A personal website is not a publication. It is a position. A publication presents work in a finalized, immutable state, and carries with it some sort of declaration - “this is my most polished and prized work!” - a position is something you inhabit, argue from, and continuously revise in public. This page explains the design decisions forming my broader position and why they took the form they did.
What follows is a colophon in the grand old sense: a printer’s note at the end of the book, recording how it was made, who made it, etc. The difference: here, the printer and the author are the same person, and the process of making is itself not only a form of argument, but the only form of argument permitted.
Living Documents
The dominant convention of academic and professional publication is that a document, once released, is finished. It carries an implicit claim: this is what I think, full stop.a I find this convention dishonest in proportion to how seldom it is actually true. Thinking is continuous; positions shift; evidence accumulates; people change their minds and rarely say so in public. This site operates under a different premise, one that I strive to operate all of my life under.
Every essay and post on this site carries an epistemic footer — a structured block that reports my current relationship to the work. We package the values into an “epistemic seal” that combines the page’s mark (an SVG illustration, more on this later) with the values; a card with more details can be obtained upon hovering over this illustration, and the page footer will be the comprehensive guide. The footer only appears when a status field is set in the document’s frontmatter; standalone pages and some short items omit it.
The vocabulary below is genre-general, but reads differently across genres. For a personal essay, confidence reflects credence in a thesis — “I might change my mind.” For an empirical research paper, it reflects expected generalization — “this would replicate.” For formal mathematics, it reflects credence in proof correctness, with a special value proved available for theorems with complete proofs (where any numeric value would be false precision). Evidence reads analogously: the strength of arguments and supporting writing in essays, the empirical base in research, the structure of the proof in mathematics.
The full set of fields:
Status — a controlled vocabulary describing where the work stands: Draft, Working model, Durable, Refined, Superseded, or Deprecated. A document marked Working model is not just unfinished — it is a position I currently hold but would not stake much on. A document marked Durable is something I expect to hold up. Superseded means I wrote a better version; Deprecated means I no longer endorse it.
Confidence — an integer from 0–100, representing my credence in the central thesis. Not false precision: a rough honest assessment is more useful than no assessment at all. When a
confidence-historylist is present, a trend arrow (↑ ↓ →) is derived automatically from the last two entries — so you can see not just what I think but whether I am growing more or less confident over time.Importance — how much I think this matters, on a 1–5 dot scale (●●●○○). Useful for orienting a reader who has limited time.
Evidence — how well-evidenced the claims are, on the same 1–5 scale. An essay with high importance and low evidence is a speculative position and should be read accordingly.
Trust score — a single 0–100 integer derived automatically from confidence (weighted 60%) and evidence (weighted 40%, with the 1–5 scale rescaled so that evidence=1 contributes zero and evidence=5 contributes the full 40 points). It is deliberately narrow: it answers “how much should you trust the central claim?” and nothing else. It says nothing about how broadly the work matters, how novel it is, or how useful it is in practice — those are separate axes (see below) that are deliberately not folded into a composite, so a high trust score on a personal essay cannot be misread as “world-shaking.” Following Gwern’s lead, the orientations are presented in parallel rather than blended into a single index. The score is not entered manually and lives only in the epistemic footer.
Scope, Novelty, Practicality — orientation fields shown as their own rows in the epistemic footer alongside confidence, importance, and evidence. Scope ranges from personal to civilizational; novelty from conventional to innovative; practicality from abstract to exceptional. These are not ratings — they are orientations, and they intentionally do not feed the trust score.
Peer status — the external review state, distinct from
status(which is my internal position). Values: unreviewed (default), under review, peer reviewed, published, retracted. A piece can be Durable (I expect it to hold up) and unreviewed (the world hasn’t checked yet) at the same time; the two axes are deliberately factored. A retracted piece renders with the field name struck through and the outer ring of the epistemic figure crossed out.Result shape — the shape of the central claim: positive (argues something works), negative (argues something does not), mixed (both, as in a double-pincer barrier paper), comparative (compares approaches), or descriptive (describes without arguing for or against). Encoded as a small glyph beside the trust score on the epistemic figure. Adds nothing to the compact row.
Stability — auto-computed at every build from
git log --follow. The heuristic: very new or barely-touched documents are volatile; actively-revised documents are revising; older documents with more commits settle into fairly stable, stable, or established. This requires no manual maintenance — the build reads the repository history and makes the inference.
