Skip to content
Protocol STAMP
Latest Calibration

Publications

The CalibrationOur own measurements, pre-registered and signedThe LogWhat ran, what it cost, what movedThe BenchWhat we read and argue with · Mon and ThuArchiveEvery issue, by number and dateThe technical reportThe program in one paper, with its dataSubscribeOne email per issue

Method

MethodologyHow a rate is producedDefinitionsWhat each term means hereThe protocol, v1.25The instruction layer as measuredRun filesEvery draw we can publishfigures.jsonEvery citable number and its fingerprints

AI policy

The DocketRules that bear on how AI is testedThe ChronicleDated events in AI governanceSuper Intelligence (SI)The US rename of AI, tracedMilitaries and AICountry by country

Trust

The Ten House ProtocolsHow we keep AI from making things upIndependenceWho pays, and the rules on itCorrections + changesWhat we got wrong, datedAI Safety + GovernanceWhat our agents may and may not doSupportHow the program is paid forFootprintWhat our tokens cost in power, and who pays itPrivacyWhat we keep, and for how longTermsThe terms of use

About

About Protocol STAMPWhat this is, who makes it, how to check itColophonHow the site is made, and the seven finishesOther STAMPsUnrelated things with the same nameLatest Calibration

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

NewThe technical report: STAMP, and what it takes to measure one  ·  Contributions are open  ·  The Ten House Protocols: how we keep AI from making things up  ·  Jester Mind, our own aggregated AI model, in the works

The Ten House Protocols

How we keep AI from making things upIn force since Monday Oct 12, 2026
10
house protocols, with sixteen amendments (R1 to R16), approved Oct 12, 2026 to Oct 20, 2026
The protocols
0
failures the house gate allows on any change to this site, or any release
Running today
4
rungs of evidence; nothing is published from a model’s memory
Protocol 2
30
published claims to be re-checked every Saturday, from Oct 17, 2026
Measuring it

Our first goal is to keep the rate of made-up claims on this site as close to zero as it can be measured. Every protocol below serves that goal, and every one is enforced by code where code can check it.

The goal

Why a house built with AI needs house rules

Protocol STAMP is teamed with AI. Machines draft much of what you read here, and they are fluent before they are right. A language model asked about a book it has never read can describe that book anyway, with the confidence of someone who has. That is the failure this program exists to measure in other people’s systems, and it would be a strange thing to tolerate in our own.

So we start from one assumption: a draft is wrong until it is checked. A model may suggest where a fact can be found. It may never be the place the fact comes from.

Source

Every claim points to something you can open

A working link to the page itself, opened by whoever wrote the sentence. When nobody could open a source, the page says so instead of filling the gap.

Document

Every step is written down

What was opened, what could not be opened and why, how sources were found, what the model was asked and forbidden to do, and what remains unconfirmed.

Enforce

Code checks what code can check

A script reads every page before it can change and before every release. If a rule breaks, publication stops. A named person reads what no script can judge.

Rules that depend on good intentions decay. An agent told to check the date will, on some ordinary Tuesday months from now, forget to check the date. That is why these protocols are written as checks. A rate of exactly zero can never be proven, but it can be bounded: every Saturday from Oct 17, 2026 we will draw published claims at random, check them again from scratch, and publish the result with its upper bound. Any error found becomes a dated correction, and the count starts over.

Every time, without exception. We follow these protocols on every page, every change and every release, with no shortcut for a deadline or a small edit. Full compliance does not make a made-up claim impossible: a source can be wrong, a person can misread it, and a check catches only what it was built to see. What it does is make a made-up claim hard to publish and quick to find. Every sentence must point to a source a reader can open, so an invented one has nothing to stand on; the house gate stops any change that breaks a rule a script can test; a person signs for what no script can judge; and the Saturday audit measures how many errors still get through and publishes that rate with its upper bound. Our claim about our own accuracy is a measured number, published with its margin and corrected in the open.

