The Record That Can Catch Its Own Author: the first signed AI gate row and the first live key-establishment path in a governed AI hub

SQUIRL OS RESEARCH DISCLOSURE — READ FIRST. This record is part of the Squirrel OS research corpus. All measurements are reported with their denominators and their limits. Where a result is a single-custodian artefact, an unverifiable corpus, or a system still under test, the record says so in the same sentence as the result. This paper is not a claim that an AI system is safe, certified or audited. It is an account of one narrow, checkable property that a system gained on 21 November 2026, and of what that property does not prove. On that date a governance row produced by a running assistant was signed with ML-DSA-87 (FIPS 204) by a key that the writing system does not hold and cannot obtain, and the hub that verifies such rows gained a confidential channel that can seal a payload but has no endpoint that can open one. Key establishment is ML-KEM-1024 (FIPS 203), stretched through HKDF-SHA384, sealing with AES-256-GCM (FIPS 197). Both were verified in the runtime they actually run in, not in a development copy: twelve self-test checks passed in the deployed function, and a payload sealed by the hub was opened offline with an independent implementation and recovered identically. The claim that follows is about separation, not cryptography. Three powers that are normally fused in one process now live in three places. The writer produces the record and holds no signing key. The signer holds the key, runs on the operator's own machine, and cannot write rows. The verifier checks signatures against a registered fingerprint and cannot amend anything. No single component can produce a record that looks signed without being signed. Two limits are stated here rather than buried. First, every private key involved is held by one custodian, so a 2-of-3 crossing certifies that one person drove three seat identities over the same bytes, not that three independent parties consented; the posture is recorded as single-custodian research and is repeated wherever a quorum result is reported. Second, the programme is under test, not complete: payload escrow is not yet wired, fault injection has not been run against this configuration, and the crossing digest remains SHA-256 until the participating boxes move. One signed row is one row. What the work buys is narrower than safety and more useful than assurance: a stranger who trusts neither the system nor its operator can take a row, a public key, a fingerprint and a message shape, and determine for themselves whether that row was signed by the holder of that key, without asking anyone for a report. The paper includes the verification recipe, and it records the row that was refused rather than signed when its digest did not match the accepted message shape, because a system that fails visibly is worth more than one that never fails. The evidence repository for this record is private. The verification recipe, the exact row fields, the public key fingerprint and the algorithms are all stated in the paper so that the check can be reproduced without repository access. No customer data, tenant identity, credential, wallet address, transaction amount or health information appears in this record; the payloads used in the confidentiality tests were structural and synthetic by construction.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24413089
Primary Topic
Cryptography and Data Security
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

The Record That Can Catch Its Own Author: the first signed AI gate row and the first live key-establishment path in a governed AI hub

Leon Calvin II long
Zenodo (CERN European Organization for Nuclear Research)
Cryptography and Data Security
preprint

The Record That Can Catch Its Own Author: the first signed AI gate row and the first live key-establishment path in a governed AI hub

Leon Calvin II long
preprint en

Abstract

SQUIRL OS RESEARCH DISCLOSURE — READ FIRST. This record is part of the Squirrel OS research corpus. All measurements are reported with their denominators and their limits. Where a result is a single-custodian artefact, an unverifiable corpus, or a system still under test, the record says so in the same sentence as the result. This paper is not a claim that an AI system is safe, certified or audited. It is an account of one narrow, checkable property that a system gained on 21 November 2026, and of what that property does not prove. On that date a governance row produced by a running assistant was signed with ML-DSA-87 (FIPS 204) by a key that the writing system does not hold and cannot obtain, and the hub that verifies such rows gained a confidential channel that can seal a payload but has no endpoint that can open one. Key establishment is ML-KEM-1024 (FIPS 203), stretched through HKDF-SHA384, sealing with AES-256-GCM (FIPS 197). Both were verified in the runtime they actually run in, not in a development copy: twelve self-test checks passed in the deployed function, and a payload sealed by the hub was opened offline with an independent implementation and recovered identically. The claim that follows is about separation, not cryptography. Three powers that are normally fused in one process now live in three places. The writer produces the record and holds no signing key. The signer holds the key, runs on the operator's own machine, and cannot write rows. The verifier checks signatures against a registered fingerprint and cannot amend anything. No single component can produce a record that looks signed without being signed. Two limits are stated here rather than buried. First, every private key involved is held by one custodian, so a 2-of-3 crossing certifies that one person drove three seat identities over the same bytes, not that three independent parties consented; the posture is recorded as single-custodian research and is repeated wherever a quorum result is reported. Second, the programme is under test, not complete: payload escrow is not yet wired, fault injection has not been run against this configuration, and the crossing digest remains SHA-256 until the participating boxes move. One signed row is one row. What the work buys is narrower than safety and more useful than assurance: a stranger who trusts neither the system nor its operator can take a row, a public key, a fingerprint and a message shape, and determine for themselves whether that row was signed by the holder of that key, without asking anyone for a report. The paper includes the verification recipe, and it records the row that was refused rather than signed when its digest did not match the accepted message shape, because a system that fails visibly is worth more than one that never fails. The evidence repository for this record is private. The verification recipe, the exact row fields, the public key fingerprint and the algorithms are all stated in the paper so that the check can be reproduced without repository access. No customer data, tenant identity, credential, wallet address, transaction amount or health information appears in this record; the payloads used in the confidentiality tests were structural and synthetic by construction.

Zenodo (CERN European Organization for Nuclear Research)
Cryptography and Data Security
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon