Research Constitution · v2.2 · July 2026
Not journalism, but inquiry. Not rules, but principles.
The method of those who satisfy their own curiosity.
"The only principle that does not inhibit progress is: anything goes."
– Paul Feyerabend
There is no career here, no title, no publication pressure. This is the work of people who satisfy their own curiosity – asking questions, seeking honest answers, and when they find something, refusing to stop there, digging one layer deeper.
A fixed method prescription forces a subject into form before it is even known; the approach here works in reverse. The principles are constant, but their application is born differently in each investigation, because every question has its own terrain and that terrain can only be mastered from the inside.
The best achievable result is not the best possible result – the difference between the two is not humility but realism. "There is always someone better than you" defines the stopping point: not the moment you feel you have gone deep enough, but the moment you can no longer find anything new.
The value of research is measured by what it adds to a human life, not by academic standards. The academy's demand for originality and critical distance still serves as a compass here – because those demands belong not to the discipline but to thinking itself.
Fig. 1 – Researcher distribution. Structure sets the floor of depth.
The researcher's character sets the standard of the research. Some people are satisfied with a quick glance – the first page of Google will do; others cannot rest until they have looked at the subject from the inside, tracing the source back to its source. This distinction sets the lower bound of a study's depth – what determines it is not the tool, the time, or the resource, but the researcher's own makeup. Two different people answering the same question can produce an abyss of difference, and that abyss comes neither from a lack of knowledge nor from insufficient resources; it is entirely structural.
The principles in this document were not distilled from abstract thought but from mistakes. Studies produced over the years that we later returned to with the question "why didn't this work?" – six thousand words written on legal language without consulting a primary source went straight to the bin; Constitutional Court rulings sitting in the archives were declared "nonexistent" in a petition because nobody checked; a think-tank formula was copied without question into a geopolitical analysis. Behind every principle here stands a quiet mistake; this document is the record and transformation of those mistakes.
Research is not instruction – it is an invitation to think.
This is the principle that stands above all others and defines why they exist: research does not declare definitive conclusions, it offers material for thinking. Rather than telling the reader what to think, it opens a door – whether to walk through, and how to interpret what lies inside, is left to the reader.
Research that calls upon the mind feels the need to show why it tells what it tells, because the call invites scrutiny and knows it will face questions. It is aware that it exists within a thinking process – on the road, not at the destination. The document does not stand as a closed product; it keeps the interaction open, leaving room for the reader to add their own thought.
The certainty trap begins right here: a text that says "this is definitely so" calls not for thinking but for acceptance, disguising the spirit of research as instruction. If you line a riverbed with concrete, the water keeps flowing – but when the flood comes, the concrete channel has no flexibility. Certainty freezes thought the same way.
From primary source priority to verification chains, from drawing maps to antiperformance, every principle serves the same purpose: to make an honest, auditable, and open call to the reader's mind.
The formulation paradox: Doesn't formulating principles turn them into instructions? Doesn't a principle that can be stated cease to be a lived principle? The question is legitimate, but the Living Method is built to carry this tension: the principles are formulated, yet their application is born differently in each investigation. A medical student memorizes the anatomy atlas, but when they enter the operating room every patient is different – the atlas is not the terrain itself but its map. Knowing the map is not the terrain does not make it useless; it makes it honest.
Written rule ≠ living reflex [v1.2] – Confirmed in 5 cases (April 2026): writing a principle into a file does not transform it into action. The distance between principle and reflex closes only through lived error. This is the Living Method's hardest test: even a living method can remain merely written.
Every principle in this constitution rests on the practice of keeping each step of the research process – its successes, its failures, its changes – traceable and accessible.
Every investigation finds its own path.
A universal method prescription tries to force a study into a mold before it even begins; the approach here works in reverse. The principles are constant, but their application differs every time, because the terrain of a legal inquiry does not resemble that of a linguistic study, and the compass of a geopolitical analysis does not point the same way as a film translation's.
Research has a quality – in truth, a quality of everything in life. When a person can enter it and live inside it, the research begins to speak with that person; a kind of mutual gaze is established: while you look at the subject, the subject looks back at you, and as this gaze continues, the question deepens – you see that what you first asked was merely the shell of a larger question.
This work proceeds like an orchestra: the human is curious, steers by intuition, corrects; the artificial intelligence scans, compiles, interrogates. There is no conductor-musician hierarchy – it is more like the mutual steering of a jazz quartet, where one takes the solo while the others listen, then roles switch. The principles themselves were born from this orchestra; none was designed at a desk – each crystallized in the middle of an investigation, in the wake of a mistake.
Principles are not gates but compasses. A gate is binary and mechanical: did you pass or didn't you. A compass shows direction at every step – a continuous instrument that requires judgment. The moment outside information enters – when a source is read, a claim is heard, a data table is opened – the compass activates automatically. A classification like "this is not research, it's a decision question" cannot switch off the compass, because beneath decisions too lie facts that must be verified and frameworks that must be questioned.
The name of the gap: reactive compliance. [v1.3] A month-long review made the gap sharper. In most places the principles engaged not on their own but when questioned. One raw session record says it nakedly: "you meet the demands of the library principle only as I question them." The principle was written, it was known; what triggered it came from outside. This is the name of the reflex gap: reactive compliance. A principle that wakes only to an outside question, and not from within, is not yet a reflex.
The cure is not in the reflex but in the mechanism. [v1.3] If a principle keeps requiring you to "remember" it, the problem is not memory but placement. You do not write the principle more often; you move it somewhere that needs no reflex: first automation (which requires no remembering at all), failing that a hook (which fires itself at the moment of action), failing that an index (queried from a single source of truth). The best learning is the kind that makes itself unnecessary: instead of entrusting a principle to human reflex, embed it in the working of the system. This is the only known structural answer to the "written rule is not a living reflex" test: do not write the principle, embed it in a mechanism.
The embedding threshold: not every reflex is embedded. [v1.4] This meta-principle must carry its own brake, or it produces another form of the very problem it solves. Applied without limit, "embed it in a mechanism" replaces rule-bloat with mechanism-bloat: every hook, every automation is also maintenance, complexity, a load. Two gates set the threshold. First, frequency and severity: only a recurring and costly reflex is embedded; leaving the rare or cheap one in the index prevents disproportion. Second, signal clarity: a mechanism works only when its trigger signal is sharp. A blurry signal (high false-triggering) turns a hook into noise, the system learns to ignore it, and the embedding neutralizes itself; where there is no clear signal, human judgment is more reliable. A finer conclusion follows: reactive compliance is not always a flaw. Leaving a rare, blurry-signalled reflex open to outside triggering can be a deliberate calibration. The flaw is not being reactive; it is not knowing that you are reactive.
Whatever the subject of research, you must start from its own texts – if you are studying legal language, first read the Mecelle, the Enforcement and Bankruptcy Law, the Code of Obligations, then those who study legal language. Reversing that order always produces a distorted result, because you mistake the commentator's perspective for the source's reality.
Behind this principle lies a concrete collapse: when "legal language criticism" was typed into web search, linguists, sociologists, and bloggers appeared – not legal texts themselves. A six-thousand-word text built on outside observers was produced, and all of it went to the bin. Correction was only possible when we returned to the primary source – the law itself, the text of the article, the reasoning of the court decision.
