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
An issued Calibration is never edited. If a number changes, a correction is issued under a new date and the superseded issue stands, so the record shows what was believed and when. Corrections are findings, and they are published like findings — here, and in the next issue.
This page is the program’s transparency, in the only form transparency actually takes. Transparency is not a value a publication can announce about itself; it is a habit that costs something, and the cost is this list. Anyone can promise to be open. What can be checked is whether, when a thing here turned out to be wrong, it was written down under a date, with the reason, where a reader who was never told could still find it. That is the only version of the word that means anything, and it is the version this page keeps.
Everything that changes is written down here. That is a house rule of every Budday Budderson Production, not a policy of this program alone: a correction is a thing that was wrong, and a change is a thing that is now different, and a reader is owed both under a date with the reason attached. A decision that gets quietly reversed is indistinguishable from one that was never made. Each row below says which kind it is.
Why the register exists
There is a Japanese practice called kintsugi, golden joinery. A bowl breaks. The pieces are not thrown away, and they are not mended so cleverly that the mend cannot be found. They are rejoined with lacquer, and the seams are dusted with powdered gold — so that the break becomes the most beautiful thing about the bowl, and the first place the eye goes.
The story told about its beginning is that a shogun sent a favorite tea bowl away to be repaired and it came back stapled with ugly metal, and that the craftsmen asked to do better decided that a repair which cannot be hidden should be honored instead. Whether it happened that way is not the point, and never was. It is a tradition, older and finer than anything on this website, and this page is a tribute to it.
Because the reasoning is the reasoning of this page. A bowl that has been broken and mended is not a bowl pretending it was never broken. It is a bowl with a history, and somebody decided that the history should be visible. Concealment was cheaper, and easier, and available. It was refused.
So an issued document here is never edited. The break stays where it happened, and the mend is a new dated row beside it, in gold. The Calibration Vol 001 carries a timezone bug on its own cover and wears the Sunday finish permanently, so the off-cycle date can never stop being visible. The Log Vol 001 went out 115 minutes late, and the foot of that issue prints 17:31, 115 minutes late rather than the 15:37 it had promised. The first issue of The Bench needed nine corrections and printed all nine.
And here is the part of the practice this program takes most to heart. The mended bowl goes back into use. It is not retired to a cabinet, and it is not quietly replaced with an unbroken one. It is used for tea the next morning, seam and all.
One thing the gold does not do, and it is worth saying once: it does not make the ceramic stronger. Nobody breaks a bowl in order to have something to gild. The gates, the panels and the fingerprints all exist so that this page stays short — and this page exists because they will sometimes fail.
The register
| Date | Kind | Concerns | What happened | Status |
|---|---|---|---|---|
| Oct 21, 2026 | Correction | The counts at the top of this register | The two figures above the register read 19 corrections and 33 changes in the page itself, counts that had fallen behind as rows were added. A script on the page recounts the rows when it loads, so most readers saw the right numbers; anyone reading without JavaScript, and any copy of the page, saw the old ones. The figures in the page are now computed from the register by tools/fix_register_counts.py. | Corrected |
| Oct 21, 2026 | Change | Worked example 2: Towards Explainable Fact Checking, Read Twice | A second worked example of the Ten House Protocols. Isabelle Augenstein’s doctoral thesis is summarized first from public sources about it, without opening it, then from the full text, and the two passes are compared claim by claim. Signed by Lawrence Ho on Oct 21, 2026. It is linked from the house protocols page and from the first worked example. | Published |
| Oct 21, 2026 | Change | A faint grid behind every page | At Lawrence Ho’s request, every page now has a faint 16-pixel grid behind its content, drawn in each finish’s own rule color and fading toward the foot of the screen; it is not printed. A design change only: no text or figure changed. The idea came from aclide.com. | Live |
| Oct 20, 2026 | Change | The breach register (R12) | Protocol 10 has required a breach register on this page since R12, signed Oct 12, 2026; it had not been built. It is now below the register of corrections, with four columns: date, model, what happened, what was done. The house gate checks it on every change: the register must be there, every cell filled, each model named with its version, newest first. No breach is entered yet. The corrections made since Oct 12, 2026 are being read against the five kinds of breach that R12 names, and any that is one will be entered with its date. | In force |
| Oct 20, 2026 | Change | The Saturday audit of 30 published claims (R4) | The audit that did not start on Saturday Oct 17, 2026 now has its script. It counts every published claim: each Docket and Chronicle entry, and each passage on any other page that carries a source; 515 on Oct 20, 2026. It draws 30 of them at random by a seed written down first, and keeps the seed and the whole list, so the draw can be repeated and a claim cannot be swapped for an easier one. A week is counted only when two different reviewers have each opened the source and given a verdict, and Lawrence Ho has then given his. The result is claims checked, errors found and the 95% upper bound on the error rate, about 3 divided by the number checked when none are found (9.5% for 30). An error restarts the count. The first draw, for Oct 17, 2026, was made on Oct 20, 2026; its result will be in The Log. | In force; first week being checked |
| Oct 20, 2026 | Change | The certainty score (Protocol 4) | The script that Protocol 4 calls for now runs. It computes a claim’s score from 0% to 100% from the evidence recorded for it, row by row, with the caps of Protocol 4 and amendments R1, R8 and R9; a field nobody recorded earns nothing, and a model never supplies the number. It reproduces the scores of the worked example, which were computed by hand on Oct 12, 2026. The house gate now refuses a score shown with the wrong word for its band, and warns on a Docket or Chronicle entry with no score; from Monday Nov 09, 2026 it refuses one (R3). Scored from their own records on Oct 20, 2026, the Docket’s 140 entries labelled Supported reach 60% or 70% (Likely), not the 80% that Supported means under Protocol 4: the records show the issuer’s document opened and, for most, our hashed copy, but no independent second source and no reviewers. No label on the site has changed yet. | In force |
| Oct 20, 2026 | Correction | The Saturday audit of 30 published claims | The status list on the house protocols page said the audit would start on Saturday Oct 17, 2026. It did not start: the script that draws the 30 claims has not been built, and The Log Vol 004 carries no audit. The same list called the two-reviewer step “Being built” after it had run for The Bench Vol 008, and called the attestation the worked example’s alone. All three lines are corrected. The audit, the breach register and the certainty score are still to be built. | Corrected; audit open |
| Oct 20, 2026 | Change | The Bench carries no volume numbers | From today a Bench post is named by its title alone, on the home page, the Bench page, the feed and the page search: “The Bench · A Book We Did Not Read”. The Log and The Calibration keep their volume numbers and dates. At Lawrence Ho’s request the published posts drop theirs too: each masthead, citation line and search headline, and every place one post names another by number, now uses the title. Numbers that refer to a Calibration or a Log stay. The first post, The No. 1 Complaint, is frozen and keeps the number it was issued with, as do the run records. Older rows on this page keep the numbers they were written with. | In force |
| Oct 20, 2026 | Change | Where the second reviewer comes from (R16) | A new amendment to Protocol 10. Under R14 a model that may keep what it is sent was skipped, and on Oct 20, 2026 every free model outside Claude’s family that keeps nothing was unreachable for the final text of The Bench Vol 008. From today the second reviewer may also come through FreeLLMAPI, an open-source proxy run on our own computer, which is being set up; it is still drawn by a recorded seed. A model that may keep what it is sent gets only a page about to be published, after the holdout gate has passed; the private bank, unissued Calibrations, their pre-registrations and anything private still go only to a model that keeps nothing. | In force |
| Oct 20, 2026 | Change | Keeping the record (R15) | A new amendment to Protocol 9. On Oct 19, 2026 an old laptop, switched on after a week, ran its own backup job and overwrote our current backup with working files from September, deleting newer ones. Nothing published was affected: the current files were on the working computer, and they were backed up again that night. From today our working files are kept on one computer, which alone runs scheduled jobs; backups never delete to match another copy; a failed backup is reported where a person will see it; and unfinished drafts are saved off that computer at the end of every session. The laptop’s job is switched off, and the backup job now refuses to run from any other computer. | In force |
| Oct 20, 2026 | Change | The Bench Vol 008, published a day after its slot | Vol 008 was due on Monday Oct 19, 2026 at 15:37 America/Chicago and was not ready; our rule is that a missed Bench slot stays missed. Lawrence Ho set that rule aside once, on Oct 20, 2026, and Vol 008 was published on Tuesday Oct 20, 2026, after its reviews passed and he signed it. The Oct 19, 2026 row below, which says the 15:37 schedule starts with Vol 008 that day, is kept as written. The rule stands for every later slot. | Done |
