A Map Drawn
on Thin Sheets

MXene (pronounced “maxene”) is a very thin sheet. Rows of metal atoms and rows of carbon (or nitrogen) atoms alternate like a sandwich. Many such sheets stack like the pages of a book. The spacing between sheets is measured with X-rays and is typically around 1 nm.

The most studied is Ti₃C₂Tₓ, made of titanium (Ti) and carbon (C). The Tₓ at the end denotes the atoms attached to the sheet surface. How the surface is treated changes the properties considerably. As Chapter 9 will show, even for the same Ti₃C₂, some values spread by more than 3 eV (a very small unit of energy used when dealing with a single electron) depending on the treatment.

Synthesis goes roughly like this. A hard crystal that exists before the sheets form (the precursor) is immersed in a chemical that dissolves out one layer (etching). This exposes the sheets, which are then peeled off one at a time (delamination). There are several etchants: hydrofluoric acid (HF), a mixture of lithium fluoride and hydrochloric acid, molten salts, and so on.

How many kinds are there? Metal–carbon combinations actually synthesized in the lab alone number more than 50. Counting mixed-element ones and ones with different surfaces, the number approaches 100. Candidates predicted only by computer calculation number more than 1,000.¹

“MXenes comprise a very large number of materials. Which composition should be used for which property — can the structures and synthesis methods be organized systematically?”— This map began on 2026-10-06 from this single question.²

MXene first made its name through its ability to store electricity, i.e. its capacitance. Capacitance means how much electricity it can hold, as in a capacitor. The landmark results published in Science in 2013 and Nature in 2014 were both about the capacitance of Ti₃C₂Tₓ. In particular, the key was how much it stores in a given volume (the volumetric value).³

After that, research branched out in every direction: how well it conducts electricity, how well it blocks electromagnetic waves, how hot it gets under light, how well it carries heat. These are called properties. Papers piled up for each property.

The person who asked the question is a researcher who once wrote papers with MXene. In their own words, however, “I didn't study the material itself; I used it to build structures.” So the materials knowledge needed for this map has been built up piece by piece during the work.⁴

The question was simple, but the answer was not. The next chapter shows why.

Numbers are scattered inside the papers. Some sit in the middle of the main text, some in figure captions, and some are bundled with several samples in one sentence via the word “respectively.” To find the value you want, you have to open the papers one at a time.¹

So a few databases that collect only the numbers (data stores that gather values into tables) already exist. The most similar one we found was a public dataset of 573 rows collected from 114 papers. It is a useful resource. But on closer inspection it had four shortcomings.²

• 91% of the values are concentrated in a single MXene, Ti₃C₂.
• The values were extracted by AI, but there is no report of how accurate they are.
• No value records “which sentence of the original it came from.”
• Above all, wrong values were mixed in.

We checked the last point ourselves by opening the originals. We compared the values in that database line by line against the original papers. Values that did not match the original amounted to 17.9% in the first 15 papers and 8.0% in an expanded set of 22 papers. Because the number of papers examined (the sample) is small, neither figure is definitive.³

The wrong values often had the same shape. The value of a composite (two materials combined), made by mixing MXene with another material, was recorded as if it were the value of pure MXene. The number itself is in the original, so at a glance it looks right. But “whose number it is” was wrong.⁴

For a flagship property such as how well electricity flows (electrical conductivity, in S/cm — larger means better conduction), record values like “24,000 S/cm” circulate without their conditions. These are numbers stripped of which sample was made how and measured how. This is exactly where a map with attached evidence is most useful.⁵

This is not a race to collect more values. It is the work of collecting values that can be trusted.— the sentence that defined “our share” of this map

Even obtaining the originals was not easy. From the start, we decided to use only legal routes, because values taken from unauthorized copy sites can neither be published to others nor re-checked by others against the same original.⁶

Yet even for papers that anyone can read openly, publisher websites often blocked programs from accessing and reading them. In the first comparison, of 60 open-access papers the full text could be retrieved by program for only 16. In the end, a paper repository that works like a public library became the legal route a program can read. (How papers blocked to programs were obtained separately and scored the same way is in Chapter 10.)⁷

It is not only competitors. There is also a public dataset of 4,488 MXene structures generated by computer calculation. This is not a competitor but a partner: a “calculated value” column can be placed side by side under the same composition name.⁸

So the task was set. Collect the numbers, but attach to each number the sentence in which it is written. And score separately whether each pair is correct.

A row of this map is not just a number. It records, as one bundle, which MXene (composition), which sample and which unit the number belongs to, and from which sentence of the original it came. The synthesis method and the measurement conditions are recorded too. Like the call number on a library book, anyone can use this row to find the way back to the original location.¹

Look at the sentence in the figure above. It reports surface area (m²/g — how many square metres 1 g of powder would cover if fully spread out). One sentence contains three samples and three numbers. Because the word “respectively” is attached, the pairs must be matched in order. The first sample (few-layer Ti₃C₂Tₓ) is 6.05 m²/g. This kind of pairing confuses people and AI alike.²

Some rows cannot be used even when the number is right. A row extracted from the sentence “extending to 16 hours, the specific surface area became 95.7 m²/g” was one of them. The value 95.7 was exactly as in the original. But the sample-name field was empty. Since there was no way to know which sample the number belongs to, it was not put on the map.³

Even if the number is right, it cannot be used when “whose number it is” is blank.

Eight steps to put one property on the map

Even when the property changes, the order is always the same.⁴

1. Fix the list of papers first. So that favorable papers cannot be cherry-picked in later.
2. Screen. Keep only papers that report the value from “their own measurement.” Papers that only calculate, and papers that copy someone else's value, are excluded.
3. Transcribe. An AI reads each paper and extracts, row by row, the composition, sample name, value, unit and original sentence.
4. Make trap questions. Secretly alter a few of the extracted rows (Chapter 4).
5. Save the scoring criteria first. Before scoring, write down and save what will count as “correct.”
6. Two judges score independently. Each opens the original and picks one of “correct · wrong · can't tell.”
7. Attach a mark. If both say “correct,” ●; if at least one says “wrong,” ✕; otherwise ○. (In Chapter 6 one more mark, ◐, appears.)
8. Decide pass or fail. If ● is 80% or more and ✕ is 10% or less, the round passes.

The fifth step is the backbone of this work. If the criteria are set after seeing the results, the criteria bend toward the results. It becomes not measuring but “fitting.” It is like sealing the answer key in an envelope before the exam. This is called preregistration (registering in advance). If we want to change the criteria, we open a new round and measure again from the beginning.⁵

This creates some inconvenience. A moment comes when, looking at the results, you think “it would look much better if only this field were excluded.” That story is in Chapter 5. Even then, we did not put on the map a number chosen after seeing the results.

Think of grading a school exam. The question paper is the original paper. A student's answer is one row the AI extracted from the original (“this sample's value is this”). There are two teachers grading: Claude and Codex, two AIs built by different companies. Neither sees the other's grading.¹

But how would we know if a teacher grades carelessly? So before trusting the teachers, we test them first; we are “grading the graders.” We secretly mix eighteen trap questions into the pile of answers. Twelve are answers deliberately altered to be wrong, and six are correct answers left untouched. The teachers do not know which ones are traps.

The reason for including the six correct answers is simple. Even a teacher who says “wrong” to everything would catch all twelve altered answers. Only by also checking whether the teacher recognizes intact answers as “correct” can we trust that teacher.

What the traps look like

Some of the traps actually used.²

• Changing the value. In the original, a sample's surface area was 16.7 m²/g. The trap sent it as 61.79. Both teachers caught it, saying “the paper says 16.7.”
• A sample name that doesn't exist. To the value 556.283 m²/g we attached a sample name not in the paper (“composite calcined with graphene, sample B”). Both said “no such sample in the paper.”
• A different MXene. The 34.21% efficiency of converting light into heat (photothermal conversion efficiency) belonged to niobium MXene (Nb₂C) particles. We sent it as Ti₃C₂. Both caught it: “these particles are Nb₂C.”
• Changing the unit. We changed 1.15 nm to 1.15 Å (a unit 10 times smaller). Both said “it is nm, not Å.”
• Blanking a field. We deliberately sent a condition clearly stated in the original as a blank. This tests whether the judge recognizes “omitted” rather than “wrong” (Chapter 5).
• Changing the property name. The number and sample were left as they were, and only “which property the value belongs to” was changed. This is a trap newly added on 2026-10-09 (Chapter 7).

Catching all eighteen traps does not mean “the judge is perfect.” Eighteen is not many. Even if a judge actually errs once or twice in ten (up to about 17%), it could get all eighteen right by chance. So we only go as far as saying “this judge has been tested enough to be trusted.”²

What one misses, the other catches. That is why there are two teachers.