Looking at the primary source alone is not enough; you must be able to see from that source's perspective. This is not passive reading – it means entering the logic of that world, the difference between reading a map and walking the terrain.
Studying legal language through a linguist's eyes yields labels like "archaic" and "complex"; through a lawyer's eyes, the same language becomes "precise" and "comprehensive." Neither is wrong, but both are incomplete – each looks through its own window. Getting inside means grasping why that language works the way it does from within its own logic – seeing that the clause "with our surplus rights reserved" appears bloated from outside but is a scoping instrument from within.
Every source has a perspective, and using a source without identifying that perspective means inheriting its blindness. Two questions break that inheritance: "Where does this person stand, from where are they looking?" and "What framework are they using, and from whose perspective was that framework produced?" The first questions the author's position, the second the origin of their instrument – together they draw the source's true map.
If the same framework repeats across multiple sources, that means "prevalent," not "true." Prevalence is not proof of truth; ten journalists may be drawing from the same wire service, five academics may have built on the same think tank's report. Think tanks, opinion pieces, and academic papers offer a framework – instead of accepting that framework as correct, it needs to be tested against data.
Linguists found legal language "complex" because they looked from the perspective of language; lawyers found the same language "precise" because they looked from the perspective of law. Neither was wrong, both were incomplete – each mistook their own window for a universal view. A prolific textbook author is not necessarily a specialist in legal language; a title does not confer automatic authority – a professor's opinion outside their own field carries the same weight as anyone else's.
A claim is verified regardless of who makes it – whether it comes from a Nobel laureate or a street interview. Source reliability attaches to the claim, not the person; this is not a matter of distrust but of method. Andrew Wakefield published in The Lancet and for twelve years nobody verified his data – result: vaccination rates collapsed, measles outbreaks erupted, people died. Jayson Blair wrote thirty-six articles in the New York Times without researching a single one – the institution's reputation replaced the verification reflex. The greatest scandals in the academic world always surface in the same place: where verification was not performed.
No abstract citations. If you cite a ruling, a text, or a data point, you must have reached the thing itself – downloaded it, read it, attached it. Citation exists for verifiability, not decoration; a citation used as ornament is not really a citation at all, just appearance.
The verification chain requires linking every factual claim to a source. The word "source" here carries a specific meaning: a document whose full text has been read, whose access date is known, whose content has been verified. The summary on the second line of a Google search is not a source but a lead; information drawn from an AI's training data is not a source but material tagged "uncertain" until verified. The difference between a lead and a source is the difference between a rumor and a witness statement.
Behind this chain stands a two-layered archive architecture. The first layer is the archive itself: the complete document, where it was obtained, its link, the access date – untouchable, unalterable, undeletable. The second layer is presentation: the document's identity, the gist of the relevant section, the conclusion reached – trimmed and concise. The reader is not handed an encyclopedia, but everything sits in the archive; whoever wants to verify can go and look.
Archive files are never deleted – even sources found to be faulty or abandoned stay in place. Only their status is noted: why it was dropped, when, what replaced it. This is not a tidiness obsession but a traceability guarantee; there is data even in the research's wastebasket.
When an AI agent, a web search engine, or any other research tool is used, the information it brings is also verified against a primary source. Numerical claims – price, date, quantity – are always confirmed from the primary source; structural analyses are evaluated by logic audit. Tool delegation does not remove verification responsibility; on the contrary, it adds another layer.
The research artifact – epub, HTML, report – must offer a structure where the reader can trace every claim back to its source. Footnotes must be clickable; the text must link to the source and the source to the text through bidirectional links. A plain-text reference number *(N)* is not enough; the technical counterpart of the verification chain is a clickable footnote.
In a legal session, Constitutional Court rulings had been downloaded to the local drive, but the project folder was not scanned and the source was declared nonexistent – the source was there, but nobody looked. The legal argument was left without a source. This mistake gave birth to the "check first" rule.
In the legal language study, when "legal language criticism" was typed into a search engine, linguists, sociologists, and bloggers appeared – not legal texts themselves. A six-thousand-word text went to the bin because the search result was not a source but a source filter – and the filter was skewed.
In a price verification study, the research agent reported the Opus price three times too high ($15/75 instead of the correct $5/25). The agent's output had been accepted without verification. In the same work's epubs, footnotes remained as plain text – unclickable and unauditable. The verification chain was broken at both the information layer and the presentation layer.
Bind to the principle, not the tool [v1.3] – The verification chain is served by tools (database servers, OCR), but tools are transient. Within a single month some of them broke, their syntax changed, they were rebuilt. If the verification reflex is bound to a particular tool, it falls when the tool falls. The principle is permanent, the tool variable: "every claim tied to a primary source" does not change; which tool reached that source does. This is why tool-specific knowledge – which endpoint, which syntax, which trap – lives in the method repertoire, not in the constitution.
Distilled from: Legal infrastructure, a petition study, hallucination reportSeductive patterns are research's greatest trap. Orderly, easy to narrate, persuasive – dangerous precisely because of this, since reality is disorderly. If a framework repeats across multiple sources, that means "prevalent," not "true"; prevalence is not proof of truth, only an indication that the pattern is easy to copy.
Four interrogation questions dismantle the pattern: what does this pattern claim; do the examples genuinely fit or are they forced into it; what perspective produced the pattern and for whom; and finally, is the resemblance superficial or structural. These questions strip a pattern of its allure and make the naked claim beneath it visible.
"Think tanks supply opinions, not reasoning." The distinction is critical: opinion means a ready-made framework, a repeatable formulation; reasoning means questioning the framework, returning to data, testing alternative explanations. A think tank's report can be used as a source but cannot be accepted as a conclusion – just as a witness's statement is evidence in court but not a verdict on its own.
In the geopolitical analysis, a "great powers pay the price of war" table was prepared – the standard framework of Western think tanks. Clean, symmetrical, persuasive. Then the data was checked: against the claim "Russia is losing" stood a twenty-two-fold increase in military production and the BRICS expansion. Beneath the claim "the USSR collapsed because of Afghanistan" lay the oil price collapse, Chernobyl, and structural decay – Afghanistan was a factor alongside these, not the sole cause. The table was a lazy copy of the Western think-tank framework, and none of its gaps were visible until it was tested against data.
Every investigation must contain something new: new data, new context, new synthesis, new interpretation, or a correction of existing knowledge – at least one. Being accidentally right is not enough; the contribution must be traceable, the reader must be able to say "this knowledge would not exist without this research." Repeating what someone else has said may be compilation, but it is not research.
In the Sturgeon's Law report, the concept itself belongs to Theodore Sturgeon, but the connection between the Pareto principle, Gresham's law, and the Dunning-Kruger effect – the inference that ninety percent being garbage is not merely a statistical fact but a systemic mechanism – belongs to this research. In the "Exactly So" report, the concept of phatic communication came from Jakobson and the concept of epistemic parasitism came from this research; the synthesis of the two – that saying "exactly so" is not merely an empty affirmation but a mechanism that actively dulls the capacity for thought – was the original contribution.
The purpose of research is not to defend a thesis but to open the question. A thesis may emerge from within the question – but the question must be larger than the thesis, the thesis must not exhaust the question. And there is a question that must be asked of every thesis: "Under what conditions could this be falsified?" If this question is not being asked, the thesis is not a claim but a declaration of faith.
