The question now
What has formed through training that lets a model interpret relations and respond to new situations? How does a momentary latent activation express that structure, and what could observations of those activations establish?
The October research review separates three obligations:
- Formation: explain what training selects across experiences.
- Realization: show how the selected structure constrains processing and responses within a declared family of situations.
- Identification: determine what observations can recover, and which differences between possible models remain invisible.
These are connected questions, but a result about one does not settle the others. An object that predicts responses need not explain how training formed it. Recovering a generative factor need not recover the network's mechanism.
A more precise research agenda
I am comparing candidate descriptions through sparse dictionaries, predictive states, algebraic realizations, representation symmetries, causal abstraction, and learning dynamics. The aim is to specify what each description preserves and what evidence could distinguish it from its rivals.
The observation contract matters: unpaired activations, input-activation pairs, trajectories, task responses, interventions, and training records are different sources of evidence. The next step is to choose a target and an observation regime before claiming that the target can be recovered.
Related mathematical work asks how exact computational certificates can support identifiability claims. Local rank calculations, generic statements, and uniqueness for a particular observed instance remain separate obligations. The current work does not establish a general global-uniqueness theorem or a new computational-complexity result.
Status and limits
This is a working research direction and literature synthesis, not a new neural result or a canonical promotion. It does not establish that passive latent observations determine arbitrary training-formed mechanisms. It also does not prove that all local observations are insufficient: the meaning of "local" and the available assumptions must be specified first.
The current ULR status remains Canon 24 · Ledger 105 · Progress 31, with M8 = NO for the declared candidate registry. The founding motivation and reopening conditions remain the reference points for any stronger claim.