Sometimes the two teachers disagree. For a row extracted from a sentence saying the spacing between layers “increased to about 1.41 nm,” one teacher said “correct” and the other said “wrong — this is the value of the sample before another chemical was added.” The rule is that if even one says “wrong,” it is ✕. That row did not go on the map.³

If the answer key is ambiguous, even the teachers fail

Once, both teachers failed the qualification test. Both caught all twelve altered answers. But of the six “intact correct” answers, one teacher called only four “correct,” and the other only three.⁴

On inspection, the problem lay not with the teachers but with the answer key. The rows we had chosen as “correct” had ambiguous definitions to begin with. Since then, before scoring, we first open the originals of the correct-answer rows. Ambiguous rows, such as a lone number from a table or a value measured in a different liquid, are removed from the correct-answer candidates. In the next round after this fix, both teachers recognized all six correct answers. However, since this is a single comparison with a different property and different papers, we do not claim it was “because of this.”⁵

Conversely, too clean is also suspicious. In one round, both teachers called all 36 rows “correct” and scored perfectly on the traps too. Rather than rejoicing, we first wrote down “were the traps too easy?”⁶

How we chose the papers

Papers anyone can open, first. Every number on the map carries its original sentence. Since other people must be able to check that sentence for themselves, at first we used only full texts that are freely available.¹

Aimed at the places where the map is empty. We chose papers to fill the spots that the usefulness test (Chapter 10) had flagged as “cannot answer — no data here.” For example, a spot such as “the capacitance of MXene etched with sodium hydroxide instead of HF.”

Choosing by title alone misses. At first we chose by the words in titles and abstracts. Then, while looking for “papers that etched with an alkali,” we caught “papers that etched with HF and then washed with an alkali.” Same words, different experiment. So nowadays we also check whether the abstract actually states that condition, and a person reads them once more to filter.²

The test's “answer papers” are kept out. The usefulness test checks whether the map can rediscover 20 “known facts.” If the papers that first stated those facts were put into the map, the map would be getting the answers right by copying the answer key.

Papers obtained separately are scored the same way. Papers obtained outside the open route include publisher full texts. Still, they are scored by the same method, and the results are reported separately for each “where it came from” stream. We do not throw data away, but we do not blur it by mixing, either.³

Write the rules first, before seeing the results. A rule changed after seeing the results gets the label “exploratory.”

How we changed the rules

① Write it down and save it first. Before running a test, we write “what will be counted · what score passes” in a file and save it (commit it). That way, no one can later change the story to “actually, the criterion was this.”

② If it fails, we do not sit the same test again. On 2026-10-09, while scoring the 45 papers a person had brought in, one judge missed two trap questions: it called rows whose name field had been deliberately left blank “correct.” Re-running the same test sheet would amount to running it until it passes. So that round was left as “undetermined,” and we re-took it with a new test sheet with freshly drawn traps. This time both judges caught all 18 traps, and 222 of 223 rows passed.⁴

③ A rule changed after seeing the results gets a label. In the usefulness test, “a Ti₂C battery has a larger capacity than Ti₃C₂” disagreed. Looking into it, the Ti₃C₂ values in our map included many composites mixed with high-capacity materials such as silicon — whereas the answer paper compared pure MXenes with each other. So we changed the rule to “compare only samples with no mixed-in material.” But because it was changed after seeing the results, we labelled it “exploratory” and do not use it as evidence that the map is useful.⁵

④ Write what improved and what got worse side by side. Re-measuring with the changed rule, the Ti₂C comparison turned into “agrees.” But “Nb₂C is larger than Ti₂C,” which had earlier been “agrees,” became “no answer”: there was only 1 paper with pure Nb₂C data. That “agrees” had been leaning on composite values. Writing only what improved would make the map look better than it is. So we wrote down both.

⑤ The real evidence comes from new data. The changed rule is now frozen, and we will test again with papers newly obtained from here on. If it passes then, that will be the evidence that “the map is useful.”

Electrical conductivity, EMI shielding (how well it blocks electromagnetic waves), batteries, interlayer spacing. Even as the property changed, the failure points were similar. The numbers mostly matched the originals. What was wrong were the fields describing the numbers, for example “should the substrate film underneath also count as a mixed-in material?” or “is the sample wet or dry?”¹

Interlayer spacing: failing, then measuring again

Interlayer spacing is the gap between one MXene sheet and the next. In the first scoring, a row counted as “correct” only if it also got “was the sample wet or dry?” right. But papers rarely reported this state. The result was ● 45%, a fail. Most of the ✕ came from exactly this state field. Rows in which the number itself was wrong were only 2.2%.²

Calculating after seeing the results gave “78% if only the state field is excluded.” A tempting number. But it is a yardstick chosen after seeing the results. This number was not used on the map.

Instead, we started over. A new criterion, “exclude the state; check only value, unit, composition and sample,” was saved before scoring, and the two teachers re-scored the same rows. The result was ● 90%. This time it passed. The rows went on the map with a “state not verified” mark.³

Thermal conductivity: failed three times, passed on the fourth

Thermal conductivity is how well heat is carried. It is used to remove heat from devices. This property taught us the most.⁴

First. ● 66.7%, fail. The cause was not the numbers but the instructions we wrote. For the field recording the direction of heat flow, we allowed only two words: “parallel to the sheets” and “across the sheets.” But the papers used other words, such as “horizontal,” “vertical” and “axial.” The transcribing AI left the field blank, and the judges gave ✕, saying “it is stated in the paper but was left blank.” There was also a row in which the unit was forgotten, writing 550 mW as 550. That was our mistake.

Second. Only the instructions were fixed and re-saved, and the same papers were transcribed again. ● 81.2%, above the pass line. But ✕ was 18.8%, above the 10% ceiling, so another fail. This time all sixteen ✕ were blanks. Not a single number was wrong.

Third. We created a new mark on the scoring side, counting “wrong (✕)” and “omitted (◐ — present in the paper, but the field is blank)” separately. Into the traps we also put four “fields deliberately blanked although an answer exists.” Both teachers correctly called all four “omitted” rather than “wrong.” This was the first confirmation that the two could be told apart.

The third result was: ● 62, ◐ 22, ✕ 0. Correct plus omitted is 98.8%. Of 85 rows, not one number was wrong. Yet it is a fail, because omissions were 26%, above the 20% ceiling set before scoring.

The numbers were right in all 85 rows. The only reason for failing: the transcription was not thorough enough.

This ceiling was a safeguard placed on purpose. Counting “omitted” separately can make the scoring lenient: even with many blanks, it is easy to wave them through with “well, it isn't wrong.” So we decided in advance that “if blanks exceed 20%, the transcription is deficient.” And that safeguard actually tripped. We did not lower the ceiling.⁵

What needed fixing was not the scoring but the transcription. We added the instruction: “Before leaving a field blank, search once more in the experimental-methods section and the figure captions.” But we did not rewrite the scores of the finished round. We opened a new round and re-transcribed the same 35 papers from the beginning.

Fourth. ● 70 · ◐ 12 · ○ 2 · ✕ 1, 85 rows in total. Correct plus omitted 96.5%, wrong 1.2%, omitted 14.1%. This time all three conditions were met: pass. Both judges also caught every trap.⁵

However, we do not say “the new instruction reduced omissions.” This time the transcription was redone, and the judges' instructions also changed slightly. We cannot separate what produced the effect. The purpose of this round was to fill the map with more solid values. The thermal-conductivity points on the map were replaced with the 82 points from this passing round.

When both omit the same thing, no one notices

Something similar happened with the battery property. Two transcribing AIs read the same papers independently and extracted values, and there were conditions both had left blank, conditions clearly stated in the experimental-methods section of the original. Two answers being “the same” cannot catch that both “omitted the same thing.” Those blanks surfaced only in the scoring against the original.⁶

And we do not keep rerunning until it passes. We decided not to revise the same question paper more than three times. If you keep revising and rerunning until it works, that becomes not measuring but “fitting.”⁷

The first comparison with a human-made answer key

Until now the test was “do two AI judges say the value matches the source?” That alone does not tell how often we are right by a human standard. So we ran the same method on a public answer key that another group built by hand: 600 text passages from perovskite solar-cell papers, with band-gap values marked by people. It is not MXene, but it can still test us against human answers.

Recall 97.4% — almost every value the people recorded, we found. Precision 83.1% — a fair number of our rows were not in their key. Judging those 45 rows again, both AI judges found 42 of them stated in the source. Opening them, some looked like values the key’s rules leave out, and some like our own wrong picks.

AI judges are good at “is this written in the paper?” Whether a value counts as the material’s own value is a human definition — two AIs cannot settle it for us.

