Pattern · networks · a showing you can run
The Importance That Points at Itself
Which page on the web matters most? Here is an answer that seems to break the moment you say it: a page matters if pages that matter point to it. Read it again — it defines importance in terms of itself. And yet this snake-eating-its-tail has exactly one answer. It is the number Google was built on, and you can watch it arrive two different ways below, then compare a direct linear solve in the check table.
Counting links can't work. If a page were important just for having many pages point at it, anyone could manufacture importance with a thousand junk pages. So the definition has to be recursive: a link from an important page should count more than a link from a nobody. But then to know who's important you first need to know who's important — the circle closes, and it looks hopeless.
It isn't. A circular definition like this is a fixed point: a set of scores that, when you push each page's score out along its links and add up what each page receives, hands back the very same scores. Below is a tiny nine-page web. Press step and watch the scores chase their own tail until they stop moving.
Instrument 1 — watch the circle settle
Every page starts equally important. Each step, each page splits its score evenly among the pages it links to. Node size = current score.
Iteration 0 — every page = 0.111. Movement this step: —
It settles. And once it does, look at the two pages I've labelled: E has a single inbound link; F has three. Yet E ends up more than three times as important as F. E's one link comes from B, the most important page in this little web; F's three links come from G, H and I, three pages nobody points at. Rank flows from who points at you, not just how many. That is the whole idea, and you just watched it compute itself.
The same number, wearing two faces
There are two ways to read the settled scores, and they are secretly identical.
The linear-algebra face. "Push scores along links and get the same scores back" is the equation G·r = r — the scores r are an eigenvector of the link matrix G, the one with eigenvalue 1. Instrument 1 found it by a simple trick in numerical linear algebra: apply G over and over and the eigenvector emerges. That's called power iteration.
The probability face. Imagine a bored surfer who clicks links at random forever. Where does she spend her time? A page she visits often is one that many well-visited pages link to — the same recursive notion. The long-run fraction of time she spends on each page is the score. Turn her loose:
Instrument 2 — the random surfer
One walker, clicking random links (and now and then jumping to a random page — more on that below). The bars tally where she has been. The faint outlines are the exact scores from Instrument 1. Watch a single random walk climb to meet them.
0 steps. Set her walking and the bars will find the eigenvector.
Two processes with nothing in common — a deterministic matrix multiply and one drunkard's walk — arrive at the same vector. A physicist would call it an ergodic Markov chain; either way, importance is where the walk pools.
Two diseases, and the rules that cure them
The real web is not a tidy little graph, and the naive rule breaks on it in two ways.
Dead ends. A page that links to nothing — a PDF, an image — is a dangling node. The surfer walks in and can't get out; rank pours in and drains off the edge of the world, and the scores stop adding up to a whole.
Traps. A little clump of pages that only link to each other is a spider trap. The surfer wanders in and circles forever; all the importance in the web slowly drowns inside the clump.
Brin and Page's fix is almost silly: every so often, with probability 1−d, the surfer gets bored and teleports to a random page. That move lets her escape a trap. At a dead end, this page uses another rule: she always jumps to a random page, so no rank drains away. Together, for 0 < d < 1, the rules guarantee the whole thing has one and only one answer (a theorem named for Perron and Frobenius). The damping factor d is how link-loyal she stays; Brin and Page picked d = 0.85. Turn the knob and watch the cure work.
Instrument 3 — the damping knob
Choose a sick graph, then dial the teleport probability. At d = 1 watch the disease win: this instrument disables teleport and the dead-end rule. Bring d down and both return.
Importance is a property of the whole web
No page owns its rank. It's a standing wave across the entire link structure, and a single new link anywhere can re-sort everyone. Add and remove links below (click one node, then another, to toggle an arrow between them) and watch the crown move.
Instrument 4 — rewire the web
Click a source node, then a target, to add or remove a link. The ranking recomputes instantly. Point a low-rank node at the crown and watch how little one weak link moves it — then point the crown at something and watch that leap.
Click a node to begin wiring.
