source

CLE measured figures, 14 October 2026

Published 2026-11-05 by the Chain Of Title LLC desk. Figures state the record as of the dates given in the text; the record itself is the source to check them against.

Measured 14 October 2026 against production. Every figure below is a count read off the live record or a stated sample, with the method beside it. This file is SOURCE for the deck, not published copy: it carries the caveats a presenter needs, several of which must survive into the slides.

House rules that bind anything built from this: counts never verdicts, no intent, no grading words, no second person in register voice, no em dashes, and artwork is never the only instrument named. A launch the register did not observe is unknown, never cleared.

Served record, 14 October 2026. Metadata documents and images are SEPARATE instruments and are never pooled into one "files" figure.

• launches on record: 1,902,056
• declared a metadata URI: 1,873,872
• metadata documents held: 1,808,348
• images held: 1,748,292
• both instruments held: 1,725,484

THE FIGURE TO USE WHEN SCOPE IS QUESTIONED. "On record" is not "watched at creation." Restricted to launches observed at creation since 2 September 2026: launches 1,798,098; metadata documents held 1,733,039 (96.4%); images held 1,673,582 (93.1%). The 103,956 difference are launches found later by the chain scanner rather than watched at birth. If asked "watched at creation?", the honest number is 1.80 million, not 1.90 million.

Median time from the creation transaction to our hashed copy of the launch document: 0.19 seconds (all-time median 0.206 s; last 24 hours 0.187 s, p90 1.8 s, p99 8.7 s; last 7 days 0.125 s). The homepage publishes 0.3 s, which is conservative.

This is the collector lane, which subscribes to pump.fun and Raydium LaunchLab event streams and covers 79.7% of launches. The chain-wide scanner lane, which reaches every other launch program, captures documents a median of 167 seconds after block time. Both are far inside the fastest host decay observed.

Sample of 1,500 launches drawn at random from launches where we hold BOTH instruments, so every one was provably alive at capture. Each original address re-fetched 14 October 2026. 194 metadata and 213 image checks returned a refusal rather than an answer (gateway throttling) and are excluded from the rates rather than counted as gone; the denominator is definite answers only.

• content-addressed (IPFS): n 1,010, definite 823, still serves 823, gone 0, rate 100.0%, byte-identical 100.0%
• self-hosted metadata: n 490, definite 483, still serves 364, gone 119, rate 75.4%, byte-identical 98.9%
• self-hosted images: n 490, definite 472, still serves 439, gone 33, rate 93.0%, byte-identical 87.9%
• metadata.j7tracker.io: n 230, definite 230, still serves 123, gone 107, rate 53.5%, byte-identical 100.0%
• pooled metadata: n 1,500, definite 1,306, still serves 1,187, gone 119, rate 90.9%, byte-identical 99.7%

SAY THIS PER HOST, NEVER POOLED. Content-addressed storage lost nothing: zero of 823. The loss is entirely on hosts the launcher controls. On metadata.j7tracker.io, which carries 266,784 of our documents, 46.5% no longer serve. Applied to that host's holdings this implies roughly 117,000 documents gone, which independently corroborates the 28 September availability exhibit's projection of about 112,000 within 5%. Two samples, two methods, two dates.

VOLUNTEER THE WEAKNESS BEFORE THE ROOM FINDS IT. Two thirds of launches put their document on content-addressed storage, where the address is the hash and the content cannot be altered or decay while anyone pins it. For those, our copy is convenience rather than scarcity. Say so first.

Exact, not a sample: every document we hold from metadata.j7tracker.io checked against the Internet Archive's complete index for that host, every path.

• metadata: ours 266,784, in the Internet Archive 110, in both 32, share of ours 0.012%
• images: ours 47,810, in the Internet Archive 3,216, in both 998, share of ours 2.09%

Images have a second public copy about 174 times more often than metadata. Crawlers follow images and rarely fetch a bare JSON endpoint nothing links to. This is why the two instruments need separate sentences.

Where both copies exist, the Internet Archive's bytes hash identically to ours. That is independent corroboration of the method and it gets stronger as more copies exist, not weaker. In all 23 overlapping cases measured, our capture was first: median 0.238 seconds against the Internet Archive's median of about 2.4 hours. Caveat: the byte-identity comparison against Internet Archive copies rests on 3 documents hand verified. Do not present that as a rate until more are checked.

Three birth artifacts produce an edge, each matched by hash:

• The launch document, matched on the hash of the whole file as served at launch. One file carries the name, the ticker, the description, the image address it pointed at, and any links. A document match subsumes all of it.
• The image, matched on the hash of the actual bytes, never the URL.
• The ticker pair, symbol and name together, matched as a unit.