Four recipe steps — precursor, etching, delamination, processing

For each paper we took one representative sample and extracted, with source sentences, how the precursor (MAX phase) was obtained, how it was etched, how the layers were opened, and how the measured form (film, electrode, powder) was made. As with values, two judges check each step against the paper, and only steps both accept appear on the map.

Two failed rounds, then people rebuilt the test sheet

Every judging round hides trap items whose answer we know (a changed number, text from another paper, a missing step). The first two rounds were void — not because the judges failed, but because the traps we generated by machine mixed “right” and “wrong” items. In the third round, the traps that are hard to generate were picked by hand, opening each source. Both judges caught all 32 traps, and 88% of real steps passed.

What looked “missing” was in the next box

Processing was the weakest step (74%). Reading the failures, many said “the heat treatment is missing.” But heat treatment had been extracted into a separate “post-treatment” box that we never sent to the judges. Asking again with processing and post-treatment joined, the same papers went from 33% to 91.5%; the round as a whole passed at 92%. (The definition of “processing” given to the judges also changed by one sentence, so the gain cannot be split between the two.)

The recipe belonged to another sample — 20–21 October

The day after release, while another AI reviewed a text we were about to send out, an odd row turned up: clicking the value of a gel made at pH 10 showed the recipe of the pH 4 gel. Counting, 1,767 of the 3,962 values with a recipe had a sample name different from the recipe’s sample. We opened 30 random pairs; 26 were truly different samples — composites with the opposite mixing ratio, undelaminated multilayer samples, even a different MXene composition.¹

The cause was ours. When extracting recipes we had taken several samples per paper, yet the map used only the first sample’s recipe for every value in that paper. The value matched the paper and the recipe matched the paper, but nobody had checked that the two belonged to the same sample.

The fix was not to extract again. For 1,384 of the mismatched values, that sample’s recipe had already been extracted. Both judges judged those steps again (each caught all 40 traps; 91.3% of real steps passed). 1,300 values now show their own sample’s recipe. We also made the name-matching rule stricter (to stop a pure MXene value picking up the recipe of a composite whose name contains it); counted with that rule, “another sample” values fell from 2,213 to 913. We then judged 182 newly extracted samples and 179 samples found again under the strict rule in the same way (in both rounds both judges caught all 40 traps; 93.6% and 88.9% of real steps passed). For 185 of the remaining values we extracted that sample’s recipe anew (102 samples) and judged it the same way (both judges caught all 40 traps; 91.6% of real steps passed). 2,110 values now show their own sample’s recipe, and 103 “another sample” values remain.²

A correct value and a correct recipe can still belong to different samples; that has to be checked on its own. For the next material we name the samples first and attach values and recipes to them.

This was not only our problem. When solid-state synthesis recipes text-mined by a Materials Project team (Kononova et al., 2019) were checked field by field against the papers on a random sample of 100, the chemistry (precursors, targets, reactions) was right 93% of the time, but every field including the conditions was right only 51% of the time. «What it was made from» is far easier than «under which conditions».

The map now

Of 4,097 values (768 papers), 3,962 now carry a recipe (756 papers). For 3,859 of them the value’s sample and the recipe’s sample have matching names; the rest are labelled “Recipe of another sample in this paper — not confirmed to be this value’s sample.” On first visit the map opens a “one question, three clicks” guide; press “Show me” and it walks through it by itself.

What we must still state plainly

• Some values still have no matching-sample recipe. They are labelled “another sample in this paper,” and that recipe may not be this value’s.
• Sample names are matched as text. A sample the paper calls by two names can show as “not matching,” and a different sample with a similar name as “matching.”
• “Matches the paper” does not mean the paper is right.
• No value has been checked by a human expert yet.
• On 20 fixed real questions, the share answered down to the recipe with source sentences is 81%; counting only cases where the recipe’s sample matches the value’s sample, it was 79% (19 October). Re-measured after the sample-linking fix, both were 80.6% (early 21 October); after more values switched to their own sample’s recipe both are still 80.6% (morning of 21 October — a brief 79.8% came from our scorer cutting a sample name at 60 characters, failing to find its map point and picking another sample from the same paper; fixed and re-measured).

1. Fixing only from the inside loses the way — look outside

The judges were poor at catching scope errors (full-cell values, compositions outside the card, values of a different material). We added a sentence to the judging instructions telling them to check scope. On our own test it looked better. On items it had never seen, two got worse: the new sentence collided with the MOF rule "record derived materials with a label".

The answer was outside. Other studies had already reported the same thing: rule files grow by appending rather than editing, contradictions follow, and AI models do not reliably follow "this rule takes priority" wording. So we moved from words to code that labels rows from fields they already carry (device, derived material, composition).

We also brought in answer keys from outside. Our test answers had only ever been checked by us against the source. With an outside hand-made key for magnetic nanoparticles (ChemX), the key itself mixed in values read from graphs, so it could not serve as a ruler. With an alloy key that records its evidence (LitXAlloy, 19 papers), 66 of our 67 yield-strength values matched — a first sign that the engine holds in a field it had never seen. The outside key even found a hole in our own number checks (a unit-reading bug that wrongly blocked 49 correct values).

A test we wrote ourselves only shows what we fixed. Direction comes from outside research and outside answer keys.

2. A new rule for every problem lowers efficiency

Several stop-gates appeared in a single day: stop if a common-core rule is missing, stop if a definition changes without a rename map, rules for splitting the test set. Moving to new materials (magnetic, alloys), the MOF-specific trap code no longer fit, and we copied it in just to pass the check — matching on the surface only.

So we separated two things. Kept firm: quality — no wrong value slips onto the map silently; a new engine version is accepted only if it improves on unseen items in a paired comparison; sealed-test numbers are never hidden. Kept flexible: procedure — a check that does not fit a material can be passed with a written reason. And we stopped appending sentences to the judging instructions; when something needs fixing, we edit the existing sentence.

Rules serve the purpose. Block firmly only what must be blocked; label and record the rest.

3. Other lessons

• Do not re-measure on the problems you fixed. The test set is now split into "practice" and "validation". Looking at practice alone, we would have accepted a version that did not help.
• Changing a definition quietly ages old answers. Splitting "pore volume" into total and micropore volume turned answers with the old name into wrong answers. A rename map is now written before results are seen.
• Errors found must reach the map. Rows our source check marked wrong were used to score the round but stayed on the map — values like 37,500 m²/g. 85 rows were removed.
• Open the source. An "impossible" 36,125 mAh/g was a Li–O₂ value normalized per catalyst mass; a single upper limit should not block it.
• The engine is a product. It grows on a planned track (specification, version numbers, sealed score sheet, next-version goals) and a field track (problems found while building the map).

4. Re-measured with the new version, and learned the limit of word rules (31 October – 1 November)

A week later we re-measured the MXene capacitance papers (251) from scratch with the new engine version. Both judges supported 1,407 of 1,493 values, and none of 20 rows checked against the source was wrong. This time the definition changed: each value now records «what mass it is divided by (active material · electrode · device)» and «measured in a 3-electrode cell or in a device». As learned above, the old-to-new mapping was written before the results; old values were kept and only the new fields were added. Re-reading the sources showed that 2 old values belonged to a different sample (a «respectively» pairing error); they now carry a warning.

The next morning we tried to attach tags such as «MOF with deliberately created defects» or «value after cycling» using word rules. The rules were written before testing, and all four fell short of the threshold (33–60% correct). Words only tell us «this sentence contains the word defect»; they cannot tell «which sample, measured when, this number belongs to». So the tags were not published, and that information will instead be extracted as fields and checked by both judges. The first such field is the one the specialist named: mass loading (Chapter 16).

A rule that counts words does not know which sample a value belongs to. Only the step that reads the source can answer that.

Evidence: test-set results v1.9 · v1.10 · hard set · outside keys X1 (ChemX) and X2 (LitXAlloy) · a review of five outside papers on rule conflicts — kept in our internal records.

The first map started with one property and one MXene, Ti₃C₂: gravimetric capacitance. We collected 351 values from 81 papers. After that we added properties one at a time, and also MXenes other than titanium ones. When work function was added, the map's points (one value = one point) reached 1,952. How it grew further on 2026-10-09 is told in Chapter 10.¹

The cards below explain the ten properties in plain words, with the scoring results attached. A green band means pass, red means fail, and grey means the sample was too small to judge.

