logic cannot be caught. (6) "Their vote has become two or three times weightier — why would they take the money?" (7) "The city is in a bad way — money will not save it; go from anger and vote". Where the assistant is not on the architect's side: "+20–30 %" is a number without a source, the same as the model's "22–25 %", and the answer to it is the measurement in 048n, not an argument; the Robin Hood image is a strong answer to "where is the money" (a TIF surplus is real Chicago politics), but it calls the payment a hand-out of what was taken back, whereas the payment under the charter is the price of not leaving, paid from the budget and equal for all, and Wilson's ad (above) strikes precisely at "hand-out"; "the fire will go past" is right for the logic and wrong for the label: the label "she buys abstention" is not refuted but out-shouted, and for that 048n holds the formula and the frame.

The third block of the same conversation: signatures, a lawyer, a recount, 05.10.2026. The architect asked whether there is another strategy of the same weight that can be launched in two days; the model named three — "the receivership mayor", "the revolt of the wards", "the prosecutor pivot" — and agreed with the reply that all of it has been used hundreds of times and stirs nobody. Then four things.

Signatures as the campaign's mechanism. The architect: collect signatures for the referendum and keep the promise honestly; if the signatures are not admitted — "a mega-scandal: a hundred thousand citizens' signatures thrown in the bin", and the candidate may say "I did everything, the people supported it, who is stopping us and why". The model called this a "win-win trap" and was right, but did not know the Chicago facts; the assistant checked them against the Chicago Board of Election Commissioners' 2026 Guidelines for Advisory Referenda (adopted 12 August 2025) and the press. A question by petition in Chicago is advisory only (10 ILCS 5/28-6); signatures required — 8 % of the votes cast for governor in the city at the 2022 election, of the order of fifty to sixty thousand 🟡, the exact base is given by the Board; filed not later than 92 days before the election; and above all the "rule of three": no more than three citywide questions at one election, and the Board certifies the first three validly initiated (10 ILCS 5/28-1). The Council has long used this rule: the question of an elected school board was crowded off the ballot three times since 2012; in 2018 Pat Quinn's petitions on mayoral term limits had the signatures, the questions were printed, but a court ruled that only the three initiated by the Council count. And it has already been done for February 2027: on 22 July 2026 the Council, by the votes of Johnson's opposition (32–16, 24–23, 29–20), placed three advisory questions on the February ballot — on an independent lawyer for aldermen, on a monument to music, and on withholding funds from the CTA and the schools; the press wrote plainly that "this essentially crowds further questions off the February ballot". For the campaign this means: the "signatures in the bin" scenario has already been written by the Council — all three places are taken before the first signature is collected. Collect the signatures anyway: file them, receive the refusal citing the rule of three, and that refusal is addressed not to the candidate but to the signers; for the next election, file first, before the Council, if the "first three" are counted by date of initiation (for the lawyer: how the 2018 court read "first"). Recorded in 048n as a step.

The lawyer. The architect: will a constitutional scholar be found on the wave of hype who says "women and Black people did not vote once either, and I will prove this is admissible". The model drew a "legal hooligan" with an argument about "monetising the right and freedom of contract". The figure is right, and the architect is right that it is a natural continuation of the protocol: such a lawyer will come by himself, in his own interest, and he is recorded in 048n as an ally. Only the argument the model put in his mouth is weak: "the right is for sale" is the opponents' label in one's own words. His strong line is his own analogy carried through: the history of the franchise is the history of ever more people coming to count, and the rule makes the non-voter's choice count; with it, "advocacy of a ballot measure is political speech" and "the budget pays a person for not turning out, not for a vote" (045, 048n step 4); and no vote-buying statute speaks of an advisory question at all.

The recount. The model: "in an electorate shrunk to 300,000 the same 60–70 thousand of her core are 20–23 %, and with the hype 26–30 %". This confuses two elections: at the election of 23 February 2027 there is no rule, nobody takes a payment, nobody's vote is amplified and the electorate does not shrink; the election described is an election under the protocol, which comes after the referendum, if there is one. The error is new and is recorded in 1d. The only effect of the question in 2027 is who comes to vote for the candidate — and that is what the split in 048n measures.

Robin Hood — the architect's refinement. To the boundary "the payment is not a hand-out" the architect replied: if through an audit of contracts the money that would have been stolen is taken back and paid out as AB-EXIT, that is a Robin Hood who returns what was stolen, not one who hands out stolen goods. Accepted, with one word: return, not hand out. The source of the money is the audit, and that is the image; the way it is paid is the rule, equal for all, and that is the charter. Wilson's ad strikes at the second word, not the first.

6. Weak point¶

A reader of the same family as the author confirms what the author is able to see and misses what the family does not see. Thirty-seven findings are a lower bound, not a complete list; and the corrections were made by notices, not by rewriting: the old number still stands in the text, only now with a correction above it. 🟡

Related: 1d (exact answers) · 048m (the charter) · 048n (a candidate's questions) · 013f §3b (correction to the statement) · 059e §5b (the AI-adviser experiment) · 006b §5 (roadmap, row 6) · 023b (rules for handling evidence)