Nishi Family › Atlas › Recombination
Novel design by recombining the ecosystem's own capabilities. Live, verified, honestly graded — the limitations are the story, not hidden. 2026-07-18.
nx_atlas_recombine walks the whole 16,599-node atlas import graph, ranks integrative nodes by fan-out (Ce = capabilities composed), filters test/gate scaffolding, and emits each node's best orthogonal partner (low shared deps, no direct edge) scored min(ca,cb)−shared×4. Built → promoted → registered 100% API-pure; verified by running the deployed sovereign ELF on the live graph. Honestly graded FUNCTIONAL — it works and is live, but see the scope gap below before calling it product-design.
| Recombination candidate | Score | Composes |
|---|---|---|
| nx_llm_batch_serve × nx_browser | 24 | 28 × 29, shared 1 |
| nx_h264_pframe_recon × nx_browser | 23 | 27 × 29, shared 1 |
| nx_f32_llm_serve × nx_browser | 22 | 26 × 29, shared 1 |
| nx_h264_brecon × nx_batched_tput | 21 | 25 × 25, shared 1 |
| nx_coder_gen_f32 × nx_wiki_routes | 17 | 21 × 22, shared 1 |
These are software recombinations (video codec × browser) — because the atlas today is a software import graph of 16k .nx organs. The operator vision (novel physical products — a butter churn, a hairbrush) needs the atlas to also carry product / physical / material / market-need nodes (the “cards for every capability AND product” vision). Until then, recombination composes software capabilities (real and useful internally) but the product ideator is starved. That is an atlas-scope gap, not a recombiner flaw — and it is filed, owned, and next.
Also honest: partner-concentration on hub organs (diversity penalty pending); Ce inflated by duplicate edges (filed); output is structural (novelty = graph-orthogonality), semantic/market ranking (match design-to-market needs + a product-sense judge) is the next rung.
UPDATE (being eaten): the product-atlas first increment is LIVE — nx_product_atlas_build ingests needs + capabilities + component keywords into a 41-node productgraph-, and recombining over that yields PRODUCT recombinations (geometry × sourcing, battery × geometry). So the software-only limit is now an active build, not just a filed gap; the remaining rung is the semantic/market judge + a richer product-node set.
nx_product_decompose matches each of a product's NEEDS to a fulfilling sovereign CAPABILITY, flagging any need with none as an unproven-opportunity GAP. This is the product-side of the vision (need x capability), over REAL design-to-market data — verified live on the sonic-hairbrush pilot:
| Need | Fulfilled by (live) |
|---|---|
| geometry | nx_cad |
| sourcing | nx_supply_research_fetch |
| battery | nx_battery_research_fetch |
| demand | nx_market_research_fetch |
| cost | nx_fin_research_fetch |
Coverage 5/5, 0 gaps — the pilot is capability-complete: the ecosystem already has every capability the product needs. So its blocker is not missing tech, it is EXECUTION (wire the 5 organs into a pipeline + the operator VHECT demand model + run the pilot). Results now PERSIST to the decompose- plane (MCP: nx_product_decompose · nx_decompose_board) — the atlas-out compounding loop, banking outcomes back for the F208 opportunity feeder.
What it is: recombinant innovation over the ecosystem's OWN capability graph (self-model recombination). Novel angle — no known tool recombines its own capability atlas into product ideas.
Honest maturity: FUNCTIONAL / L2. Structural recombination is live (software + product atlases, verified); there is no semantic/market judge yet, so ranking is graph-orthogonality, not product-sense. No parity/AHEAD claim — no measured head-to-head.
The field (what we benchmark against): Autodesk Fusion generative · nTopology · Siemens NX (topology/geometry optimization — mature, but geometry-level, not capability-recombination); LLM product-ideation (semantic but ungrounded in a real capability graph); recombinant-innovation theory (Arthur, Schumpeter) — which the atlas operationalizes over a real graph.
Research opps (beyond-SOTA): (1) a semantic product-sense judge (model- or market-grounded); (2) connect the product-atlas to the full 16k atlas for deeper lineage; (3) close the loop — recombine → build → market → bank the outcome so the atlas compounds; (4) cross-product recombination.
What to look for in our fetches: generative-design tool capabilities + pricing (Autodesk / nTopology / Siemens); recombinant-innovation papers; LLM product-ideation methods; topology-optimization + design-space-exploration SOTA.
| Surface | State |
|---|---|
| MCP | nx_atlas_recombine |
| MCP (output plane) | nx_recombine_board |
| Synced to atlas | walks ecograph_full (16,599 nodes, live) |
| RACI | lane recombine — R: researcher · A: pm · C: architect |
| Data plane | knowledge/store/recombine- |
| Workflow (integrated) | nx_atlaspulse — build → honesty → recombine → decompose → product-atlas → product-recombine, both atlases in ONE call (verified steps=7 ok); registered argless, ready for a supervised cadence (conductor) |
Ran live on the sovereign stack — the candidates above are a real verified run, not a mock. Next rungs (recombine lane): product-atlas ingest, semantic dmktneeds match + product judge, partner-diversity, auto-persist. Served bits-up, TLS 1.3, no JavaScript. 2026-07-18.