Agentic AI Governance
Before an agent runs a tool, flux7 checks whether it may. A rules engine answers in under a millisecond, and when the rules leave the case open, past decisions kept in memory are consulted, then an arbitration, and finally a human when it matters.
Four layers, from rule to human
flux7-mesh says yes or no, flux7-memory recalls what was decided before, flux7-supervisor arbitrates the unclear cases and flux7-console hands them to a human. Every decision is written to a hash-chained log.
SIGNAL BOX · flux7 one tool call, one route
---------------------------------------------
lever post answers in
[1] \ flux7-mesh <1ms policy
[2] \ flux7-memory ~100ms precedents
[3] | flux7-supervisor ~350ms rules + judge
[4] | flux7-console minutes human
---------------------------------------------
agent ==[1]==[2]==[3]==[4]==> tool
\
==x refused, route locked
---------------------------------------------
register · every lever pulled, hash-chained
Research · Believe and act
An agent that investigates must decide what to check next and when to stop. Here both decisions stay out of the language model, which only reads the evidence. Belief and choice are computed by code and logged, so every run can be replayed. The first pre-registered studies ran on Kafka incidents, and the next testbed will be a world of AI personas.
LOG · dead reckoning on a Kafka fault illustrative --------------------------------------------- check reading belief, top cause consumer lag rising ###..... 0.38 broker cpu normal #####... 0.61 group events 3 rebalances #######. 0.89 --------------------------------------------- next check worth less than its cost -> STOP --------------------------------------------- reading: language model · belief, choice: code
Music · Human × Machine
On a MIDI piano, an AI follows the chords and the key as they are played and answers on the instrument. It also reads finished tracks, from the audio file or straight from the Renoise project.
Reads chords, key and cadences live from any MIDI piano; the AI answers through the instrument, turn by turn.
Key, chord grid, melody and timbre from any mp3, flac or wav. CLI and MCP, stem separation.
Exact notes from .xrns files without Renoise running: song map, notes in beats, playable sequences.
Side projects
Small MCP servers in daily use, horoscopes computed from the real sky, and a history of Madagascar told by the evidence.