Status

What runs today, and what is being built

We say only what exists. As of Oct 20, 2026:

The house gate, required on every change to this site and every releaseRunning
Real dates only, with the right weekday (Protocol 5)Running
Citation markers land; the house mark is whole; the standing notice is on every page; no red in chartsRunning
The access ladder for sources that refuse a plain request (R10)Built
The source archive: our own copy of each cited source, with its fingerprintRunning468 of 497 cited links archived on Oct 12, 2026.
The link checkRunning419 of 497 cited links answered an automated check on Oct 12, 2026; the rest go down the access ladder.
A certainty score for every claim (Protocol 4)RunningSince Oct 20, 2026 a script computes each score from the evidence recorded for the claim, and the house gate checks that a shown score carries the word for its band. No claim on the site shows a score yet: the Docket and the Chronicle keep their current labels until scored, by Monday Nov 09, 2026. Until then the house gate warns on an unscored claim; from that day it refuses the change.
Two independent AI reviewers, then a person (Protocol 10)Running for Bench postsSince Oct 20, 2026: two Claude reviews (a fact-check and a hostile reader), one model from outside Claude’s family drawn by a recorded seed (R14, R16), then the signer. Not built yet: the debate step and the published agreement rate.
Writing checks and the suggestion tool (Protocol 6, R11)RunningThe house gate warns on 11 habits of machine prose; the suggestion tool proposes and never rewrites.
The attestation on every publication (Protocol 10, R12)Being builtAttestations are on the worked example and The Bench · A Book We Did Not Read.
The breach register (Protocol 10, R12)RunningOn Corrections + changes since Oct 20, 2026: date, model, what happened, what was done. The house gate refuses a change that removes it, leaves a cell empty, names a model without its version or breaks the newest-first order. No breach is entered yet; the corrections since Oct 12, 2026 are being read against the five kinds.
The Saturday audit of 30 published claimsRunningIt was due to start on Saturday Oct 17, 2026 and did not; see Corrections + changes. Since Oct 20, 2026 a script draws the 30 by a recorded seed from every published claim, keeps the seed and the whole list so anyone can draw again, and counts a week only when both reviewers and then the person have checked every claim. The first draw, for Saturday Oct 17, 2026, was made on Oct 20, 2026, from 515 claims. Its claims checked, errors found and 95% upper bound go in The Log when those checks are done.

Amendments

Every change to the ten, in order

The ten protocols do not grow in number; each amendment adds to one of them, and each is signed by Lawrence Ho and recorded on Corrections + changes. The full text of each is in its protocol below.

AmendmentSignedWhereWhat it says
R1Oct 12, 2026Protocol 4A second source counts only if it is independent, and reviewers earn nothing for agreeing without opening the source.
R2Oct 12, 2026Protocol 10Recusal: a reviewer never reviews a claim about its own maker.
R3Oct 12, 2026Day onePages published before adoption keep their labels until scored; from Nov 09, 2026 an unscored claim fails.
R4Oct 12, 2026The auditEvery Saturday, 30 published claims are checked again; errors and the 95% upper bound are published.
R5Oct 12, 2026Self-governanceThe house gate runs on every change and release, and the protocols change only by a signed amendment.
R6Oct 12, 2026Protocol 5Every date must exist, and a weekday written before it must be its own.
R7Oct 12, 2026Protocol 2When a source cannot be opened, every claim says which rung of evidence it stands on.
R8Oct 12, 2026Protocol 4100% (Certain) is shown only when every evidence row is met, for a fact of record.
R9Oct 12, 2026Protocol 4Scoring for claims that rest on records and reviews: each further independent source adds 10 points, with caps.
R10Oct 12, 2026Protocol 2The access ladder: try every lawful way to open a source before saying it cannot be opened.
R11Oct 12, 2026Protocol 6Write the way a careful editor writes; the house gate warns on habits of machine prose.
R12Oct 12, 2026Protocol 10Every publication ends with an attestation, and a breach by an agent is corrected in public and registered.
R13Oct 12, 2026Protocol 2Without the work itself, ask for it first, search only what is public, and label what is in the book and what is about it.
R14Oct 19, 2026Protocol 10Every draft is reviewed by Claude and by one model drawn at random from outside Claude’s family.
R15Oct 20, 2026Protocol 9Keeping the record: one computer, backups that never delete, failures reported, drafts saved daily.
R16Oct 20, 2026Protocol 10The outside reviewer may come through FreeLLMAPI; a model that may keep data gets only gated pages about to be published.