How to read the numbers — with some care

A few interesting numbers show up. The median below is the value in the very middle when the values are lined up by size. However, every one of them was computed by mixing values measured under different conditions, so we can only say “this is roughly what is typically reported.” They are not values to compare against each other to say “which is better.”²

• Work function. The median of 40 rows of Ti₃C₂ not mixed with other materials is about 4.53 eV. But the range is 2.85–6.0 eV, spreading by more than 3 eV depending on the surface treatment. This very spread is the raw material for answering “made which way, what is the work function?”
• Surface area. MXene not mixed with anything had a median of about 12.9 m²/g, and composites mixed with other materials about 43.5 m²/g. That does not mean “making a composite gives four times more.” Measurement temperature, gas and synthesis method all differed. The morphology field was also empty in as many as 86 of 152 rows.
• Photothermal conversion efficiency. Median about 39% under 808 nm light, about 42% under 1064 nm light. Concentration, light intensity and the calculation formula differed from paper to paper.
• Gravimetric capacitance. It also contains unusually large values such as 4,000–12,000 F/g. These are values converted from measurements of a very small spot on a single sheet. On the map they are recorded exactly as written in the original.

Do not read the number of points on the map as the amount of research.

The number of points must also be read with care. In the round that collected non-titanium MXenes, 62 rows of Nb₂C and 36 rows of Mo₁.₃₃C came in. But 84 of those rows came from just two papers. Many points do not mean “niobium and molybdenum MXenes have been studied a lot.”³

Still, considering that the first map started with “Ti₃C₂ alone,” the fact that the tiles in the figure above now sit together on one map is itself progress. MXenes mixing several metals, and MXenes using carbon and nitrogen together, also came in with their original sentences attached.⁴

Up to Chapter 6, most of the map's points were titanium (Ti) MXenes: 82.3% of all points. Only one side of the map was dense. So on 2026-10-09 the effort went into filling in MXenes of other metals such as niobium, vanadium and molybdenum.¹

The first batch — failed, but not one number was wrong

First we opened 27 not-yet-used papers among the open-access papers already retrieved by the program. Excluding 12 (papers that only calculate, papers without the desired properties, and so on) left 15. From these, 96 rows were transcribed.

Scoring result: ● 81, ✕ 15. ✕ was 15.6%, above the 10% ceiling, so not passed. But on opening the 15 ✕ rows one by one, the number of rows with a wrong number was 0. All were label problems: “which property is this value of?”

• 6 rows — the length of two sheets' worth (the c lattice parameter) was put in the “spacing between sheets” field. The instructions given to the transcribing AI said “allowed,” but the instructions given to the judges said nothing about it. Our two sets of instructions differed. Both judges said “wrong.”
• 9 rows — supercapacitor (a capacitor that charges very fast) values were put in the “battery capacity” field. We had not created a suitable field for those values. Only one judge said “wrong.”

Still, the verdict was not changed. One can calculate “excluding these two causes, all 81 rows are correct,” but it is a number made after seeing the results, so it is not used as evidence (the envelope rule of Chapter 5). On the map, 80 points went up with the mark “round not passed · both judges say this row is correct.” And one rule was fixed: the term definitions given to the judges are copied verbatim from the sentences given to the transcribing AI.²

The second batch — papers obtained separately

There were 137 papers that could not be retrieved because publishers blocked the program. Of these, 56 were obtained separately.

Papers obtained separately got no special treatment. They were screened, transcribed, mixed with traps and scored in exactly the same way. This time two new “property-name change” traps were added. Remaining: 29 papers, 253 rows.

Result: ● 253/253. Both ✕ and ◐ were 0. As said in Chapter 4, when it is too clean we suspect first. So we checked the judges again first. Both judges caught all 18 traps, including the newly added “property-name change.” They were not simply calling everything “correct.” However, why this differs so much from the first batch cannot be separated out: the instructions had been fixed, and the batch of papers itself was different.³

“Correct (●)” means “the original was transcribed correctly.” It does not mean “the original is right.”Example: a value whose unit looks odd (62.5 A h g⁻¹) was also “correct,” because it appears exactly so in the original. This value was not in a field that goes on the map, so it was not added.

From this batch, 234 points were added to the map. There are still gaps. 77 paywalled papers have not yet been obtained. Papers that could be obtained separately are mostly “free to read” ones, so the set may be skewed in that direction.

On the same day, thermal conductivity also passed in its fourth round, as noted at the end of Chapter 6.⁴

17 “candidates for a second look”

We swept the whole map once with an eye to “the usual range.” 17 points came up that are exactly as in the original but far outside the common range: an interlayer spacing of 0.354 nm (usually around 1 nm), a work function of 1.7 eV, an electrical conductivity of 34,520 S/cm, a gravimetric capacitance of 4,000–12,000 F/g, and the like. By category: interlayer spacing 9, two-sheet length (c lattice parameter) 3, capacitance 2, battery 1, electrical conductivity 1, work function 1.

This does not mean they are wrong. They may be values measured under very special conditions. For example, 4,000–12,000 F/g is a value converted from a measurement of a very small spot on a single sheet (Chapter 9). So instead of deleting them, we only attached a “candidate for a second look” mark. And these points were excluded from the test below.⁵

Testing the map — 20 known answers

When you get a new navigation system, you first try a route you already know. If it suggests a wrong route, you cannot trust it for routes you don't know either. The map was tested the same way.

From papers already established in this field, we picked 20 “known facts,” for example “etching with hydrochloric acid and lithium fluoride gives a wider interlayer spacing than etching with hydrofluoric acid.” They were chosen without looking at the map and checked down to the page in the original. The rules for comparing them with the map were also saved before seeing the map's numbers. And the test was run exactly once.

Of the 7 it could answer, 5 agreed with the known facts. The median interlayer spacing on the HCl/LiF-etched side was 13.7 Å, and on the HF-etched side 12.1 Å, exactly as known. The median electrical conductivity of the titanium and vanadium families was 3,612 S/cm, and of the niobium family 31 S/cm, also in the known direction.⁶

1 disagreed. It asked how strongly a “film thin enough to be transparent” resists electricity (sheet resistance), but our map could not separate transparent thin films from thick films, so the two were mixed. It seems the conditions did not match, rather than the map's numbers being wrong. Even so, we did not fix it after seeing the results. 1 was inconclusive, because our map put “polymers added as binders” and “materials mixed in to add function” into the same field.

The remaining 13 could not be answered at all. A comparison needs papers on both sides, but one side had only 0–3 papers. For example, for the capacitance of MXene etched with sodium hydroxide (NaOH) instead of hydrofluoric acid, the map has 0 papers.

The 13 unanswered are not failures. They are the list of what to fill next.

What this test showed is clear. The map's first weakness right now is not “wrong values” but “thin data for comparison.” Even the 5 that agreed go only as far as “no contradiction,” not “proven.” Conditions differ from paper to paper, and 7 is a small number. So showing the map to a leading authority in this field has been postponed. After filling the gaps, it will be tested again with the same 20 questions and the same rules. How far the 13 drops is the first yardstick for “has the map become useful?”

Perovskites — seven rounds, and «30 papers wobble»

We compared our extraction with a public answer key (the Perovskite Database · values copied by hand from papers · CC BY 4.0), drawing 30 new papers per round. In the rounds where we changed the instructions, the «how it was made» score passed the 80% bar (83.4%, 87.4%); re-measured on 30 fresh papers with the instructions unchanged, it fell to 79.1% and 71.8%.

At this size we could not tell whether a score rose because one line of the instructions was better or because those 30 papers were easier. Pooled over four rounds: 81.2% with the first rules, 79.3% with the revised rules — no visible effect. So we closed the rounds without writing «passed».

A single 30-paper round swings by about 15 points. A rise inside that swing does not show that a change helped.

What we learned went into the tool that opens the next field: a round does not start unless there is a document saying how the field’s answer key was recorded, and with fewer than 60 papers it runs only as a «test round». It is a place where the tool stops, not a promise someone has to remember.

Sensors — the map’s first application property

For the first time we collected values of «what it is used for» rather than properties of the material itself. A 30-paper test round first: of 76 lines, 97.4% were found in the source by both AI judges. The two misses sat on definition boundaries — a fluorescence detection limit filed as «electrochemical», a recovery time written as response time — so before the main round we added two definitions (response time = rise time only · optically measured detection limits kept separate).

Main round, 60 papers: 175 lines from 48 papers. Both judges caught all 18 planted traps; 92.9% of the real lines were supported by both, 3.5% wrong — passed. The map gained six properties (pressure sensitivity · gauge factor · response time · gas response · gas detection limit · electrochemical detection limit), 214 values from 62 papers. Papers mix ppm and ppb, μM and nM, s and ms; the map shows one unit and keeps the source value and unit alongside.

Our own mistakes — caught before release, and one after

The unit code got two things wrong: it missed the Unicode minus «−», so 8 negative values came in as positive, and it read «10⁻¹⁴ M» as «10 M» (one character turned into another while code went through the shell — a trap already written in our rules). Both were caught before release by scanning value ranges.

That night, while testing a new «compare list» in the 3D explorer, we found one more: 7 older conductivity values had lost the «×10⁻⁵» of «4.77×10⁻⁵» and were shown as 4.77 (already public). We fixed them in the map build and logged it in the map’s «🕒 How it grew» for that day.