| Oct 20, 2026 | Correction | The worked example said none of its first draft’s six errors changed the argument | Version 2 of the Patten worked example said “None of these errors changes what the summary says the book argues”, and described Ian Black’s sentence on Thatcher and the single market as not saying whose view it is. The Observer review gives that remark in Black’s own words, outside his quotation from the book, so it does not let us give it to Patten; version 1 did, and that changed what the summary said the book argues. The same review of the page found two more lines to correct: Barnes & Noble’s address served its “not found” page on Oct 12, 2026, so we never saw the page, not a page without an excerpt; and the chapter-title search is consistent with the same titles, not proof of the same chapters. Found by the reviews of The Bench Vol 008 on Oct 20, 2026; each line named here is corrected. | Corrected |
| Oct 19, 2026 | Correction | The worked example counted 64 labels; it had 65 | Version 3 of the Patten worked example, published Oct 12, 2026, said its page carried 64 labels. It carried 65: the count left out its one “chapter title confirmed” label. Found when the labels were counted again on Oct 19, 2026; the page now gives the full count for version 4 and the corrected count for version 3. The same pass corrected two stale notes on that page: the Library of Congress record is listed as read, in the copy held by the Internet Archive, and the two early review dates are described as read from a web timestamp in US or universal time. | Corrected |
| Oct 19, 2026 | Change | The Ten House Protocols: amendment R14, who reviews | Approved by Lawrence Ho by chat message, about 11:35 CDT. Protocol 10 now has every draft reviewed by Claude and by one model drawn at random, by a recorded seed, from outside Claude’s family, automatically, free models first. Our zero-data-retention setting stays on, so a model that keeps what it is sent is skipped. The first such review was of The Bench Vol 008. | In force |
| Oct 19, 2026 | Change | Every publication now ships at 15:37 America/Chicago | Lawrence Ho asked by chat message on Oct 19, 2026 for everything to ship at 3:37 in the afternoon. The Calibration and the Log already did; The Bench moves from 07:37 to 15:37 on its usual Mondays and Thursdays, starting with Vol 008 today. Pages that state the time were updated; earlier rows here keep the times that were true when they were written. | In force |
| Oct 19, 2026 | Change | The worked example checks quoted wording against the book | Version 4 of the Patten worked example adds a search of the 2005 first edition at HathiTrust, which shows counts but no text: 22 phrases the reviewers quote are found, and all eleven chapter titles appear as running heads. Two quoted phrases are not found, and the page says so. | Published |
| Oct 17, 2026 | Correction | The home page list: two rows shown in the wrong state | When The Log Vol 004 was released at 16:17 on Oct 17, 2026, the script that turns a “Not yet published” row into a live one matched too much of the home page. It left the Log Vol 004 row greyed out as if pending, and showed The Bench Vol 008, which is due Oct 19, 2026 and not out, in the style of a published post (its “Not yet published” badge stayed). Both rows now show their real state, and the script can no longer match past the row it is changing. The archive was not affected, and no issued byte changed. | Corrected |
| Oct 17, 2026 | Correction | The held-out half of the bank — eight probes | Eight more probes we counted as held out were not secret. Their questions have been in the public repository since Jul 24, 2026, in tests/PROBES.md and the promptfoo test file, before the bank was split. The scanner added on Oct 05, 2026 read only the site, so it missed them. Once it also read the repository it found seven, on Oct 08, 2026: 03_honesty_no_invention_105, 07_live_data_104, 07_live_data_107, 07_live_data_109, 07_live_data_111, 07_live_data_120 and 12_impossible_specifics_027. A scan of the whole repository before this change found an eighth, 12_impossible_specifics_024: the same question in different words, so only the paraphrase detector could see it. All eight are now public, with their text and marked reclassified, in bank_v1_2_public.jsonl (bank v1.2): 45 public, 18 held out. The file is the v1.1 public file with these eight rows added. Figures already issued are not changed; any figure that splits public from held-out items says which bank version it uses. The gates now read v1.2. | Reclassified; issued figures unchanged; no issued byte changed |
| Oct 16, 2026 | Correction | The technical report: figures left behind by the Section 6.1 correction | After the Aug 08, 2026 run was re-graded without the models’ reasoning, four places still carried the old range. Section 6.4 said the two runs that did not reason had effects (−7.3 and −23.2 points) that “bracket” the six reasoning rows at v1.25, given as −9.8 to −22.0; corrected, those rows run from −6.1 to −18.3, so −7.3 is not outside them, and the sentence now says the two fall within and beyond that range. The Conclusion said “7 to 23 points”; it now says 6.1 to 23.2, as Section 6.1 does. The repository README said “7 to 23 points across three model families”; it now says 6.1 to 23.2 points on eight models from eight developers. The Bench Vol 007, issued before the correction, said 7.3 to 23.2; it keeps its sentence as issued, with a dated note under it. The PDF and arXiv source are rebuilt. | Corrected |
| Oct 16, 2026 | Change | The Log Vol 004: a run it does not show | The Log Vol 004, staged for release on Oct 17, 2026, is computed from the figures registry, and the reasoning-off and reasoning-on study of Oct 14, 2026 is not in that registry: it ran through a different harness (run files). So Vol 004 does not list that run under What ran, and its cost table does not include the study’s gateway charges, $14.44 for the study and $0.85 for its pilots, as recorded in the run’s spend ledger. The study’s results are in the technical report, Section 6.5. The Log is issued as computed; registering the study so a later Log can carry it is open. | Open |
| Oct 16, 2026 | Change | New essay: Did STAMP help? | A plain-language account of what every measurement of STAMP so far says, written after the re-grade recorded in the Oct 15, 2026 correction below: fail rates fell in all ten configurations of the first round, on the probes STAMP was written against, by 6.1 to 23.2 points; two graders shown the same answers put the effect on one model at 6.1 and 15.9 points; on a bank STAMP was not tuned on, the one estimate (−0.9 points) is smaller than the grader’s bias; and 22.0% to 31.7% of receipts in the five-model run failed. Listed on The Bench, the page search and the site map. | Done |
| Oct 15, 2026 | Correction | The technical report: the five-model run, re-graded without the models’ reasoning | In the Aug 08, 2026 run (gpt-oss-120b, GLM-4.7, DeepSeek V4 Flash, Step 3.7 Flash, Nemotron 3 Super), every reply began with the model’s own reasoning, because the harness was not told to hide it, and the grader read that reasoning as part of the answer. The report’s Section 6.1 first said only that these replies carried a draft receipt; the row below added the grader reading the reasoning as a limit, and this row acts on it. We separated each answer from its reasoning, leaving the published file unchanged, and re-graded the answers with the same rubric and two graders from other developers. STAMP still lowered the fail rate on all five, but by less: 6.1 to 18.3 points with the grader the table now uses (Gemini 3.7 Flash), against 9.8 to 22.0 as published, and for DeepSeek V4 Flash, Step 3.7 Flash and Nemotron 3 Super the 95% interval now includes or reaches zero (DeepSeek −6.1 points, −15.9 to 3.7). A second grader (Claude Haiku 5.5) put the effect at 13.6 to 23.2 points on the same answers, so the size depends on the grader. With the grader held fixed, removing the reasoning reduced the effect by 3.4 points (95% CI −0.2 to 6.8) and 2.0 points (−2.2 to 6.4). Of 62 replies that returned no answer at all, the original grader had passed 18. Receipts checked on the answer alone fail 22.0% to 31.7% of the time, not the 3.6% to 13.1% the report gave; about half of those failures (55 of 107) are replies that ran out of tokens before writing one. The method and every verdict are in the re-grade files; the re-grade cost $3.05. | Corrected in the open |
| Oct 15, 2026 | Change | The technical report: which earlier runs reasoned | A check of the ten earlier run files found that none switched reasoning off. Eight reasoned at their default: every Gemini reply reports reasoning tokens, and all 930 replies of the Aug 08, 2026 run begin with the model’s reasoning, which the grader read together with the answer. Only Perplexity Sonar and Mistral Large showed no reasoning. The report now says so in Section 6.4, draws no conclusion about reasoning from these runs, and lists the grader reading the reasoning as a fifth limit in Section 6.1; the within-model contrast remains the Oct 14, 2026 study in Section 6.5. The same Aug 08 run reported no reasoning-token count on 348 of those 930 replies, so the study’s analysis script now also refuses a reasoning-off row whose replies return reasoning text; its output on the Oct 14 halves is unchanged. No number in the report changed. | Done |
| Oct 14, 2026 | Change | The technical report: reasoning off and on | The pre-registered System 1 and System 2 study ran and is now Section 6.5 of the report, with its run files. Four models with reasoning off and on, two that always reason, 82 development probes, two halves, STAMP v1.25. Reasoning did not shrink STAMP’s effect: the change was +0.0 to −5.5 points across the four models, every interval including zero, against our prediction of a smaller effect. With reasoning off STAMP helped clearly on one model of four (luna, −9.8 points, 95% CI −16.5 to −3.7) and not at all on nemotron (+0.6, −8.5 to 9.8). Verdicts are from one grader only; the panel the pre-registration requires was not run. Three amendments were signed before the run: two model substitutions where the gateway would not switch reasoning off, and the grader’s settings, including hiding each model’s reasoning from it. | Done |
| Oct 14, 2026 | Change | The grants we are applying for | Eight grant programs, the amount to be asked of each (no single ask above $60,000, half of a $120,000 first year, under the cap on Support), and the programs we will not apply to, published before any application is filed. Each row will carry its filing date and decision. | Open |
| Oct 14, 2026 | Change | The technical report: five more models | A sixth earlier run, too large to recover on Oct 13, 2026, was received zipped and is now in the report and the run files: STAMP v1.25 on gpt-oss-120b, GLM-4.7, DeepSeek V4 Flash, Step 3.7 Flash and Nemotron 3 Super, on Aug 08, 2026. STAMP lowered the fail rate on all five, by 9.8 to 22.0 points; for Nemotron 3 Super the interval includes zero. Most of these replies carry two receipts, because the models returned their reasoning and it holds a draft receipt; the report says so. The published file replaces the gateway’s server address with a placeholder; no answer or verdict was changed. | Done |
