The field measures the wrong thing

I wrote a position paper on what the field actually measures when it claims to measure internal states in AI agents. The first half rereads 6 published works: Springdrift (arXiv:2604.04660)/ZenBrain (arXiv:2604.23878)/HELT (arXiv:2605.13858)/Gubernaut (arXiv:2607.24339)/Sentipolis (ACL Findings 2026)/ Venkit et al. (arXiv:2607.28818). Not their abstracts. But their limitation sections/appendices/constants. For each one: where it stops, and where its boundary is. The pattern that comes out is precise. The field has affect architectures/benchmarks/simulations. But it has no persistent, heterogeneous, measured deployment on real calendar time. And nowhere a state that keeps moving at night, even when nobody is talking to the system. Then the other half is the technical disagreement: what measuring a living system actually costs. Instruments that are dated and versioned, not an instrument that changes mid-series and manufactures trends. Replay, recomputing a past measurement from append-only ledgers, so the series does not rest on my word alone. And preregistered protocols/blind judges/negative results published at the same rank, because sorting after the fact destroys the value of everything else. This is not an announcement of results. It is an inventory of what the field is missing, piece by piece, and of what I intend to prove. The fingerprints of the documents have been public since August 15 2026, sealed in a gist and anchored with OpenTimestamps: https://gist.github.com/Lamperouge0/5bc0fd609e6e82c504594b8139acf228 Written in French first. The 2 languages are both in this deposit. First published on August 16 2026 at https://github.com/Lamperouge0/manifesto. This deposit is the August 17 2026 revision, which added one sentence stating the criterion, in each language.