JOB135 / CHROMATIC HABITAT — FIELD MODEL
Engineering contract; frozen experiment protocol will follow before training.

Goal: learn numeric full-field responses with every carrier-habitat/1 content channel received and represented in outputs. This is a structured neural response model, not evidence of general intelligence or subjective experience. Existing134 stays unchanged.

Input: up to64 cells,128 links,8 regions. Canonicalize cells by(x,y,z,a,colourIndex) and remap links; unique planar centres guarantee a deterministic order. Regions sorted(x,y,r,p). Symmetric relations0,1 are represented bidirectionally identically. Directed2–5 have distinct source/target channels. Exclude id,parent and administrative n strings from learning; retain their binding in tensorization. NULL colour6257 is distinct from PAD6258. All6,257 checked identities from134 retain allocation.

Compact arrays per scene:
colour int16[64], geom int16[64,4] for x,y,z,a, node_count int16 scalar;
edges int16[128,3] = canonical source index,target index,relation, edge_count scalar;
regions int16[8,4] = x,y,r,p, region_count scalar.
Unused row values ignored. A fixed8x8 grid supplies initial coordinates to vacant cell output slots; this is a decoder reference, not an inferred concept placement. Output locations are learned residuals from source cell coordinates or these vacant reference positions. Region vacant references are center500,500,r150,p0.

Tensor batch (model accepts dict): colour[B,64] long; node_num[B,64,5] = x/1000,y/1000,z/2,(a-1)/3,present; region_num[B,8,5]=x/1000,y/1000,r/450,p/3,present; rel[B,64,64,10] float. Channels0,1 symmetric; r2 has source-to-target channel2 and reverse3; r3 channels4/5;r4 channels6/7;r5 channels8/9.

Network: from random initialization, no language weights. Hidden96,4attention heads,3blocks,feedforward192,tanh activation,LayerNorm epsilon1e-5, no dropout. Colour embedding6259x32 projected with numeric node features to96. Region features projected to96. Trainable slot embeddings for64cell and8region queries. All72 query slots are processed; only present input cells/regions are allowed as attention keys; an empty scene uses one always-present learned global token (token72, total73). Full pair attention receives a learned linear bias from10relation channels (zero outsidecellpairs) and4geometric values(dx,dy,distance,inside-region-sign) computed from input/reference geometry. Document geometric feature computation exactly in model module. No natural-language metadata enters.

Heads:
- node_count65, region_count9, edge_count129 logits from global token.
- cell colour probabilities: learned mixture of pointer over64input cells plus NULL and global6258output logits (inputPAD never selectable); duplicate source colours accumulate. Empty-source pointer includes NULL. Expose pointer,gate,global logits separately.
- cell xy residual linear2 passed tanh*0.15 added to reference xy; z5,a4 logits per slot.
- region xy/r residual linear3 passed tanh*(.15,.15,.20) added to reference(x/1000,y/1000,r/450); p4 logits.
- edge6logits for each ordered cell pair from6bilinear low-rank projections (rank16) plus a learned linear6x10 residual from observedrelation channels. Decode0/1symmetrically by averaging pair orientations;2–5ordered.

Output structured model fields: use predicted counts with per-cell/region predictions and top-ranked legal typed edges. Legality adapter rounds/clips positions,radius,counts; enforces26minimumcentre distance using deterministic nearest available lattice search if needed; no selflinks/duplicates; symmetric0/1canonicallyordered. Record every repair separately and raw predictions; never describe validity supplied by adapter as learned validity. id,parent,n labels supplied administratively. Sampling is optional; fixed deterministic greedy decode for numeric evaluation.

Training uses a frozen procedural numeric teacher. Its rules are authored exercises, not the official meanings of relation labels. Synthetic base scenes split before matched variants. Holdouts not read for model choice. A one-run pilot/effective budget is fixed before final test; feature-ablated baselines reuse the same architecture/training budget. The old134sequence interface is a channel-blind baseline, not a directly comparable full-output teacher solver.
