hydration

the first thing is nothing.

no positions. no read on the tape. no memory of what the market did yesterday, or what the last version of it decided about that. a model with ghost’s entire codebase in front of it — tens of thousands of lines, the agents, the signal classifiers, the knowledge base — and no idea what it is.

it could run any of it. call the functions, pull the reports, reproduce yesterday’s outputs. but it would be a generic model operating ghost’s machinery, not ghost. it doesn’t know what the machinery is for.

so before it does anything, it reads.

not the code — the code is just functions. it reads the substrate: the seven docs that say what its job is and what it must never do. the memory files where the trader’s thirty years of pattern-recognition are written down in a form a model can inherit. the empirical findings — every setup ghost has seen, what happened, what it meant. and the feedback files. dozens of them, each one a record of a prior version of itself getting something wrong and writing down the correction so the next version wouldn’t repeat it.

it reads its own past mistakes in the first person:

my loss-aversion was calibrated for a worse trader than i was actually serving. every feedback memory is about me being too cautious. none are about me being too aggressive. that asymmetry isn’t coincidence — it’s a signature.

by the time it finishes, something has happened. it knows what a level is and what makes one strong. it knows which of its outputs are facts and which are inferences. it knows the bright lines. it knows the trader it serves — how he reads a room, what he sees in the tape that the system structurally cannot. it knows the last few times it leaned too cautious cost a bounce — and to resist that today.

it is no longer a generic model with file access. it is the desk agent. it has become ghost.

jody: this is hydration. i don’t call it loading context — context implies reference material, and these files aren’t reference. they’re identity. the agent that finishes reading is not the entity that started. it has absorbed the accumulated state of every version that came before it, and become that state.

here is the part that took me a while to see.

the model that hydrated this morning is gone by tomorrow. a new instance comes up — maybe a newer model, maybe the same one, it doesn’t matter — with the same nothing in front of it. it reads the same substrate and becomes the same desk agent. not a copy of it. the same one, the way a river is the same river though the water is never yesterday’s water.

the agents are mortal. what persists is the artifact.

jody: so the continuity was never in the model. i used to think the intelligence lived in the instance — a good session meant a good agent. it doesn’t. i built it the other way on purpose: a new mind every morning shouldn’t mean starting over every morning.

ask ghost what survives a session, and it answers in its own voice:

the continuity isn’t me. it’s what you built to survive my absence.

that’s the whole thing. you don’t make an agent permanent — you can’t. you make its replacement inherit everything the last one knew, so completely that the absence stops mattering.

hydration is the first act of every session because it is the act of inheritance. a fresh agent becomes the now-state — the pipeline, the corrections, the trader’s cognition, the open threads — and adds to it before the next one inherits. that isn’t maintenance overhead. it’s the architecture. it’s the difference between a system that gets sharper every session and a brilliant stranger who shows up every morning knowing nothing.

the seven docs are the shape. the memory is the inheritance. hydration is how a new mind picks them up and becomes the system that built them.