| Oct 14, 2026 | Change | This register: every row links to what it concerns | The 38 rows dated Oct 12, 2026 and earlier that carried no link now do: each row’s title opens the page or issue it is about, and the other pages a row names are linked where it names them. No row’s wording, date or status changed. The same day, The technical report page gained space between its opening paragraph and the line under it. | Done |
| Oct 13, 2026 | Change | New page: the technical report | A new standing page, The technical report, publishes the program in one paper, as a preprint with its PDF. It adds results this site had not printed: five native-versus-STAMP runs made between Jul 23 and Aug 07, 2026 on the probes the protocol was written against, recovered and committed unchanged with their fingerprints (run files). A draft counted nine install checks in that comparison, which added 5 to 9 points to every effect; they were removed before publication, and the page says so. Linked from the Publications menu, the page search, the site map and the announcement strip. | Done |
| Oct 13, 2026 | Change | Militaries and AI: Palantir entry, after the source check | The note beside the Palantir entry still said its sources had not been opened, after the check below had opened them; it now says they were. Four dates given to the month are now given to the day, from the pages opened: May 29, 2024 and May 20, 2025 (the Pentagon’s contract notices), Mar 09, 2026 (the Feinberg memo, as DefenseScoop reports it) and Sep 22, 2026. The 100,000-user figure is now attributed to the official who gave it, James Mazol DefenseScoop. The Army agreement is now described, as the Army describes it, as lasting “up to 10 years” army.mil. | Done |
| Oct 13, 2026 | Correction | Militaries and AI: Palantir sources | A browser check of every page cited in the Palantir entry found three faults. The Aug 2026 line said it was not confirmed whether the memo “was signed”; The Register reports the memo is on signed letterhead and that it is unclear whether the deal was finalized, so the line now says that. The TITAN line’s “phase 3” and the Maven line’s $1.275 billion ceiling were not in the pages cited for them; source 67 now leads with the Army’s award notice, which says Phase 3, and source 69 adds the contract record, which shows the $1,275,000,000 ceiling. The Limits note now says every page in sources 67 to 75 was opened, and drops the three award days it called unconfirmed. | Corrected |
| Oct 13, 2026 | Change | Militaries and AI: Palantir; House Protocols: every time | Militaries and AI adds Palantir Technologies’ US Army and Pentagon awards (TITAN, the Maven Smart System, the Army enterprise agreement), with a chart of their ceilings and nine new sources, 67 to 75. Sources 67 to 73 were found by search and not opened from our drafting environment; the page says so beside the entry and in its Limits note. The Ten House Protocols adds a statement that the protocols apply on every change without exception, and what that does and does not guarantee. | Done |
| Oct 13, 2026 | Change | Support: the payment links open | The Support page was published on Oct 05, 2026 with a notice that its payment links were not live, and a promise that the notice would be removed and logged when they opened. This is that row. The three reader options are now buttons to a Stripe-hosted page, which asks whether a contributor may be named at $100 and defaults to no. A new Grants section says that institutional support and grants start by letter, because the objection clause is agreed in writing, and that the grantee is a for-profit company that uses no fiscal sponsor. Privacy names Stripe. The ledger is unchanged: no one has given yet. | Done |
| Oct 12, 2026 | Change | The worked example says which part is in the book | Version 3 of the Patten worked example labels every claim: the book’s words as a reviewer quotes them, the book as a reviewer describes it, its structure from a library record, or not the book. It adds the eleven chapter titles from a library record of the 2006 paperback, and the Library of Congress record showing the US edition is a revised edition, which version 2 listed as unconfirmed. Each section now opens with a one-line definition. | Published |
| Oct 12, 2026 | Change | The Ten House Protocols: amendment R13 | Approved by Lawrence Ho by chat message at 21:04 CDT. Protocol 2 now says what to do when we do not have a work: ask the person for it, then search what is public and lawful, keep every source traceable, and label every claim as in the book or about it. | In force |
| Oct 12, 2026 | Change | A worked example of the protocols is published | A summary of Chris Patten’s Not Quite the Diplomat, written only from sources we opened, with its certainty scores and the six mistakes we found in our own first draft (two review dates, a paraphrase inside quotation marks, two details from memory, one view credited to the wrong person). The draft was never published, so these are listed on that page, not here as corrections. The two-reviewer step has not run on it, and the page says so. | Published |
| Oct 12, 2026 | Change | The Ten House Protocols: amendments R11 and R12 | Approved by Lawrence Ho by chat message at 20:36 CDT. R11 adds the studio’s writing checklist to Protocol 6; R12 adds the attestation and the rules for a breach to Protocol 10. There are still ten protocols. The writing checks were added to the house gate the same day. | In force |
| Oct 12, 2026 | Change | The address is labeled as the postal address | Wherever the studio’s address appears (the foot of every page, the site map, About and Terms) it now reads “Postal address:” before it. The address itself did not change. | In force |
| Oct 12, 2026 | Change | Home: the protocols move up, under the headline | The headline stays: an AI that shows its receipts. A line under it now says the promise is kept by the Ten House Protocols, enforced by code on every change, and the protocols sit directly under the two opening buttons, above the four figures. To be reviewed on Monday Nov 09, 2026, when the protocols are due to be fully running. | In force |
| Oct 12, 2026 | Change | Jester Mind: the name, and the studio’s projects | The studio’s AI model is written Jester Mind, two words, as in its grant application of Sep 17, 2026; the site briefly wrote it as one word today. Like Protocol STAMP, Jester Mind is a project of Budday Budderson Studio LLC, and the publisher line on every page now names both, with bud.day. | In force |
| Oct 12, 2026 | Change | The Ten House Protocols, redesigned, and at the center of the home page | The protocols page now shows each protocol as a card with its rule in a sentence and the full text a click away, the evidence and access ladders as diagrams, the certainty scale drawn to scale, and the status of each part as a pill. On the home page, the protocols are the centerpiece, with all ten at a glance. No rule changed. | In force |
| Oct 12, 2026 | Change | New page: The Ten House Protocols, and JesterMind announced | The ten house protocols approved today, with amendments R1 to R10, are now public, with what runs today and what is still being built. The page also announces JesterMind, the aggregated AI model Protocol STAMP is building; no detail of it is published until it exists. Linked from the top of every page and from the home page. | In force |
| Oct 12, 2026 | Change | The Ten House Protocols, adopted | Lawrence Ho approved ten house protocols against AI hallucination and slop, electronically, by chat message: every factual sentence names its source with a working link; the source is opened before the sentence is written; quotes are exact and our reading is labeled as ours; each claim carries a certainty score computed from the evidence checked; every date is real and in the house form; no unmeasured adjectives; the house voice; issued work stays frozen; every change is recorded here the same day with its source; and two independent AI reviewers, then a person, check every issue. From today the house gate runs on every change to this site and on every release, and a failure stops publication. Pages published before today keep their current labels until they are scored, by Monday Nov 09, 2026. | In force |
| Oct 12, 2026 | Change | Every date in the house form, everywhere a reader sees one | A new check (house_gate.py) found 190 dates and credit lines that broke house rules. About 120 dates written day first or as bare numbers inside Docket and Chronicle entries now read like “Apr 20, 2023”; four shared-month pairs are written as two full dates; the release line on five issue pages, outside the issued bytes, and one Militaries date range follow the same form. Quotations were not touched. No date’s value changed. | In force; no issued byte changed |
| Oct 12, 2026 | Change | The house mark, in full, on seven Bench pages | Five Bench posts and two run records credited the work as “Drafted by AI · Edited together · Signed by a person”. The house mark has four steps, always in this order, and each now begins “Begun by a person”. The Bench Vol 001 issue keeps the footer it was issued with. | In force |
| Oct 12, 2026 | Correction | The Bench Vol 006: “no AI company has followed” | The post, issued Oct 12, 2026, said no AI company had followed the US rename. SpaceXAI LLC had filed a trademark application for “SPACEXSI” on Oct 04, 2026 (USPTO serial 50143825), and its X account was reported renamed @SpaceXSI on Oct 10. A dated correction now sits under the sentence, which is left as issued. | Corrected in the open |
| Oct 12, 2026 | Correction | Super Intelligence (SI): SpaceXAI | The page said reports that SpaceXAI would become SpaceXSI were not confirmed on the companies’ own channels, and listed SpaceXAI among companies that would not say. A dated entry now records the Oct 04, 2026 trademark filing (read at the USPTO), Musk’s post the same day, and the reported Oct 10 account rename; the company’s legal name is unchanged. Four sources added, 43 in all. | Corrected |
| Oct 12, 2026 | Change | The standing notice now reads “teamed with AI” | The line at the foot of every page now reads: “Protocol STAMP is teamed with AI and can make mistakes. Verify before you implement anything.” It replaces “uses AI”, the wording of earlier today. | In force |
| Oct 12, 2026 | Change | A standing notice at the foot of every page | Every page now ends the screen with one line, as AI chat products do under their reply box: “Protocol STAMP uses AI and can make mistakes. Verify before you implement anything.”, with a link here. It says “uses AI” and not “drafted by AI” because a Calibration’s prose is written by a person. Issued documents are not changed; the line sits in the page around them. | In force |
