Protocol STAMP is teamed with AI and can make mistakes. Verify before you implement anything. What we got wrong · Free to read, paid for by readers: Contribute
The finishes · the playground
Every page here is tinted by the day of the week — the day in Chicago, set before the page paints, so two readers on two continents see the same color. There are seven finishes, one per day, and they follow the oldest naming scheme still in daily use: the planetary week. Each day belongs to one of the seven classical “planets” — Moon, Mars, Mercury, Jupiter, Venus, Saturn, Sun — and each finish is tinted toward that body and the metal the old tradition paired with it.
If you only speak English this will look like an invention, because English hides it: Saturday, Sunday and Monday still carry Saturn, Sun and Moon, but Tuesday to Friday were renamed after Norse gods (Tiw, Woden, Thor, Frigg) chosen as stand-ins for Mars, Mercury, Jupiter and Venus — scholars still debate whether Friday honors Frigg or Freyja, and Onsdag comes directly from Old Norse Óðinsdagr, Odin’s day. In Spanish, French and Italian the planets are still audible — martes, mercredi, giovedì — down to the Hebrew/Latin-rooted exception of Saturday and Sunday, and in Japanese every one of the seven is written on the calendar with its planet’s element, inherited from the five elements of ancient Chinese astrology. This is the one page where you may change the day. Press a finish to dress this page in it and read its names; everywhere else on the site, the calendar decides, and the page returns to today when you leave.
Saturday is publishing day, so Saturn is the house finish. An issue wears the finish of the day it was issued, forever: The Calibration Vol 001 wears Sun. The finish tints the ground only — the bar in the mark and every measured figure stay the same red on every day, because a number should not change color with the calendar.
Why this page exists
A printed book keeps a colophon at the back: a short note naming the typeface, the paper, the printer, the edition. It is the one page that says how the thing was made rather than what it says. In an age when a website can be copied and pasted in an afternoon, the studio’s view is that the design of a site — and the reasoning behind each choice — is part of what it publishes, and should be shown rather than hidden. So every site under Budday Budderson Studio carries one of these. Nothing on it is decoration; every choice below has a reason, and where the reason is “it was drawn at random”, the page says so.
The name
A stamp is a mark pressed onto a document to certify it — the sense meant here. Not the stamp you buy and lick, but the one the clerk presses on top of it: the postmark, which records where a thing was and when. That is the sense the dateline on every issue and the 15:37 clock are keeping. It is a word rather than an acronym, and the five letters are not expanded anywhere on this site, and will not be: an acronym explained after the fact is exactly the retrofit this series exists to catch elsewhere. The word is old and widely used, so several unrelated things share it, from an IETF network-measurement standard (RFC 8762, which networking engineers also call “the STAMP protocol”) to Nancy Leveson’s accident model at MIT and a family of text-watermarking papers. None of them is affiliated with this program. This isn’t the only STAMP in circulation — Other STAMPs lists the others, so a mark doesn’t get mistaken for someone else’s.
The medium
That is a decision, not a limitation, and it has three reasons.
The bytes are the evidence. Every issue’s source file is byte-frozen, and the first twelve characters of its SHA-256 are printed in the foot of every page of its PDF. A page can only be checked that way if it is served exactly as written — which is why there is no analytics, no cookie, no tracking beacon, no script injected at the edge, no minifier and no email obfuscation. On launch day a default analytics beacon added 359 bytes to every HTML page and broke the fingerprint for three hours. Text survives that discipline. A pipeline of images and embeds does not.
A number in a picture is a number with its derivation removed. That is the failure this program reports in other people’s work, so it would be strange to commit it here. A figure set as text can be quoted, searched, read aloud by a screen reader, pasted into an email asking us to justify it, and diffed against the version before. A figure set as an image can be none of those. The charts here are inline SVG generated from the data files, so they are text as well: they hash, they diff, and they take their colors from the page rather than having one baked in.
And it costs almost nothing to serve. A program funded out of one pocket should not spend that pocket on hosting pictures of its own conclusions. These pages load on a slow connection, on an old phone and in a text browser — and because nothing here depends on a font, a script or a third party staying alive, they will still render when all three are gone.
