Palimpsests / Notes
NotesNotes
Short notes on audit logs for AI systems, the EU AI Act and running language models on your own hardware. New notes also go out through the RSS feed.
Protecting an audit log from being rewritten from scratch
A hash chain catches an edited record but not a log rebuilt from the first record. What closes that gap, from which record on, and what stays open.
Can a coding agent implement a binary format from its spec alone?
A small, strict test for coding agents. Give one a frozen specification and its test vectors, no reference code, and ask it to reproduce eleven values byte for byte in a language it chooses.
Anatomy of a tamper-evident audit log: PALA-1, a PKCS#11 anchor and SCITT
The three layers of a Palimpsests audit log taken apart: the PALA-1 hash chain, the chain head on a PKCS#11 token and SCITT receipts. What each one proves alone, and what they prove together.
AI logging standards in draft: ISO/IEC 24970, prEN 18229-1 and how we map them
Where the AI logging standards stand in October 2026, why they matter for the EU AI Act, and how Palimpsests maps their requirements to code while the text is still moving.
Keeping the audit log on your own machine when your agents use cloud models
Agents that call hosted models can still keep their tool-call record locally. How Palimpsests records Claude Code, Codex, OpenTelemetry and LiteLLM tool calls, and what that record proves.
OpenSSF Gold software and the Cyber Resilience Act
What the EU Cyber Resilience Act asks of products built from open-source components, what an OpenSSF Gold badge shows about a component, and what it doesn't.
Tamper-evident audit logs: how they work and how to build one
What makes an audit log tamper-evident: hash chains, anchors and witnesses, what each one proves, how to build such a log, and what to ask before trusting the claim.
PALA-1: three design principles for verifiable audit logs
Verification without disclosure, integrity through a hash chain, and checking a log with no remote service: the ideas behind the PALA-1 format and where each one stops.
EU AI Act Article 12: what it asks of AI you run yourself
Which systems the logging rules cover, what the articles require, when they start and how the fines work, in plain language.
How to check an AI log without reading it
Why the person who checks an AI system's log often shouldn't read it, and how a hash-chained format makes that possible.
Recording the tool calls your agent already makes
Adapters for LiteLLM, MCP servers and OpenCode that put tool calls on a tamper-evident log, and what such a record proves.
Running a language model on an isolated network, with a log you can prove
What changes when an LLM runs with no network at all: installing it, timestamps, anchors and checking the log offline.