Separate from the content layer: creator or signing wallet; funder of that wallet (the money graph, which is blind to an operator funding every wallet from a clean source, which is why the content layer exists); curve buyers, and whether buyers across launches trace to a common funder.

Context carried on a connection: the exact second, the creator's share of supply in the first block, graduation where confirmed against the curve account, and the launch program.

ADMISSION RULES, which belong in the deck because they are the credibility:

• Two or more shared artifacts, or two different instrument kinds. One shared artifact is a lead, never an edge.
• An artifact shared by more than 20 distinct creators is a template, a default or a widely circulated image, and contributes nothing.
• A connected group larger than 100 creators keeps its edges, each with its evidence, but receives no cluster label, and that exclusion is stated rather than silent. At that size the instrument stops resolving individual hands and saying so is the honest output.
• No pair rule separates a coordinated operation from a popular image at pair level. This was measured and stricter rules were rejected: they dissolved the cultural component but fractured known true cases with it. What separates them is group size. The largest genuine component held 30 creators; the cultural component held 12,287; nothing sat between.

Report all tiers with their composition stated. Rank none of them: which tier matters is the reader's question, not ours.

• declared a document byte-identical to another launch's: 682,229 (37.4% of documents held)
• from 2 or more distinct wallets: 476,567 (26.1%)
• carrying 2 or more distinct tickers: 29,527 (1.6%)
• different ticker and different wallet: 28,829 (1.6%)

Method note for the deck, stated as a property and never used to hide rows: where the name, ticker and launch tool are identical, a byte-identical document is expected, so the match confirms rather than reveals. Where the tickers differ, the file shows a connection the names do not.

1. A match needs both files. One document proves nothing. 2. You can only compare what you captured, not what you can find later. 3. The files do not wait. On the largest self-hosted launch-metadata host, 46.5% are already gone from the only address they ever had. 4. Therefore these connections are findable only by someone who captured every launch at the moment it happened. Partial coverage cannot do it. The Internet Archive holds 110 documents from that host; a group of 42 launches cannot be found inside 110 documents.

Measured: of the 119 documents in the sample now gone from their only address, 23 carried a connection to other launches, tying 245 launches together through files that are no longer served anywhere. Treat the 23 of 119 as roughly one in five rather than a precise rate; the 682,229 and the specimens below are exact counts from the full record.

USE THIS ONE. $CASHTAGUU, "FIRST COIN ON CASHTAG". 30 launches declared this byte-identical document from 6 distinct wallets between 02:38 and 02:43 on 22 September 2026. The document was served at metadata.j7tracker.io/metadata/c4ae927565a04d53.json and now returns 404. We read it 0.214 seconds after the creation transaction. sha256 0da573c2c1dfa7b17fd358f8ed741bff99917d42af6c33b6c27942c4304d5f97

DELETION WITH MONEY BEHIND IT. $LONG, launched 22 September 2026, completed its bonding curve. Its launch document is no longer served at metadata.j7tracker.io/metadata/K6YRBVVbiF.json. We read it 0.373 seconds after creation.

A THIRD, available if a larger group is wanted: 27 launches from 10 distinct wallets inside seven minutes on 16 September 2026, document now 404.

DO NOT LEAD WITH THE LARGEST GROUP FOUND. It is 42 launches from 14 wallets inside 42 seconds on 18 September, and its launch name is an accusation against an unnamed third party. Publishing launch names as recorded is settled policy for the register's own pages, but as the lead example in front of counsel it amplifies an unproven allegation, which is the one thing the register never does. $CASHTAGUU makes the identical point carrying no accusation.

• The timestamp has no external anchor. The capture time and the hash both live in our database. Nothing public pins them to a moment nobody controls. A daily published hash root anchored in a public ledger would convert this from a business record into something closer to self authenticating. Not built, not costed, and whether it changes how the record would be introduced is a question for counsel.
• The image instrument needs a look before image hashes carry weight. Where a launch declares its artwork on a trading terminal's content delivery network, our stored copy may be that network's re-encoding rather than the launch bytes. Observed image capture lags of 518 s and 557 s on such launches against 0.19 s for documents. A measured 11.6% byte mismatch on self-hosted images was traced almost entirely to one terminal's transcoding, not to anyone altering artwork, and that figure must not be used.
• The availability exhibit was corrected on 14 October 2026. Its earlier statement that a public second copy was something "we could not find" is withdrawn on the page itself with a dated note; the host-wide check found 32. The projection it sat beside is unaffected and is now corroborated twice.
• The register's own coverage window is 2 September 2026.

All published documents