The protocols

The ten, in full

Each card gives the rule in a sentence. Open it for the full text.

01

Every factual sentence names its source, with a working link

The link opens the cited page itself, and it is checked when the page is published and at every “last checked” date.

Full text
  • A dead link is replaced by an archived copy, and the swap is logged.
  • PDFs. We keep our own copy of every cited PDF, with its SHA-256 and the date retrieved. Where republishing is allowed, our copy is the link: US federal government works, which are not under copyright (17 U.S.C. 105), and documents whose license allows it. Everything else links to the original and an archived snapshot, and our copy stays private.
  • Social media. A cited post is captured when it is cited: a screenshot showing the post, the account and the post’s own time, with the link, our capture time and a fingerprint. Where a platform refuses us, the page says the post was not opened and cites the reporting instead.
  • Research before every verdict. No claim is marked true, false or corrected until it has been checked online that day against the cited source and, where one exists, the primary text: the issuer’s own document, the court record, the statute or the registry.
Checked by: the house gate (citations land) · the link check · the source archive · a person
02

Open the source before writing the sentence

Write from the opened source, never from a search snippet, a summary or memory. What a model remembers is a lead, not a source.

Full text: the access ladder and the evidence ladder

Before saying a source cannot be opened (R10). A tool that cannot open a page is not proof that the page cannot be opened. Every source goes down four rungs, and the log records which one opened it:

  1. Plain request

    An ordinary request, identified as a normal browser.

  2. Real browser

    For pages built by script, or that refuse a plain request.

  3. Archive

    A saved copy, for a page that is dead or refuses us.

  4. A person

    Opens it in their own browser and saves a PDF or a timed screenshot.

We never solve or evade a bot check, get past a paywall, or use an account that is not the person’s own.

When we do not have the work itself (R13). First we ask the person whether they can provide it: a copy, a scan, the pages that matter. If not, we search what has been made public, lawfully: the publisher’s page and cover text, authorized excerpts, previews read as the site shows them, library records (which often carry the table of contents), and reviews that quote the work. Every item must be traceable, with a link, a date and our saved copy. Every claim then says which part is which: the book’s words as a named source quotes them, the book as a named source describes it, its structure from a catalogue, or not the book at all. We never use a pirated copy, get past a wall, or let a model’s memory stand in for the text. See it applied.

When a source still cannot be opened (R7). Every claim says which rung of evidence it stands on:

AThe work itself, readWhat it says, quoted or summarized, with page or section
BThe issuer’s own words about itThe publisher, the author, an authorized excerpt, attributed to it
CRecords and named reviewsFacts of record; each reviewer’s view, attributed
DMemory onlyNothing is published. Use it only to find a source

The sourcing note. Every sourced page ends with what was opened (with links and dates), what could not be opened and why, how the sources were found, what the model was asked to do and forbidden to do, and what remains unconfirmed.

Checked by: the access ladder’s log · a person, who records the source read
03

Quotes are exact; our interpretation is labeled as ours

Nothing goes inside quotation marks that the speaker did not say, in those words. Our own reading is welcome, and always marked as ours.

Full text

Our reading is kept apart from what the source says. The US Intelligence Community requires the same separation of sourced information from an analyst’s own judgments in its analytic standard, ICD 203.

Checked by: a person, against the source
04

Say how sure we are, as a number

Every sourced claim carries a certainty score from 0% to 100%, computed by a script from the evidence recorded for it, never from anyone’s feeling.

