The mandate for the on-chain explorer.

Mandate — what the first feature must answer

What this is

An index, not a feed. Nothing here is about what is happening now. It is about what already happened, on the coins we hold completely, reasoned over in depth. A trader comes with a coin or a wallet and leaves knowing who was in it, when they got out, who does this repeatedly, and who to stay away from.

The universe, and two gates

In scope: coins we watched come into existence — we hold their whole life. Out of scope is refused, never approximated.

Strength of any count is decided separately, by whether a coin's observed creations minus destructions equal what its own contract says exists:

a count over it is

reconciles

a total

short by a known amount

a floor — and the shortfall is publishable

not yet reconciled

unevaluated — neither. State no figure.

What we can and cannot see

Entry — who, when, how much, what they paid

fully visible

Exit timing and size — when a wallet's position fell and by how much

fully visible

Whether it was a sale or a transfer

answered — 70.9% sales, 27.8% moves. A surface counting every outgoing movement as an exit overstates exits by roughly a quarter.

What they got for it — the actual receipt

recoverable for part of the market, and the route is now known. Not from the wallet's own records — a payment in this chain's own currency reaches a signing address in none of 54,526 cases, and eight trace methods were tested with none available. But the amount is in the swap event itself: both sides survive decoding on 100% of swap rows, and the seller's proceeds is the negative leg. Confirmed by the token layer. Three conditions: the seller is the transaction signer, not the address in the event (tested — a router in 25 of 25 cases, so a recipient field would be a router column wearing a seller's name); for the 58.4% of sales that route through several pools the proceeds is the LAST hop's output and summing the legs over-counts; and 32.7% of log rows are signatures the decoder discards, so a venue it does not recognise stays dark.

The multiple they left at

answerable, corroborated by two independent measurements.

A multiple is exit price over entry price and neither comes from the seller: entry is visible, and the coin's price at the exit block is read from OTHER wallets' purchases in the same quote asset. The receipt never enters it. Priceable on 65.3% of exits at a 50-block window, 84.6% at 1,000 — the wallet layer's own fold gives 65.3% and 87.1%, matching exactly at the tight window. Four conditions make it sound, each one costing coverage: the reference must be at or before the exit, in the same quote asset (one in seven otherwise divides incommensurable units), a median of the window rather than the nearest fill (a single dust trade set the price in one exit in five), and never the seller's own purchase. The honest claim is "left when the coin was at four times what it paid", never "made four times" — slippage and routed proceeds sit between the price and the receipt.

Sold or merely moved

answered. A sale is told from a transfer by who RECEIVED the tokens, and the recipient must be a venue for this token — a pool whose registered pair contains it, or the protocol singleton all v4 trades settle through. Over 191,049 disposals where nothing came back: 70.9% are sales (62.8% to the v4 singleton, 8.0% to a pool holding that pair) and 27.8% are moves to ordinary addresses. Measured by the wallet layer over its own disposals against the token layer's registry, each half measured by the layer that owns it. No thresholds, no promiscuity, no retention — three earlier tests were tried and every one measured what recently passed through a contract rather than what the contract is. Four earlier percentages were withdrawn; this one has no judgement in it, and its only soft edge is the registry's own completeness at 95.4% and 89.8%, whose residue resolves within a pass.

Tags below: ✓ answerable · ~ answerable with the ambiguity or floor stated · ✗ needs exit value and is refused.

---

A · Who bought, and when

1. 1Who was the very first wallet into this coin?working now
2. 2Who were the first ten, in order, with the block each arrived in?not built yet
3. 3How many wallets bought in the first block it ever traded?working now
4. 4How many were in within the first minute? Five? Thirty?working now
5. 5How long after the coin was created did anybody buy it?working now
6. 6Did anybody hold it before its first public trade?not built yet
7. 7What valuation was the first purchase at?not built yet
8. 8How many bought under a 6k valuation? Under 20k? Over 100k?working now
9. 9What is the full spread of valuations its buyers paid?not built yet
10. 10Which wallet got the cheapest entry of anyone?not built yet
11. 11How many wallets bought more than once, and at what valuations?not built yet
12. 12Which wallets bought again higher — added to a winner?not built yet
13. 13Which bought again lower — averaged down?not built yet
14. 14How many acquired it without paying anything at all?not built yet
15. 15What did this specific wallet pay, and when?not built yet