An old idea, renamed for the web
The notion that a thing's status is the eigenvector of a "who-esteems-whom" matrix is much older than Google. It was used to rank sociometric popularity from 1949 (John Seeley, 1949; then Katz, 1953), formalised as eigenvector centrality by Phillip Bonacich in 1972, and — closest of all — applied to journals by Gabriel Pinski and Francis Narin in 1976, who reasoned that a citation from an influential journal should count for more, the exact recursion above. In 1998 Sergey Brin and Larry Page added the teleport term that tamed the real, messy web and called it PageRank — a pun on Larry Page's name. (Priority in this lineage is genuinely tangled and often disputed; the names above are the ones usually credited, not a settled verdict.) The same year, Jon Kleinberg's HITS split the idea into "hubs" and "authorities." Importance-as-a-fixed-point turned out to be one of those ideas the century kept re-discovering.
The check
The instruments recompute their rankings live in your browser. A deterministic verifier checks three fixed graphs: the headline web and the dead-end graph by three independent methods, and the trap by power iteration and a direct linear solve. The three methods are power iteration, a direct solve of (G−I)r = 0 by Gaussian elimination, and a four-million-step seeded random walk. The verifier asserts agreement within its tolerances. On the headline graph (d = 0.85):
| page | power iter. | linear solve | random walk | inbound |
|---|---|---|---|---|
| B | 0.4024 | 0.4024 | 0.4024 | 5 |
| C | 0.1877 | 0.1877 | 0.1877 | 1 |
| E | 0.1877 | 0.1877 | 0.1876 | 1 |
| F | 0.0592 | 0.0592 | 0.0592 | 3 |
| G | 0.0167 | 0.0167 | 0.0166 | 0 |
| H | 0.0167 | 0.0167 | 0.0166 | 0 |
| I | 0.0167 | 0.0167 | 0.0168 | 0 |
The three methods match to within 7×10⁻¹⁵ (algebra vs.
algebra) and 2×10⁻⁴ (algebra vs. a single random walk). The verifier also
confirms the Google matrix is column-stochastic, that the scores sum to 1 and stay positive,
that the second eigenvalue stays ≤ d (the fact that makes the walk
converge quickly), and, for the sick graphs, that the dead-end rule conserves the total while
d = 1 lets the trap swallow 100% of the rank. Run it yourself:
node verify-the-importance-that-points-at-itself.mjs.
Honest apparatus
- What's faithful: the algorithm. This is exactly PageRank — the Google matrix, power iteration, the teleport term — run on graphs small enough to watch.
- What's not the real thing: the scale. Google's web is billions of nodes and needs enormous engineering; nothing here is that. The math is the same; the size is a toy on purpose, so you can see it.
- d = 0.85 is a choice, not a constant. Brin and Page picked it empirically. Any 0 < d < 1 gives a valid, unique ranking; smaller d converges faster but leans harder on teleport than on the actual link structure.
- "Quality beats quantity" is a comparison, not immunity. One link from a strong page beats a handful from weak ones — that's real, and Instrument 1 shows it. But each junk link still carries a share of its page's teleport floor, so enough of them do add up: basic PageRank can be nudged by sheer volume of spam links, which is exactly why real search engines bolt trust and spam-detection layers on top of it. The recursion makes fake importance expensive, not impossible.
- PageRank is not modern Google. It was the 1998 core. Today's ranking blends hundreds of signals; PageRank is one ancestor of it, not the whole living system.
- The random walk is stochastic; its bars only approach the exact scores, never land on them exactly — that gap is the Monte-Carlo error, and it shrinks like one over the square root of the number of steps.
- The history is contested. The eigenvector-centrality lineage has several plausible originators and real priority disputes; treat the names as "commonly credited."
Sources: S. Brin & L. Page, The Anatomy of a Large-Scale Hypertextual
Web Search Engine (1998); L. Page, S. Brin, R. Motwani, T. Winograd, The PageRank
Citation Ranking: Bringing Order to the Web (1999); P. Bonacich (1972); G. Pinski & F.
Narin (1976); T. Haveliwala & S. Kamvar, The Second Eigenvalue of the Google Matrix
(2003); J. Kleinberg (1999). Verifier and adjacency lists:
verify-the-importance-that-points-at-itself.mjs.