The version history block, directly above the epistemic footer, uses a three-tier fallback: authored history: notes when they exist (written by me when the git log alone would not convey what changed), then the raw git log, then the date: frontmatter field as a creation record. make build auto-commits any changed content files before the Hakyll compilation runs, so the git log is always current.
The Current page extends this premise from essays to ongoing work. Every research project listed there carries its own updated: timestamp and a status drawn from a controlled vocabulary — in-review, revising, drafting, building, early-stage, paused — and the page itself wears a masthead “last updated” date as its thesis. The page has no epistemic footer because it isn’t an argument; it is, rather, the closest thing this site has to a publication, and yet by design it is the part of the site most committed to being out of date the moment you finish reading it.
The point of all this is simple: when you read something on this site, you should know what kind of claim I am making. The date a document was last modified is not decorative. A 40% confidence rating is not self-deprecation. The system is an attempt to make explicit something that most writing leaves implicit — where the author actually stands.
Marks
The Marks are an SVG illustration that I devise for each page. They are intended to visually say what the piece is about, both on the page itself and in the preview popups. Different marks have been devised over different matters, but in general I try to come up with a list of ideasb and then have an LLM mock them up so I can compare. The final mark is then created through an iterative process with the LLMs and manual tweaking.
In the future, the website might use other illustrative means for specific pieces, but at present I have not decided to make this a priority, and the engine does not support it.
Typography
You are reading this sentence in SPECTRAL, which is not only a font with particular personal importance to me, but also an exceedingly pleasing font to read. OpenType features — smcp, onum, liga, calt are used throughout the website, which necessitates our self-hosting setup.c
The UI and Headers are also Spectral. I previously used Fira Sans, but came to strongly dislike that look on the website, and have moved to unifying the bulk of the text in Spectral. Code uses IBM Plex Mono; we previously used Jetbrains Mono, but I have since come to decide that IBM Plex Mono matches the feel of the website and is a bit more readable at a glance for speedy eyes.
Dropcaps
This website uses dropcaps, though currently sparingly - I am rolling out dropcaps incrementally as I would prefer to take the time to get them right. A personal project of mine, formschneider, is the primary system used for creating the dropcaps, their animations, and subtle variations - see that project’s repository for more. Other dropcaps that are used or adapted into variations in formschneider are freely licensed.
The Rest of the Look
The website avoids color where possible, preferring a “paper and ink” look. The palette is an application of restraint grounded in my studies of Tools for Thought. A prominent example is that there is no accent color on the website, for instance when one hovers over a hyperlink. I experimented with many different potential accent colors and found that all of them felt they took away from the paper and ink look, rather than contributing something to it. Color is often used to do work that typography should do, such as demonstrating hierarchy, creating emphasis, etc. When those functions are handled by weight, size, and spacing insteadd, color becomes available for the things it cannot be substituted for — and on a site with no data visualizations requiring color encoding, those things turn out to be very few.e Previously, colors weren’t even used in syntax highlighting of inline code, but I have since decided that this is another useful exception.
A major redesign of the website’s visuals in November 2026 brought a few new features and variations. Firstly, a previous topbar was replaced with a masthead in the left column that scrolls away to reveal an empty margin. The top bar in the previous iteration of the website was originally conceived as a mechanicism for quick navigation of the website; I’ve since realized that I rely on hyperlinks and otherwise and that the bar was simply taking up space rather than serving this purpose.
Secondly, the background of the website allows some variation now, between a plain color and two variations. I experimented widely with many different potential variations and chose two that are subtle and stay out of the way of the content. Others are certainly possible, but in aiming for the look described above, I found these to be the most effective candidates.
Thirdly, a subsystem for special themes was introduced. The first run of this was the Halloween 2026 theme, and more special themes will come from time to time. These operate with a predefined date range in which they supersede the default, but can always be opted out of.
Finally, several options are configurable, and some are automatically configured. Text size can be increased or decreased; we automatically adapt it in certain instances, such as for mobile devices. We also provide a focus mode and reduced motion as configurable options at all times, and more configurability may come at some point.f
Nothing is watching you
The site has no analytics, no visit counters, no fingerprinting, and no third-party scripts.g The Hetzner VPS that provides this content runs only open source software, and my machines use almost exclusivelyh the same. The code is licensed under the GPL and hosted on a self-hosted Forgejo instance at this domain, with a GitHub mirror; you are welcome to inspect it, fork it, or build on it under the terms of its license.