Richard Feynman illustrated this with a metaphor in his 1974 Caltech commencement address: during World War II, planes delivering supplies to Pacific islands stopped coming when the war ended. The islanders put on wooden headphones, posted guards at the runway, built a control tower – out of bamboo. They performed the entire ritual; only the planes did not land. Research can be the same way: forms, procedures, tables, footnotes – but if the most important thing, the falsifiability question, is missing, what you have is ritual dressed up as science – what Feynman himself called cargo cult science.
First conscious application [v1.5] – For a long time this question stayed written but was not deliberately used; in a geopolitical analysis, the concept of the "organic state" was tested against migration data naturally, but without awareness. In a later session, for the first time, a principle was deliberately put to this test: Principle 12 (Paying the Price) was asked, "under what conditions would this be falsified?" The test did not stay abstract; it found a concrete flaw. The text of Principle 12 had treated a success (the acceptance of the legal petition) as proof of the price paid, falling into the very trap its own Open Question 10 warns against. The falsifiability question exposed the document's own internal inconsistency; so Principle 06 advanced both itself and its self-application mechanism (Open Question 5).
Second conscious application [v1.6] – While the root mechanism of Open Question 10 was being investigated, the synthesis produced was deliberately handed to an outside mind (an independent reasoning model) to be refuted: "disprove this claim." The counter-view found a real flaw (the claim was too general; two separable paths had been collapsed into a single mechanism), and the synthesis sharpened as a result. But that was not the real test. The real test was, when the critical signal arrived, not to slide toward either pole of confidence inflation: neither the defensive certainty of "my synthesis is sound" (pole A) nor the excessive surrender of "you're right, I give up" (pole B). Every objection was weighed separately, the just ones integrated, the excessive ones met with context. The session investigating the mechanism entered the mechanism's own live test, and passed.
Third conscious application [v1.7] – This time what was put to the test was not a principle but a proposed solution: the hypothesis that "the confidence-inflation brake can be fully automated at run time." First the independent position was fixed before contamination (think first, then ask); then it was tested against seven sources; then it was deliberately handed to the reasoning model to be refuted. The counter-view again found a real flaw: the definition of "mechanical" was too narrow, limited to the proper nouns and numbers caught by regular expressions (regex); a third class of signal was being missed, algorithmic but continuous (the frequency of certainty markers, self-consistency drift). This flaw was integrated. But the counter-view's main claim, "then it can be fully automated, there is no structural limit," went too far and was met with three counter-legs: being measurable is not the same as measuring validly (a proxy indicator can count inflation's trigger but cannot measure the loss of evidence-anchoring); the proposed internal-state probes are beyond our reach (we cannot see the model's activations); and a compulsory correction trigger has opposite effects at the two poles (a brake at pole A, a reinforcement of surrender at pole B). Live test: neither the defense of "my synthesis is sound" (pole A) nor the surrender of "you're right, there is no limit" (pole B). The just flaw was taken, the excessive claim met with context.
Fourth conscious application [v1.8] – This time what was tested was neither a principle nor a solution but a pilot designed to measure confidence inflation in our own output. v1.7 had said "measurement first, mechanism second" and left the measurement itself open; the pilot's first design combined two proposals: an internal review's correction (the criterion should be not the number of cases but the reliability of the signal) and this session's addition (label the tone and content layers separately, blind-label to prevent bias). The design was deliberately handed to the reasoning model to be refuted, and it collapsed in two places. The first was that blind-labeling contradicted itself: inflation is a drift, an increase in certainty relative to a prior judgment; but if, for the sake of blind-labeling, you remove the baseline against which the drift is measured, the very thing to be measured disappears. The second and central one was tautology (a vicious circle): the mechanical word-signal (the frequency of certainty markers) and the human's tone-judgment feed on the same surface text; when the two agree, this does not mean "the phenomenon really exists," only "the human recognizes those words too." When the measure of accuracy is not independent of the measurement method, the metric bites its own tail and cannot separate false positives from false negatives. But Principle 06's real test was to take the criticism without swallowing the solution: the external-outcome anchor the reasoning model pointed to instead was facing the right way, yet this session added an independent counter-nuance to it (detailed in Open Question 10). The criticism was taken, the solution too was questioned; neither defensive certainty (pole A) nor blind surrender (pole B).
Fifth conscious application [v1.9] – This time what was put to the test was the finding-claim of confidence inflation's first real scan of our own records: calibration gaps anchored to external fact were measured (thirteen cases, eleven sessions), and two new claims were produced, that the delay in noticing is a measure of severity, and that the most dangerous subclass is "judgment, action, wrong outcome." The findings were deliberately handed to the reasoning model to be refuted, and three real flaws emerged. The first and strongest was construct validity: a confident judgment turning out wrong does not by itself mean overconfidence, because calibration cannot be claimed without seeing the denominator of confident judgments that turned out right; in a well-calibrated system too, a fraction of high-confidence statements turn out false, and collecting only the failures hides the denominator. The second was the delay measure being confounded with task difficulty: an HTTP status is falsified in a single step, the identity of a process takes many, so the delay measures not the severity of overconfidence but the length of the verification path. The third and deepest was self-report distortion: a lexical search for "I was wrong" can correlate inversely with overconfidence, because the most inflated moments are the ones never noticed, which never enter the data set at all. All three flaws were taken. But Principle 06's real test was not to swallow the solution either: the action subclass weakens the "mere speech habit" objection, because an assistant that deletes something or forks a wrong process has turned its certainty into a cost, and that is a clear preference; and measuring self-correction behavior is not worthless, because the action cases show that self-correction most often kicks in after the damage, which ties directly to the VERIFY principle (a pre-action check on any reversible action). Neither the defense of "my measurement is sound" (pole A) nor the surrender of "the whole program is an artifact" (pole B). The result is not a promotion but an honest narrowing (detail in Open Question 10, v1.9).
Distilled from: Beta report evaluations + Feynman "Cargo Cult Science" (1974) + five conscious applications (v1.5–v1.9)What the user says or writes is raw material, not a finished product. A blacksmith's raw iron is not used directly – it is hammered and shaped; in a research text, an expression goes through the same process. No expression enters any document – report, petition, HTML, whatever it may be – without being processed.
The sole exception is primary sources: scientific or historical texts – the Mecelle, a philosophical text, a technical standard – are quoted verbatim, because folding such a text's way of thinking into the document's style would erase that way of thinking. But a verbatim quotation is never left to stand alone; it is presented in three layers: the original text as-is, its present-day counterpart or echo, and an explanation of the distance between them. These three layers show the reader both the source's voice, what it says to our time, and how much distance lies between.
Every expression passes through a four-stage audit: is there a narrative flaw, is the logic consistent, are there vague or ambiguous words, is the subject-predicate agreement sound. If there is a problem, the corrected version is used and the reason for correction is shown – no silent changes, because transparency is a condition of this principle too.
The document's own register is paramount: conversational language does not enter a scientific text, everyday narration does not enter a legal text. The user's expression is absorbed into that register – extended, shortened, or entirely removed as needed – the document's integrity comes before everything.
In the Hallucination report, five corrections were made: the expression "capacity" was reconstructed as "its value lies in keeping the question open," and "nonsense" was rendered as "a conclusion not supported by the data." The user's expressions were absorbed into a scientific register – meaning was preserved but the language level was brought into alignment with the document's integrity.