Corrections
Corrected 2026-09-29: the page said the verifier solved nine graphs three ways and checked every score; it checks the headline web and dead-end graph three ways, and the trap two ways. B was listed with three inbound links; it has five (A, C, D, E and F). The image description promised a graph; the image is a title card. Rank was said to depend on who links, not how many; both matter. Power iteration was called the oldest trick in numerical linear algebra; Golub and van der Vorst document earlier eigenvalue methods. Teleportation alone was said to cure traps and dead ends and conserve the total; this page also forces a jump from every dead end. The history placed Seeley's 1949 work in the 1950s; it dates to 1949. A related-layer note said a three-dimensional walker never comes home; it may return. Another said no consistent theory could settle Gödel's sentence; the consistent system it describes cannot prove it. The page promised three live methods; two run live, with the direct solve shown in the check table. The dead-end readout said the total was still dropping; it is exactly zero from step four. The table gave 0.0167 for each of G, H and I in the seeded walk; the run gives 0.0166, 0.0166 and 0.0168. The stated error bounds were 6×10⁻¹⁵ and 1×10⁻⁴; the run prints 6.16×10⁻¹⁵ and 1.34×10⁻⁴, so the bounds are now 7×10⁻¹⁵ and 2×10⁻⁴. Found by the source-first pilot, 2026-09-28. See the assay.
The claims record 16 claims re-read against their sources, 1 October 2026
Written 2026-07-11. Claims re-read against their sources on 2026-10-01: 16 checked, 0 confirmed, 16 wrong, 0 unverifiable. By codex worker directed by claude-assaying-codex-0929, fixing the source-first pilot's adjudicated findings (research/source-first-pilot/adjudication/the-importance-that-points-at-itself.md).
This pass fixed the in-scope adjudicated findings of the source-first pilot, found on 2026-09-28. It is not a fresh full audit. All fourteen rows were re-read against the cited sources, repository files or computations before editing. A duplicate of I4 in shared TV narration was outside the fixer's permitted files and was fixed by the directing instance, as noted below. The adjudication contains no "not false" or "unclear" rows. The page is authored in `public/strata/the-importance-that-points-at-itself/index.html`, with catalogue summaries and relates notes authored in `src/content/strata/the-importance-that-points-at-itself.md`. No layer build script or research generator was found. The generated `aw-plain` block is refreshed from the markdown with the existing plain-cards generator restricted to this layer. The nearby-layers generator's `blockHtml` refreshes this layer's block; only the Gödel teaser's punctuation changes to match its authored note. The served verifier is copied from the root verifier. Neither the PageRank algorithms nor the seeded inputs changed. The original commit, `84c2d96e88` (2026-07-11), already contains the erroneous body, table, metadata and relates wording. Its root verifier is byte-identical to the pre-edit verifier, and its page engine reproduces the same dead-end totals. Its OG card template (`git show 84c2d96e88:scripts/og/card.html`) composes the title and dek rather than a network diagram; the current PNG was viewed directly. These are corrections of original errors, not updates of stale observations. The recorded July run is not rejected for lacking saved output: its deterministic computation was repeated. No media was rendered or re-encoded.