The two judges check that a value matches its source sentence. Turning that value into the number on screen is our code, which the judges do not see — so we scan value ranges once more before release.

Next — desalination, then MOF·COF

• Desalination — of 57 open-access MXene desalination papers, all 34 not yet on the map are running as one round (CDI · solar evaporation · membranes). It is small, so results will be stated as «in these 34 papers», not «in desalination».
• MXene + MOF·COF composites — after that. A map of MOF·COF themselves will be opened with checks from experts in that field.

Two things finished first — desalination and composites

We finished the two rounds named at the end of chapter 11. All 34 desalination papers ran as one round, and both judges supported 96.9%. The 34 papers on MXene mixed with MOFs or COFs passed at 95.3%, adding 61 values. These composite values became the bridge to the new islands.

The MOF island — same name, a hundredfold apart

We collected MOF BET surface areas in four rounds. All four passed (91–99%), giving 856 passed rows. But for the same MOF name, values differed tens to hundreds of times between papers.

When we looked again at the name groups, a large part turned out to be «different materials sharing a name». MIL-53 split into an aluminium form and an iron form (×244 → ×4.7 and ×7.5), and MOF-74 is a different material for each metal. ZIF-8 had another crystal phase, ZIF-L, mixed in (×152 → ×43). The spread that remains comes from samples and conditions, such as samples soaked in buffer or different synthesis times.

A number can match its source sentence while the name still points to a different material. Name groups need checking too.

Publishing without waiting for the experts

We asked experts in the field two questions, and the answers have not come yet: whether a «surface area» with no model name counts as BET, and where to place a framework that links metal units by imine bonds. We published the islands anyway. The defaults we chose are written on the page, and the values are marked «passed AI judging · not yet expert-checked · may change». When answers arrive we will change them and record the date and the reason. On the first-screen 3D view they are galaxies far from MXene.

The second MOF property — CO₂ uptake failed once

The CO₂ uptake round (770 rows) failed, with 53% supported and 45% wrong. The cause was one line of the judging instructions, not the values. The line said «isosteric heat, selectivity → null (recorded but not mapped)». One judge read it as «not put in the map column»; the other read it as «the property field must be empty». As a result, 283 correctly recorded rows were all marked wrong.

Both judges scored almost perfectly on the planted traps. Traps test «is the value and sample right», not «did both judges read the instructions the same way». So a perfect trap score could not, in principle, catch this flaw.

We added one clarifying paragraph to that line and re-ran only the judging, registered before seeing the results. It passed with 89.5% supported and 8.6% wrong, and judge disagreement fell from 45% to 6.4%. We recorded it as «passed after fixing one line of the instructions». The island now shows 370 CO₂ uptake values from 65 papers, grouped by condition.

Taken alone, the «CO₂ uptake» rows were 12.5% wrong, and the page says so. When we counted those 61 rows, 39 were not static isotherm values (thermogravimetry, flowing gas, time-dependent) and 16 were per volume. This was a gap in the extraction definition, and one of the two judges caught it.

When the two judges split widely, suspect the instructions, not the values.

What the engine learned — two upgrades

First (the same evening). We added three things.

• The word «null» now says in plain words what it empties.
• When the judges split on more than 20% of rows, scoring raises «instruction wording suspect». Tested on past judgement files, it rang on the failed round (42.9%) and stayed quiet on normal rounds.
• A new trap, «next column»: a value swapped for the number in the neighbouring column of the same table row.

The second COF round was the first to run on this engine, and it passed (96.7%). Both judges caught both next-column traps.

Second (that night). In the 20 rows sampled after the second COF round passed, 3 were values given only as «surface area», with no «BET» in the sentence. Our default does not count these as BET, but that rule was not in the instructions; it was only being caught in the sample check.

So we put the rule into the definitions. In the next round's extraction, 13 rows were set apart as «surface area, model not stated». The same upgrade fixed four more things:

• two measurements at different temperatures were being merged as duplicates
• a trap was picking a «±» error value
• units were compared only by their first two characters
• control approvals and judging batches now carry fingerprints, so scoring stops if they change midway

Links inside the galaxies

At first most stars (papers) on the islands floated alone, and COF had no name groups at all. The MXene map looks connected because it has a linking rule: same material, same property column. We built such rules for the neighbouring families.

• COF name groups are set by the two molecules a framework is made from (TpPa-1 · TpBD · TAPB-DMTP …), and cousins built from the same molecule are linked.
• For MOFs, we normalised name spellings (six kinds of hyphen, letter case) and re-matched the groups. We also linked «cousins with different metals» that share a linker (BDC · BTC …).

The rule was «better unlinked than wrongly linked». While testing, we caught and fixed a mistake that treated pyridyl porphyrin as the same molecule as carboxyphenyl porphyrin. Lone stars fell from 130 to 74 for COF and from 136 to 102 for MOF. The rest cannot be linked from the name alone; the next rounds need a «made from» extraction field.

Finished that night · still open

• Third COF round — 120 papers, widened to include papers without «BET» in the abstract. It is the first real run of the second upgrade.
• The third COF round passed (96.7% · COF island 466 values · 178 papers). With the new rule, extraction set apart 13 rows as «surface area, model not stated». Two borderline rows, with BET only in a neighbouring sentence of the same paper, are recorded as «verdict withheld».
• MOF «same-sample» pairs — we extracted the BET of the same samples in the 65 CO₂ papers (94.5% passed judging). The pairing rule (same DOI and the sample name identical to the letter) was fixed before seeing results. 142 pairs from 31 papers resulted. At room temperature and 1 bar the relation between surface area and CO₂ uptake was weak (rank correlation 0.13 · 66 dots · 19 papers · dots from the same paper cluster, so this is a shape, not evidence). The island page shows them as scatter plots per condition.
• A third engine upgrade — four fixes from that night went into the next round: BET computed from another gas kept apart · pore volume split into total and micropore · when the sentence lacks «BET», the supporting sentence must be quoted · a «made from» field. In its first real run (fourth COF round · 98.3% passed · COF island 567 values · 222 papers) three of them had almost nothing to act on, so their effect could not be measured. The fourth field stayed empty because the output schema blocked it — our mistake of testing only the instruction text, not the real output path; after fixing it we re-tested one paper through the real path. «Registered and working» is not the same as «effective», so this upgrade has only reached the first part. The first signal came from 120 MOF CO₂ papers, where such rows are common (99.4% passed · MOF BET 1,135 values): in 100 rows checked against the source, «CO₂-derived BET recorded as BET» went 3→0 and «micropore volume recorded as total» 6→0. But splitting the field created a new reverse mistake, «total recorded as micropore», in 3 rows. One round only, so we say «fell», nothing more.
• When the engine changes, old results must move with it. The same day we first ran a fixed cross-material test (140 correct values · 48 deliberately altered ones). The judges refused 18 MOF/COF values we had checked as correct. It was not an engine error: that very upgrade had split «pore volume» into «total pore volume» and «micropore volume», so answers carrying the old label had quietly become wrong. The worry was «losing old rules while learning a new material»; the opposite — «a new rule silently discarding old results» — turned out to be real too. Now the engine itself blocks both: a round will not start if any of the 11 common-core rules is missing, or if a definition was changed without an «old name → new name» map written before seeing results. With the definition-changed items set aside, values accepted as correct were 97–100% per material. Catching altered values was 48/48 — a ceiling — so the next test will be rebuilt from rows that were actually wrong. Running that «hard test» (105 rows that were actually wrong) showed the judges check «is this number in the sentence» well but rarely «does this value belong in this table» (per-area units, full cells, out of scope) — 2/35 for MXene. The control review had been catching those — and then we found something worse: rows the control review marked wrong were never removed from the map. An impossible surface area of 37,500 m²/g was on the islands. We removed 85 rows (MOF BET 1,084 · COF 533), and every map build now drops control-review «wrong» rows.
• Human expert checks are still zero, so we do not use the word «verified».

1. The engine still runs on another material

We took the frame of the MXene hydrogen-evolution (HER) round and changed only the «which material» layer. We wrote down that amorphous MoSₓ, MoO₃ and MoP are not MoS₂, and that Mo₂CTₓ is an MXene even though its name contains Mo. From 76 open papers we extracted 459 values; both judges supported 96.3% of them, and all 20 rows in the post-pass sample matched the source. The three extraction mistakes we found while reading the controls (amorphous MoS₃ written as MoS₂, a value measured with carbon nanotubes attached to the plain sample, and a potential sentence read as an overpotential) were also filtered out by the judges.