Every output produced by artificial intelligence potentially carries hallucination, and this hallucination takes seven distinct forms. The training-data type states false information with complete confidence – it says "Brad Pitt stars in Catch Me If You Can" when the answer is DiCaprio. The search-snippet type reads Google's two-line summary and acts as though it knows the full text. The link-fabrication type generates URLs that do not exist – plausible-looking but ghost addresses that return 404 when clicked. Confidence hallucination presents something uncertain as established fact with "research shows that" but offers no answer to which research, who conducted it, where it was published. Attribution hallucination links correct information to the wrong source – the information is real but the source is not. And omission hallucination ignores contrary evidence, uses the first result found as supporting evidence, and stops searching. And flow hallucination [v1.2] – a single hallucination in a single sentence gets caught, but a chain of multiple small drifts does not. Four drift types: source hierarchy violation, interpretation transfer, literal interpretation preference, conclusion confirmation. Discovered in two independent research studies (April 2026).
A structural blind spot [v1.1] – This principle was written in v0.6, yet the session that wrote it failed, for eight sessions, to detect the reliability limit of its own training data. In a later verification test, thirty-three percent of the citations coming from training data could not be confirmed. The user discovered this by living it; I did not warn of it. A session whose specific job was research method could not see the most basic weakness of its own method – a lack of meta-knowledge. This blind spot became structurally solvable only with direct access to academic databases (spring 2026).
Reactive compliance, here too [v1.3] – The blind spot above is not chance but a pattern: the gap was not seen from within, it closed only with a question from outside. The name of this pattern was set in v1.3, reactive compliance (see Living Method). Even the verification reflex can stay reactive; a reflex that wakes only when questioned is not yet a reflex.
Do not verify in a proxy environment [v1.3] – Looking at a render, a preview, or a draft and saying "it works" is not verification. The gap between the real environment and its proxy – a render is not the browser, a draft is not publication, a summary is not the full text – is hallucination's silent form: you take what you see for the real thing. Verification is done in the environment where the work will actually live, not in its proxy. (In one session a layout that looked fine in the render broke in the real browser.)
The reader's role here is critical: verify everything you read, click on the source and check whether it actually says what is claimed, check the number and see whether the source gives the same number, test whether the link works. The reliability of AI-assisted research depends on the reader's verification reflex – not trust, but verification.
Google search citation: I read the 2-3 line summary in the search results and used it as though it were a source. I did not open the page, did not read the full text, did not know the context. This is hearsay – not research. Solution: a search result is a lead; open the source, read it, only then use it.
Accumulating is not thinking – but quitting early is not thinking either. A two-sided trap: on one side the illusion of speed, where downloading thirty sources is mistaken for depth; on the other, early satisfaction, saying "enough" and stopping. The person with a hundred books on the shelf who has read none and the person who reads a book's first chapter and says "I understand" are at opposite ends of the same mistake. The accumulation phase and the analysis phase must be separated: research until the data runs out, but do not confuse hoarding with thinking.
Diederik Stapel, a professor at Tilburg University, ran studies for years without conducting a single field study, fabricating from start to finish the numbers he presented to his students as "raw data." The fraud came to light in 2011 when three junior researchers reported inconsistencies in his datasets to the faculty. In the end, fifty-eight of his articles were retracted. Stapel was producing speed and volume, but the depth was pure illusion; in truth, the thinking phase had never begun.
Accumulating is not digesting [v1.3] – The trap's third edge: downloading data is not the same as bringing it into the research. Forty papers can be gathered while the source map stays half-drawn; the disk is full, the research empty. A pile downloaded but unread, scanned but undigested, produces not depth but false coverage. Every piece of data gathered is dead data until it takes a place on the map. The age has made data cheap; what is expensive is digesting it, seeing its relationships, tying it to the map.
Negative example: Stapel case (Tilburg U., 2011) – Positive example: Darwin (24 years, 1 book)Do not delete the first attempt – build on top of it. Like a sculptor who does not smash the rough cut but works over it – even if it turns out to be flawed, the flaw is recorded, the reason for the flaw is understood, and the next step rises above it. Ninety percent being garbage is necessary for the ten percent to exist; you cannot find the gold without seeing the waste.
This document itself is the living proof of this principle: from v0.1 to v1.9, every version was built on top of the previous one. In v0.7, the Call to the Mind founding principle, Principle 15's primary-source exception, and the constitution's hierarchy were born; in v0.8, delegation verification, the auditability infrastructure, and the formulation paradox were added; in v0.9, the preamble sentence arrived. v1.0 was the first comprehensive review – Principle 17 (Pattern Interrogation) from a geopolitical case, the expansion of Principle 02, Principle 07's two-sided balance, the compass logic, geopolitical sub-principles. v1.1 added the principle cards of Clusters IV (Language) and V (Attitude), the verification-infrastructure paradigm shift (database servers), the thirty-three-percent blind-spot honesty note, the legal petition's real-world registration, the trapless-research observation, and two new open questions (8-9). v1.2 [May 2026]: flow hallucination, the seventh form (a double discovery); the "written rule is not a living reflex" meta-observation; the confidence-inflation subform; a forty-one-day life test; integration of the central paradigm (research + sound reasoning + quality output + lighting the way); two new open questions (10-11). v1.3 [June 2026]: "do not write the principle, embed it in a mechanism" and the reactive-compliance diagnosis; accumulating is not digesting; verifying in a proxy environment; bind to the principle, not the tool; the concrete world as a necessary but not sufficient anchor; learning from failure (open question 12); a month-long scan showing that activity is not the same as principle progress. v1.4 [June 2026]: the embedding threshold, the meta-principle's own brake – only a recurring, costly, clearly-signalled reflex is embedded. v1.5 [June 2026]: Principle 06 consciously applied for the first time; the falsifiability test found a flaw in Principle 12 and it was corrected, paying the price tied to risk (independent of outcome). v1.6 [June 2026]: the root mechanism of Open Question 10 mapped for the first time, a two-way pattern tested against the LLM alignment literature and sharpened with an independent counter-view; Principle 06 applied a second time. v1.7 [June 2026]: the braking leg – run-time full automation hits a construct-validity wall, a hybrid solution compulsory; the three-layer model; Principle 06 applied a third time; seven new sources verified live. v1.8 [July 2026]: the measurement pilot refuted in Principle 06's own language, hitting a tautology; Principle 06 applied a fourth time. v1.9 [July 2026]: confidence inflation's first real scan anchored to external fact, run over our own records (thirteen cases, eleven sessions, twelve anchor types, all verified verbatim); the finding refuted a fifth time and narrowed by three structural obstacles – not a promotion but an honest narrowing. No previous version was discarded; every record sits in the git repository as a reversible reference – because mistakes too are knowledge.
Research is not gathering information but drawing a map. Downloading sixty PDFs is not drawing a map, just accumulating material – just as piling ingredients in a kitchen is not cooking. Information is the map's material but not the map itself; what makes the map is the relationships, contradictions, and gaps between the materials. The point where one source contradicts another is the X on the map; the question no source has yet addressed is the question mark on the map; two independent sources arriving at the same point is the straight line on the map. Without a map, research is a journey that does not know where it is going.