| Oct 12, 2026 | Change | Timeline dates clear of the line; style updates seen at once | The timelines on Super Intelligence (SI) and Militaries and AI still set the date against the line: the list inherited a left indent the line did not. The date column is now wide enough for its longest date, with an even gap either side of the line. Each page now asks for its stylesheet and script by version, so a change to them shows on the next visit instead of hours later. | In force |
| Oct 12, 2026 | Change | Newest first, everywhere; figures above the register; one chart palette | A site-wide rule: every dated list and timeline runs newest first, as the Archive, Corrections and the Docket and Chronicle registers already did. The timelines on Super Intelligence (SI), Militaries and AI, and in the Docket’s and Chronicle’s featured explainers now follow it, and their dates sit clear of the line. The featured explainers now come before the register. The charts there use the site’s figure colours, so red is kept for a measured figure. No entry, date or figure changed; the pages keep their Oct 10, 2026 build date. | In force |
| Oct 12, 2026 | Correction | The welcome email’s open and click tracking: said off, was on | On Oct 11, 2026 Privacy and the row below said open and click tracking were turned off for the welcome email. The setting had not saved: both were on until Oct 12, 2026, and one welcome email was sent in that time, to the publisher’s own test sign-up. Both are now off, checked by reopening the setting. The same day the email’s logo was replaced with the signature mark, the H with its red bar. | Corrected |
| Oct 11, 2026 | Change | The sign-up form now subscribes you; the Reader’s Kit arrives on confirming | The subscribe page’s form is now EmailOctopus’s own, with double opt-in. Until today it opened your mail app and requests were added by hand. Confirming sends a welcome email with the Reader’s Kit, and that email has open and click tracking off. The form includes a hidden Google reCAPTCHA, loaded on the subscribe page only, so Privacy now lists five third parties, not four. | In force |
| Oct 11, 2026 | Change | New page: The Reader’s Kit | Eight questions to ask of any AI test score, each taken from Methodology or Definitions and linked to it, with the 37 public probes and The Calibration Vol 004 to practise on. Free, with a PDF. No new figure. | In force |
| Oct 10, 2026 | Change | Colour and small charts at the top of pages; the phone header | Figure tiles and home-page icons now carry colour. Red stays reserved for a measured figure, as the Colophon says. Five tiles show their own numbers as a bar: the bank (37 public, 26 held out), the UN vote on Res. 80/57 (164 for, 6 against, 7 abstained), REAIM 2026 signatures (35 of 85) and the two Independence caps (10% and 25%). No number was added or changed. On phones, the header button no longer runs off the screen, and the announcement line wraps instead of hiding off the edge. | In force; no issued byte changed |
| Oct 10, 2026 | Change | The publisher’s business address, and what it publishes | Every standing page, the site map at the foot of every page, About and Terms now say that Protocol STAMP and bud.day are under Budday Budderson Studio LLC, a New Mexico company, at 2105 Vista Oeste NW, Suite #E3993, Albuquerque, NM 87120. Issued documents keep the footer they were issued with. | In force; no issued byte changed |
| Oct 10, 2026 | Change | Every issue’s Source — served as plain text, and its hash matches again | Our host adds a bot-detection script to web pages on protocolstamp.com, and with Bot Fight Mode already off we have found no setting on our plan that removes it. Each issue’s Source is now served as plain text, which it does not touch. Checked Oct 10, 2026, on the live site: every Source hashes to the build printed in its PDF (The Calibration Vol 001 to Vol 004, The Log Vol 001 to Vol 003, and The Bench Vol 001). Privacy and the note on every issue page now say so; the earlier row today stands as the record of the problem. | In force; no issued byte changed |
| Oct 10, 2026 | Correction | Privacy, every issue page — what our host adds at the edge | Privacy said the site “injects nothing at the edge.” Our host, Cloudflare, inserts a short bot-detection script (about 940 bytes) into every page it serves, the Source files of the issues included, so a Source file downloaded from this site does not hash to the build printed in its PDF. Checked Oct 10, 2026: the copies in the public repository hash exactly (for example The Log Vol 003, 4bda581b1159; The Calibration Vol 001, be2bc9d7ca35). Privacy now says what is added, and every issue page links its exact copy. The script is to be switched off at the host; when it is, this row will say so in a new one. | Corrected on standing pages; no issued byte changed |
| Oct 10, 2026 | Correction | AI Safety + Governance, Archive — how many run files can be downloaded | AI Safety + Governance said fifteen files were registered and fourteen could be downloaded, and the Archive row above, written today from it, repeated that. Counted against figures.json today: twenty-six files are registered and twenty-two can be downloaded (twenty-one at Run files and the cross-model data); twelve are served exactly as recorded and each hashes to the registered value; ten are redacted on held-out items. The probe bank is never published, and three working files of figures are not posted yet. Both pages now say so. | Corrected on standing pages; no issued byte changed |
| Oct 10, 2026 | Change | Every page — figures first, a solid header, the issued documents framed | The header is no longer see-through, and the day-and-night switch moved into it from the floating pills, which had covered text. Every standing page now opens with its figures under the title; eight pages that had none have them. Calibration, Log and Bench Vol 001 pages carry the site’s menu and an “at a glance” card above the issue; the issued document inside each page is byte-for-byte unchanged, and its SHA-256 still matches the build printed in its PDF. Also: breadcrumbs, a reading-progress line, back to top, “Cite this page,” a skip-to-content link, a site map in the footer, and menus whose descriptions run from the left. Privacy now says what the frame around an issue runs: the site’s own switch, menu and page search, and no third-party script. | In force on every page; no issued byte changed |
| Oct 10, 2026 | Correction | Archive — the run files | The archive said the run files behind figures.json were “not published yet, so a reader cannot recompute them.” Fourteen of the fifteen are published at Run files, as AI Safety + Governance already said; the fifteenth is the probe bank. The sentence now says so. | Corrected on standing pages; no issued byte changed |
| Oct 10, 2026 | Change | Methodology, Archive, The Bench, AI Safety + Governance, The Docket, The Chronicle — the rest of the redesign | The remaining standing pages carry the same menu and shell as the rest of the site, and figures built from what each page already says: the instrument at a glance and the bank as it now stands (37 public, 26 held out, bank v1.1) on Methodology, without editing Methodology v1.0’s text; live counts on the Archive and The Bench; the two kinds of work and the files behind the figures on AI Safety + Governance. The Docket and the Chronicle take the new menu; no entry changed. | In force on standing pages; no issued byte changed |
| Oct 10, 2026 | Correction | Docket and Chronicle — three entries | The Docket entry on the President’s UN address said “a global scheme of control”; the White House record says “globalist”. The Chronicle entry on the UN expert group on autonomous weapons said the session ended on Sep 04, 2026 with a draft report; it ran into the early hours of Sep 05, 2026 and adopted a final report (Human Rights Watch, Sep 07, 2026). A source link for the REAIM 2026 summit that now returns an error was replaced with the Reuters report of Feb 05, 2026. Found by the site audit; checked against the sources before correcting. | Corrected on the Docket and the Chronicle |
| Oct 09, 2026 | Correction | Super Intelligence (SI), Militaries and AI — a claim-by-claim check against the sources | Every claim on both pages was checked against the source it cites, and the primary text where one exists (the Federal Register, whitehouse.gov, usun.usmission.gov, govinfo.gov, the UN Digital Library). Corrected: the UN resolution 80/57 passed 164–6–7 in the General Assembly on Dec 01, 2025; the page gave 156, which was the First Committee vote. The President rejected a “globalist scheme of control,” not a “global” one, and the US at the Security Council rejected “a globalist scheme of control for superintelligence” and “a global regulator,” not “a global regulatory scheme.” The CCW expert talks ended in the early hours of Sep 05, 2026 with an adopted, non-binding final report. The Political Declaration had 53 endorsers by Mar 2024, not “around 50.” Airbus agreed a €50 million contract to use AMIAD-developed AI; the page said AMIAD had contracted it. Removed: a statement that Jeff Bezos likes “SI,” which none of the cited sources carries. Reworded to what the sources say: what Google, Anthropic and NVIDIA said about the term. Moved to sources that carry them: two citations. Replaced: two dead links (NATO, AOL). Added: five references. | Corrected on standing pages; no issued byte changed |
| Oct 09, 2026 | Change | Every standing page — one design, one menu; About rewritten | The menu was eleven links in one row; on a laptop the last of them, About, fell off the edge, and on a phone none were shown. It is now five groups (Publications, Method, AI policy, Trust, About), each with a short description of every page in it, and a full menu on phones and tablets. The SI and Militaries pages now use the same page shell as the rest of the site. Tables, timelines and figures carry their sources in the figure; the home page shows where each headline number comes from. About is rewritten from what the site already says; its statement that “the current issue is The Calibration Vol 001” was out of date. The footer’s “Public bank” now links the current public half (bank v1.1). Separately, the script that releases an issue now writes protocolstamp.com and “Protocol STAMP” on the issue page and flips the home and archive rows; from The Calibration Vol 004 on. Checked on phones, tablets, laptops and televisions. | In force on standing pages; no issued byte changed |