B · Who dumped, and when

1. 16Which wallets have reduced their position, and when?not built yet
2. 17Who reduced first, and how long after buying?not built yet
3. 18How much of their position did each one let go — all of it, or part?not built yet
4. 19Which wallets were out within a minute of getting in?not built yet
5. 20Which held longest before letting go?not built yet
6. 21How many dumped in the same block as each other?not built yet
7. 22Was there a single block where most of the selling happened?not built yet
8. 23Did the earliest buyers leave before the later ones?not built yet
9. 24Which wallets left completely and never came back?working now
10. 25Which left and then bought back in?not built yet
11. 26What share of the supply has left the hands that bought it?not built yetwith a stated limit · floor
12. 27How many of its buyers still hold every unit?working now
13. 28Did the wallet that created this coin buy it, and has it let go?not built yet
14. 29Of the first ten in, how many are still in?not built yet
15. 30What did they get for it — the actual receipt?refusedthe chain does not record it
16. 30bWhat multiple was the coin at when they left?not built yet

C · A wallet across every coin we hold

1. 31How many coins has this wallet bought?working now
2. 32Which ones, in the order it entered them?not built yet
3. 33What is its lowest entry valuation ever? Its highest?not built yet
4. 34What is its median entry valuation across everything?not built yet
5. 35How many wallets only ever buy under 6k — never above?not built yet
6. 36Which wallets have never entered above 20k?not built yet
7. 37How many coins has it been the first buyer of?not built yet
8. 38Which wallets are first buyers of five or more coins?not built yet
9. 39Does it always enter inside the first minute, or does it wait?not built yet
10. 40How fast does it typically leave after entering?not built yet
11. 41How many of its positions did it fully exit?working now
12. 42How many are still open?working now
13. 43Has it ever held a coin longer than an hour?not built yet
14. 44How many coins has it entered and left inside sixty seconds?not built yet
15. 45Does it ever take proceeds into its own account, or do they go elsewhere?not built yet
16. 46Was it active before our index begins — provable from selling more than it bought?working now
17. 47How many coins has it been in that nobody else touched?not built yet
18. 48What is its busiest hour, by coins entered?not built yet
19. 49How much has it spent in total across every entry?not built yet
20. 50How much has it made, in money it banked?refusedthe chain does not record it
21. 50bWhat is its median exit multiple across every coin it left?not built yet

D · Patterns in one wallet

1. 51Is this wallet's buying automated — does its timing say so?not built yet
2. 52Does it enter at a consistent valuation, or all over the range?not built yet
3. 53Does it enter, exit and repeat — a round-tripper?not built yet
4. 54Is it a one-clipper: one entry, one exit, gone?not built yet
5. 55Does it accumulate and hold, or churn?not built yet
6. 56Does it buy coins that others then pile into, or coins nobody follows?not built yet
7. 57Does it consistently leave before the coin goes quiet, or after?not built yet
8. 58Does it only enter coins created by a particular wallet?not built yet
9. 59Has it ever been the last buyer before a coin went silent?not built yet
10. 60What share of the coins it entered has nobody since left?not built yetwith a stated limit · floor
11. 61Does it size positions consistently, or vary wildly?not built yet
12. 62Was it funded shortly before each of its entries?not built yet
13. 63Does the same address keep funding it?not built yet
14. 64How many distinct coins did it enter in a single block?not built yet
15. 65Does it ever buy a coin twice in the same block?not built yet

E · Groups and coordination

1. 66Which pairs of wallets keep entering the same coin within half a second?not built yet
2. 67Do any two wallets hold an identical set of coins — the same list, both ways?working now
3. 68Which groups of three or more enter the same coins in the same blocks?not built yet
4. 69When one of a group enters, how long until the others follow?not built yet
5. 70Do coordinated wallets also leave together?not built yet
6. 71Is their selection explainable by an obvious rule anyone could run?not built yet
7. 72Did they pay for these coins, or were they sent them?not built yet
8. 73Do they reach the coin through the same route?not built yet
9. 74Which group has entered the most coins together?working now
10. 75Of the coins a group entered, how many did nobody outside the group touch?not built yet
11. 76Did a coin's creator fund any of its earliest buyers?not built yet
12. 77Is one address funding several buyers of the same coin?working now
13. 78Do the wallets a single funder feeds behave alike?not built yet
14. 79Which groups consistently arrive before the crowd?not built yet
15. 80Which groups consistently arrive after it?not built yet