2. Whose story is a bridge?

The MXene map already held many values measured on composites of MXene and another material. At first we planned to open islands for the materials with the most of these «bridges». Tae-Eun Song, who makes this map, pointed out that those bridges come from MXene papers: MXene is the protagonist and the partner is a helper, so they cannot be presented as the partner's own results. An outside view (Codex) gave the same answer: a value belongs to the system that was measured. So an island starts from the material's own papers, and bridges only connect islands. We borrowed the words «active material · support · conductive additive» that battery (BattINFO) and catalysis (NOMAD) schemas already use.

Before moving a number, ask who did the work.

3. Protagonist and helper — a mirror image

First we checked whether papers state roles at all. Two judges read four groups (MXene catalysis · MXene energy storage · MoS₂ catalysis · MXene HER) without seeing each other's answers. Both judges found a stated role for 82–94% of the samples and agreed on «protagonist or not» for 91–97%. So roles became a common field of the map.

The result came out as a mirror image. In catalytic composites, MXene was the protagonist in only about one sample in ten; mostly it was an electron path or a support. In MoS₂ composites, MoS₂ was the protagonist in nearly eight in ten, and the partner (graphene and others) was the electron path. In energy storage, MXene was its own protagonist far more often. A role does not belong to a material; it changes with the sample and the property.

4. What one common scale showed

We compared the HER overpotential of MXene and MoS₂ (at 10 mA cm⁻², lower is better) on the same definition. With all values, MXene composites looked better than MoS₂ at 104 mV. But those values were mostly the performance of other catalysts (Pt and others) that MXene supports. Comparing protagonists only, MXene as the catalyst had medians of 463 mV in acid and 227 mV in alkaline; MoS₂ as the catalyst had 250 and 198 mV. The acid MXene cell has only 8 values and conditions are not matched. Still, one thing was clear: judging a material by a mixed value praises the wrong one.

5. Other things we learned

• Tools are wrong too. The recipe table counted catalytic overpotential as «higher is better», and the public page showed it that way for a day. We fixed the list of lower-is-better properties.
• Pieces that don't fit are set aside, not thrown away. That morning we put the «role» piece in a ledger; it fitted the same evening. We don't cling to pieces that won't fit, but we don't forget them either.
• Look for answers that already exist outside first. When stuck, we first listened to standards from other fields and to a different model's objections, and corrected our plan there.
• Next. We will upgrade the engine to extract roles together with values. The next island is graphene: the name covers single layers, few layers, reduced graphene oxide and graphite flakes, so it opens as an island that judges «what this sample really is».

Judge a material by a mixed value, and you praise the wrong one.

Evidence: T1 MoS₂ HER round · role readings R1, R1b, R1c, R1d (two judges) · direction note (roles and bridges) · outside view (Codex) · ledger of set-aside pieces — in the internal records.

1. One name, many materials

"Graphene" is used for a single carbon sheet, a few-layer flake, reduced graphene oxide (rGO), graphene oxide (GO) and nitrogen-doped graphene. So this island gives every value an «identity» column: what the paper calls the tested sample, checked against the source by two judges. Carbon nanotubes, carbon black, graphene quantum dots and generic carbon shells were written down in advance as not graphene. Before seeing any results we asked Codex for holes in these rules and fixed eleven. One of them was a contradiction: quantum dots were excluded in one line and listed as allowed in another.

From 223 open papers we drew 80 at random and extracted 403 values. Both judges supported 87.8%, and all 20 rows in the post-pass sample matched the source. The judges filtered out a row that called graphene quantum dots graphene and a row that called «nitrogen-doped carbon» nitrogen-doped graphene. They missed two things: three rows that wrote nitrogen-doped rGO (NrGO) as nitrogen-doped graphene, and one row that gave a value measured after ten thousand cycles as if it belonged to the fresh sample. Both are now examples in the next round's instructions.

The same name does not mean the same material. Match the identity first.

2. Two lines I got wrong

While reading the controls I marked one paper as having «no graphene at all», yet both judges supported its values. On a second look the paper used a «graphene-based conductive filament»; my search had missed it. I also read «GC» as glassy carbon in a paper that defines it as a «graphene catalyst support». Both rows were among those I had removed from the controls, so the score did not change. Still, the lesson came back: before reporting that the judges were wrong, reopen my own side first.

3. Graphene, too, was mostly the helper

For 114 composites, two judges separately read what the paper says graphene does. Both judged that the paper stated it in 87% of samples, and they agreed on «protagonist or not» in 97%. Graphene itself was the active site in only 9 composites; in the rest it was a support or a conductive path. That is the same shape as MXene and the opposite of MoS₂. Catalysts made only of graphene-family carbon had a median acid overpotential of 536 mV, weaker even than pure MXene (491) measured with the same definition. The good graphene-composite values were mostly the scores of other catalysts that graphene supports.

4. A number measures more than the material — the electrode support

A talk on two-dimensional materials we listened to that day made one point: for 2D materials, the substrate a value was measured on is part of the value; change the substrate and the band gap can shift by more than half. The talk was about optical and electronic properties, but hydrogen evolution has the same spot: the electrode the catalyst is loaded on. Nickel foam produces hydrogen by itself.

Splitting the three islands' values by support showed large differences. 41% of MXene overpotential values were measured on nickel foam, against 6% for MoS₂ and 2% for graphene. A gap in the comparison table that looked like «MXene is better in alkaline» mixed material differences with support differences. The support detector was tested on 12 random papers: 11 right, one caught «Ni Foam» in a reference title. The analysis was written on the card before the results, and the passing criteria were not changed. From the next electrochemical round on, «electrode support» and «state after use» are added to the instructions automatically.

A number does not measure only the material. It also measures where the material sits.

5. Judging without seeing each other

Tae-Eun Song, who makes this map, asked a question: if the second judge can see the first judge's results, won't their agreement rise by itself? We checked and found one side closed and one side open. One judge cannot open any file at all. The other ran in a temporary folder but could, in principle, read the first judge's results already sitting in the repository, and its log did not record which files it opened. So «it did not look» could not be proven.

There were signs against copying: only 70% of samples got exactly the same role name, the distributions differed, and agreement was in the same range as earlier rounds. We changed the method anyway. Both judges now start at the same time. Results stay sealed outside the repository until both finish. The full run log is kept, and code checks whether either judge opened the other's results or the repository. We also tested that this check can catch something, starting with whether the log records the file names a judge opens.

Two judges agreeing is evidence only if they could not see each other.

6. Other lessons

• Check every place that shows the data. After «I can't see the TMD links» I fixed one screen and replied that the lines were there. The person meant the front page. The same data has to be checked on every screen that draws it.
• Adding something new makes old text awkward. Adding an island left seven pieces of text that assumed «MOF and COF only»: buttons, a summary line, the «new» box, audit numbers and the method note. When an island opens, those lines are fixed with it.
• Next. The phosphorene (black phosphorus) island: battery capacity on the same definition as the MXene map, with layer number, substrate and mass basis as columns from the start.

Sources: the G1 graphene HER round (card results and extra analysis) · role reading R1e (two judges) · electrode-support analysis (substrate_split) · sealed judging tools (blind_roles · run_round judge) — in the internal records.

1. What the person decided — where a machine cannot answer

AI is fast and good at reading a paper and copying a number. It could not decide what the work is for, how far to go, or what comes first. A few decisions by the person making this map (Tae-Eun Song) changed its direction.

• Put it out first. With no human check at all, publishing felt premature. «I won't hesitate» settled it: the map went public on 11 October labelled «passed AI judging», not «verified». Questions and answers from outside researchers started after that.
• Whose value is it. When the plan was to open the neighbour with the most composite «bridges» first, the person pointed out that those bridges came from MXene papers — they are not that material's own results (Chapter 13).
• The size of an island. The next material is chosen in proportion to how much literature it has, not how many links it has. Graphene has roughly ten times as many papers as MXene.
• Judging without seeing each other. «If the second judge can see the first judge's answers, won't agreement be inflated?» That question produced sealed, simultaneous judging (Chapter 14).
• What to ask a person. We stopped asking experts to check numbers one by one — two AIs already copy numbers well. Instead we ask rules and common sense only a person can answer, such as «would you count this as an MXene value?»

The AI does the «how» quickly. The person decided «what for» and «how far».

2. What the AI got wrong — and what caught it

This book promises to keep its mistakes on the page. Here are mistakes made by the AI that builds the map, each with what caught it.

• Counting the same paper twice. Ten papers came in under two different IDs and five values overlapped. A duplicate check published by another research team found it (map 4,102 → 4,097).
• A scorer that cut text. Sample names were cut at 60 characters before comparing, so correct values were scored as wrong. Re-measured, 79.8% became 80.6%.
• Good and bad, reversed. The recipe table treated catalyst overpotential as «higher is better», and the public page showed it that way for a day (Chapter 13).
• My own test sheet was wrong. One of the trap questions I made to test the judges was itself wrong, so one round was left as «cannot judge». Before saying «the judge was wrong», I had to reopen my own side first (Chapter 14).
• What both judges missed. A paper wrote only «d-spacing of 1.17 and 1.21», with no unit. Both judges passed it and the map showed 1.17 Å. Working from the same paper's X-ray peak (7.55°) gives 11.7 Å — that is, 1.17 nm. Both also missed a lithium–oxygen cathode capacity (36,125 mAh/g) mixed into the ordinary battery-capacity column. A flag added after judging, on values outside the usual range, caught these four (fixed 27 October).

