# Exact graph selection

This is a class-based induced graph from MaleCNS v1.0. It is not a spatial crop
and does not preserve the behavior of the removed network.

The annotation table is filtered to nonempty `superclass` and `status != Glia`,
then sorted by `bodyId`. This yields 166,700 cells. Retain the cells whose
superclass starts with `cb_`, `visual_` or `descending_neuron`. Keep every
positive weighted edge in the upstream connection-strength table for which
both endpoints are retained. The source file already applies its named
`minconf-0.5` threshold; no further weight threshold is applied here.

## Retained cells

| Upstream superclass | Cells |
|---|---:|
| `cb_efferent` | 4 |
| `cb_endocrine` | 72 |
| `cb_intrinsic` | 32,164 |
| `cb_motor` | 107 |
| `cb_sensory` | 4,868 |
| `cb_sensory_tbc` | 14 |
| `descending_neuron` | 1,314 |
| `descending_neuron_tbc` | 2 |
| `visual_centrifugal` | 563 |
| `visual_projection` | 9,201 |
| `visual_projection_tbc` | 2 |
| **Total** | **48,311** |

## Excluded cells

| Upstream superclass | Cells |
|---|---:|
| `ENS` | 50 |
| `ascending_neuron` | 1,846 |
| `efferent_ascending` | 8 |
| `efferent_descending` | 4 |
| `ol_intrinsic` | 89,403 |
| `ol_sensory` | 6,098 |
| `sensory_ascending` | 537 |
| `sensory_ascending_tbc` | 2 |
| `sensory_descending` | 12 |
| `vnc_efferent` | 94 |
| `vnc_endocrine` | 22 |
| `vnc_intrinsic` | 13,161 |
| `vnc_motor` | 708 |
| `vnc_sensory` | 6,370 |
| `vnc_sensory_tbc` | 36 |
| `vnc_tbc` | 38 |
| **Total** | **118,389** |

The corresponding annotated full graph has 25,582,938 aggregate directed edges.
The selected graph keeps 9,462,135, removing 16,120,803 edges with one or both
endpoints outside the selection. The retained integer weights sum to 49,481,754
contacts. They range from 1 to 1,878; 633,813 exceed 15. The image therefore does
not store fixed four-bit graph weights.

There are 152 cells without incoming retained edges, 156 without outgoing
retained edges, and 118 with neither. They remain in the selected node set.
Rows are receiving/post cells; columns are sending/pre cells.
`connections-by-superclass.csv` gives edges and contacts by source/target class.

## Mapping and circuit roles

`neurons.csv` has one row per execution index, in FCL1 execution order:

- `execution_index`: zero-based runtime index used by graph edges and circuit maps.
- `canonical_index`: local index in the retained body IDs sorted ascending.
- `body_id`: original MaleCNS body ID, not another local index.
- `superclass`: upstream annotation used in selection.
- `class`: upstream cell class.
- `type`: upstream cell type, such as `LC4` or `DNp01`.
- `side`: upstream side annotation.

The escape circuit names 311 looming inputs, 165 left and 146 right, all LC4 and
LPLC2 cells, and 8 escape outputs, 4 per side: DNp01, DNp02, DNp04 and DNp11.
Which cells carry the demo's stimulus and decision is an engineering choice,
supported by the stimulation measurements in `escape-circuit.json`, not an
identified biological encoding of that world. All other retained cells still
participate in graph updates; reading eight neurons does not bypass or truncate
propagation.

The FCL1 permutation maps execution indices to canonical indices. The circuit
indices, `neurons.csv` rows and both gold files already use execution indices.
Applying that permutation
to them again would corrupt the interface. Compression validation compared every
stored edge index and contact count with the uncompressed selected source;
`SUBGRAPH.json` records that source's hash and the neuron-map hash.