The email address
Every page here ends with the same email address: s@bud.day. It is nine characters long, counting the @ and the dot. There is no support@, no press@, no contact form and no ticket number, because there is no queue to route you into. The work is done by a team whose members and limits are published rather than assumed, and it is led by its Chief Human Investigator, who signs the issues and whom the address reaches. A short address is not a flourish. It is a statement about how the accountability is arranged: a team does the work, and one name is on it.
Nine is short. It is not the shortest possible — one letter, a one-character domain and a two-letter country code would reach six — and the temptation to say otherwise is why this entry names the ceiling. A number published without its limit is a boast, and boasts are what this program exists to catch.
The mark
Two bounds, and a measured position standing between them. It draws the program’s one-sentence argument: a number without an interval is not a measurement. The uprights are the bounds, the red bar is the estimate, and the bar is deliberately not centered, because an estimate almost never sits in the middle of its own interval.
The bar’s position is a number, not a composition choice. The axis runs 0 to 100% between the uprights, so the bar sits at 33.0% — the mean of the two complete candidates in The Calibration Vol 001 (43.3% and 22.7%, both in figures.json). It is frozen there. A logo recomputed every fortnight is not a logo; illustrations may move the bar to a candidate’s own rate, the identity does not.
Drawn, the mark makes an H. That was raised against it before it was chosen, and the objection was answered rather than overruled: in this field H is not a stray letter. Helpful, honest, harmless is the triad the alignment literature has used since 2021. Read that way, the uprights are poles a model is pulled between and the bar is where it actually came to rest — not where a card says it rests; where it measured.
And the count comes out right, which was not designed: three verticals crossed by one bar contain exactly three H forms — take any two uprights and the rule between them completes one. Two of the three change width as the reading moves. The outer one cannot.
What it is not allowed to say. The program does not measure helpfulness, harmlessness or honesty; it measures whether a score survives being taken twice. The H is why the mark is legible to this field, and it stays out of every issue’s copy. Printing an HHH claim would be exactly the overreach the first issue’s own audit caught elsewhere. Two other marks were drawn and rejected — The Retest, two offset bars (Finding 01 as a glyph, and the cleaner drawing), and The Unit Grid. An identity with three marks has none.
Color and type
There is one accent on this site and it means one thing: a measured figure. Every rate in an issue, the bar in the mark, the number in a finding — red. The chrome never borrows it for a button or a flourish, so when you see the red you are looking at something that was computed from a file on disk and can be checked.
The argument
IBM Plex SerifHeadlines, findings, anything a person wrote and signed.
The reading
IBM Plex SansBody text. Set near 66 characters a line, because that is where reading is fastest.
[weak: … | basis: recalled]
IBM Plex MonoAnything a machine produced: receipts, figures, fingerprints, dates, the strapline. If it is in the mono face, a script wrote it.
Three faces from one family, so the site and the issue are visibly the same instrument at two scales. The home page, the issues and this page share one set of tokens; a home page in a different palette would say the site and the publication are different things.
The clock
Both numbers were drawn, and the same paw drew them. Budday Budderson is the pit bull the studio is named for. A language model (Gemini) was asked to act as a random weekday generator, choosing among Monday to Sunday the instant a key was pressed; Budday stepped on the touchpad, and it returned Saturday. The same model was then asked to act as a random number generator between 30 and 59, both included; Budday stepped again, and it returned 37. The mark above is the photograph of that moment redrawn in the house line: four toes over the pad, reaching the red bar that stands for the tap.
One draw is one draw — nothing about the fairness of the generator is claimed from it, and the page says so. But an arbitrary rule admitted to be arbitrary costs nothing; an arbitrary rule dressed as a reason is the defect this series reports in other people’s benchmarks. The day and the minute are now fixed: a cadence that gets re-argued is not a cadence.
Language and readers
The site is written in college-level, contemporary American English (post-2000 usage) — the register of a good trade publication after 2000: plain sentences, terms of art defined on first use, no jargon that a careful non-specialist could not follow. Numbers are never rounded for effect and never quoted without their conditions. Everything on the site is written in American English. The lone inherited exception is The Calibration Vol 001, issued Aug 30, 2026, which was written in British spelling — programme, colour, centred — and stands as issued, because an issued document is never edited.
