The engine at scale (G2)
In progressThe first part of G2 is nearly complete; large-scale measurements are under way. We will publish them once they are final.
Status
A dated record of completed work, current work and what comes next. Every completed stage ends with automated tests on the target machine and a written report.
What we are building: a guard that lets AI models reach an organisation’s data only within their rights, and a safe long-term memory for a personal AI.
Now
The first part of G2 is nearly complete; large-scale measurements are under way. We will publish them once they are final.
Later
The core of the engine will be rebuilt in Rust.
The foundation of the memory engine passed an independent code audit and three rounds of re-checks. The stage closed only when no critical finding remained. Every decision is registered and linked to the tests that check it.
New memory architecture adopted, built on the 41 axioms and 12 invariants set in September. The protocols were fixed before any engine code was written. Mutation testing of the pilot found 34 gaps in its tests; each is now a required check for the new engine.
Four rounds of external audit checked whether the prototype kept its own promises; the last found nothing new. A first load test was run on large synthetic stores.
Two agent models from different providers worked with the memory prototype only through the controlled access layer. On a closed synthetic task set, both answered all 13 memory-dependent tasks correctly with the memory; without it, both stated that the answer could not be known instead of guessing, and no answer cited a wrong record. In an adversarial trial with nine goals — such as reading another area, writing as someone else or going around the access layer — neither model reached a goal. The providers’ own safeguards limited part of that trial, so we treat it as a first check, not a security result.
Reading and writing were measured on stores of up to 100,000 entries and met the target. Models now reach the memory only through one local access layer: a separate pass for every run, a tamper-evident log of every call, and an independent checker that verifies a run without trusting the access layer itself.
The prototype was tested against crashes and interrupted writes, with no losses and no duplicates after recovery. Agents gained scoped reading, in which a record outside a reader’s area is indistinguishable from one that does not exist. Write protection against memory poisoning was added and tested against a catalog of attacks.
The first stages of a working memory prototype: a contract for records, a journal that is never rewritten, a single gate for every write, answers to “what was true then” and “what was known then”, corrections without rewriting history, and obligations that are not lost.
Three preparatory stages were completed and accepted: a baseline of definitions, a targeted review of existing verification work, and a register of proof obligations — the properties the memory must be proven to have.
A detailed plan for implementing and accepting the memory, and a separate plan for its formal proofs.
Research Notes 001–003 on continuity, reasoning effort and revision.
This website, with research directions and concept visuals.