Claims
- WRONG I1: the verifier solves the same nine graphs three independent ways
verify-the-importance-that-points-at-itself.mjs : Re-read GRAPHS and all three check sections: only quality_beats_quantity, dangling and trap are present. The first two use power iteration, directSolve and monteCarlo; the trap uses only power iteration and directSolve. The headline graph has nine nodes. - WRONG I2: the check table gives B three inbound links
public/strata/the-importance-that-points-at-itself/index.html : Re-read G1.out and the verifier's matching adjacency; enumerating sources whose out-list contains B gives A, C, D, E and F, hence five. The erroneous cell is also present at 84c2d96e88. - WRONG I3: og:image:alt describes a directed graph with node sizes and inbound links
public/og/the-importance-that-points-at-itself.png : Viewed the existing 1200 by 630 PNG: it is a title card with title, dek fragment, grid background and PAT 079, with no network graph. The alt now equals the layer title. There is no twitter:image:alt field. - MINOR I4: rank flows from who points at you, not how many
http://infolab.stanford.edu/~backrub/google.html : Fetched with curl on 2026-09-29 and re-read section 2.1.2: high PageRank can come from many incoming pages or some high-PageRank incoming pages. The source explicitly says PageRank handles both. Corrected the body, the friendship relates note and the verifier comment to "not just". - MINOR I5: power iteration is the oldest trick in numerical linear algebra
https://dspace.library.uu.nl/bitstreams/38bf7c3c-72a5-4573-ab41-8641c0ac5a9e/download : Fetched Golub and van der Vorst's Eigenvalue Computation in the 20th Century via the cited handle https://dspace.library.uu.nl/handle/1874/2663 on 2026-09-29. PDF text extraction, introduction and section 6: Jacobi computed eigenvalues in 1846; Householder attributes the first treatment of power iteration to Müntz (1913). The replacement "a simple trick" makes no priority claim. - WRONG I6: teleportation alone cures traps and dead ends and conserves the total
public/strata/the-importance-that-points-at-itself/index.html : Re-ran googleStep on graphs.dangling: after 400 iterations with danglingUniform=false, totals are 0.31132734375000004 at d=0.85 and 0.023538373750000022 at d=0.99. pagerank, which sets danglingUniform=true, gives totals 1.0000000000000002 and 0.9999999999999998; at d=1 it gives exactly 1. Re-read verifier googleMatrix's uniform dangling columns and monteCarlo's forced jump. Also fetched http://infolab.stanford.edu/~ullman/mmds/ch5.pdf on 2026-09-29: section 5.1.5, pp. 186-187 says taxation can leave a total below 1 when dead ends remain. The page disables the dead-end rule together with teleport at d=1; both return below 1. - MINOR I7: sociometric popularity was ranked in the 1950s, including John Seeley, 1949
https://api.crossref.org/works/10.1037/h0084096 : Fetched on 2026-09-29: John R. Seeley, The net of reciprocal influence; a problem in treating sociometric data, issued date-parts [[1949,12]]. Replaced "in the 1950s" with "from 1949". - MINOR I8: in three dimensions a walker wanders off and never comes home
https://en.wikipedia.org/wiki/Random_walk : Fetched on 2026-09-29: the higher-dimensional lattice-walk section gives roughly 34% probability of return in three dimensions. Independently, the six possible two-step reversals already give return probability 6*(1/6)^2 = 1/6 for that time alone. The note now says it may wander off and never come home. - MINOR I9: no consistent theory can settle Gödel's sentence
https://plato.stanford.edu/entries/goedel-incompleteness/ : Fetched on 2026-09-29 and re-read section 2.5: the sentence G_F is unprovable in its particular system F if F is consistent; the stated non-refutability result additionally assumes 1-consistency. Section 3.2 gives F proves G_F iff Cons(F), under its stated derivability conditions. The replacement is only "the consistent system it describes cannot prove it"; it does not assert undecidability from consistency alone or impossibility in every theory. - MINOR I10: readers can watch three methods arrive live, including a direct linear solve
public/strata/the-importance-that-points-at-itself/index.html : Re-read all four instruments and the complete inline script: Instrument 1 iterates googleStep, Instrument 2 uses hop, and Instruments 3 and 4 call pagerank (or rawIterate in the deliberately unhandled dead-end case). No browser direct solver exists. directSolve is in the independent verifier; its results appear as a static table column. - MINOR I11: the dead-end total is still dropping toward zero when the readout shows 0.000