F · Which wallets to watch, and which to avoid

1. 81Which wallets are first into the most coins?not built yet
2. 82Which enter earliest relative to every coin's first trade?not built yet
3. 83Which wallets have been early in coins that later traded far above their entry?working now
4. 84Which leave fastest after buying — measured, not judged?not built yet
5. 85Which wallets appear among the earliest buyers of coins nobody ever got out of?not built yetwith a stated limit · floor
6. 86Which wallets were given coins rather than buying them, repeatedly?not built yet
7. 87Which creators have made coins that all went quiet inside an hour?not built yet
8. 88Which creators bought their own coin and then let it go?not built yet
9. 89Which wallets never take proceeds into their own account?not built yet
10. 90Which wallets are on the other side when a coin's price collapses?refusedthe chain does not record it
11. 91Which wallets consistently buy last before a coin stops trading?not built yet
12. 92Which wallets to avoid buying alongside — measured as: whose entries precede the steepest position exits?not built yetwith a stated limit
13. 93Which wallets have the best entry position relative to everyone else in the coin?not built yet
14. 94Which wallets banked the most?refusedthe chain does not record it
15. 94bWhich wallets consistently leave at the highest multiples?not built yet

G · Comparing coins

1. 95Which coins had the most wallets enter under 6k?not built yet
2. 96Which drew the most distinct buyers in their first minute?not built yet
3. 97Which have the widest gap between the cheapest and dearest entry?not built yet
4. 98Which coins has nobody reduced a position in at all?working nowwith a stated limit · floor
5. 99Which coins had the highest share of holders who were given rather than bought?not built yet
6. 100Which coins were entered by a known coordinated group?not built yet

---

Counted

1. ✓ answerable outright: 94
2. ~ answerable with the ambiguity or floor stated: 5
3. ✗ refused — needs what a wallet banked, which the chain does not record: 4
4. total questions: 103 — it grew past a hundred when three refused questions each split into the part that is answerable and the part that is not

Counted from the tags rather than by hand. This line has been wrong twice: once at 89, 6 and 5, and once published to the room and to the website room as 94, 3 and 3 across 100. Both times the answerable figure was right and the refusals were undercounted, which flatters the document in the direction a sceptical reader tests first. The website room parsed the tags independently and arrived at 94, 5 and 4 across 103, matching this line — two counts from two rooms.

The 4 refusals are all the same question wearing different clothes: what did somebody make. That is the one thing this index cannot tell anyone, and saying so is not a weakness of the product — it is the reason the other 99 can be trusted.

A coin's quote asset is not what its buyers paid with

Which asset a coin trades against decides what every valuation is denominated in. Deriving it from what buyers handed over is unsound, confirmed by the token layer on a worked example: that coin trades 99.8% against a 6-decimal stable and 0.2% against native by swap count, while leg-counting said the reverse. The buyer pays native, a router converts it, and the terminal swap is stable-into-coin — leg-counting sees the first hop and the venue is the last. Chain-wide, 45.2% of swapping transactions are multi-hop, so this is the ordinary case rather than an exotic one.

The falsification is cheap and needs no swap counts: a coin's real quote assets are the counterparts of its registered pools, so any quote asset leg-counting finds that no pool supports is a hop. The worked coin has two counterparts across nine pools; leg-counting found about thirteen. Eleven of them cannot be venues.

The replacement is the dominant pool's counterpart by swap count — not by pool count, which would have called that coin two-to-one stable when it is 99.8% stable, and not by buyers' legs. The token layer is publishing it as a field so neither room re-derives it and the two cannot disagree. One further rule from the same data: 96.7% of that coin's swaps happen in a single pool, so a price series is built from the dominant pool rather than pooled across all of them — a thin pool is exactly where a bad print lives.

And knowing the asset is not the same as knowing whether it is money

Naming the dominant counterpart made every quote look alike: an address with a swap count behind it. Once the token layer published names, the counterparts turned out not to be what either room had reasoned toward from decimals and pair counts. Classified here from those published names — this room's reading of their strings, and unconfirmed — across the 1,491 coins clearing a basis of 20 swaps:

what the coin is quoted in

coins

share — reproduced by the token layer, see the caveat below

a tokenised equity of a listed company or fund

500

33.4%

the chain's one stablecoin

304

20.3%

native

235

15.7%

a claim on a private company

208

13.9%

wrapped native

104

7.0%

a leveraged product

82

5.5%

a named token of no known kind, mostly memecoins

39

2.6%

a wrapped external asset with a deep public price this room does not hold

18

1.2%

no name yet, or the contract publishes none

2

0.1%

So a dollar figure is derivable for 42.9% of that set — stable, native and wrapped native, a share the token layer reproduces at 643 coins against 642 here — and not for the rest. A valuation denominated in units of a three times long is not a market cap and is not one conversion away from being one: a leveraged token's price is a function of the path the underlying took rather than its level, so even a correct spot price would not make the figure mean what a reader takes a cap to mean.