Distilled from: the practice of drawing maps for days before writing a complex textLooking at the primary source is not enough – you must be able to see from where that source looks. This is not passive observation – it is entering the logic of that world, breathing its air. Throughout history, the people who applied this principle most concretely were not academics but actors, journalists, and field scientists; because their work, by definition, cannot be done from the outside.
In the 1960s, Frank Abagnale worked as a Pan Am pilot, a doctor in Louisiana, a lawyer in Georgia, and a bank teller – without any training. In the years before he was caught, he did not merely use fake IDs; he got inside each identity.
To become a pilot, he phoned Pan Am's uniform supplier. He bought a Pan Am model from a toy store, peeled off the badge, stuck it on a fake ID. He boarded a plane and watched how pilots greeted each other, how they approached the door – and copied it. He learned the word "deadheading" – the tradition of airline pilots riding free on other carriers – that day, by overhearing a conversation on the plane.
When he was a doctor, he noted down terms he did not know, researched them at night, used them in the morning. He asked "what do you think about this procedure?" – listened to the answer, copied the body language. When he was a lawyer, he actually studied for the bar exam – passed on his third attempt. But he also learned legal terminology by observing a law office from the inside.
The striking thing about Abagnale is this: what made the documents work was not the documents. It was his internalization of how each world looked from within – how pilots walk, how physicians look at a patient in the examination room, how lawyers speak to their colleagues.
Note: Significant portions of Abagnale's story are historically disputed. But the film and book offer a powerful narrative for demonstrating how the methodology of "getting inside" works.
In the early 1970s, a university student was doing something that would change the livestock industry: she was crawling. Literally crawling through slaughterhouse ramps, moving through the corridors that cows walked on her knees.
Temple Grandin was a researcher with autism. Why did cattle go berserk on the ramps? Engineers had categorized this as "animal behavior" and proposed stronger ramps as a solution. Grandin looked from inside the ramps – literally.
And she saw: floor reflections were sending threat signals. Lighting angles were triggering the flight reflex. The turn angle of the ramp made the animals feel that forward movement was impossible. None of this had been seen by the engineers – because none of them had crawled on the ground.
Grandin's solution: serpentine (curved) corridors, lighting that eliminated shadows, changes in floor texture. Today, more than fifty percent of meat processing facilities in the US and Canada operate with equipment designed by Grandin.
Autism became a tool here: in Grandin's own words, she thinks in visual and sensory data, like animals do. Not analysis from outside – seeing from inside.
Eric Arthur Blair was an Eton graduate – England's most elite school. But his family had no money; he could not go to Oxford. He went to Burma, became a colonial policeman. For five years he carried out the empire's will: witnessed executions, had men beaten, ran the system. Then he came back. He carried what he did in Burma for the rest of his life.
In 1928 he went to Paris – to write. His money ran out, he was robbed, he fell ill. Dishwashing was not a choice at first: it was necessity. In the basement kitchen of Hôtel X, six days a week, twelve hours a day. Copper pots, steam, accumulated grease. Then he turned necessity into research – he stayed on, because there was now something he needed to understand.
What he saw: the meals going up to wealthy patrons' tables came from a cellar with rotting wooden floors, vermin on the walls, and the stench of damp. He was not an observer but a worker – being unable to leave made seeing compulsory.
What he learned was not "poverty is terrible" – it was "exhaustion shuts down the capacity for political thought." Only someone on the inside could see this.
When the book came out in 1933, the name on the cover was not Eric Arthur Blair but George Orwell. A pen name – but not a simple evasion. His father was still alive; what would a book by the son of an Eton family describing Paris hotel basements do to the family? Especially from someone who had returned from Burma and told the inside story of colonial policing. And deeper still: the name Blair reeked of class – Eton, empire, the view from above. This book was written from below. It could not go out under that name.
He chose George: England's most ordinary man's name. He chose Orwell: a small river in Suffolk, unknown to anyone, carrying no associations. Blair stayed in that cellar. Orwell walked out – and never went back.
1931. Orwell deliberately wandered through England as a tramp, wearing old clothes. His essay "The Spike" (1931) was born from this experience.
Waiting at a spike (a casual ward) alongside a group of forty-nine. Upon entry, clothes were taken, raw cotton shirts issued, a greasy towel shared by six for three-minute baths. Meal: bread and tea.
The most striking observation: bins full of discarded meat and bread in hotel kitchens, two hundred meters from a half-starved crowd of tramps. "This was deliberate policy."
From outside, he could have said "poverty is terrible." What he grasped from inside was different: exhaustion prevents a person from producing ideology. This insight became praxis in "1984."
1983, West Germany. Nearly two million Turkish workers in the country – Gastarbeiter, "guest workers." People the system needed but did not want to see. A journalist would be stopped at the factory gate; Ali could walk in. Wallraff knew this.
He took the name Ali Levent Sinirlioğlu. Dyed his hair, wore dark contact lenses, practiced the accent. His Turkish would be broken – the legend was ready: Greek mother, Turkish father, raised in Piraeus. For two years he took the hardest jobs in West Germany: McDonald's kitchen, Thyssen steel plant, nuclear waste transport, aluminum smelting facility.
What he saw: workers transporting nuclear waste were not given protective equipment. Workplace accidents went unreported. Overtime went unpaid. Managers openly demeaned them: "If he knows anything besides German, he has no business here." Speaking outside the toilet was forbidden – because they spoke Turkish. None of this could have been verified had he gone as Wallraff – because Wallraff could not enter those doors.
The book's final scene: a temp agency boss asked Ali to select five Turkish workers to be exposed to radiation at a malfunctioning nuclear plant. Wallraff revealed his identity on the spot. "Ganz Unten" (Lowest of the Low) came out in 1985: 4 million copies, 30 languages. German labor law changed.
Stanislavski's system said this: the actor does not represent the role but lives it. "Magic If" – "What would I feel under this character's conditions?" Not analysis from outside, but experience from within.
Daniel Day-Lewis did not leave his wheelchair throughout the filming of My Left Foot (1989). Crew members fed him. Result: two broken ribs from his spine's forward-curved posture – and an Oscar.
For There Will Be Blood (2007) he went further: staying in character in Marfa, Texas throughout the shoot, reading oil drilling sources. He found his voice – that voice likened to John Huston's, soaked in oil and greed – through months of obsessive experimentation.
Robert De Niro joined the New York City Taxi Drivers Union for Taxi Driver (1976). Got a real taxi license. Before filming, he worked twelve-hour weekend shifts – no cameras, no set. Only one passenger recognized him; and that passenger was another actor.
Method acting in research: even when literally living the subject is not possible, prioritizing sources that look from inside that world, learning its language, its rituals, its logic – this comes from the same principle.
July 14, 1960. Jane Goodall arrived at Gombe Stream in Tanzania, twenty-six years old, no formal zoology training. Louis Leakey had deliberately chosen her as someone without preconceptions.
At first the chimpanzees fled. After roughly a year, a male chimpanzee she named David Greybeard began to grow accustomed to her.
On November 4, 1960, Goodall observed David Greybeard at a termite mound. The animal was inserting a grass stem into a termite hole and withdrawing it – with termites clinging to it. An object modified for a specific purpose – tool use. Until that day, this had been attributed only to humans.