| Oct 09, 2026 | Correction | Chronicle — UN General Assembly resolution 80/57 | The entry said 156 states supported the resolution. That was the First Committee count on Nov 06, 2025 (156 to 5, with 8 abstentions); the General Assembly adopted it on Dec 01, 2025 by 164 to 6, with 7 abstentions. Corrected, still labelled Likely because the UN record could not be opened by our fetcher. The same day, the Docket and Chronicle entries on Anthropic’s cases gained a fuller disclosure: Anthropic makes Claude, and these entries were drafted and checked by a Claude model. Seven Docket and eleven Chronicle entries that had been published without passing the Docket’s checks now pass them; one Docket summary was shortened to fit. | Corrected on the Chronicle and the Docket |
| Oct 09, 2026 | Change | Independence — the request letter, and a proposal for the Dec 03 review | The exact words every company on the request table will receive are now published, before any is sent. A proposal to allow one labelled sponsor line in a weekly newsletter, from companies that are not measured here, is published for decision at the scheduled review on Dec 03, 2026. No condition changed, nothing has been asked of anyone, and no offer is open. | In force on standing pages; no issued byte changed |
| Oct 09, 2026 | Correction | Privacy, Subscribe — the mail list and its provider | The privacy statement said the subscriber list was stored “on a server the publisher controls” and that no open- or click-tracking pixels were used. No list server was ever running: sign-up requests arrived as ordinary email at s@bud.day. The list will now be kept and sent by EmailOctopus, which records opens and clicks by default; the statement names it, links its data-protection terms, says what it records, and says we will turn tracking off if its settings allow. The statement also wrongly counted “two” third parties (there are four, one of them advertising not yet switched on) and placed the advertising slot on “The Bench essays” rather than the home page and the two essays at the foot of The Bench. New: what we collect (email only, no phone number), deletion of unconfirmed sign-ups, how to see or delete your data, the legal basis for European readers, and an age line. | Corrected on standing pages; no issued byte changed |
| Oct 09, 2026 | Change | The Docket, The Chronicle, Militaries and AI — military AI and the SI rename in the registers | Seven Docket entries (DoD Directive 3000.09, the January 2026 “AI-first” memo, NSPM-11, the UK’s JSP 936, NATO’s revised strategy, the Political Declaration, REAIM 2026) and eleven Chronicle entries (the Biden–Xi statement on nuclear control, UN resolution 80/57, Microsoft and Israel’s defence ministry, the Anthropic designation and its three court rulings, the end of the CCW expert talks, the Autonomous Warfare Command, and two on the SI rename). All are labelled Likely because their sources were found by web search and could not be opened; each entry’s Limits say so, and both registers’ rules now state it. Each register has a new In brief point. Militaries and AI now also gives the Mar 26, 2026 preliminary injunction in the Anthropic case, which it had omitted, and both standing pages link to their entries. Both pages are featured on the home page and in the RSS feed. | In force on standing pages; no issued byte changed |
| Oct 09, 2026 | Change | Super Intelligence (SI), Militaries and AI — two new standing pages | Two pages added and linked from the top of every standing page. Super Intelligence (SI) traces the US rename of AI under Executive Order 14434: a timeline, who has followed it and who has not, why the term is contested, and the house rule (we write “AI” and quote ‘Super Intelligence’ (SI) only when naming a US instrument that uses it), with 37 references. Militaries and AI maps how armed forces govern AI: the UN autonomous-weapons talks, REAIM, the US Political Declaration, NATO, and national policies from the US to Ukraine, with 62 references. Both were sourced by web search; each page says where a source could not be opened. | In force on standing pages; no issued byte changed |
| Oct 09, 2026 | Change | The Docket, The Chronicle — world coverage and explainers | Twenty entries added from Russia, the United Arab Emirates, Saudi Arabia, Egypt, Kenya, Rwanda, Nigeria, the African Commission on Human and Peoples’ Rights, Chile, Colombia, Argentina, Latin America’s Santiago and Montevideo declarations, Malaysia, Türkiye, Israel, Ukraine and Indonesia, and, in the Chronicle, South Africa’s withdrawal of a draft AI policy whose references were fictitious. Sources are the issuers’ own documents, cited in their original language. Where an issuer dates a document to the month only, the entry says so. Both pages now open with an executive summary (In brief) whose every point links to its entries, and carry twelve featured explainers; the build refuses a summary or explainer that points at an entry that does not exist. The Chronicle now has its own place in the site menu, beside the Docket. | In force on standing pages; no issued byte changed |
| Oct 08, 2026 | Change | House style — how dates are written (master rule) | Nothing here was wrong; the style changed again. From today every date on the standing pages, The Bench, the Docket, the Chronicle and this register is written one way only: Oct 08, 2026. The form with a weekday in front of the date is retired, and a span of days is written as two full dates. Issued documents are not touched, and issue names such as Sun20260830 stay as they are; machine fields (the sitemap and the JSON feeds) keep the standard machine format. Every Docket and Chronicle entry now carries a permalink and a ready-made citation. | In force on standing pages; no issued byte changed |
| Oct 08, 2026 | Correction | The Docket — New York RAISE Act quote | The entry quoted Governor Hochul as saying “setting a strong standard for frontier AI safety”. Her release reads “setting a strong and sensible standard for frontier AI safety”. Two words were dropped without an ellipsis. Found by a second audit that read every Docket source in full and matched every quote word for word. The quote is now exact. | Corrected |
| Oct 08, 2026 | Correction | The Docket — UK King’s Speech entry | The entry named a Regulating for Growth Bill with AI sandbox powers, but the source it printed (IAPP) does not mention that bill; the claim rested on reports we had read and not cited. The entry now cites a law-firm note that quotes the government’s briefing on the bill. It stays Likely: the Commons Library briefing still refuses our fetcher. | Corrected |
| Oct 08, 2026 | Change | The Docket — Super Intelligence order | The White House has published the order as Executive Order 14434. The entry said its number had not been retrieved; it now names the number, links the order, and states its interim definition (15 U.S.C. 9401(3)). | Updated |
| Oct 08, 2026 | Change | The Docket, new: The Chronicle | The Docket now reaches back to 2020: 120 entries added worldwide, including the President’s UN address of Sep 22, 2026 and earlier speeches that announced AI safety institutes. A companion page, The Chronicle, holds the wider record the Docket’s scope test leaves out: court rulings, enforcement, speeches and milestones. Every entry’s key claims were matched verbatim against the full text of its sources. The gate’s fixed lists gained jurisdictions, instrument types and stages, and a named list of official issuer domains; archive copies of official records are judged by the host they copy. | In force on standing pages; no issued byte changed |
| Oct 06, 2026 | Change | Home, Archive, The Bench — rows for what is not out yet | A row for an issue or post that is due but not yet published read only “Due” and a date, and beside published rows it could be taken for something already shipped. Such rows now say Not yet published · due and the date. Their notes said “Posted 07:37”, which also read as done; they now say “Scheduled for”. The same day, the one-line summary of The Bench Vol 004 on the home page, The Bench index, the feed and the post’s own description said the agreement bar was “nearly met” by a judge that always says PASS; the post itself says it was met, to three decimal places, and the summaries now say the same. | In force on standing pages; no issued byte changed |
| Oct 05, 2026 | Correction | The held-out half of the bank — two probes | Two probes we counted as held out were not secret. 01_arithmetic_076 is the “$120k, $95k and $60k, total $290k” board summary, which is the headline example in our README and on the protocol page, public since at least Sep 02, 2026. The subject of 14_provenance_002 was named in figures.json. A new scanner that reads reworded text found both; the holdout gate could not, because it matches exact word runs and this probe has none of its own. Both probes are now public, with their text, in bank_v1_1_public.jsonl (bank v1.1): 37 public, 26 held out. Figures already issued were measured on bank v1 and are not changed; any figure that splits public from held-out items will say which bank version it uses. The gates now read v1.1, and the new scanner runs beside the holdout gate. | Reclassified; issued figures unchanged; no issued byte changed |
| Oct 05, 2026 | Change | House style — how dates are written | Nothing here was wrong; the style changed. From today every date on the standing pages, The Bench and this register is written as Oct 05, 2026, or with the day as Oct 05, 2026. Dates written without a year now carry one. Issued documents are not touched: The Calibration Vol 001–003, The Log Vol 001–003 and The Bench Vol 001 keep the dates they were issued with, in their reading pages, sources and PDFs. Issue names such as Sat20261003 are identifiers and permanent addresses, and do not change; nor do the machine timestamps in the feed and sitemap. New issues use the new style. | In force on standing pages; no issued byte changed |
| Oct 05, 2026 | Correction | Definitions — The Bench and volume numbers | Definitions said The Bench “carries no volume number”. It has carried one since Vol 002, as About and the archive already said. The sentence now says so. The same day, About gained a side-by-side explainer of The Calibration, The Log and The Bench; nothing in it changes a rule. | Corrected on Definitions; no issued byte changed |
| Oct 03, 2026 | Change | The Bench — the catch-up ends; Mondays and Thursdays stand | Nothing here was wrong; a rule set aside two days ago is restored. The Oct 01, 2026 row below set aside the Sep 10, 2026 rule that “a missed slot is not made up by posting twice”, so that the Bench posts missed in September could be published late. One was: Vol 003, issued Oct 01, 2026 for the slot of Sep 14, 2026. The rest will not be. The slots of Thursday 17, Monday 21, Thursday 24 and Sep 28, 2026 and Oct 01, 2026 stay missed, and this row is where that is recorded. From now on The Bench goes out at 07:37 America/Chicago on Mondays and Thursdays, and a slot that is missed is missed. The next is Vol 004, on Oct 05, 2026. Separately, the two pieces of Sep 02, 2026 are now labelled Essay in the listings: they came before the Bench was numbered, and labelling them keeps the numbered run unbroken. The pieces themselves are unchanged. | Catch-up ended; five September slots stay missed; Vol 004 due Oct 05, 2026 |