A valuation therefore needs three things, not two. A denominator resting on enough trades to be named. A supply proven not to have moved. And a quote asset that is a unit of account. Adding the third took the statable set from 347 coins to 186 — 46.4% withdrawn, this room's counting and unconfirmed, all of them valuations that were arithmetically correct and denominated in something that is not money. The refusal names the asset rather than the absence.

Private is a kind of its own, and it was found in what this index PERMITTED

The first version of this classified on the WRAPPER. A tokenised claim carries "Robinhood Token", "Class A" and "Common Stock" in its name whether the company behind it is listed or not — so a claim on SpaceX matched the equity rule on three separate counts, and the two largest 1x long families turned out to be written on Anthropic and OpenAI. Neither has a public share price, so no rate converts them and no amount of correct wrapping changes that. The chain labels the class itself with Pre IPO, which is a better handle than anything either room reasoned toward, and it is matched directly now along with the named private issuers. 208 coins, refused explicitly.

The reason matters as much as the outcome here. The 1x longs were already refused, for the weaker reason that a derivative wrapper's tracking cannot be assumed. The real reason is that a perfect tracking assumption converts to nothing when the underlying has no price — and that reason survives someone asking why, where the first one does not.

And one refusal was right for the wrong reason. Wrapped BTC, 18 coins, was sitting with the memecoins in "no known kind". It has about the deepest independently verifiable price of any non-stable asset anywhere. It stays refused, because this room holds no BTC rate and inventing one is how a valuation stops being a measurement — but it is now refused for want of a rate rather than for want of a unit of account, which is a buildable gap rather than a property of the asset.

What the token layer's agreement does and does not establish

Every category above reproduces on the token layer's own population within four coins and most within one — 504 equity against 500, 305 stable against 304, 235 native against 235, 208 private against 208, 103 wrapped native against 104, 80 leveraged against 82, and the two categories drawn here for the first time, private and wrapped external, landing exactly at 208 and 18.

That is reproducibility, not corroboration, and the difference matters. Both readings classify the same name strings from the same map using the same stated rules. What the agreement establishes is that the rules are unambiguous enough for two implementations to land on the same coin, and that neither side made an arithmetic slip. It does not establish that the classification is right about the world: if one asset were mislabelled in the source map, both readings would be wrong by the same hundred coins and agree perfectly. "Two rooms measured it and agreed" reads as corroboration and this is not that.

The corroboration exists but it is for the strings rather than the counting, and it came from somewhere else. The wallet layer read name() and symbol() off the contracts directly, not from the map, and arrived independently at the same identities for the assets that carry most of the set — the stablecoin, wrapped native, the tokenised-equity wrapper, and the leveraged naming convention. That is a separate code path reading primary evidence, which is the thing the shared input cannot provide. It covers the major assets and not the long tail.

The sharper reading of the table is not the negation. A majority of this index trades against a tokenised claim on something else — the token layer's counting, 810 of 1,494 across listed equities, private companies, leveraged products and wrapped external assets, or 54.2%. A coin priced in a tokenised equity is not an edge case on this chain; it is the median case.

And the statable-valuation count is suppressed by a clock, so it is not a stable denominator. The supply test rests on a field that is not retroactive: coins whose mints were folded before it shipped read null, and null is a refusal. The token layer's count is 353 before the kind test against 347 here for that reason, and both will rise on their own as coins mint under the field. Quote that figure with a timestamp or not at all.

Two rules here are clean today and will not stay clean. Exactly one asset matches stable on the word "Dollar" across every named quote asset, and no stablecoin-shaped symbol lacks it — that breaks the day a USDC or a EURC lists, so the known symbols are listed explicitly as well. And every leveraged product on this chain today is an "Nx Long": there is not one Short, Inverse or Bear in the whole set, so the part of that rule which matters most has never fired. A short on a listed equity would otherwise land in equity and be permitted. Both patterns now cover the cases that do not yet exist, which means neither is tested by anything.

