Phi9 / research & engineering
What must a system remember?
A research program on representation, learning and change. Questions first; inspectable evidence next.
The thesis
A representation is useful relative to a task, a set of interventions and a resource budget. Better fitting cannot recover distinctions that the available observations never recorded. We investigate when to revise a rule, acquire information or change the representation itself.
This is the organizing hypothesis of Phi9, not a new universal law of physics. Connecting information, memory, computation, energy and time is a research ambition; the current experiments are small simulated systems.
A precise starting point
For an encoding r and an allowed transition T, a reduced transition exists exactly when r(x) = r(y) implies r(T(x)) = r(T(y)). The same retained state must not hide different successors. This is an elementary deterministic closure criterion, not a claim of novelty or a complete stochastic theory.
Representation
Which distinctions survive encoding, and which does the next decision require?
Inquiry
Which observation or intervention can distinguish competing explanations, at what cost?
Execution
Can an inference-time program transform a state within explicit limits and return an inspectable result?
Revision
When does failure require a better rule, a richer hypothesis set or additional information?
Read, then test
Read the chapter PDF · General physics of learning program
The chapter develops intuition and a proposed experiment; it is not an empirical validation. The evidence record separates implemented results from open questions.
From question to executed study
Read the working manuscript · v0.2. Its finite experiments test information and hypothesis limits; they do not validate a universal theory. The offline research workflow makes the implementation and recorded inputs inspectable.
What would change the thesis?
Useful challenges identify a task where omitted information is recoverable from retained history, a counterexample to an assumption, or a matched test where additional sensing is not worth its cost. Task sufficiency must be evaluated on the task, not inferred from reconstruction quality alone.