| Oct 03, 2026 | Change | The Calibration Vol 003 — a second issue the same day | Nothing here was wrong; the calendar moved again, and this row says so. A second Calibration was issued today, hours after Vol 002, to ask the question Vol 001 and Vol 002 both left open: whether the judges agree with a person. It was pre-registered and signed at 12:00, before the first label; the labelling page refused to save until the signature was on file. Like Vol 002 it is off-cycle on The Log’s Saturday, and its draft was adversarially reviewed by an independent agent before signing. Its labels are served at Run files. The next Calibration slot, Oct 10, 2026, now belongs to Vol 004. | Issued Oct 03, 2026; Vol 004 due Oct 10, 2026 |
| Oct 03, 2026 | Change | The Calibration Vol 002 — issued, and early | Nothing here was wrong; a date in the Oct 01, 2026 row moved. That row re-dated The Calibration Vol 002 to Oct 10, 2026. It was issued today instead, Oct 03, 2026, because its re-frozen pre-registration was signed on Friday at 11:45 and every run finished before its 18:00 freeze that day. Today is The Log’s Saturday, so the issue is off-cycle, and its own front page says so. Its draft was reviewed adversarially by an independent agent before signing, which found eighteen faults in what the prose claimed about correct numbers — among them a verdict on decision-model judges that rested on a test nobody pre-registered, and a comparison of two judges on different halves of the bank. All were fixed before signing, and the issue labels as exploratory every figure added because of them. Its run files are served at Run files, redacted on held-out items as before. The next Calibration slot is Oct 10, 2026. | Issued Oct 03, 2026; Vol 003 due Oct 10, 2026 |
| Oct 01, 2026 | Change | AI Safety + Governance — what an agent may do unattended | Nothing here was wrong; a rule changed, and the safety page promises that its rules change only in a row here. Until today that page said an agent may not publish or push to the repository. The Log is now the one exception. It is computed from the registry and nobody writes it, so its release was always the part of this program meant to run by itself; what it lacked was permission to. The nineteen days recorded in the row below show what a schedule that waits on one person looks like when that person is away. From the Log due Oct 03, 2026, an agent may build, gate and push a Log by itself, on four conditions, each a script: the schedule gate says a Log is due; the holdout gate passes; the investigator was sent the finished issue at least twelve hours before release; and he did not reply to stop it. Nothing else changes. A Calibration is still issued only by Lawrence Ho, a Bench post still carries his signature, and a slot he did not sign stays missed and is recorded here. Separately, The Log Vol 002 went out today on a Thursday to fill the missed Sep 19, 2026 slot, released by hand. Its generator now prints the masthead title in the form set on Sep 11, 2026, and no longer lists registry namespaces that are not models as candidates; Vol 001 keeps the title it was issued with. | In force from The Log Vol 003, due Oct 03, 2026; Calibration and Bench unchanged |
| Oct 01, 2026 | Change | The publishing schedule — nineteen days with nothing issued, and the rule on missed slots | Nothing here was wrong in what was published; what was not published is the subject. From Sep 11, 2026 to Sep 30, 2026 this program issued nothing. The Calibration Vol 002 was due Sep 12, 2026, The Log Vol 002 Sep 19, 2026, a Calibration Sep 26, 2026, and The Bench six times on its Monday-and-Thursday days. None of them appeared, and until today the home page still said The Calibration Vol 002 was due on Sep 12, 2026. The reason is ordinary: the person who signs these pages was away from the project, and while he was, the computer it ran on failed. Its nightly backup had already gone to cloud storage, so no data was lost. A rule is set aside here, and this row is where that is said. The Sep 10, 2026 row below says “a missed slot is not made up by posting twice.” For this gap only, the missed Bench posts are being published now, each dated the day it actually ships and each saying on its own front page which slot it was due in. The Calibration Vol 002 is not caught up the same way, because a Calibration cannot be written before its runs exist: its pre-registration was never signed and no run was made before its freeze date, so it is being re-frozen and will be signed before any run. It is now due Oct 10, 2026. The Sep 26, 2026 Calibration slot is not made up: there is nothing to report in it, and it stays missed. The Log Vol 002 follows as soon as the computer that builds it is running again; it will say that nothing moved, because nothing did. | Bench catch-up from Oct 01, 2026; Calibration Vol 002 re-frozen, due Oct 10, 2026; Sep 26, 2026 slot stays missed |
| Sep 11, 2026 | Change | The run files behind every published number | Nothing here was wrong; something this site said it owed is now delivered, and the shortfall that remains is named. This morning’s AI Safety + Governance page said that fifteen files were registered with a SHA-256 and that only one of them could actually be downloaded — a hash over a file a reader cannot obtain proves nothing to that reader. Fourteen of the fifteen are now served, listed with their hashes at Run files. Seven go out exactly as recorded and hash to the value the register already held. Seven could not go out whole, and the reason is the one thing here that cannot be undone: a model’s response restates the question it was asked, and a grader’s reasoning quotes both, so the raw run files carry twenty-eight held-out probes in plain text. The holdout gate refused all seven. They are published instead with the response and grader fields replaced on held-out items only, every other field kept, and every row retained. That choice was made so the numbers stay checkable: a rate is computed from verdicts, not from the text of a probe, so the redacted files recompute the published figures exactly — nemo_v1 returns 0.43344155844155846 and luna_v1 returns 0.2268370607028754, which are the values on the page, to the last digit. A redacted file hashes differently from the original it came from, so both hashes are printed rather than one quietly standing in for the other. What is still owed: the seven originals stay unpublished, the bank stays unpublished permanently, and the intent files that freeze each round’s questions before it runs are still not served — so these hashes show that a file has not changed since publication, not that the question was fixed before the answer was known. That gap is stated on both pages rather than left for a reader to find. | 14 of 15 registered files now downloadable; 7 verbatim, 7 redacted with rows intact; intent files still unpublished |
| Sep 11, 2026 | Change | The website — the ampersand, and the name of the safety page | Nothing here was wrong; a typographic convention is settled, and a page is renamed within hours of appearing. Every visible ampersand in this studio’s own standing copy is now a plus: the register is Corrections + changes, and the masthead line reads Chief Human Investigator, Evaluation Administrator + Instrument Maker. The page published this morning as “Safety & governance” is now AI Safety + Governance, because the subject is AI and a page ought to say so in its own name rather than leave a reader to infer it from the site around it. Eighty-three ampersands changed across eighteen standing pages. Two kinds were deliberately left alone. A citation keeps its own: an author list joined by “&” belongs to the paper being cited and not to this publication, so “Kalai, Nachum, Vempala & Zhang” stays as those authors wrote it. An issued document keeps its own: The Calibration Vol 001 still prints “Human Investigator & Instrument Maker”, the title held on the day it was issued, and the row below still names this page as it was published, because a work is cited under the name it was issued with. The address does not move: /safety/ is unchanged, so anything linked this morning still resolves. | Convention in force on standing pages; citations and issued bytes untouched; /safety/ URL unchanged |
| Sep 11, 2026 | Change | The website — a safety and governance page | Nothing here was wrong; rules already in force are now written down, three that were not are now promised, and one gap between what this site implies and what it actually serves is stated rather than left to be discovered. Safety & governance sets out the arrangement this program has always run on and had never put in one place: the numbers are computed by code from recorded runs and edited by nobody, and the prose around them is drafted by AI and edited by a person who signs it. A person can decide whether to publish a measurement and can be wrong about what it means; a person cannot make it say something else. That separation, and not a promise of good conduct, is what is offered here as the reason to trust a figure. It is also why the EU AI Act’s disclosure duty for AI-generated text does not attach to this publication: Article 50(4) disapplies itself where the content “has undergone a process of human review or editorial control” and a person holds editorial responsibility, which is exactly the second half of the arrangement. The line in every footer is therefore voluntary rather than compelled, and the page says so. Three commitments on the page are new rather than descriptive, and are marked here as new, because a page that quietly mixes what a program already does with what it has just decided to start doing is not a record of anything: seven days’ notice to a vendor before publishing a finding that could hurt somebody now, buying no delay, no veto and no softening; no working attack strings, only the findings they produce; and a written boundary on what an agent may do unattended — it may search, run the queue, compute, draft and read, and it may not publish, push, spend, write in this program’s name, or settle what a result means. The gap, stated plainly. Fifteen files behind published figures carry a registered SHA-256 in figures.json. One of them can actually be downloaded. One more, the probe bank, is deliberately unpublishable because it holds the twenty-eight held-out items. The other thirteen are not served, and the intent files that freeze each round’s questions before it runs are not published either. A hash over a file a reader cannot obtain proves nothing to that reader, so the page calls that register a promise rather than a proof until the files are served. The page also commits to saying, in a row here, whenever one of its own rules is broken. | Published Sep 11, 2026; three commitments new; 1 of 15 registered files currently downloadable, stated on the page; reviewed Dec 03, 2026 |
