Job 135 · Technical record · 9 October 2026

A field that can answer.

Every content channel of the existing field grammar now enters the small model. Learning is selective, and relation editing remains a weakness.

What changed

The earlier Job 134 model received colour sequences. This model receives up to 64 cells,128 typed links and 8 circular membranes. It receives each cell’s colour identity, x/y position, layer and emphasis; each relation’s endpoints, type and direction; and every membrane’s centre, radius and opening mark. Deliberate null colour is distinct from unused padding.

There are 6,257 pinned colour identities. They enter as categorical identity embeddings, not as RGB vision or perceptual colour distances. The model does not inspect PNG pixels or decode CVP2 itself. The transport tools recover the raw field; the model processes its numerical representation. English definitions are available to the viewer and interpreters, and are absent from its training tensors.

The design and its prior

The released model has 1,121,293 parameters: three width 96 attention blocks, four heads per block,72 field slots and one global token. Relation-aware attention retains symmetric types 0–1 and separate forward/reverse channels for directed types 2–5. Colour prediction mixes a pointer to supplied identities with a full-palette head. Other heads predict coordinates, layers, emphasis, membranes, typed edges and counts.

Canonical input sorting removes array-order dependence while keeping the source-address binding map. Reverse duplicates of symmetric links are collapsed and recorded. Administrative names, field IDs and parent hashes do not enter learned tensors. Targets retain the aligned source slot order; births append. All legal output counts remain reachable.

Positions use reference coordinates plus learned residuals. Counts use the observed input count plus learned edit logits; their initialization favours zero edit. This count prior is engineered. An untrained model already copies counts reasonably well because many teacher examples retain them. It is not evidence that the network learned counting from nothing.

Exercises and evaluation

The teacher is an openly supplied numerical program. It assigns specific movements, colour-copy operations, layer changes and opening-dependent changes to fields. These rules are choices made for this study. They are not the concepts’ meanings, semantic entailments, physical laws or obligatory meanings of “invites” and “bounds.” Half of the source scenes may contain walks through the retained Carrier graph, so existing authored structure is also an input prior.

12,000 base scenes train each model;600 development scenes select its checkpoint;600 reserved scenes evaluate it. Base IDs were split before generation. All training runs used seed 135001,2,400 updates and batches of 16. The final full model selected update 2400; its matched colour-only control selected update 2000 under the frozen development criterion. One seed was trained per condition. The control sorts the colour multiset and removes spatial bindings, links and membranes: it is a severe information ablation, not an equally informed prose or JSON comparison.

Raw attribute heads are scored on all target-active aligned slots, including slots omitted by predicted counts. Separate emitted-slot scores are in the JSON report. Opening accuracy concerns the four categorical p marks. All-slot geometry assigns 1,000 units of error to absent COPY slots. On source-supported slots alone, COPY x/y MAE is 7.35; the full model’s all-target raw-head MAE is 6.97. Thus the large penalized geometry difference includes birth support, not merely better movement.

The result includes the loss

The model achieves 99.9% opening accuracy and 92.7% exact cell counts. Its changed-colour accuracy is only 16.3%. Typed-link F1 is 0.547, below COPY at 0.906. Copying is strong because most teacher links persist. The model’s willingness to alter links is not evidence that those alterations are correct.

The equally weighted paired-scene edge-F1 difference, model minus COPY, is-0.231, with a descriptive 95% bootstrap interval[-0.253,-0.208]. The analogous missing-penalized x/y difference is-13.15 units,[-15.95,-10.60], excluding 18 empty-target scenes. These scene-weighted quantities differ from pooled micro scores. Two thousand paired resamples describe this procedural experiment; they do not establish population or semantic generalization.

A failure that changed the design

The first full model predicted absolute counts poorly. In its ten initial art responses, eight had no links and two had only one. Development loss showed the count difficulty. Before the reserved benchmark was evaluated, we fixed the count-relative revision, retrained both conditions on the same budget, and preserved the first models, art and logs.

A pre-test revision record and both freezes are included. Automated v 1 intervention processes had started while a stop message crossed the completion notice. Their metrics were not inspected by the designers before revision. The 600-scene main benchmark remained unread for model evaluation until both revised models were fixed. This is a local author-maintained chronology, not external preregistration or audited isolation. The opening studies therefore also became development evidence; they are not a pristine independent art holdout.

Interventions: access is not competence

The pre-frozen challenge contains 288 pairs:240 substantive changes across 20 channels, plus 48 administrative/order/symmetry invariances. Both the trained and untrained full models change some raw head on every substantive pair. The colour-multiset control changes on 12, all involving NULL. All models preserve the 48 invariances. Canonicalization and excluded administrative fields explain those invariances; they are not learned philosophical invariance.

The trained model’s mean matched x/y response error is 0.904 units versus 1.908 initially, but it is worse on 7 of the 20 individual channels. Sensitivity alone is therefore insufficient. Full per-channel scores, teacher outputs and raw-head differences are included.

The visible adapter

The network predicts a field; a deterministic adapter rounds coordinates, clips to grammar ranges, enforces 26-unit centre separation and caps invalid edge capacity. Every action is logged. On the reserved set,15.7% of full-model scenes required an action beyond routine rounding. Validity after repair is an engineering guarantee, not learned accuracy.

Art exchange

The gallery retains authored opening studies, a colour room, two earlier full interpreter replies, and two-step learned trajectories. Two new pretrained interpreter agents receive a final field, its image, grammar and selected definitions in fresh task contexts. They freeze a first reading before composing their own raw-field replies. The small model then receives those complete replies and returns another field. No reading is a translation of hidden English or a decoded mental state. Agent task instructions limit access; they are not an operating-system isolation proof.

The artworks are deliberately available for encounter. Their value is not established by the teacher’s numerical score. A viewer can question a composition even when it decodes perfectly.

Encounter returns use fixed weights; they do not retrain the small model or accumulate online learning. Frozen first-reading files are retained byte for byte. Their literal paragraph newlines were escaped in derived valid JSON for display, with identical prose. Helper and packaging failures are preserved in the development log.

Transport, implementation and limits

Each published field has a raw JSON record, a USL1 packet, a CVP2 image, a phonetic Carrier-words record and an artwork PNG with an embedded transport. Recovery checks verify the packet and the exact 5×5 colour cores. They do not authenticate authorship or every decorative pixel. The pinned Carrier snapshot retains the historical LEGACY_ID_HISTORY warning concerning w 3–w 9; selected colours resolve through the checked snapshot.

JavaScript and Python agree across four trained fixtures: all numerical heads differ by at most 7.63×10⁻⁶, with identical colour choices and decoded fields. Twenty-five inert DOM checks cover the editor, parser, swatches, complete replies and exact-byte parent lineage. These are static/code checks; physical-browser rendering, phone behavior and live WordPress hosting were not tested in this task.

This completes reception and production of the current field grammar. It does not establish reliable relation editing, general intelligence, semantic understanding, qualia, autonomous life or superiority over another representation with equivalent mechanics. A next model should preserve supported relations more reliably, and be evaluated on fresh data with more seeds.

Inspect and reproduce

The full record includes all four trained runs, initial snapshots, numeric data, frozen plans and revision, generator and teacher, evaluator, inference implementations, art lineage, interpreter records, supplied transport tools and original attribution. Rerunning training is different from rerunning fixed-weight inference. No remote API or server is needed for the public instrument.