Your browser will make external requests, of course, when visualizing popups for hyperlinks and in the event that you click any outbound hyperlink. At present, these requests can be made to arXiv, Crossref, GitHub, Open Library, and Wikipedia, among others. Requests can also be made if you peruse the photography map tiles. In the website repo, nginx/security-headers.conf is the place to look if you are curious.
The .git subdomain of this website runs my personal Forgejo instance. Unfortunately, due to an influx of automated traffic, Anubis is used alongside some analytics and statistical analysis software that I have written entirely from scratch. Nginx logs and CSP reports, along with analysis of failed Anubis challenges, are monitored and reviewed for the subdomain. I have tried to find a middle ground that avoids analytics on the main domain (which I feel quite strongly about; I do not want the typical analytics and privacy invasions that come with most websites!) with success at the present moment.
The Build
This is a Static Website. For the purposes of this website, though the content is highly dynamic and iterated upon, the medium of expression is rather stable. There are numerous advantages to using a static webpage, many of which are focused at the Hetzner box from which this webpage is served.
The engine of this website, Ozymandias, uses Hakyll for reasons of performance, extensibility, and, of course given the underlying language Haskell, elegance! We refer you to the upcoming Ozymandias article for further technical information. I had been wanting to do a project in Haskell ever since I took my undergraduate programming languages course,i and Hakyll was more extensible and thus suitable than the alternatives I reviewed. The philosophy of a static website is that the website is a program and the content is the source codej. This engine is licensed as free software and you are welcome to use it for your own website if you wish; it is highly extensible and configurable, both with features beyond those present on this website and the ability to condense the enabled features to a subset of what is used here.
Mathematics on this website is written at build time and drawn by the browser. This is a tradeoff; we previously shipped rendered mathematics that was not drawn by the browser, but have since switched, and use MathML.
A full accounting of what this build process has actually produced is available at the Build page. It is generated automatically at each compile: corpus word counts, length distributions, tag frequencies, link analysis, epistemic coverage, repository metrics, and build timing — all computed from the same source tree that produces the content. Think of it as the build system reporting on itself.
AI and This Site
Ac proinde haec cognitio, ego cogito, ergo sum, est omnium prima & certissima, quae cuilibet ordine philosophanti occurrat.
I will never have an AI system of any type autonomously write anything that goes out under my name, whether in my personal communications or in pieces on this website. This is a boundary of authority: I keep the final say and every direction call, without exception, and the argument a piece makes is mine. For substantive writing — on this website in particular, essays, letters — I work entirely by hand. Writing is religious in severity to me, and I do not intend to outsource it. Doing so would also constitute outsourcing of my thinking, and though I may not agree with everything he says, I won’t go against Descartes so easily.
I hope it is clear that this is a narrower claim than “I never touch the tools.” I use AI extensively as something to ideate with, typically through arguments. I will use it to get relentless pushback on an idea, to surface objections and similar writings, and to draft what I call “variant outlines,” which serve to help me consider multiple various structures for a project before I commit to trying onek. I am quite averse to having AI generate entire drafts for me, and will absolutely under no conditions let it “hold the pen” on something I sign.
I do treat technical write-ups as an exception. Where a piece exists to document an artifact — say, for example, a case study about code anyone can read, results anyone can rerun — I am more willing to work with AI on the prose, because the artifact is the claim and the write-up is secondary to it. Where the writing is the artifact, the paragraph above applies without qualification.
The majority of the code that comprises the build system of this website was created in collaboration with AI. The initial versions of the website and Ozymandias were much more human implemented, with roughly equal contribution from myself and LLMs in terms of pure code output. As models have progressed substantially since that time, a point has been reached where it is no longer efficient for me to write the website’s code myself, and as of late 2026, I no longer contribute very much code at all to the website. All code that makes it into the repositories of the website and engine, however, is reviewed by me — and in that case, I have taken more of a higher level role, making administrative decisions about directions, doing extensive code review, and occasionally implementing a bit by hand, usually at this point constrained to test coverage.