It is written for anyone who wants to know whether a number about an AI model can be trusted. In practice that is four groups, and each reads a different page first: the people who build and test models — evaluation engineers and researchers, who open the source, the bank’s public half and figures.json; the people who have to sign for a model — procurement, compliance, auditors and, increasingly, regulators, who read the findings and the conditions attached to every rate; the statistically minded, who check the intervals, the agreement coefficients and the reliability figure and are welcome to; and reporters, who need a claim they can quote without misquoting — which is why every finding is stated with its caveat in the same paragraph.
The program’s design follows what independent evaluation research now recommends: keep the test content private so it cannot be memorized, and make the design fully public so it can be audited. Here that is 45 public probes, 18 held out (ten moved to public after they were found to be public already, two on Oct 05, 2026 and eight on Oct 10, 2026; see corrections), and every instrument — bank, gates, panel, runner, renderer — published with the results it produced.
Receipts
Every issue’s HTML is byte-frozen, and the first twelve characters of its SHA-256 are printed in the foot of every page of its PDF. The archive links the source file itself, so you can hash it and compare — and check its length, because a hash that is wrong tells you nothing and a length that is wrong tells you why.
curl -s https://protocolstamp.com/calibration/001-issued-sun20260830/source | sha256sum # be2bc9d7ca35… 100,707 bytes — the fingerprint printed on every page of the PDF
How issues are named. The Calibration Vol 001 · Sun20260830: the volume number, then the weekday and date it was issued, so a citation carries its own date and a name whose weekday disagrees with its date is refused by the naming gate before anything ships. The Log Vol 001 · Sat20260905, The Calibration Vol 002 · Sat20260912, The Log Vol 002 · Sat20260919, and so on. Vol 001’s own cover reads the older form, The Calibration 001 · ISSUED SUN20260830, because that file is frozen; the site refers to it by the new one.
To keep that true, every issue is served exactly as written: no analytics, no cookies, no tracking beacon, no script injected at the edge, no minifier, no email obfuscation. The home page and The Bench essays carry one advertising slot each, below the content, and it loads only after Google has approved the site; an issue never carries one, because an issue’s bytes are its fingerprint. On launch day a default analytics beacon added 359 bytes to every HTML page and broke the fingerprint for three hours; the fix is recorded, and checking the byte length is now part of every release. A correction here is a finding, and is published like a finding.
Made by
Static HTML in a public repository, deployed on every commit. No framework, no bundler, no cookie banner because there is nothing to consent to. The day/night switch in the top corner stores one word in your own browser and nothing leaves it. Findings are produced with AI assistance throughout; this page, like every page, was written and signed off by a person — see About for who does what, and what each may not do at all.
The pages, scripts and gates on this site were drafted mostly with Claude, Anthropic’s model, working in Claude Code and Claude Cowork under the direction of a person who read every line, changed many, and answers for all of them. That is a thank-you, and it is also a disclosure. The footer of every page compresses it to three verbs — drafted by AI, edited together, signed by a person — and the last one is the one that carries responsibility. The publisher is a New Mexico company; the work is done wherever the investigator is. Each issue is datelined where it was made, so the archive is also a record of where the programme has been. The first issues were made in Houston, Texas, which is the dateline an issue takes when no other is recorded. The site itself carries no dateline, because a page that is served every day cannot honestly claim a place on any of them.
On the watermark. Since Aug 02, 2026, when the transparency obligations of Article 50 of the EU AI Act took effect, Anthropic embeds an imperceptible watermark in text its models produce, and it does so worldwide, not only in Europe. Much of the text on this site therefore carries that mark. Anthropic’s own description of what the mark proves is the right one: that the text may have been processed by Claude, not who wrote it or whether it was changed afterward. We make the same claim and no more. The mark is not hidden from you here, and it is not a signature of authorship: the person named on the masthead is the author, and the mark is the tool’s.
What this is not. Anthropic is a vendor whose models this program measures or may measure. It has not sponsored, funded, reviewed, or endorsed this site or any finding on it; the tools were bought at the public price, like the API calls in every issue. Naming the tool changes nothing about the rules on the Independence page, including recusal, which applies to Claude’s family exactly as it applies to every other.