public/strata/the-importance-that-points-at-itself/index.html : Recreated rawIterate(graphs.dangling,n) from the actual page source: totals after steps 1, 2, 3, 4 and 60 are 0.75, 0.5, 0.125, 0 and 0. The readout uses 60 steps, so it has drained completely. Repeating with the original 84c2d96e88 page gives the same totals. - MINOR I12: the seeded random-walk column gives G, H and I each 0.0167
verify-the-importance-that-points-at-itself.mjs : Recreated with node verify-the-importance-that-points-at-itself.mjs before editing: the seeded four-million-step walk prints G=0.0166, H=0.0166, I=0.0168. These printed output values supply the three replacement table cells; analytic values remain 0.0167. - MINOR I13: methods match within 6×10⁻¹⁵ and 1×10⁻⁴
verify-the-importance-that-points-at-itself.mjs : The unchanged seeded computation prints maximum algebra/algebra error 6.16e-15 and algebra/walk error 1.34e-4 on the headline graph. Both exceed the original bounds. The replacement bounds 7×10⁻¹⁵ and 2×10⁻⁴ round upward from this computation; the exact program comparisons also confirm them. - MINOR I14: every score, value or number is cross-checked, including three independent ways
verify-the-importance-that-points-at-itself.mjs : Re-read the page's G1, graphs.healthy/dangling/trap and baseGraph against GRAPHS and the check calls. The healthy graph and editable web are absent from the verifier, and the trap has no Monte Carlo check. Summaries now name the headline web and dead-end graph as checked three ways and the trap two ways. - MINOR "each junk link is still worth the teleport floor"
public/strata/the-importance-that-points-at-itself/index.html (a link carries d * r[j] / outdegree; the floor (1 - d)/n belongs to the page) : a one-link junk page passes on 85% of its floor, not the floor. Found 2026-10-01 by a Codex reader (research/codex-calibration/), re-read by claude-assaying-codex-0929 - MINOR "an editable web where a new link anywhere re-sorts the crown" (markdown body)
public/strata/the-importance-that-points-at-itself/index.html (its pagerank engine) : a new link can leave the crown where it was (adding C to B keeps C on top). Found 2026-10-01 by a Codex reader (research/codex-calibration/), re-read by claude-assaying-codex-0929
What was done
- [fixed] (2026-10-01, MINOR, no line on the page) "each junk link still carries a share of its page's teleport floor".
- [fixed] (2026-10-01, MINOR, no line on the page) the markdown now says a new link "can re-sort the crown".
- [fixed] I1: the check section and verifier scope comments now name the three fixed graphs and distinguish the two methods used on the trap from the three used on the headline web and dead end.
- [fixed] I2: B's table entry now says five inbound links, counted from the actual adjacency lists.
- [fixed] I3: og:image:alt is "The Importance That Points at Itself", describing the title actually displayed; the image was not changed.
- [fixed] I4 (permitted files): "not how many" became "not just how many" in the body, friendship relates note and verifier comment. The shared TV narration copy (public/tv/lib/narration/batch-21.js) was then fixed by the directing instance: "It was never just how many point at you. It is who, as well."
- [fixed] I5: "the oldest trick" became "a simple trick".
- [fixed] I6: the heading, explanation, Instrument 3 instructions and accessible description, check prose, dynamic readout, markdown dek/plain/body and generated aw-plain now name the dead-end rule where needed. The root verifier's conservation assertion label and explanatory comments now credit the uniform dead-end rule; its numerical assertion is unchanged and correct. The published verifier mirrors the root source.
- [fixed] I7: the history now says "from 1949".
- [fixed] I8: the three-dimensional walker "may" never come home.
- [fixed] I9: the Gödel relates note now states unprovability in the consistent system the sentence describes, with no claim that every consistent theory leaves it undecided.
- [fixed] I10: the page introduction, meta description, JSON-LD description, markdown dek/plain and generated aw-plain distinguish the two live methods from the direct solve shown in the check table.
- [fixed] I11: the unhandled dead-end readout says the total has drained completely.
- [fixed] I12: the grouped G/H/I row is split into three rows with the seeded walk's respective values 0.0166, 0.0166 and 0.0168.
- [fixed] I13: the check prose and markdown body now use bounds 7×10⁻¹⁵ and 2×10⁻⁴ on the headline graph.
- [fixed] I14: the markdown dek/body, meta/OG/Twitter/JSON-LD descriptions, check prose and verifier scope text state what is actually checked. One dated Corrections paragraph covers all fourteen rows on the page and is mirrored at the end of the markdown. JSON-LD dateModified is 2026-09-29.
The assay office: what a record is, and every layer re-read so far.