Almost every catch came from looking again from somewhere else: another team's method, another AI's review, a range flag, a person's question. Looking again from the same place with the same eyes misses the same things.

3. Facing a human answer key for the first time

Two AI judges saying «this matches the source» may differ from a person saying «this is right». The first time we compared against a hand-made answer key (600 perovskite devices), we recovered 97.4% of the values in the key. But only 83.1% of the values we extracted were in the key. Opening the 45 rows not in the key, 41 had been passed by both judges — values that are in the paper but that the people did not put into their key. So «both judges support it» measures «it is in the source», not «would a person put this value in this column».

One MXene expert answered too. For a pure Ti₃C₂Tₓ film conductivity of 11,869 S/m, the answer was «hard to explain physically for a pure film». We did not delete it; it became a «look again» value. It is one person's answer, but it was the first time a person stopped a value that two AIs had passed together.

Two AIs agreeing gets you as far as «it is in the source». Whether the value belongs in that place is still for a person to see.

4. Could we use one judge instead of two? They cover each other

Running two judges (Claude and Codex) every time takes time. So we measured «skip Claude whenever Codex says yes». Over the last twenty rounds, of 6,561 rows Codex supported, Claude rejected 105 — 1.6%. It looked small.

We checked 20 of those 105 against the source. In 6, Claude was right: the number was in the paper but it was the value of something else — a fixed potential written as an overpotential, a control sample's value given to the main sample, a range covering two samples written as one value. In 6, Codex was right: Claude was too strict about how a material was named. Of the other 8, 6 were papers that themselves gave two different values for the same sample, and only Claude flagged the contradiction. So both judges stay. They do not save each other's time; each sees what the other cannot.

5. Outside eyes — what a first-time visitor sees

The people who make a map are used to it. So three «first-time researcher» reviewers (a freshly started AI · an AI from another company · the AI that built the map) scored the site. The average was 6 out of 10. All three pointed to the same place: the strength — every value carries its source sentence — is clear, but the first screen does not say what to do. A guide page that redirected to itself forever was also found then; from the builder's side it went unnoticed for almost ten days.

6. What grew from the exchange

• Questions become rules. The person's questions («what if it sees the earlier answers?» «whose result is that?» «why not publish yet?») became tools and rules the same day.
• Mistakes become test sheets. The shapes of the AI's mistakes (same sample, different condition · missing unit · different definition) went into the next round's traps and instruction examples. Mistakes are not thrown away; they go back into the engine.
• A person's time is the scarcest thing. So we never ask a person what a machine can answer. We ask one rule that takes ten seconds, and let that answer apply to thousands of values at once.

A person's single remark becomes a rule; the AI's mistake becomes a test sheet. That is how the two grew together.

7. How to ask a specialist — rules, not numbers; one minute, not homework

We sent a researcher in the field six rule questions, such as «should values above 3,500 F/g be deleted or only flagged?» and «what should one look at first when comparing capacitance?». The answers came back one word each («flag», «mass loading»), and those six words became rules of the map. One answer applies to hundreds of values at once.

There was also the opposite experience. To put a «human check» into the paper, we built a page that asked, for 100 released rows, «does this sentence report this value?». The person building this map stopped it at once: rows taken from tables cannot be answered from the sentence alone, so the reader would have to reopen the paper, and unpaid work should not be handed to others as homework. That was right. The page was not sent. In the manuscript, the check is described as what it is: an AI re-reading the source.

Ask people only what a machine cannot answer, and briefly: definitions and rules.

Basis: decision records (DECISIONS · direction documents) · duplicate check (R-157) · scorer correction (R-161) · recipe-table correction · first human answer-key measurement (T1c PV600) · expert answer (question queue, third reply) · flag source checks (27 Oct) · one-judge test (20 disagreement rows · 27 Oct) · site first-impression scoring (26 Oct) — in the internal records.

What we still don't know

Mostly only open-access papers were examined. Nearly every round used only full texts that anyone can legally open. The papers obtained separately were also mostly free to read (Chapter 10). From the afternoon of 2026-10-09, 30 publisher full texts also came in, but most of it is still a story “within open-access papers.” Including paywalled papers could change the picture.²

Boundaries with ambiguous definitions remain. For example: “is a chemical added to separate the sheets an etchant, or a substance that intercalates between the sheets?” One such question was answered by obtaining all 33 foundational papers of this field and consulting their original texts. Boundaries not yet answered carry only a “judgment deferred” mark.³

How much we missed — measured for the first time (1 Nov, first estimate). What we had measured was mainly “does an extracted value match its source sentence” (precision — the unit, property and sample as well as the number). On 1 November we measured the reverse for the first time — “what share of the values in a paper did we capture” (recall). From papers that have capacitance (F/g) on the map we drew 40 at random, and two AIs from different companies, without seeing each other’s answers, read each paper to the end and listed every F/g value (eight altered copies, with a value deleted or changed on purpose, were mixed in; both AIs got all eight right). Of the values both AIs listed with the same number, 265 belonged to samples that contain an MXene or were made from one, and 233 of them (87.9%) were on the map. None of the 37 values of MXene-free comparison samples were on the map — by design, the map does not carry them. The 32 missing values were spread over 13 papers; one paper (a table row where only the MXene column was taken and the composite column was missed) accounted for nine. Limits: values that both AIs missed cannot be counted, so 87.9% is a first estimate that may be too high; it is one sample of 40 papers measured once. So this map still does not claim to be complete; measuring recall properly needs a reference set in which a person lists every value of the sampled papers.

To be supplemented — what we will fill in when data arrive

“Undetermined” does not mean “wrong.” It means “cannot be said yet with this data.” In such cases we did not lower the criteria to force an answer. Instead, we wrote down what was missing. Below is that list.

• Thickness and electrical conductivity. Some films made by the same method had very different conductivities. We wanted to see whether thickness explained this, but of 300 open-access papers, only 4% reported thickness and conductivity together. The usable papers numbered 12, short of the preset 15. → Measure again when more papers reporting thickness have been gathered. On 15 October, recounted on the larger map, films meeting the same conditions number 14 — still one short. Next, as in the “mass loading and thickness” round below, we will have the thickness of each value's own sample found in the source text.⁴
• The “description fields” of capacitance. This round was “undetermined” not because of the number of papers but for another reason. The “correct rows” of the judge qualification test themselves had ambiguous definitions, so the two judges recognized only 4 and 3 of the 6 correct answers (Chapter 4). Since the judges could not be trusted, the results were not used either. → Re-select the correct rows and measure in a new round. → On 31 October the new engine version re-measured 251 papers and added «what mass it is divided by · 3-electrode/device» fields, so this field now stands on a new definition (Chapter 8).⁵
• The 13 unanswered in the known-facts test (11 when re-measured after filling the empty spots · Chapter 5). The papers on the other side of the comparison numbered only 0–3. To fill: electrical conductivity of HF-made multilayer and powder samples · capacitance of MXene etched with an alkali such as sodium hydroxide · capacitance of composites mixed with a polymer (e.g. polypyrrole) · battery capacity of Ti₂C and Nb₂C · light-to-heat conversion of Ta₄C₃ · interlayer spacing of MXene made with another fluoride chemical (NH₄HF₂). → Find open-access papers, put them through the same scoring, then test again with the same 20 questions (Chapter 10). On 15 October, with the map grown to 3,530 values, we asked the same 20 questions under the same rules: questions that could be answered rose from 9 to 14 · agree 9 · disagree 1 · inconclusive 4 · no answer 6. At 64% agreement it is still below the preset 75%.⁶
• Splitting two fields of the map. Split the “mixed-in material” field into “polymer added as a binder” and “material mixed in to add function.” Add a “transparent thin film” mark to the morphology field. These caused the two questions in the known-facts test that were inconclusive or disagreed.⁶
• Routes of synthesis (13 October). When we first built the map, only the hydrofluoric-acid (HF) family had its own columns; everything else went into “other”. A researcher in the field then pointed out that MXene is no longer made with HF alone — molten salts, milder etching with ammonium chloride, organic and inorganic etchants are all in use. So we split the 105 “other” points again: alkali (e.g. sodium hydroxide) 38 · fluoride-free etching such as iodine or iron chloride 14 · electrochemical etching 4. Growing MXene directly from gases (vapor-phase synthesis) had no values at first; that evening, 45 open-access papers on newer routes went through the same checks and the first 14 vapor-phase values (Ti₂CCl₂, Ti₂NCl₂) came in. Most of the 27 papers received that day also used the HF family. → We will collect more papers that measure MXenes made by the newer routes.
• Volumetric capacitance. The very value that made MXene famous, yet only 15 values in 7 open-access papers were directly measured. Even with a perfect score, “insufficient evidence.” → More papers with direct measurements are needed (Chapter 9). By 15 October the map holds 105 volumetric values from 30 papers — we will first sort out how many were measured directly, then measure.
• Papers not yet obtained (about 110 as of 2026-10-15). The 14 October list held about 220. That night 85 more papers were obtained and went through the same scoring (173 rows · ● 165); the map now holds 3,530 values from 757 papers, and the share of titanium-only compositions fell from 72.8% to 69.5%, below 70% for the first time. What remains are papers for which no legal open copy has been found yet. → Keep looking for legal ways to obtain them.⁷

The map is not a discovery. Discovery happens when a lab actually makes what a gap on the map points to.