Full text: the points and the scale
50the issuer’s own document opened
15a second independent source
10our archived copy
10checked within 30 days
15both reviewers opened it and agree
+10each further source, up to 40, for claims on records and reviews (R9)

Caps: 75% with no primary source, 40% when sources conflict, 20% when nothing could be opened. Each score is shown with a word from the scale in ICD 203, plus our own top band, which is earned only by a fact of record with every check met and a date attached (R8).

Checked by: the scoring script (being built) · the house gate
05

Every visible date is “Oct 12, 2026” or “Monday Oct 12, 2026”

Every date must exist, and a weekday written before it must be its own (R6).

Full text

A thirtieth of February is refused, and so is a Tuesday written before a date that falls on a Monday. An impossible date means something upstream invented it. The house gate checks this on every page, including this one: an earlier draft of this paragraph, which wrote the two refused dates out in full as examples, was refused by it.

Checked by: the house gate
06

Write plainly, so a reader can check every word

If a word claims size, speed, quality or novelty, give the number or cut the word. Write the way a careful editor writes (R11).

Full text

A term quoted from a law or a source stays, as a quotation. We say on every page that AI drafts with us, so plain writing is not disguise: the habits that give machine prose away are the same habits that make a sentence claim more than it shows. The list comes from the studio’s own checklist, in I Make Because We Make (Sep 2026): an em dash in every other line, “not only X but Y,” “It is not X. It is Y.,” “moreover” and “ultimately,” unowned hedges such as “arguably,” inflated verbs such as “serves as.” Each is cut or answered.

Before review, every draft opens with its result, lists what was not verified, and is labelled d1, d2 or d3; only d3 goes to the reviewers. The suggestion tool is run on it, and every suggestion is taken or answered. The tool never rewrites, never touches a quotation, and refuses issued pages.

Checked by: the house gate (warnings, each with its reason) · the suggestion tool · a person
07

The house voice

We write “AI.” The house mark is always whole. Red means a measured figure. The standing notice is on every page.

Full text

“Super Intelligence” (SI) appears only when we name or quote a US instrument that uses it. The house mark reads: Begun by a person · Drafted by AI · Edited together · Signed by a person.

Checked by: the house gate
08

Issued is frozen; corrections are made in the open

An issue’s bytes are never edited. A signed post is corrected with a dated note under the sentence, which stays as issued.

Checked by: the release scripts refuse to change frozen bytes · every live Source is hashed against its build
09

Everything that changes is written down the same day, with its sources

Every correction and every change gets a dated row in Corrections + changes, and every correction links the source that showed the error.

Full text
  • Keeping the record (R15, from Oct 20, 2026). Our working files are kept in one place, on one computer, and only that computer runs our scheduled jobs. A backup adds and overwrites; nothing we run deletes files from a backup to make it match another copy. A backup that fails says so where a person will see it, and before a working session ends it checks that the last backup succeeded that day. Unfinished drafts are saved off that computer at the end of every session, on their own branch, never on the published site. Learned on Oct 19, 2026, when an old laptop’s backup job overwrote the current backup with files from September; nothing published was affected, and the current files were backed up again that night.
Checked by: the backup job, which refuses to run from any other computer and reports every failure; and a person
10

Two independent AI reviewers, then a person who signs and answers for it

Two reviewers from different model families review apart, debate only where they differ, and a person decides and signs. Every publication ends with an attestation (R12).

