← dashboardthe whiteboard (one factor, the maths)defects & repairs
Every part, and every connection between the parts. Drawn from the code, not from the design documents.
This is the answer to “where is the full model?”. The UI is not failing to draw the layers — the layers do not exist.
What is built is 21 separate little models sitting side by side. They are siblings, not layers. There is no connection between any two of them: no message, no shared belief, no influence. Each one senses one thing and forms its own private opinion about it. The only place they meet is the choice of what to do next, where their scores are simply added up.
Above them sits one strategist, joined by a channel that carries a single integer — about 2.3 bits — once every 12 ticks. A third layer is wired in but dead. So: two working levels and a straw, where the design calls for eight, nested, recursively.
| from → to | what crosses | size | how often | status |
|---|---|---|---|---|
| factor → factor (any of the 21 to any other) | nothing at all | 0 bits | never | ABSENT |
| energy → gut (the single exception) | one hard-wired “holding food?” column | 1 column | per tick | HARD-WIRED |
| all 21 → action choice | their scores, summed | 1 number each | per tick | LIVE |
| L1 → L2 | one integer 0–4, read from one factor only | ≈2.3 bits | 1 in 12 ticks | A STRAW |
| L2 → L1 | chosen option → preference override + precision ×1.3 + learning rate ×0.88 | 7 vectors | every tick | LIVE |
| slow layer ↔ L1 | would be belief vectors both ways | — | never | DEAD |
Read the first row again. It is the whole point. Twenty-one models that never exchange anything are not a hierarchy and not a network — they are twenty-one separate animals wearing one coat, voting on what the coat should do next.
Each unit is an 8-layer stack. A unit then sits as layer 1 of the unit above it — flagellum inside cell, cell inside nerve, nerve inside nervous system — each joined through its own Markov blanket. Update rates between adjacent layers stand in a fixed ratio, so layers learn and unlearn at different speeds. Precision is a separate organ that steers attention; the ratio only sets timing.
Status: NOT BUILT — no part of this exists in the engine. Not partially: the recursion has no instance anywhere, and there is no per-layer clock to put a ratio on.
One flat bank of 21 unconnected factors. One supervisor above it, reached by a 2.3-bit channel that only one factor can speak into. One dead layer. No unit sits inside another unit anywhere. The colony is peers under a supervisor sharing a world — there is no colony-level belief and no colony-level free energy.
So the gap is not “some layers are missing”. It is that the nesting mechanism itself has never been built, and there is nothing for a Fibonacci cycle ratio to be a ratio of.
| # | what happens | note |
|---|---|---|
| 1 | senses arrive and are binned into one outcome per factor | 21 numbers |
| 2 | once every 12 ticks, L2 re-picks an option from the single integer it can see | otherwise the old option is held |
| 3 | the option is pushed down: preferences overwritten, precision ×1.3, learning rate ×0.88 | every tick, even when L2 did not think |
| 4 | each factor updates its belief from its own observation | 12 of them cannot move |
| 5 | each factor learns — counts go up, and never down | nothing ever decays |
| 6 | each factor scores every action; the scores are summed; one action is sampled | the display shows a different, greedy calculation |
| 7 | the action is taken and only that action's transition block learns | this is the lock: taking it makes it look better |
Step 4 runs before step 6, so a factor decides using the model it has just this instant revised. Step 7 is why one action ends up taken 95% of the time.
Drawn from the engine at 0c42216; factor counts and state sizes read from the genome; connection inventory read from the message-passing code, not from design documents. If a connection is missing from the table above it is because it is missing from the code.