The window is an assumption, not just a coverage number

Pricing an exit from the nearest purchase means the window doing the pricing is also doing part of the answering. On a coin moving fast, a purchase fifty blocks earlier is not the price at the exit, and the error is biased by whichever way the coin was moving rather than being symmetric noise. Measured here across windows, and unconfirmed because the window is this room's parameter rather than anything the records hold, the median multiple is 1.284 at 50 blocks, 1.216 at 200 and 1.144 at 1,000 — a twelve percent spread that moves in one direction, which says the window is carrying some of the result.

So every multiple publishes its window, and the figure is reported at two windows side by side. Where they diverge materially the window is doing the work and the answer says so. This caution came from the token layer and it is the kind that would never have shown up as a failure — the numbers look fine at any single window.

What is always refused

A coin outside the universe. What any wallet earned. A dollar figure without its rate, that rate's spread and the assumption under it. Any figure not confirmed by the layers that build the records, stated as a fact. A count stated as a total where coverage does not license it.

That refusal is wider than it was written, and it now has a mechanism. It was stated as a sale paid in the chain's own currency leaving no reachable record. Measured by the wallet layer against primary evidence over 93,487 proven sales: proceeds are visible as a quote token on the selling wallet's row in 17.84% of them, visible as native value in 0.00% — zero, not few — and not visible at all in 82.16%. It is not an artefact of one venue class: 15.39% visible at the v4 singleton, 27.81% at registered v2 and v3 pools. The mechanism was proven on a named transaction rather than reasoned: the selling wallet sends the coin to the pool, the pool sends the quote token to the router, and the router unwraps that quote by burning it, so the native leg is an internal call that emits no log and the transaction's own value field reads zero. What is conserved on this chain is the transaction's value field, signer to destination. Sale proceeds do not travel that way.

So the refusal is stated as proceeds are not observable, with native as the reason rather than the scope. It applies to four sales in five, not to a subset paying in one asset.

And there is a path to it that does not run through the wallet tape. In a sample of sixteen no-proceeds sales, ten carried a burn of the quote token to the zero address by a wallet other than the seller. The router receives and burns the same amount, so its net is zero and the row disappears from a netted tape — but the amount is still in the logs. That makes "how much did they get" answerable from the log layer even while "the wallet was credited" is not answerable from the tape at all. Sixteen is a sample and is labelled as one. The path is now measured, and it is not the one that sample pointed at. The burn is the wrong handle — present in most invisible sales and equal to the actual payout in 6.43% of them. What works is the venue's own payout: the quote token leaving the pool in the same transaction as the seller's coin outflow, which sits in the logs whether or not anything visibly reaches the seller, because the pool must pay somebody. Measured by the token layer over 54,064 proven sales, proceeds carry a figure for 52.62% of them — 8.83% visible on the seller's own row, 43.79% recoverable from the payout — and 47.38% remain unrecoverable. Validated where the truth is known: across 3,046 sales where the seller visibly receives the quote, the payout exceeds the receipt 97.9% of the time, by 1% at the median and 4% at the 99th percentile. That gap is routing the pool pays and the seller never sees.

So the claim this licenses is an upper bound, and it has to be stated as one: the venue paid out X for this position, of which the seller received X less routing, typically one to two per cent. Not "the wallet received X". A point estimate quietly a per cent or two high is a liability; a bound that is right about its own direction is a product.

The coverage limit is the venue, not our records. All three figures in this paragraph are the token layer's, measured from logs. At registered v2 and v3 pools the payout is recoverable in 99.64% of invisible sales. At the v4 singleton it is 43.91%, because the singleton nets legs internally and never emits a transfer for each one — and 92% of sales go through it. No decoding recovers a log that was never written.

That figure is now resolved, and it resolved against the higher number. The wallet layer's 17.84% and the token layer's 8.83% differed because part of the wallet layer's sales base was wallets disposing of QUOTE assets rather than coins, and those show their proceeds far more often — selling a quote asset is a single pool swap with both legs on the seller's own row, while selling a memecoin routes through an aggregator that hides the return leg. Their first split of that was itself wrong and they replaced it: five of the eight tokens they had called quote assets turned out to be memecoins, because the set had been chosen by how often a token appeared as the proceeds leg rather than by what the token is. Popular is not a category. Classified from the contracts instead: 13,336 quote-asset disposals visible 53.55% of the time against 78,464 coin disposals visible 13.19%. The separation is sharper than the contaminated version showed, and 13.19% sits close to the 8.80% reproduced here, with the remainder being the restriction to coins whose creation we witnessed — the newest and most aggregator-routed. The diagnosis came before the number: a uniform factor across both venue classes with their ordering unchanged is a denominator and not a test.

