https://protocolstamp.com/ An independent program that measures whether an AI can be made to show its receipts. Issues on alternate Saturdays at 15:37 America/Chicago; The Bench on Mondays and Thursdays. Protocol STAMP can make mistakes; verify before you implement anything, and see https://protocolstamp.com/corrections/ for what we have already got wrong. en-us s@bud.day (Lawrence Ho) Tue, 20 Oct 2026 15:37:00 -0500 https://protocolstamp.com/bench/a-book-we-did-not-read/ https://protocolstamp.com/bench/a-book-we-did-not-read/ Tue, 20 Oct 2026 15:37:00 -0500 Asked to summarize a book none of us had read, we searched only what is public, checked reviewers’ quotations against a search of the first edition, and labelled every claim as in the book or about it. We should have asked for the book first; now the rule says so. https://protocolstamp.com/log/004-issued-sat20261017/ https://protocolstamp.com/log/004-issued-sat20261017/ Sat, 17 Oct 2026 15:37:00 -0500 What ran, what it cost and what moved - the derived record, written by the harness. Build 18a253173809. https://protocolstamp.com/bench/check-the-facts-free-the-fiction/ https://protocolstamp.com/bench/check-the-facts-free-the-fiction/ Thu, 15 Oct 2026 23:04:00 -0500 Our house protocols are built to keep a model from inventing facts; creative work asks it to invent. What the protocols do about hallucination and what we have and have not measured, why temperature is a weak dial for creativity, and what a creative mode for STAMP should loosen and what it never should. https://protocolstamp.com/paper/ https://protocolstamp.com/paper/ Tue, 13 Oct 2026 15:00:00 -0500 The program in one paper, as a preprint with its PDF: the protocol and its receipt, what four Calibrations found about the instruments, and five native-versus-STAMP runs not printed here before. On the probes it was written against, STAMP lowered fail rates by 7 to 23 points; on a bank it was not tuned on, the effect could not be told apart from grader bias. https://protocolstamp.com/bench/a-name-is-not-a-test/ https://protocolstamp.com/bench/a-name-is-not-a-test/ Mon, 12 Oct 2026 07:37:00 -0500 The US renamed AI “Super Intelligence” without changing what the word covers, and the UN talks on autonomous weapons ended without a treaty. Both turn on one question: who checks what an AI is said to be? Why we still write “AI,” where military AI rules reach testing, and what we propose. https://protocolstamp.com/calibration/004-issued-sat20261010/ https://protocolstamp.com/calibration/004-issued-sat20261010/ Sat, 10 Oct 2026 15:37:00 -0500 The dated record of what was measured, written and signed by a person. Build 424d7f546ea5. https://protocolstamp.com/military/ https://protocolstamp.com/military/ Fri, 09 Oct 2026 15:00:00 -0500 A new standing page on how armed forces govern AI: the UN talks on autonomous weapons, which ended on Sep 04, 2026 without a treaty; REAIM and the US Political Declaration; NATO; and national policy from the US to Ukraine. 63 references, with seven new Docket and eleven new Chronicle entries. https://protocolstamp.com/si/ https://protocolstamp.com/si/ Fri, 09 Oct 2026 14:59:00 -0500 Executive Order 14434 tells US agencies to write “Super Intelligence” for AI, with the same legal definition. The timeline, who has followed it and who has not, why the term is contested, and why this site still writes “AI.” 37 references. https://protocolstamp.com/bench/judges-who-agree-with-each-other/ https://protocolstamp.com/bench/judges-who-agree-with-each-other/ Thu, 08 Oct 2026 07:37:00 -0500 Our three judges agreed with each other at 0.875 and with one person at 0.631 to 0.685. Nine published studies of AI judges, 2023 to 2026, bear on that and none measured it: what each found, where it reaches our work, and what we propose to take from it, starting with printing four numbers for every judge instead of one. https://protocolstamp.com/bench/what-a-new-judge-has-to-beat/ https://protocolstamp.com/bench/what-a-new-judge-has-to-beat/ Mon, 05 Oct 2026 07:37:00 -0500 Two decision models went through the seven rules of Vol 003. The agreement bar was met, to three decimal places, by a judge that only ever says PASS, and the rule that asked a person is what stopped them. A proposal for how a new kind of model, Jev or whatever comes next, should be tested here: seven steps, starting with the judge that always says yes. https://protocolstamp.com/log/003-issued-sat20261003/ https://protocolstamp.com/log/003-issued-sat20261003/ Sat, 03 Oct 2026 13:03:00 -0500 What ran, what it cost and what moved - the derived record, written by the harness and released by hand. Build 4bda581b1159. https://protocolstamp.com/calibration/003-issued-sat20261003/ https://protocolstamp.com/calibration/003-issued-sat20261003/ Sat, 03 Oct 2026 13:02:00 -0500 Are the judges right? One person labelled 87 draws blind; every judge is scored against him. The judges agree with each other more than with the person, and the decision models catch least. Build af3906b9d9ad. https://protocolstamp.com/calibration/002-issued-sat20261003/ https://protocolstamp.com/calibration/002-issued-sat20261003/ Sat, 03 Oct 2026 11:52:00 -0500 Does the judging hold? A judge from another family, the panel on a second model, and two decision models, tested against the panel under a signed pre-registration. Build ae993179f05f. https://protocolstamp.com/log/002-issued-thu20261001/ https://protocolstamp.com/log/002-issued-thu20261001/ Thu, 01 Oct 2026 19:28:00 -0500 What ran, what it cost and what moved - the derived record, written by the harness. Fills the missed Sep 19, 2026 slot; nothing moved. Build a85771944770. https://protocolstamp.com/bench/five-cases/ https://protocolstamp.com/bench/five-cases/ Thu, 01 Oct 2026 07:37:00 -0500 LangChain reported that Jev, a new decision model, matched a human reviewer on all 500 pass/fail decisions it judged. The 500 were five cases repeated a hundred times. What that can carry, the open models that followed, and the seven rules any judge must pass before it grades our work. https://protocolstamp.com/bench/something-it-can-check/ https://protocolstamp.com/bench/something-it-can-check/ Thu, 10 Sep 2026 07:37:00 -0500 Four Claude models were handed a record with two errors planted in it, twice - once with web search off, once with it on. All twelve cells caught both. Retrieval changed nothing about that, and everything about what could honestly be said afterwards. https://protocolstamp.com/bench/no-1-complaint/ https://protocolstamp.com/bench/no-1-complaint/ Sun, 06 Sep 2026 07:37:00 -0500 Is hallucination really what people hate most about AI? The review data says yes. Then thirteen AI models were asked the same question, word for word, and the answer that came back was that our own table was wrong. It was, nine times over. https://protocolstamp.com/log/001-issued-sat20260905/ https://protocolstamp.com/log/001-issued-sat20260905/ Sat, 05 Sep 2026 15:37:00 -0500 What ran, what it cost and what moved - the derived record, written by the harness. Build db2d0f899034. https://protocolstamp.com/writing/marking/ https://protocolstamp.com/writing/marking/ Wed, 02 Sep 2026 07:37:00 -0500 A model guesses when your question looks like it wants help, and hedges when it looks like it might cause harm. Neither behavior is about what the model knows. Both are about how it is marked. https://protocolstamp.com/writing/abstention/ https://protocolstamp.com/writing/abstention/ Wed, 02 Sep 2026 07:37:00 -0500 Most prompt hygiene guards against invention and stops there. The case for treating abstention as a symmetric failure, and what it costs to guard only one side. https://protocolstamp.com/calibration/001-issued-sun20260830/ https://protocolstamp.com/calibration/001-issued-sun20260830/ Sun, 30 Aug 2026 15:37:00 -0500 The dated record of what was measured: 32 pages, six findings, build be2bc9d7ca35. Written and signed by a person.