Influences and Attributions
The amount of influences on this website is immense, and cannot be detailed in the fullest extent. Every other webpage that I have visited, whether beautiful or pitiful, has evoked some type of reaction or response in me, and that response has played some role, even if minute, in the design of this website.
Code
Certain code on this website or in the Ozymandias engine underneath it was adapted from other websites or systems; I stand on the shoulders of giants. Pandoc and texmath, both projects of John McFarlane, are core components of this website’s toolchain, and the website would not possibly exist without them; Hakyll has already been mentioned, so thanks to Jasper Van der Jeugt; the website’s search is built on transformers.js by Joshua Lochner and all-MiniLM-L6-v2 by Reimers and Gurevych; Lilypond and MuseScore are currently used for engraving musical compositions displayed on this website (but will be superseded soon by my upcoming Epiphany; thanks is also due to Mozilla for PDF.js.
Ideas
Many others influenced the ideas that make this website what it is. The originall sidenotes script was inspired by the website of Gwern Branwen, and, in fact was a fork of that website’s script (authored by Said Achmiz); the build telemetry and much inspiration on design principles was taken from Matthew Butterick’s Practical Typography and, by extension, the course websites implemented in Racket of Shriram Krishnamurthi, whom I was fortunate to study under at Brown; the influence of tools for thought can be directly traced to CSCI1377 at Brown and Will Crichton, and more broadly attributed to Andy Matuschak and Michael Nielsen; the /now page is inspired by that of Derek Sivers; of course, I should probably also name Tufte once more here to close.
The tradition of the personal website is one that is built on a sense of community and interaction. I am thankful to everyone else who has a personal website and shares their content with the world. I am also particularly greatful to the open source and broader open culture movements, who have given me and the world so much. This website would not exist without you - and I wouldn’t be the person I am without your influence - what a role model!
The Future
This site is unfinished, deliberately! The nature of (and beauty of) living documents is that they can never be finished in the absolute sense. Thus the website has a future for as long as I have one. I think that is something to appreciate.
This is particularly problematic in academia, where there is a long tradition of researchers whose work was eventually disproven taking an extreme defensive stance, usually rooted in confirmation bias.↩︎
unless I have one clear standout; an example of that is the Crawler Chronicles, where I knew exactly what I wanted from the onset!↩︎
Google Fonts strips OpenType features during subsetting for bandwidth. Self-hosting with
pyftsubsetpreservessmcp(small capitals),onum(old-style figures),liga(common ligatures), and the full optical size range. The difference is visible: old-style figures sit on the baseline rather than hanging above it, small capitals are drawn to match the x-height rather than being shrunken full caps, and ligatures prevent collisions in letter pairs like fi and fl.↩︎Color and saturation/hue are actually well known to be less effective than other means of distinguishment. I refer you to Tufte’s The Visual Display of Quantitative Information for more.↩︎
Ones exception is the heatmap on the statistics page, which uses a four-step greyscale scale. Even there, the encoding is luminance rather than hue.↩︎
I promise you, much more has been prototyped, but finding what’s actually appropriate for the website’s feel, which, if you couldn’t tell, is quite important to me, is a very difficult balancing game.↩︎
This is enforced at the nginx level via CSP headers, not just by omission. The Content Security Policy prevents any script not explicitly whitelisted from executing. The whitelist is short.↩︎
It is nearly impossible to run an entirely free system, but in approximation, it is actually wonderfully easy.↩︎
The programming languages course at Brown, somewhat infamous, is taught entirely with Racket, which is essentially a dialect of Lisp. The course itself is extremely focused within the functional paradigm as far as implementations go. I am aware that Racket itself, curiously, has some means by which static webpages can be built - the course infrastructure was produced this way, and this website has in a few places taken minor inspiration from it!↩︎
Source code here chiefly distinct from mere markup language, like HTML.↩︎
This is something that I have long struggled with. I can vividly remember being in fourth grade and getting feedback on my essays that I would probably still get now, along the lines of “this could be organized to be more conducive to the point you are making!” The process of struggling against organizational challenges, which at a high level are perhaps the challenges involved in any task of considerable complexity, is a process I still greatly enjoy and will not fully automate. AI thus serves to augment me in the way that I propose it should augment us all in other writings on this website.↩︎
This is no longer the case, as the system has been redesigned from the ground up.↩︎