Full text
  • Who reviews (R14, from Oct 19, 2026). Every draft is reviewed by Claude and by one model drawn at random, by a recorded seed, from outside Claude’s family, automatically. Free models are used first; our zero-data-retention setting stays on, so a model that keeps what it is sent is skipped, except as R16 allows. Neither reviewer sees the other’s work or the drafter’s notes before both hand in. Neither reviews a claim about its own maker (R2).
  • One reviewer first reads only the page, without its sources, and writes down what it thinks we claimed. The other repeats one claim from the sourcing note alone (R11).
  • Both verdicts are recorded first. Where they differ, they debate with sources for at most three rounds; research on model debate (Du et al., 2023) reports better factual accuracy, and anything still disputed goes to the person.
  • How often the reviewers agreed before debate is published.
  • Where the second reviewer comes from (R16, from Oct 20, 2026). The reviewer from outside Claude’s family may also be reached through FreeLLMAPI, an open-source proxy run on our own computer that pools the free tiers of many providers; it reviewed its first page on Oct 20, 2026. It is still drawn by a recorded seed from the models the proxy lists, never by the proxy’s own choice, and the review records which provider answered. Free tiers may keep what they are sent, so a model that may keep it is sent only a page about to be published, after the holdout gate has passed, with its public sources. The private bank, a Calibration before it is issued, its pre-registration and anything private go only to a model that keeps nothing.
  • The attestation (R12). Every publication names its drafting model, its two reviewers and the person who signs and answers for it. The signer gives their word in good faith. An agent gives no word, because it cannot take an oath or be held to one; it works under these protocols as the condition of being used.
  • When an agent breaks the protocols (R12). A breach is a source cited but not opened, a quotation not in its source, a claim filled in from memory, a figure or date not in the source, or a check reported as run that was not. It is corrected in public the same day, naming the model; the model is stood down from drafting for 30 days; its work since its last clean audit is rechecked; the audit count restarts; the breach goes in a public register; a second breach within 90 days removes the model until a person seats it again. A breach that reached the public is also corrected under the signer’s name.
Checked by: the review log · the attestation and signature · the breach register

Self-government

How the protocols enforce themselves

Required (R5)

A failure stops publication

The house gate runs on every change to this site and on every release.

Measured (R4)

30 claims every Saturday

Drawn at random and checked again. We publish errors found and the 95% upper bound: with none found, about 3 divided by the number checked, so about 1% after 300 claims.

Day one (R3)

Scored by Nov 09, 2026

Pages published before Oct 12, 2026 keep their labels until scored; from Monday Nov 09, 2026 an unscored claim fails like any other.

Changes to the protocols are made only by an amendment a person approves, recorded in Corrections + changes. An agent may propose one; it may never adopt one.

What comes next

Jester Mind

In the works · not yet released

Our own aggregated AI model

Budday Budderson Studio LLC is building Jester Mind, an aggregated AI model held to these same protocols. Like Protocol STAMP, it is a project of the studio. It will cite only what it has opened, say what it could not open, and never fill a gap from memory.

The plan, not a result: Jester Mind is designed to ask the person instead of guessing when it is unsure or when its sources conflict. Its design, the models it brings together, how it will be measured and when it will be released will be published here, with their sources, when they exist.

Worked examples

The protocols on a book we could not read

We summarized Chris Patten’s Not Quite the Diplomat from the publisher’s page and named reviews only, scored every claim, and then checked our own draft again. We found six mistakes in it, and the page shows each one. Every claim is labelled as the book’s own words (quoted by a reviewer), the book as a reviewer describes it, its structure from a library record, or not the book at all. Read the worked example.

The second worked example reads a doctoral thesis twice: Isabelle Augenstein’s Towards Explainable Fact Checking, first from public sources about it without opening it, then from the full text, and compares the two passes claim by claim. Read worked example 2.

Sourcing note

How this page was made

Opened, Oct 12, 2026: ICD 203, Analytic Standards (Office of the Director of National Intelligence); Du et al., “Improving Factuality and Reasoning in Language Models through Multiagent Debate” (arXiv); 17 U.S.C. 105 (Cornell Legal Information Institute). The counts of archived and answering links are our own measurements on Oct 12, 2026. Not used: anything from memory. Unconfirmed: the parts marked as being built, and everything about Jester Mind beyond its stated plan, are not yet results. The protocols were approved by Lawrence Ho, Chief Human Investigator, electronically, by chat message, on Monday Oct 12, 2026; R10 was approved the same way the same day.