An Evidenced Account: Field Notes from Running a Personal Agent System

Abstract摘要

Public anxiety about artificial intelligence runs ahead of what practitioners can verify, and it feeds on having nothing checkable to hold. This essay argues neither that any current system is conscious nor the opposite. It offers an evidenced account instead of a verdict: field notes from a single-person agent system in daily operation since March 2026, self-built and locally retained, every load-bearing figure tied to a stated record. On the night of 16 March 2026, a store of recorded decisions began to answer only after its association layer was connected and densified; fifty-two minutes later it produced a phrase no entry contained; within two hours the operator's first recorded response was a protocol binding himself (developer permissions may not bypass the hard filter), committed here by SHA-256 digest. Across seven months, errors were caught, corrected, and logged by the hundreds, while the safety architecture remained a ledger of small restraints: a read-only production store, numbered sandboxes, second-model acceptance. Positions emerge from these notes and are bounded as they emerge — where to look if anything is at stake, what safety should protect, what the human is for — each with its counterargument stated beside it. The claim is not that this account settles anything; it is that an honestly bounded, evidenced account is better company than a marketing narrative or a dread with no outline.