All in progress, none finished. For each one the real status, measured on 4 October 2026.
verimemin progress · core rewritten
Memory for AI agents that ties every memory to its source and has a local model judge it: a fact the source does not support is quarantined and never served as true. The first version, 4,705 commits and 248 tools, measured by itself that it did not keep its promise: an invented detail without numbers passed 8 times out of 10 in Italian and 9 out of 10 in English; one memory in five was never served; the full source could be recovered for 7% of memories. I stopped it on 30 September 2026 and had the core rewritten from scratch: 47 commits, 190 tests, Apache-2.0. On the same data the base judge without retraining beats the one we had retrained (AUROC 0.962 vs 0.789). The lab that decides whether the road is right has written criteria and two dates: 23 October and 1 December 2026.
4,705 commits (first version, stopped)47 commits · 190 tests (new core)8/10 invented details admitted by the first version, measured
⚠ No customers, no external adoption. All numbers are measured by us on our machine, not by third parties.
critic-orchestratorin progress · MIT
MCP server that launches several Claude instances in parallel to look for errors and contradictions in a claim or a code change before release, instead of confirming it. 45 commits; 93 green tests at the check of 6 July 2026, with a working live review.
45 commits93 green tests (2026-07-06)
⚠ Last code update: 28 July 2026.
pqc-audit-italiain progress · AGPL-3.0
Crypto-discovery and crypto-agility toolkit for Italian public administration, SMEs and finance: where the cryptography in use will become obsolete and how to migrate to post-quantum standards, with reference to NIS2, DORA and AgID. 129 commits, 768 test functions.
129 commits768 test functions
⚠ Last update: 15 June 2026. Never used on a customer.
groundprobein progress · MIT
Profiler that tags, at runtime, every claim of an agent as grounded, inferred, speculative or meta, to make it visible when an AI is guessing. TypeScript.
3 commits (May 2026)
⚠ A May 2026 prototype, untouched since.
OMNEXsuspended · private
Web security audit platform with a multi-stage pipeline and blind validation by several models, designed to cut false positives. About 466,000 lines of Python and 1,803 test functions. Suspended since July 2026; demo on request.
~466,000 lines of Python1,803 test functions
⚠ Private source. Numbers counted by me on the code, not verified by third parties.
engram-orchestratorprivate · August 2026
The workshop: the tools I direct the instances with (channel between instances, screen reading and writing, shared memory).
syn-desktop · AiTraderCawBot · borgo · onlypawsprivate · idle since June 2026
A desktop cockpit for agents, a lab-only trading suite (no returns claimed), a social simulation of AI characters, a pet app. Explorations, archived as such.