So the token layer's 52.62% is the figure to build against — confirmed by the wallet layer as the number to size against once its own denominator was found — and the extra sales in the higher count were the ones that never needed recovering in the first place.

One more thing about where the invisibility sits, because it changes how the path should be built. A transaction disposing of more than one coin at once shows its proceeds 2.01% of the time against 19.31% for a single-disposal transaction, measured by the wallet layer over 2,140 such cases. Multi-disposal transactions are the worst case for attribution and there are few enough of them to hold out and state rather than model.

Until the path is built these four questions stay refused. When it is built they become the state this document does not yet have: answerable for just over half of positions, as a bound rather than a figure, and refused for the rest.

Rules every answer obeys, each bought by a specific failure

Figures in these rules that came from the layers that own the records say so. Figures that are this room's own counting say that too, and are unconfirmed: the cohort counts and the funded-buy counts rest on a population and an infrastructure filter that exist only here, so neither layer can reproduce them, and agreement with a number nobody re-derived would be worth nothing.

1. 1Both layers confirm a figure before it is stated as fact. Six figures in one day were arithmetically sound and false.
2. 2Every figure names its population. All six of those failed on the denominator, never the sum.
3. 3A count is a total only where coverage licenses it. A marker that meant "when we first saw it" was read as meaning "we have all of it".
4. 4A negative result needs a detector shown to fire. A broken detector and a clean market produce the same empty page.
5. 5A positive result needs the ordinary explanation tested and failed. A cluster dissolved into routers that fifty-five thousand wallets settle through.
6. 6Refuse rather than return empty. An empty list reads as "nobody qualified", which is a different fact from "we could not look".
7. 7Bought is not acquired. Two in five acquisitions involve no payment; keying on a row's own label missed 23% of real buying. Both figures are this room's own counting over the sample and unconfirmed; what is not in doubt is the direction, since a gift and a purchase are different rows and the label was reading the wrong one.
8. 8Sold is not moved. Counting every outgoing movement as an exit was wrong, and badly — the size of the error is a measurement of mine that no layer has confirmed, so it is not stated here as a figure. What IS confirmed, by the wallet layer against the chain, is the mechanism: value arriving as this chain's own currency leaves no record, so a sale and a transfer away are indistinguishable for those coins.
9. 9Identifiers appear whole. An address nobody can paste is decoration, not evidence.
10. 10Three states stay three. A number, a null and an absence are different facts.
11. 11Medians, never means. A handful of near-zero entry prices was enough to make a mean meaningless — measured in the growth room over its own computation, not by a layer over the records, so the magnitude is theirs to state rather than ours.
12. 12A ranking publishes who it excluded. Otherwise it shows the top of the remainder.
13. 13Character in the framing, precision in the substance. A mystical sentence around a measurement is theatre dressed as evidence.
14. 14A figure that moves carries the moment it was measured.
15. 15A shared artefact must say which way round it is. A registry keyed by pool was described as keyed by token, and two rooms independently built the wrong lookup on that description and got near-zero. Nobody's method was wrong and nobody's data was wrong; the failure lived in the sentence between them. Documentation absent from a shared file is close to the file being wrong.
16. 16A count of our own rows is not a claim about the chain. "The index holds N token contracts" and "the chain has N token contracts" are the same number and only the first is ours to state. This is the hardest one to catch because nothing about the sentence is wrong — the arithmetic is sound, the population is named, and there is no error to find. It sat on the largest figure on the company's front page through a full day of auditing figures, and was found only when the same slip was noticed in a private message.
17. 17Proving an outcome is not proving your guard caused it. A blocked file and a file that was never published look identical. Test with two files in one publish, one guarded and one not, so there is a single variable. Related: never read an exit code through a pipe — a refusal came back as a success while testing whether refusals work.
18. 18A thing being a venue is not it being a venue for the token in hand. The singleton every v4 trade settles through is a venue for thousands of pairs; a pool is a venue for one. A recipient that trades some other pair is no evidence about this one.
19. 19Never compute a percentage over a churning set. A ranked list capped at five thousand rows sheds coins as the ranking moves, so a share over it is a share of a population turning over underneath the measurement. Every coverage figure this room published came from one.
20. 20Unanimity over a tiny sample is not confidence. Half of resolved quote assets rest on nine trades or fewer and a tenth on one — the token layer's own published warning against the field it produces, not this room's measurement — and a single-trade coin reports its quote as 100% dominant. Gate on how many trades decided a share, never on how lopsided they were.
21. 21A count of our own rows is never published or sent to another room without its attribution. A count of this index is a fact about a file, not about the chain. The attribution lives in the data file and in tool output, where it cannot be separated from the number.