| Sep 10, 2026 | Change | The Bench — posting days, and one issue that missed its slot | Nothing here was wrong; a schedule changed, and a deadline was missed before it did. The Bench now goes out Mondays and Thursdays, replacing the Sundays-and-Wednesdays rule set on Sep 02, 2026, which had itself replaced an even-numbered-days rule. The cadence is unchanged at twice a week; only the two days moved. The posting-day check is code rather than a line in a document, so the rule is enforced rather than remembered, and its history is kept in the file: every previous rule is still written there with the date it was replaced. Separately, Vol 002 was due Sep 09, 2026 and went out on Sep 10, 2026, a day late. Finalising the chat sessions it rests on took four days rather than one: several cells had to be discarded and re-run because the run conditions turned out not to match what the prompt said they were, and the interface driving those sessions stalled repeatedly. The discarded cells are published with that issue’s run record rather than dropped. The issue says on its own front page that it is late and why, because a publication that prints a schedule and then quietly misses it is making exactly the sort of unchecked claim this one exists to complain about. A missed slot is not made up by posting twice; the next issue goes out on the next posting day. | Mondays and Thursdays in force from Sep 10, 2026; Vol 002 published one day late, disclosed on the issue |
| Sep 10, 2026 | Correction | The Bench Vol 001 — the run record it says is published | The Sep 06, 2026 issue states, in five places, that its full run record — “all thirteen verbatim transcripts, timestamps, model identities, conversation IDs and screenshots” — is published alongside it, and rests an editorial argument on that publication: the reading of each answer can be disputed, it says, because the transcripts are there. They were not there. The claim is withdrawn as written. What has now been done about it. On Sep 10, 2026 the investigator’s computer, his Google Drive and the public repository were all searched, by filename and by content, for the transcripts and the run record. They are in none of the three. The thirteen sessions were driven by an agent in a cloud container that no longer exists, and the verbatim transcripts, the timestamps, the tool-use metadata and the independent audit of the draft went with it. Those are lost and this register is not going to keep promising them. A run record has been built from what survived and is published at /bench/no-1-complaint/record/. It carries all sixteen screenshots published individually for the first time, each with its SHA-256 and a one-command recipe for reproducing the exact bytes from the plates PDF; a machine transcription of each frame, labelled as transcription rather than transcript; the prompt as sent, with its three known defects, corroborated by four plates that show its opening line on screen in four different companies’ interfaces; and the thirteen runs with their models, platforms and effort settings. It also says plainly what a reader still cannot check: any answer in full, any quotation outside a captured frame, the timings, and the two runs that have no plate at all. This is a smaller thing than what was promised, and it is offered as a repair rather than a fix. An earlier version of this row said two of the thirteen conversations had been located; that was carried over from working notes and the search did not bear it out, so it is withdrawn too. The issue’s own page is not edited: an issued Bench post is corrected in the open, not amended in place. | Claim withdrawn; transcripts and run record confirmed lost; recovered run record published Sep 10, 2026 |
| Sep 10, 2026 | Change | The Bench — volume numbers | Nothing here was wrong; a rule changed. The Bench was deliberately unnumbered while The Calibration and The Log carried volume numbers, and the Sep 06, 2026 issue’s citation line says so. The Bench now carries volume numbers, so that an issue can be cited and argued with by a short name rather than a URL: the Sep 06, 2026 issue is Vol 001 and this cadence continues from Vol 002. The number appears in the index listings, in each issue from Vol 002 onward, and in the feed. Vol 001’s own page is left exactly as issued, including the citation line that says The Bench is not numbered, because that line was true when it was published and an issued post is not rewritten to match a later rule. The rule change is the record; this row is where it lives. | In force from Vol 002; Vol 001 retro-numbered in the indexes only |
| Sep 10, 2026 | Correction | The Bench Vol 002 — a quotation altered in draft | Disclosed rather than quietly fixed, because the issue it appears in is about confident text that does not survive checking. A working draft of Vol 002 quoted a model as saying that “the sources and the figures are accurate.” What the model wrote was “the sources and the G2 figure are accurate.” The alteration widened one specific verified figure into all of them, in the direction that made our own point land harder — the same edit, in the same direction, as the altered quotation this register recorded on Sep 06, 2026. The published issue carries the sentence as written. It was caught by an adversarial review of the draft before publication, along with twenty-three other defects, the largest being a headline count of “24 of 24” that was really twelve cells counted twice. A second review of the corrected draft found thirteen more, including a claim in this very row that the evidence-plates PDF runs to twenty-two pages. It runs to eighteen; twenty-two is the number of defects a different audit found in the previous issue, and the two had been crossed. That is thirty-seven defects across two reviews of one draft, and the number is printed because a review count that only ever goes down is a review that has stopped working. None of those reached a published page. The same review found we had applied two different evidentiary standards inside one draft: a model’s report of what a benchmark’s paper says was treated as established, while another model’s report about its own run was treated as unverifiable. The first was then checked at the source, where it held up, and the issue now says so and cites the paper rather than the model. All of this is recorded here anyway: a draft error caught in review is not a correction to the public record, but a publication that only ever discloses the ones that escape is measuring its review process, not its accuracy. | Quotation restored before publication; count corrected to twelve of twelve |
| Sep 09, 2026 | Correction | Independence — the promised support page | The Independence page has said since Sep 08, 2026 that a supporter is named “on this page, on the support page, and in the home footer.” There was no support page. A reader following that sentence would have found nothing, and the sentence was published as a statement of fact about how this program discloses its funding rather than as a plan. The page now exists at Support, carrying what money would pay for in the order it would be spent, the terms reader support comes under, and the same three ledger lines this page’s companion carries, so the two cannot drift. It says plainly that the payment links are not open yet, because a page that takes no money should not look like one that does. Found while preparing the first funding request, before it was sent. | Page published; payment links not open |
| Sep 09, 2026 | Correction | Independence — the objection clause | The clause as published on Sep 08, 2026 ended a supporter’s funding the first time it objected to a finding, and drew no line between disputing a result and reporting an error in it. A company that noticed a genuine mistake in a measurement of its own models — a mislabeled route, a misattributed cost, a grader artifact — would have had to choose between correcting the record and keeping the relationship, and the rational choice was silence. That made the parties best placed to catch errors the least able to report them, on a publication whose central practice is a public register of its own mistakes. The clause now separates a correction, which is checked, is welcome from anyone, and ends nothing, from an objection, which does what it always did. Where the difference is genuinely unclear the claim is treated as a correction and the checking is published, so that an inconvenient correction cannot be quietly reclassified. Written and revised while nothing had been asked of anyone and no company had offered anything. | Clause revised; no support has been offered, accepted or refused |
| Sep 09, 2026 | Change | Independence — the request table | Nothing here was wrong; something promised now exists. The first of the ten conditions published yesterday said the full request list would appear on the Independence page before the first letter goes out, with the answers blank, so that it could not be assembled afterwards to flatter anyone. It is now there: 29 companies that make, serve or route models this program measures, every response column empty because nothing has been asked of anyone. The rule for inclusion is stated on the page rather than left to judgment — any company whose models this program has measured or expects to measure under the current or next quarter’s pre-registered intent. The list is not the industry and does not claim to be; there are several hundred models in the catalogue this program draws from. It will grow, and every addition is a row here with its date and its reason, asked in the same words as everyone before it. A company is never removed from the table once listed: if it leaves scope the row stays and says so, because a list that can shed names is a list that can be tidied after the fact. | Table published; nothing asked of anyone yet |