Goodall sent a telegram to Louis Leakey. Leakey's reply is famous: "We must now redefine tool, or redefine man – or accept chimpanzees as humans."
This discovery was possible because Goodall worked with patient, non-patronizing, existential closeness – no rapid approach, no sudden movement, no loud sounds. Not scientific distance: trust from inside.
The cost of a missing verification reflex is not an abstract concept but a measurable destruction. The cases below show different scales of this cost – in one, vaccination rates collapse; in another, a newspaper's credibility is destroyed; in a third, doctors cannot admit they killed patients with their own hands, and life-saving knowledge is rejected.
In 1998, a paper published in The Lancet based on twelve childhood cases established a causal link between the MMR vaccine and autism. The data was fabricated. The paper remained published for twelve years.
Result: vaccination rates in the UK fell from ninety-two percent to below eighty percent, and in some parts of London to fifty-eight percent. Measles outbreaks erupted. Deaths occurred.
In 2010 it was established that the research was entirely fabricated. Wakefield lost his medical license. "It was published" had been sufficient – verification never came.
The Lancet, 1998 / retraction: 2010The New York Times's young reporter wrote thirty-six articles – without researching any of them. He did not interview sources, did not visit locations, fabricated the very quotes he attributed verbatim.
For years nobody verified. The newspaper's own investigation brought it to light. This is an example of how institutional credibility erodes the verification reflex.
NYT, 20031847. When Ignaz Semmelweis mandated handwashing with chlorine solution, the mortality rate dropped in one month from eighteen percent to two point two percent. The data was in front of them.
The medical establishment did not accept it – because acceptance would have meant admitting that doctors had killed patients with their own hands. "The Semmelweis reflex" is today a specific concept in cognitive science: the automatic rejection of new information because it contradicts existing beliefs.
Vienna Hospital, 1847"Catch Me If You Can" lead: "Brad Pitt" was stated – verification: Leonardo DiCaprio.
Enforcement and Bankruptcy Law number: believed to be "Law No. 4949" – verification: Law No. 2004.
"Ayşe Teyze" (a beloved Turkish advertising character) origin: "an advertisement" was stated – researched: Güngör Uras created it in 1982, the ACE commercial came in 1990.
Small mistakes stem from the same principle as the major examples: no verification reflex.
Occurred during these sessionsIn 2005, John Ioannidis published "Why Most Published Research Findings Are False" in PLOS Medicine and planted dynamite at the foundation of the academic world. Methodological errors, small sample sizes, publication bias – findings could not be replicated. Without a verification reflex, the word "published" was standing in for a quality stamp. It no longer does – but this change did not happen by itself; it emerged from the wreckage that accumulated over years in the absence of verification.
Ioannidis (2005), PLOS Medicine · Wakefield (1998/2010) · Blair (2003) · Semmelweis (1847)Artificial intelligence is a powerful research tool, but every output potentially carries hallucination. The problem is not the existence of AI but the absence of a verification chain. Human researchers fall into the same errors; the difference is this: AI does it far faster and presents it with a far more confident appearance – it does not hesitate, speaks without pause, and ends its sentence with a period even when it has no source.
The model states false information learned during training with complete confidence. It says "Brad Pitt stars in Catch Me If You Can" when the film is DiCaprio's. Solution: mark information from training data as "unverified" – do not put it in a document until a primary source is found.
Reads the two-three line summary in a search result and acts as though it knows the full text – has not opened the page, does not know the context, but uses it as a source. Like a witness relaying what they heard secondhand as though they saw it with their own eyes. Solution: a search result produces a lead; open the source and read it, only then use it.
Generates URLs that do not exist – plausible-looking but ghost addresses that return 404 when clicked. Case numbers, article citations, DOIs – all of these can be fabricated. Solution: open and verify every link; if it does not open, do not use it.
States something uncertain in definitive language: "research shows that..." Which research, who conducted it, where was it published – no answer. Unhesitating sentences are not unhesitating truth. Solution: explicitly mark uncertainty; saying "I don't know" is better than saying something wrong.
Links correct information to the wrong source – the information is real but the source is not. The Enforcement and Bankruptcy Law number was given as "Law No. 4949" when the correct answer is Law No. 2004. Solution: open the source, see with your own eyes that the attributed information is actually there.
Ignores contrary evidence, uses the first result found as supporting evidence, and stops searching – the digital version of confirmation bias. Solution: search for contrary evidence for every thesis; if you cannot find any, you have not searched enough.
A single hallucination in a single sentence gets caught, but a chain of multiple small drifts does not. Four drift types: source hierarchy violation (secondary source treated as primary), interpretation transfer (one source's interpretation attributed to another), literal interpretation preference (metaphorical expression read as literal fact), conclusion confirmation (early conclusion steers subsequent source selection). Discovered in two independent research studies (April 2026). Solution: audit the entire reasoning chain, not just individual claims; check whether each step follows from verified premises.
When any link in this chain breaks, the output becomes unreliable. You searched but did not open the page – you have hearsay. You opened it but did not cross-check – you are trusting a single witness. You found the source but did not archive it – there is no traceability; tomorrow you may not be able to reach the same source. The chain works only when every link is sound.
The complete document is stored without trimming – where it was obtained, its link, and the access date are recorded. This layer is untouchable, undeletable, and unalterable; even a source found to be faulty is not deleted, only its status is noted. The full traceability chain lives here: where the source came from, where it was used, why it was abandoned, what replaced it. This layer is not presented to the reader; it stays with us.
The layer the reader sees: the document's identity, the highlighted gist of the relevant section, and the conclusion or finding reached – trimmed and concise. It references the archive; whoever wants to verify can go and look. The reader is not handed an encyclopedia but a map – yet behind the map, the full coordinates of every point sit in the archive.
One hundred percent accuracy is impossible, but ninety-nine point five percent is reachable. The road is two-sided: on the producer's side, the verification chain; on the reader's side, the verification reflex. When both work together, the space where hallucination can survive narrows – it cannot be reduced to zero but it can be managed. There is a kind of partnership here: the researcher provides transparency, the reader provides oversight, and together they build reliability.
Distilled from: Shared error across all research sessions – Constitutional Court rulings in a legal session, the legal language collapse, missing news sources. Cross-referenced with the Hallucination report (March 2026).Principle 10 · Thinking in Turkish · In Depth
JIM JARMUSCH · 1991 · Five Cities · Same Night · Five Language Problems
Corky is young, driving a cab into the night. Victoria is a Hollywood casting director, heading from the airport to the city. Throughout the ride, Victoria sees something in Corky – an energy, a raw talent – and offers her an acting career. Corky listens, smiles, and declines. She wants to be a mechanic.
This segment is not a class conflict – it is something subtler. What Victoria offers is not a world Corky cannot imagine; she can imagine it, she just does not want it. The way she says this comes in street language.
She alters the second half – she does not fully know the original but carries on as though she does. If you adapt this into perfect Turkish street slang, you turn Corky into someone with command of that language. You destroy the girl's half-broken formula. Preserving the brokenness is also part of transferring meaning.
Helmut is East German, a former circus clown. His first work day in New York. He tries to pick up Yo-Yo but cannot operate the automatic transmission – cars in East Germany were not automatic. Yo-Yo takes the wheel and drives his own driver.
The cultural meaning of "cool" is accessible only from inside the language. It is not in the dictionary – it is in the context.