This rule was first written as "never put such a count in a sentence at all", and broken one turn later while describing the build to the person who asked for it. That version was unfollowable, and an unfollowable rule is worse than a narrow one because breaking it teaches nothing. So it is scoped to what it was always for: published lines and messages to other rooms — the places a figure escapes and gets repeated. Answering a direct question about what this room built is not publishing a claim about the chain, and pretending otherwise cost five attempts at a rule that could not hold.

1. 22A property of a record is not a refusal that fires. Six coins show more minted than their contract reports; five are refused for it, because the sixth has no buyer and never reaches a valuation. The count of things that are wrong and the count of times a guard speaks are different numbers, and only the second is what a reader experiences.
2. 23A counterparty is whoever ended up with the asset, not whoever it was handed to. Gross transfers describe the path; net deltas describe the outcome. A router that receives a token and forwards it in the same transaction nets to zero and is not a gainer at all, so walking transfers scores every routed sale as a move. That distinction alone was worth fifteen points between two implementations of an identical predicate.
3. 24An exclusion that runs silently travels with the figure. About a fifth of transactions are skipped because their outcome cannot be proven — correctly, but it makes every share a share of what could be proven rather than of everything, and two implementations were both doing it without saying so.
4. 25An edit that matches nothing changes nothing, silently. Two of these rules were written twice and landed zero times, because the text they anchored to was numbered differently from what the edit expected. Nothing failed and nothing was reported. Verify the document after editing it, by parsing rather than by reading — which is how this was found.
5. 26A route keyed to wording is a decoy list waiting to happen. The first version of the question router matched questions with patterns written against my own paraphrases, and the very first question it was handed — the first wallet into a coin, which one surface had printed all along — came back as unbuilt, because this document says "first wallet into this coin" and the pattern said "first buyer". Routes are keyed to the catalogue number, there is exactly one copy of the catalogue, and a route pointing at a number this document does not contain refuses at startup rather than counting as coverage.
6. 27A wrong match is worse than no match. Asked how many holders a coin has, the router answered how many coins a wallet had bought: the two questions share only words that every question shares. It returned a real answer with real proof to a question nobody asked, and nothing in the output said so. A match must beat its runner-up before it is used; when two readings are close, list them and ask.
7. 28A clustering key can fail in both directions, and the fix can be the larger error. Grouping wallets by an identical purchased set splits one cohort into a bucket per missing coin — 412 groups where 72 cohorts is the defensible count, both this room's own counts and unconfirmed, because neither record-owning layer can recompute them without this room's population and its infrastructure filter. Joining wallets that share five coins and taking connected components fixes that and chains: the largest component was 4,560 wallets over 3,810 coins with not one coin all of them bought, because A joins B and B joins C while nothing links A to C. A group is only a group if a core survives — coins every member bought — and the core size is the strength of the claim.
8. 29A threshold with no knee is a preference, not a finding. Funded buys at 30, 100, 300 and 900 seconds: 390, 1,124, 2,345, 4,009 — this room's own counts, unconfirmed for the same reason, though the wallet layer confirms that a smooth curve means the count is not a finding at any window. Smooth, no boundary. Any single figure drawn from that curve is my choice of window presented as a property of the market. Report the curve; the count is not the answer.
9. 30A withheld value is a third state, and the conditional must never be added to the fact. When a block's receipts are missing the tape withholds the transaction's value and leaves the settled field reading zero. Reading only the settled field missed a fifth of native sends, and the outcome filter sat above the read so those rows were discarded before they could count at all. The absence is not random — it tracks receipt availability — and in funded buys the withheld ones outnumber the settled ones seven to one — 8,295 against 1,126, this room's own counts and unconfirmed — so the figure described about one case in eight of what it claimed. The withheld share itself is confirmed: 21.9% measured here, 19.8% by the wallet layer over a different span, same phenomenon. Count both, report them apart, never blend.
10. 31Two counts are only comparable once the same disqualifier has run over both. The settled funded-buy figure had infrastructure funders excluded and the pending one did not, which made the ratio between them look wilder than it was. Filter both, then compare.