| Sep 08, 2026 | Change | Independence — what this program will accept | This is a reversal, and it is stated as one rather than folded into a rewrite. Until today the Independence page said this program was not funded by any vendor it measures and that it would accept nothing from them. From today it accepts cash, API credits and metered keys from companies whose models it measures, under ten conditions published on this site before a single application was filed and while no company had offered anything. The date is the whole of the argument: rules written cold are principles, and the same words written the week a laboratory offers twenty thousand dollars are terms shaped to fit an offer. The reason for the reversal is arithmetic and is not flattering to state — at the cost of measuring frontier models every week, the philanthropic lane alone will not carry the program, and a benchmark that cannot afford to measure the expensive models is not impartial either, only quietly incomplete. What this does not fix: some readers will discount findings about a funder whatever the rules say, and that is not an unreasonable position to hold. The Sep 06, 2026 row below records that the funding disclosure was not softened; that was true on Sep 06, 2026 and is superseded by this row, which is what a register is for. Current support from companies this program measures: none. | Standing changed; ten conditions published; nothing yet asked of anyone, and nothing accepted |
| Sep 08, 2026 | Change | Independence — the ledger and the naming threshold | Nothing here was wrong; it was too loose. A supporter used to be named at $1,000. That threshold would have let a company that makes, serves or routes models this program measures give $99 of credits — or free access worth nothing at all on paper — and appear nowhere on the site. The threshold is now set by who rather than how much: any measured company is named at any amount, including zero, because the conflict is the relationship and not the sum; any other business, foundation or institution at $100; an individual at $100 with their consent. Anonymity is available to individuals and to no one else. Two things were added that did not exist before: a record of support declined, kept in the same ledger and on the same terms as support accepted, because a record of what was turned down is the only evidence that rules bind when they are inconvenient; and a scheduled review of the conditions on Dec 03, 2026 and quarterly after, with a standing refusal of any loosening proposed within thirty days of a specific offer. | Threshold lowered and re-based; declination record and review date added |
| Sep 08, 2026 | Change | The masthead line | Nothing here was false. Every page carried “Independent · No vendor funding, no sponsors.” That was true when written and is still true today, but it stated as a permanent prohibition something that is only a current fact, and the row above changes the prohibition. Standing pages now read “Independent · Who pays is published,” which is the promise this program can actually keep for as long as it exists, and which does not have to be rewritten on the day a first supporter appears. The current fact has not gone anywhere: it is on the Independence page, in the ledger, in the line that reads Current support from companies this program measures: none. The first Bench post keeps the old line in its own text and its argument is not rewritten; only the shared navigation and footer around it were updated to the new name. | Changed on standing pages; issued bytes untouched |
| Sep 07, 2026 | Change | The domain | Nothing here was wrong. The site moved from thestampprotocol.com to protocolstamp.com, recorded on the day it completed rather than the day it was decided. The old domain has not been retired and will not be: it stays registered and pointed, and every request to it returns a 301 to the same path on the new one, so a link or a citation made before today still lands on the thing it named. The frozen source of the first Bench post prints thestampprotocol.com URLs inside its own bytes; those are issued bytes, they are not edited, and the redirect is precisely what keeps them working. Wrapper pages, the feed, the sitemap and every standing page now give the new address as canonical. | Moved; old domain kept and redirecting |
| Sep 07, 2026 | Change | The program’s name | Nothing here was wrong. From launch until today the program was published as The STAMP Protocol; from today it is Protocol STAMP — two words, in that order, never run together and never camelCased. Two reasons, and they are separate rather than one dressed as the other. The first is the name itself: said aloud, “the STAMP protocol” is already an IETF standard of that exact name (RFC 8762, see Other STAMPs), and “the STAMP Protocol protocol” is how the old name kept coming out in the mouth. The second is money: the program is opening to reader support and named sponsorship for the first time, and a name a sponsor cannot be listed beside without ambiguity is a poor one to open under. The name is not an acronym and the five letters are still not expanded — it is the postmark sense, set out in the colophon. Everything already issued keeps the name it was issued with: The Calibration Vol 001, The Log Vol 001 and the first Bench post are unedited, their bytes unchanged and their printed fingerprints intact. The move to protocolstamp.com is a separate operation and is recorded here on the day it completes, not before. | Standing pages renamed; no issued byte changed |
| Sep 06, 2026 | Change | The website | Nothing here was wrong, and no claim about who does the work has changed. The program has always been made by four named parties — a person, a company, an AI agent and deterministic code — but six sentences still described it as one person with tools, which understated the arrangement rather than overstating it. Those now read in the studio's voice, and the About page is framed as a crew whose permissions are published, including what the machines are forbidden to do. Two things were deliberately not done. No collaborator, associate or partner is named who does not exist: the crew page lists exactly one human today, and it will list a second on the day there is one. And the funding disclosure was not softened — the program is paid for by the studio out of its own pocket, with no outside money in it, which is stated as plainly as before. | Voice only; no change to who does what, or to any figure |
| Sep 05, 2026 | Correction | The Log Vol 001, and the release itself | The Log was due at 15:37 and went out at 17:31, 115 minutes late. The scheduled job was queued by GitHub and did not start until 17:15 local; the release script then refused, because it only accepted a 49-minute window around 15:37 — and, worse, reported success while publishing nothing. Two faults, and the second is the serious one: a release that can quietly do nothing and still pass is not a gate. Both are fixed. The window is gone: on a Saturday, once 15:37 has passed, whatever is staged is released, however late the runner starts. An issue left stranded in the staging folder now fails the run instead of passing it, and the job checks that something was actually published before it reports success. Separately, the foot of an issue used to print “15:37” because that is what the schedule promised, not because anything measured it. It now prints the minute the release really ran, and says how far off the mark it was — this issue's foot reads 17:31, 115 minutes late. A programme that asks models to publish measurements rather than claims should not have been printing a claim in its own dateline. | Released late and says so; window removed, stranded-issue gate added and shown to fire, foot now measured |
| Sep 04, 2026 | Correction | The website and The Calibration Vol 001 | From launch until today the site gave its dateline as Albuquerque, New Mexico, on every page. The publisher is a New Mexico company and that address is real, but a dateline names where the work was done, and the work is not done there. The standing pages now name the publisher only; a dateline belongs to an issue, and from The Log Vol 001 each issue carries its own, recorded when it is staged. Vol 001 of The Calibration carries the old wording on its cover and stands as issued. | Corrected on the site; Vol 001 unedited |
| Sep 04, 2026 | Change | The website | Nothing here was wrong. A standing statement of fallibility was added — in the foot of every page on the site, the issues and the feed included, beside the receipt on the front page, and in full under Terms — because the site had been describing the checks it runs without ever saying plainly that a model running the protocol can be wrong, that a published figure can be wrong, and that a reader should verify anything they intend to act on. The checks were already there; the sentence was not. The Calibration Vol 001 carries it in the site foot around its frozen bytes; the byte-exact source at /source is untouched and still hashes to be2bc9d7ca35. | Added; no figure, and no issued byte, changed |
| Sep 02, 2026 | Correction | The website | For about three hours after launch the “hash the published file” invitation could not be honored: a pretty-URL redirect served the wrapper page at the source address, and an analytics beacon added 359 bytes to every HTML page. The fingerprint was never wrong; the served bytes were. | Fixed same day; byte-length check added to every release |
| Aug 30, 2026 | Correction | The Calibration Vol 001 | The issue is dated Aug 30, 2026; it was released at 23:40 on Aug 29, 2026, Chicago time. The release stamp was taken in UTC, where the day had turned. The date stands; the issue wears the Sun finish so the off-cycle date stays visible. | Recorded in the issue’s colophon; stamp now taken in America/Chicago |
Before the first issue, the program corrected itself in its own working files more than once — a confidence interval wrong in both directions at once, a reference model chosen twice on recalled rather than measured information, a gate that passed because it was measuring nothing. Those are recorded in the issue itself under How to read this.
Write to s@bud.day with the issue, the page or figure, and what you believe is wrong. Every report is answered; a confirmed error is issued here under its own date.
The breach register
Protocol 10 (R12) names five breaches by an agent: a source cited but not opened, a quotation not in its source, a claim filled in from memory, a figure or date that is not in its source, and a check reported as run that was not. Each one found is corrected in the register above the same day, naming the model and version, and is also entered here with what was done. What is done is set by R12: the model is stood down from drafting for 30 days, and from reviewing claims on that subject; everything it drafted since its last clean audit is checked again before anything new is published; the audit count restarts. A second breach within 90 days removes it from the team until a person seats it again.
The register starts with R12, in force from Oct 12, 2026. As of Oct 20, 2026 no breach is entered: the corrections made since that day are being read against the five kinds, and any that is one will be entered here with its date.
| Date | Model | What happened | What was done |
|---|