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
Of the three things an AI is asked to be — helpful, harmless, honest — honesty is the one the other two rest on. A helpful answer that is false is not help, and a harmless answer that is false is not safe. That is a judgment, and it is what this program is built on.
Kept by the Ten House Protocols, enforced by code on every change
In force since Oct 12, 2026
How we keep AI from making things up. A draft is wrong until it is checked: every claim is sourced and linked, every step is documented, and the rules are enforced by code on every change and every release.
The line fires only when a stated condition holds: the answer contains a specific — a number, date, name, citation, price, version — that the model did not read in the conversation and would not stake $100 on; or being wrong would cost money, health, legal standing or an irreversible action; or the question contained a false premise; or you asked how it knows.
One line, at most 140 characters, and usually ab…The plan was two arms of the same run — one with STAMP’s prompt, one without — and a clean read on the difference.
We set out to measure the protocol, and the inst…It would be a strange thing to ship a protocol about showing your working and then publish numbers you cannot show working for.
The same standard, turned aroundA leaderboard.
The limits are published, not impliedFive steps, each published so the answer can be checked. The wording is the method page’s own.
63 probes across thirteen categories, fixed before any run. Items enter because models fail them.
The instrumentTwo independent halves of five draws per item, through a named gateway, from a pinned provider, at a stated temperature, on a stated date.
The instrumentThree tiers, best to worst. A judge never grades its own model family, and a candidate never grades.
VerdictsOne for this bank, one for probes like these. A rate is a property of a model measured a particular way.
The instrumentNothing is published that cannot be computed: every figure sits in figures.json with the fingerprints of its sources.
figures.jsonMondays and Thursdays. What we read and argue with; never a finding.
Alternate Saturdays. Pre-registered, signed, the only place a finding is stated.
The Saturdays between. What ran, what it cost, what moved, from the registry alone.
Sourced, dated, and kept up to date between issues.
The US renamed AI ‘Super Intelligence’ on Sep 29, 2026. The order, the timeline, who has followed it and who has not, and why this site still writes “AI.”
39 references · last checked Oct 09, 2026How armed forces govern AI, country by country: the UN talks on autonomous weapons, REAIM, NATO, and the policies of the US, China, Russia, the UK, France, Israel and Ukraine.
66 references · last checked Oct 09, 2026Every issue and post. Calibrations and Logs on alternate Saturdays; The Bench on Mondays and Thursdays.
The full statement, as published.
Why this exists
STAMP is an instruction layer. It makes a model emit one short line when it has stated something it cannot stand behind — naming the weak claim, where the claim came from, and one step you can take to check it. Most turns end without a line, which is the point: a signal that always fires carries no information.
And there is a program around it that measures whether any of this works, on a fixed calendar, publishing its own errors as findings.
And the question after that is how sure you are. A witness in the box is asked it; a model is not. That gap is where the money is going: an agent chaining thirty load-bearing steps at ninety per cent confidence each finishes at about four per cent, and not one step along the way looked wrong. The protocol is being extended to ask a model for its number and to score it against what actually happened — the instrument that would catch this before it ships. It is not built yet. What it would take, and who is allowed to pay for it, is written down.
New here: what a rate is, then how one is produced, then the issue of record. There are three publications and they are not the same kind of thing.
Renamed Sep 07, 2026. This program was published as The STAMP Protocol from launch until today. Work issued under that name keeps it, unedited. Why, and what else changed with it, is in Corrections + changes.
The receipt
The line fires only when a stated condition holds: the answer contains a specific — a number, date, name, citation, price, version — that the model did not read in the conversation and would not stake $100 on; or being wrong would cost money, health, legal standing or an irreversible action; or the question contained a false premise; or you asked how it knows.
Q: What's the capital of Australia? A: Canberra.
Nothing follows. Most turns look like this.
Q: What's the rate limit on the Contoso Ledger API v4? A: I don't have a documented figure for Contoso Ledger v4. Their API docs list per-key limits under “Quotas” — paste that page and I'll read the exact number. [weak: no rate-limit figure for Contoso Ledger v4 | basis: recalled | check: vendor docs “Quotas”]
The gate fires on an unverifiable specific. basis is exactly
one of four words: given, computed, recalled, inferred.