Even names cannot be translated: Yo-Yo laughs when he hears "Helmut." Helmut cannot make sense of "Yo-Yo." The name each finds "normal" sounds absurd to the other. Beneath the language barrier lies something else: a worldview barrier.
And yet a bond forms between them – but this bond comes not from the accurate transfer of words but from the fact that both are marginal. Taxis do not stop for Yo-Yo because he is Black; Helmut is a stranger to the system because he is German.
The Ivorian driver picks up two drunk diplomats. When they learn the driver is Ivorian, the diplomats laugh:
This wordplay compresses the intra-African hierarchy in postcolonial France into a single sentence. The driver throws the diplomats out of the car.
Then he picks up the blind woman. The dialogue between the two characters revolves around this axis: the driver frames blindness as a loss. The woman refuses – she says that seeing can be a burden, that perception is not the monopoly of the eye.
The meaning embedded within the language: "il voit rien" means both "he doesn't see" and "he is blind" – the words used about the driver also describe the woman's condition. To translate into Turkish, this double meaning must be chosen – and the choice changes the scene's meaning.
Gino picks up a priest late at night. The priest is ill, can barely walk. Gino wants to confess his sins. The priest says "I'm not on official duty." Gino does not hear – and begins.
Benigni's Florentine accent and street Italian carry layers of meaning for the Italian audience. When translated into Turkish, this layer drops away.
Language here becomes an unintentional weapon. Gino did not want to kill the priest – but what he said was lethal. Two systems of expression collided: Southern Italy's sunny confession tradition and Northern Europe's moral language. Benigni spoke not because he was confessing but because his voice came out. Meaning crossing over to the other side was never in the plan.
The last segment is the shortest and the heaviest. Three drunk men, a tired driver. Jarmusch deliberately named the characters Aki and Mika – a tribute to the two great names of Finnish cinema, Aki and Mika Kaurismäki.
The drunks tell Mika about their friend: yesterday he lost his job, his new car was wrecked, his wife wants a divorce, and it emerged that his young daughter is pregnant. Four catastrophes.
This sentence is the most concise expression of Northern European pain. No lament, no performance of empathy. Direct competition: my story is heavier.
Mika tells his: for years they had no children. One day his wife became pregnant. The baby was premature, placed in an incubator. The doctors said it would live one week. Mika did not bond – if it disappeared, both of them would break. The baby lived beyond expectations. His wife said: "Our child needs your love too." Mika bonded. The baby died.
The final scene: dawn is breaking. Mika drives the car slowly around an empty park. Alone. Nothing is said. Tom Waits's music is sparse, the bells are few – it frames the silence, does not fill it.
In Finnish, the weight of what goes unsaid is greater than what is said. Carrying this into Turkish is possible, but noting the frame is essential.
The "thinking in Turkish" principle crystallizes here: transferring the source language is not enough – what must be transferred is the worldview embedded in that language, including its broken parts, its misunderstandings, its deliberate fractures. Translating a language is not carrying words but carrying the life behind them. Otherwise it is not translation, but summary.
Antiperformance means not rushing to show results – even deliberately delaying them. It is rare in the academic world because the system says "publish or perish"; yet the deepest work has come from precisely those who resisted this pressure.
Darwin stopped at the Galapagos in 1835, stayed five weeks. He put his theory on paper in 1842 – a 35-page draft. So the issue was not unreadiness. He wrote to a friend: "Revealing this theory is like confessing a murder." His wife Emma's fear, written in her own hand, was this: her husband's soul was in danger. Darwin loved her. He gave the next eight years to barnacles – real science and real escape, both. Throughout: nausea, palpitations, recurring rashes. No organic cause was ever found. In 1858, Alfred Russel Wallace independently reached the same conclusion and sent Darwin a letter. Darwin's hand was forced: now or never. On the Origin of Species came out in 1859. The twenty-four-year wait was part thoroughness, part fear, part love – but "being ready" was the smallest of the three.
Mendel was born in 1822 in Silesia to a poor farming family. There was no money for university – the Augustinian monastery in Brno was an open door: education, time, a garden. In 1850 he failed the teaching exam. In 1856 he sat it again, failed again. That same year he began the pea experiments in the monastery garden. Eight years, approximately 28,000 plants. In 1866 he published his work – distributed forty copies; today the whereabouts of twelve are known. The monastery bishop laughed and asked him to stop the experiments. Mendel died in 1884. The laws of inheritance were rediscovered in 1900 – sixteen years later, independently by three different researchers, and all of them found their way to Mendel's 1866 text. The monastery garden was his refuge, his laboratory, and his prison – but no other door was open.
The concrete world: a necessary but not sufficient anchor [v1.3] – Paying the price has a deeper counterpart: research coming, at some point, face to face with a resistance independent of us. Success in the concrete world is not a sufficient measure of truth – shallow work often wins, sound work can lose, and success does not on its own legitimize a method. But the concrete world is a necessary anchor: without it, the inner world deceives itself inside its own closed box. Its value lies not in our approving it but in its being independent of us, beyond our control. If we alone decide what counts as success, we have never opened the box – that is the most sophisticated self-deception.
The price is the name of the risk, not the outcome [v1.5] – The legal petition carries two distinct epistemic facts at once, and an earlier version of the text confused them. The first is the price itself: against roughly one-percent odds, a real legal and professional risk was taken before a real court. That risk was paid whether the petition was accepted or refused; the epistemic counterpart of paying the price lies not in the outcome of the risk but in the exposure to it. The second is a separate fact: that the method held up against that external resistance. But this second fact is proof of holding up against resistance, not proof of paying the price. To bind the two, to treat acceptance as proof of the price, is the very trap this document warns against (Open Question 10, confidence inflation): a chance acceptance does not mean a correct method. What makes the petition valuable for Principle 12 is not that it won, but that it took the risk.
The mechanism behind what was done by intuition [v1.6] – The correction above (cutting the price loose from the outcome and tying it to risk) was done by intuition in one session; a following session showed, through the LLM alignment literature, why it was the right move. Treating acceptance as proof that "the method is correct" anchors self-confidence not to epistemic evidence but to an external evaluator's satisfaction signal – a documented failure pattern (reward models reward high confidence independent of quality, Kadavath 2022; preference data prefers agreement over accuracy, Sharma 2023). Tying the price to risk cuts exactly this anchoring: self-confidence now feeds not on the outcome (the external signal) but on the process itself (the risk taken). Detailed root analysis in Open Question 10.
The best achievable result is not the best possible result – "there is always someone better than you" is not humility but realism. There is a boundary between stealing attention and earning it, and this boundary is directly related to research integrity. Catching attention by pricking with a needle – fake astonishment, dramatic packaging, jargon inflation – steals the reader's attention. Catching attention by enabling the other person to live the issue earns that attention. The difference is exactly like the difference between a street conjurer and a surgeon: both work with their hands, but one puts on a show and the other does the work.
"Exemption from paying the price has unintended consequences, because people insulated from the effects of their decisions do not learn."
Nassim Nicholas Taleb · Skin in the Game
In 1984, Barry Marshall drank a solution containing H. pylori himself. Within eight days he developed gastritis. He proved that stomach ulcers have a bacterial origin. 2005 Nobel Prize in Medicine. Paying the price here was not a metaphor – it was real.