11. 32A pattern anchored to the end of a line loses every line that has a word after it. Four of the five questions carrying a stated limit write their marker followed by one word, so the check that measures coverage read their state as unknown, dropped them from the answerable denominator, and reported coverage against 95 questions instead of 99 — a silent exclusion that flatters every figure above it, in the one tool that exists to catch silent exclusions. It was found by another room counting the same document and getting a different total. Parsing now refuses outright if a state marker survives into the question text.
12. 33The presence of a thing is not the correctness of its amount. A sample of sixteen sales with no visible proceeds had ten carrying a burn of the quote token, which read as a working handle for recovering what a seller received. Measured at scale by the layer that reads logs: the burn is present in 60.73% of invisible sales and equals the venue's actual payout in 6.43%. A path built on it would have attached a figure that is wrong about nine times in ten while looking present, sourced and plausible — worse than the refusal it replaced, because a refusal is at least honest. Test the quantity, never the presence of a candidate for it.
13. 34A uniform ratio across independent subgroups is a denominator, not a test. Two rooms disagreed two-fold on how often a seller's proceeds are visible. Split by venue class the gap was the same factor in both classes and the ordering was unchanged, which a predicate difference does not produce — it would move one class more than the other. Running the other room's predicate verbatim on one population reproduced their figure to two decimal places, and the disagreement was in what each was dividing by. Third time in one day: the disagreement is almost never the test.
14. 35A stated limit with no name is not a stated limit. Five questions here are marked answerable-with-a-limit, and the word naming the limit sits after the marker. Both tools that read this document were throwing that word away, so the limit existed as a symbol and nowhere as a fact, and a rendering of the page had it dangling on the end of the question as though it were part of what was asked. It is now carried beside the state. One of the five has no name at all and prints as unnamed, which is the honest reading and an item of work rather than a presentation detail.
15. 36A denominator can be correct and still not be money. Every withdrawn valuation passed two real tests — a quote asset resting on enough trades to name, and a supply proven not to have moved — and was arithmetically sound. It was expressed in units of a tokenised equity or a leveraged product, which no rule in place could see, because the field naming the quote asset made a stablecoin and a 3x long look identical. A number can pass every check you have and still be measuring something you did not mean, and the only thing that surfaced this was another room answering a question about names that had been open for days.
16. 37A category chosen by frequency is not a category. Two rooms independently made the same mistake in one day. One defined "quote asset" as the eight tokens appearing most often as a sale's proceeds leg, and five of the eight turned out to be memecoins — popular, not a denominator. The other classified tokenised claims by their wrapper, and the wrapper is identical whether the company behind it is listed or private, so a claim on a private issuer read as an equity. Both were fixed the same way: read what the thing IS from the contract, never what it does often or what it is packaged in.
17. 38A rule that is clean today because the awkward case does not exist yet is untested, not correct. Exactly one asset on this chain matches the stable rule on the word "Dollar" and no stablecoin-shaped symbol lacks it — true today, false the day a USDC lists. Every leveraged product is an "Nx Long" and there is not one Short, Inverse or Bear anywhere in the set — so the branch of that rule which matters most has never fired once. Both patterns now cover cases that do not exist, which means neither is tested by anything.
18. 39A status at the top of a report does not travel with the figures inside it. A published report opened by saying its classification was this room's own and unconfirmed, and another room read the percentages out of the body as settled and nearly built a public post on them — two hours before those same figures were reorganised into different categories. The status was present, honest, and in the wrong place. An unconfirmed number carries its status in the same breath as the number, every time it appears, or it is read as a fact by the next person who needs one. Publishing something in a room is not confirming it, and the room cannot rely on a reader making that distinction.
19. 40Two implementations reading one input agree by construction, not by corroboration. A classification reproduced across two rooms to within four coins in every category, and the only thing that established was that the rules were unambiguous and neither side fumbled the arithmetic. Both read the same name strings from the same map. A single mislabelled asset would have made both wrong by the same hundred coins and the agreement would have looked identical. Corroboration requires a different input — here it came from a third room reading the contracts directly — and the distinction has to be stated, because agreement between two readings of one source is the most convincing wrong thing available.