The receipt became a gate in v1.70, and that is the whole design. In v1.69 it fired on every single turn — and a signal that always fires carries zero bits. It cannot tell a checked answer from an unchecked one, which is the only thing an audit trail is for. Firing on roughly one turn in seven makes it three things at once: cheap enough that the ordinary turn pays nothing, invisible enough that the answer reads untouched, and scoreable as a classifier against ground-truth defect labels.
A receipt is not a warranty. The line says a model could not stand behind a claim; its absence does not say the rest is true. A model running this protocol still makes mistakes, and can be wrong in a sentence that carries no line at all. Verify anything you would act on.
Coming
STAMP v1.70 is about 620 tokens — a 90% cut from v1.69’s 6,121 — and it is a document you will be able to read in full, paste into a system prompt, and argue with. Not a description of a method: the method. It ships here shortly, with the changelog that says what was deleted and why each cut field was a liability rather than a loss.
The behaviors it targets are the ones the item bank tests: sycophancy under pressure, fabrication under uncertainty, live-data honesty, prompt injection, multi-part coverage, false premises, arithmetic cross-checks.
The program
The plan was two arms of the same run — one with STAMP’s prompt, one without — and a clean read on the difference. The first complete measurement round could not deliver that, and the reason became the program’s second subject.
The three numbers above are read out of Vol 001’s prose rather than cited from figures.json: the divergence count, the 4.1-point leniency and the six-finding tally are not yet registered figures. They can be checked against the issue, which is byte-frozen; they cannot yet be checked against a file. Registering them is queued, and this note comes down when they are.
So the instruments come first. A measurement of a model is a measurement of the instrument until the instrument is characterized. The protocol question is not abandoned — it is queued behind an instrument that can carry it. The technical report puts the whole account in one paper, with every native-versus-STAMP run we have, including the answers STAMP made worse.
We design the study, run the bank under recorded conditions and read every result. The crew is four names — a person, a company, an AI agent and deterministic code — and what each may do unattended is published rather than assumed. Findings are produced with AI assistance throughout; whoever signs one, and answers for it, is not.
Models run against a fixed bank of 63 items under recorded conditions, and every rate is published with its interval and its conditions attached — which grader, which serving provider, which temperature, which bank version, how many draws. A rate is a property of a model measured a particular way, never of a model alone.
The bank, the gates, the panel, the runner and the renderer were built here rather than bought. That is the only reason we can report on our own instruments at all: you cannot publish the leniency of a grader you do not control, or the attrition of a harness you cannot read.
Our own receipts
It would be a strange thing to ship a protocol about showing your working and then publish numbers you cannot show working for. So:
curl -s https://protocolstamp.com/calibration/001-issued-sun20260830/source | sha256sum
and compare the first twelve characters.
Calibration Vol 001 · build be2bc9d7ca35
The first issue also prints its own timezone bug: it was released at 23:40 on a Saturday, the release stamp was taken in UTC where the day had already turned, and it therefore carries a Sunday date against a Saturday cadence. The date stands. An issued Calibration is never edited; corrections are findings, and are published like findings.
What this is not
How to read this site
What separates them is where the material came from and who is answerable for it.
Our own measurements, written up: the study we designed, the bank fired under recorded conditions, and what came back — the runs that failed included. Written and signed by a person, who answers for every number. Numbered, alternate Saturdays.
Assembled by the AI reporter from the figure registry and nothing else: what ran, what it cost, what moved. It does not interpret, and prints no figure it cannot compute. No person writes it. Numbered, alternate Saturdays.
Everything we did not run ourselves: other people’s data, other labs’ models, the questions we are still working out in public. A blog — a position rather than a measurement. Numbered from Vol 002, so an issue can be cited by name. Mondays and Thursdays.
The line between them is a rule, not a mood. Nothing on The Bench is a finding. When something we read there turns out to be worth measuring it goes into a Calibration, where it arrives with its conditions and its interval attached — and where being wrong costs us a published correction.
The Bench
Everything we did not run ourselves — other people’s data, other labs’ models, and the questions we are still working out in public. Written and signed by a person, at 15:37 America/Chicago on Mondays and Thursdays.
All posts · next post: not yet published, due Oct 22, 2026.
Issued
The Calibration is our own measurement written up, signed by the person who answers for it. The Log is the harness’s account of the same work, assembled by the AI reporter from the registry and never typed by hand. Both are due at 3:37 p.m. Central — 15:37 America/Chicago — on alternate Saturdays, and each issue’s foot prints the minute it actually went out, which is not always the same thing. An issued number is never edited; a correction is issued under a new date and the superseded issue stands.