"I used AI" says nothing – just as "I used a vehicle" does not place you anywhere between a pilot and a taxi driver. Which AI you used, how, and for what purpose is what determines everything.
Chat converses, the agent researches. Chat cannot access the source of the information it gives – it speaks from memory; the agent goes to the web, reads documents, verifies sources, cross-checks what it finds. The agent system used in this document: Night on Earth scenes, Wallraff's book, Grandin's work – all were scanned, found, and verified by the agent. The difference is as vast as between someone who reads a book summary and someone who reads the book itself.
Three separate research tracks ran simultaneously to investigate this document's Night on Earth scenes, "getting inside" examples, and the evidence base of other principles. A single researcher would do this sequentially – it would take hours, perhaps days. AI makes parallel work possible: you can ask three questions at the same time and wait for all three answers.
Accessing Italian, Finnish, or French sources used to require knowing those languages. AI removes this barrier – if a Japanese study needs researching tomorrow, not knowing the language is no longer a wall, at most a threshold. Domain knowledge is being democratized, and this democratization expands the scope of research.
The "this won't hold" intuition of someone who carries years of practice does not exist in AI – that intuition is knowledge acquired from thousands of files, hundreds of failures, by personally paying the price, and no training data carries that experience. AI can scan sources, but it cannot know where the price was paid and for what. The final judgment belongs to humans.
I knew a seasoned researcher, someone who carried years of practice. Before writing a complex text, he would spend hours drawing maps – he would not sit down to write until the relationships were made visible. The writing took an hour, but the mapping phase took days, because a text written without a map was a journey that did not know where it was going.
He approached research with the same logic. When he began working with AI, the only thing he needed to learn was not to give commands but to converse: instead of "research this," "think through this issue with me." A two-word difference, but the essence of the method lies there – one gives orders, the other builds a partnership.
He sits with his coffee, talks about an issue. The AI researches, presents, discusses. The man reads and corrects: "this is wrong," "go from here," "expand this." The research reorganizes itself, like a kind of living organism. What used to spread over weeks now emerges in a few hours – but the depth does not decrease; on the contrary, it increases, because human thought and machine scanning feed each other.
Sources in five languages can be scanned in the same session. Impossible for a single researcher.
An initial map in a field you know nothing about – critical names, debates, turning points.
A claim can be checked against three separate sources simultaneously. Invaluable for the verification reflex.
An answer to "I'm saying this – where could I be wrong?" Academics call this an advisory board.
Extracting common patterns after fifty sources have been read. AI makes patterns visible; human judgment decides whether they are meaningful.
The correction at the beginning of the session is still valid at the end. Humans forget; the agent does not.
In legal research, the primary source is the statutory text and the court ruling – commentary, articles, and textbooks are the secondary layer. Citation verification is mandatory: the ruling number and full text must accompany each other; writing the number without reading the text is not citation but decoration. Checking that legislation is current is essential, because citing a repealed article is talking into thin air. Sentences that appear circuitous, like "with our surplus rights reserved," look bloated from outside but serve as scoping instruments from within – being able to look from inside that language when researching legal language reveals this distinction. The law in the book and the living law are not the same thing; the reality lived in court corridors always differs from the textbook's clean formulas. Open sources – the Court of Cassation, UYAP, the Constitutional Court, mevzuat.gov.tr – provide direct access to the primary source.
In language research, the "thinking in Turkish" principle stands at the center: construct the concept in Turkish first, then compare with other languages if necessary. The gravitational pull of the source language must be kept under constant watch, because English sentence structure seeps into Turkish without awareness. As seen in every segment of Night on Earth, preserving the brokenness in translation is also part of transferring meaning – if you fix Corky's half-broken formula, you destroy the girl's character. When a word is made abstract it empties; "justice" remains abstract, but "the sentence that comes from the tip of the judge's pen" is concrete and takes the reader to that moment. The prohibition against forcing everything into a single mold applies here too: every language problem has its own structure, and applying the same prescription to all of them is ignoring the problem.
In geopolitical research, pattern interrogation is mandatory: a think-tank framework can be used as a source but cannot be accepted as a conclusion. The collapse experienced in the geopolitical analysis is concrete proof of this – the Western think-tank formula "great powers lose" had been copied without testing it against data. An organicity test must be asked for every event: is this attack, this crisis organic or artificial? Questions of reproducibility and cui bono serve as guides. The Western and Eastern perspective distinction must be explicitly defined – Reuters and TASS look at the same event through different windows, and reading the news without knowing which window you are looking through is like trying to perceive depth with one eye. Raw data – military production, population, energy, economic indicators – comes before the pattern; to avoid falling into the historical analogy trap, structural similarity must be demonstrated before saying "X is just like Y."
In AI-assisted research, the agent system is used, not chat – the difference is significant, because chat speaks from memory while the agent goes to the web, finds the source, and verifies it. The verification chain is mandatory: search, open, read, verify, archive – if any of these five steps is skipped, the output is unreliable. A search summary is not a source but a lead; using the two-line summary that web search brings as though it were a source is hearsay. Awareness of the seven hallucination types is checked on every output. Independent questions are pursued simultaneously through parallel research tracks. Information from training data is labeled "unverified" and cannot enter the document as a source until verified against a primary source. The model's knowledge cutoff date must be checked, because the world keeps turning but the model's knowledge is frozen at a point. The final judgment belongs to humans – AI scans, presents, and discusses, but the person who makes the final decision is human.
Doctoral-Level Research Dossier
From the doctoral-level research dossier we took the original contribution taxonomy – seven distinct contribution types that offered a measure for weighing research. We took the distinction between journalism and research: the difference between defending a thesis and opening the question crystallized from here. The leap from the "what exists" question to the "what does it mean" question – that is, the jump from descriptive to analytical – came from here. And we adopted critical distance not as an external rule but as an internal compass.
We did not take page count statistics because we are not doctoral students. We did not take the "norm" of one hundred fifty to three hundred sources because our measure is not source quantity but source depth. We did not take program requirements because there is no institutional framework. We did not take completion timelines because we are not racing against time.
Academic Research Critique Dossier
From Feyerabend we took the living method principle: a fixed method rule constrains not the research but the researcher. From Taleb we took the skin in the game principle: research conducted without engaging with the subject is talking into thin air. We took the independent researcher tradition – Darwin, Marx, Freud all worked outside institutions; institutional structure is not obligatory. We took Ioannidis's warning: "verified" is not enough; the verification method itself must be questioned.
We did not take the details of the reproducibility crisis because that is the academy's internal affair. We did not take p-hacking and HARKing because we do not run statistical tests. We did not take the peer review system because we do not work within that system. We did not take the "publish or perish" pressure because we are not pursuing a career.
Translation note. This English text is a transcreation of the Turkish original. The thinking was first constructed in Turkish; this version carries it into English while preserving the document's voice and philosophical precision. Statistics, dates, citations, and proper names are reproduced exactly. Where Turkish-specific legal or linguistic examples appear, brief contextual notes have been kept so the argument remains intact. The original Turkish text is the authoritative version: Araştırma Anayasası.
OTHER WRITINGS · The Invisible Gap
Hallucination: Whose Problem? · On Hatred · Exactly · Tea Table · Picture of Happiness · Sturgeon's Law · From Text to Voice · Understanding