Gaps are starting to show

Building the map revealed a “gap” for the first time. For the capacitance of MXene made by exposing the sheets with molten salt (molten-salt etching), there were 0 papers within the open-access papers we examined. Either no one has measured it, or if someone has, it is not in open-access papers. Finding and showing such gaps is what this map ultimately aims to do: to be “a place that helps discovery.”⁸

Human checks — colleagues first, the leading authority last

If the two AI judges make the same mistake together, no one notices. So the final check is done by people. Rules to follow have been set as well.⁹

• No one is asked about their own papers. Values from papers the professor co-authored will not be sent to the professor for checking.
• The AI's answers are not shown. So that the checker is not pulled along by the AI scoring.
• Keep the burden small. 30–60 rows per property, three buttons: “correct · wrong · ambiguous.” If 60 rows are checked and none is wrong, we can say the actual error rate does not exceed about 5%.

The first plan was a screen where people check 30–60 rows per property. In practice, asking people about individual values cost them much and gave us little. Now we ask a specialist only for short rules (Chapter 15), and values are re-checked by an AI, which we state as such. Asking the leading authority is postponed until the map has shown it is useful.

Anyone, once — “was the number copied correctly from the sentence?”

Separately from expert checks, anyone looking at the map can look at a value and answer with one click. The question is not “is this experimental value right?” — that needs other evidence, such as a repeated measurement. This question checks transcription only: “did the AI copy the number (and the sample and conditions) correctly from this source sentence?” The number is highlighted in yellow in the sentence, and when the sentence holds several numbers the page says which one was copied. There is no need to read the paper. Answers (✅ · ❌) are left as GitHub issues in the public repository (a GitHub account is needed) and are used only as a reason to look again, not as “verified” — no value changes on one person’s answer; a ❌ means we reopen the source.

Recipes are counted per paper

When we tabulated “which recipes (etching · delamination · form) most often reached high values,” we first counted every value. A few papers that reported many values then drove the result — a row with 40 values from 5 papers rose to the top. So for each recipe we count, per paper that used it, only that paper’s best value (a paper using several recipes is counted under each). The table shows the share of those papers whose best value falls in the top 25% of all papers’ best values. Four papers is only the threshold for a row, not a sign of reliability, and papers that measured many conditions may still be favoured. With the same data, changing how we count changed the top row. The table still shows co-occurrence, not cause.

Same method, different values — mass loading and thickness next to the value

MXenes made by the same method often differ several-fold in capacitance. A researcher in the field pointed to one reason: above about 1–2 mg/cm² of material on the electrode (mass loading) capacitance drops, and the drying method alone can change it two- to three-fold. So on 15 October we had the source text searched, for each of the 730 capacitance values on the map, for “the mass loading · thickness · treatment of the electrode that produced this value.” Each value got a number and the answer had to name that number, which removed the step of deciding which value a number found in a paper belongs to. The two judges were first tested with 26 traps (changed numbers · a value from another electrode · a deleted field · a field that never existed · unchanged). Of the 653 filled values, 632 were ● and 20 wrong. Only fields both judges supported went on the map: mass loading for 272 values · thickness for 95.

The limits showed at once. Few loadings are written per area, like “mg/cm²”; many are written without an area, like “2 mg per electrode”, and those cannot be compared with each other — not yet enough to ask statistically whether loading explains the differences (0 comparison groups). Treatments (annealing · alkali · plasma) are shown on the map; drying, which almost every electrode goes through, does not tell values apart and is left off the screen. These fields too are “before human checking.”

Looking again on 1 November, that loading field was attached to 280 of 1,371 capacitance rows (38 papers), and a few rows held a «mixing ratio» such as «45 wt%» rather than a loading. So we opened a new round that reads to the end of each paper, including the experimental section, and records the electrode loading per paper. Copies with the loading sentence deleted and copies with only the number changed are mixed in as traps, to check first that the judges read the file rather than remember the paper. The thresholds were written down before the results.

Next steps

1. Fill the gaps. Find the data in the “to be supplemented” list above in open-access papers and add them through the same scoring.
2. Split two fields. Split “mixed-in material” in two, and add “transparent thin film” to morphology.
3. Repeat the same test. Ask the 20 known facts again with the same rules. “No answer” has dropped 13 → 11 → 6 (15 October). Fill the other side for the remaining 6.
4. Pair with calculated values. Place the 4,488 computer-calculated structures side by side under the same composition names.
5. A person decides on release. Whether to release the map to the outside is decided by the person who commissioned this work. Originals are included only as short quotations and paper identifiers.

We decided not to inflate the goal. The realistic goals are two: to become a widely used tool, and to become a door to collaborative research.¹⁰

There was already one small signal. That reply was a motivation for this work, not its purpose. The purpose is the same as at the start.¹¹

A map that answers “if I need this property, which MXene should I make, and how?”, with every single value confirmed by its original sentence.

Where it came from — four research notes I wrote earlier

All four are preprints on arXiv. They share one theme: how to check the work an AI has done so that it can be trusted.¹

CCR — When an AI reviews text it produced within the same conversation, it is poor at catching its own mistakes. So I had it review in a new conversation, and it caught more. → This is where the map's design came from: separating the AI that extracts values from the AI that checks them, and not letting the checker see the extraction record.

D-CCR — Then does exchanging reviews over many rounds make things better? Instead, false alarms increased (the title is “More Rounds, More Noise”). → This is one of the reasons we promised “we do not sit the same test again until it passes” (Chapter 5).

CCV (HCCA) — I used the method of checking in a separate, isolated conversation to tell whether an AI gets test questions right by memorizing the answers. → It is the same idea as keeping the “answer papers” out of the map in the usefulness test.

XMR — When does it help to have an AI from a different company do the second check? → This is why the map has two judges, Claude and Codex.

It was not carried over as is. The four papers dealt with errors in text and code; this map deals with scientific values. I only borrowed a few ideas. Conversely, what was newly learned from this map — testing the judges first with trap questions, and writing down the rules before seeing the results — can be given back to the next verification study. The two lines of work correct each other.

Call a lake a sea, and you will find the sea in the lake; call a rock a jewel, and you will carve the jewel out of the rock.

Similar efforts, and what is different

Attaching the source sentence to every value is not new with us. The band-gap database built in 2022 by the Cole group at the University of Cambridge tags each of its 100,000 records with the sentence it came from. It holds about 30 times as many records, and it also estimated, on sampled papers, what share of the values it captured (recall). What differs here: this map gathers experimental MXene values across many properties, ties each value to fields on how the sample was made (etching · delamination · form · electrolyte · thickness — as far as the paper states them), publishes the verdicts of two different AI models for each value and the judges’ own test (planted-error controls) for each round, and shows side by side the groups where “the same recipe gave different values.” The two projects measured accuracy with different rulers, so we do not say which is more accurate.

Where it is going — a dream

For now it is MXene alone. But the same method can be used for other thin-sheet materials (2D materials). Further out, could even far more complex subjects, such as the brain, be drawn as 3D maps that AI gathers and people check? For now it is a dream — a dream to be pursued, as this book has shown, by honestly separating “verified” from “not yet” at every step.

Why do it

The words above are my own. They are the spirit of believing something is there, even though it may not be, and carving it out to the end. And one more thing — a stuck problem gets solved in an unexpected place. The numbers of a wholesale market, the numbers of a material, someday even a map of the brain: things that seem unrelated are connected like a spider's web. Finding those connections is science, and it is the reason to keep trying. That is also why I made “Songjarvis,” the AI colleague that built this map with me — not as a tool that only does what it is told, but as a colleague that looks widely and comes to realize things on its own.

My heart in doing all of this is simple: to leave traces that are useful to many people, and to be someone the next generation can look at without shame.