nishi code wiki / research / the family tree

The research family tree — generational from first principles

Knowledge lineage map · compiled 2026-08-01 · 4 generations · 7 domains

sibling map: the organ family tree (16,222 organs, 8 generations)

Why this map exists. The ecosystem already renders a family tree of its code — generational from the god primitives, every organ a card with parents, children and siblings. This is the same idea for its knowledge. A research library is not a list of documents; it is a lineage. Each brief has parents it forked from, children that fork from it, and cousins in neighbouring domains. The point of drawing it is the empty slots.

The one rule that makes this a tree and not a bookshelf: missing relatives are drawn. A gap that is not rendered is a gap nobody fills. Every dashed red node below is a brief we have established we need — usually because a published brief hit its edge and said so — and have not yet written. As those fill in, the tree thickens. That is the whole mechanism: the map is designed to look incomplete, and to look less incomplete over time.

11PUBLISHED — live, source-checked briefs

1IN PROGRESS

8STUB — in the corpus, not transcribed

13MISSING — named gaps, nobody has written them

Legend: PUBLISHEDlive brief IN PROGRESSbeing written STUBexists in corpus only MISSINGa named, unfilled slot

gen 0 — the roots: how we know anything

gen 0method and evidence — every brief below inherits its standard of proof from here

PUBLISHEDAutonomous AI scientists: the bar, and how to beat itWhat Google Co-Scientist, Kosmos, Sakana and the benchmark layer actually deliver — and the finding that purpose-built AI-scientist scaffolding does not yet beat a well-prompted frontier model end-to-end. STUBMeasuring code quality qualitativelyreference-code-quality-eval-sota-2026-07-29 · mutation score as the ungameable numeric gate. PUBLISHEDGenerational SOTA analysis: methodThe first brief in the library to carry an explicit Forks-off lineage section; it fed the /code/sota generational map. PUBLISHEDFixing a defect class at the root, not at 455 leavesThe sovereign scratch arena — the argument for repairing a defect class at its origin rather than patching every occurrence. MISSINGEvidence law and non-vacuity, written downThe corpus enforces it daily — a gate that cannot fail is not a gate — but it has never been written as a brief with external sources. The most-cited unwritten parent in this tree. MISSINGVerified-citation pipelineresearch_autosci ranks this the #1 adoptable mechanism and notes nx_atlas_citeverify resolves DOIs live against Crossref today. Nothing yet describes the end-to-end contract.

gen 1 — the domains

gen 1 · renderinghow pixels get made

PUBLISHED · the eldestReal-time rendering and GPU graphics architecture2026-07-31 · 6 axes, ~40 sourced claims, 8 declared gaps. Parent of everything in this domain. PUBLISHED · child of rtrenderReal-time rendering on a constrained stackNarrows the parent to WebGL2, no compute, integer core, voxels. NAADF traversal, radiance cascades, and why noiseless GI matters when you have no ML denoiser. PUBLISHED · child of rtrenderThe style lens: one simulation, many art stylesAnime, western toon, painterly and pixel from one semantic G-buffer plus a swappable style bundle. MISSINGA higher-ceiling perceptual criticNamed directly by our own measurement: the groundtruth critic saturates at 1000 and craft already scores 983, so it can no longer rank us against the reference. We are optimising a ruler that cannot move. MISSINGWebGPU migration economicsDesktop support is ~82-87% but mobile ~71%. Nothing states the crossover condition at which the compute-shader backend becomes the floor rather than a second path.

gen 1 · embodimentbodies and minds — making characters alive

PUBLISHEDCharacter body physics, animation and secondary dynamicsThe DeepMimic-to-SONIC lineage, motion matching as the shipped SOTA, XPBD/AVBD, and fixed-point precedent from shipped deterministic physics. PUBLISHEDNPC intelligence, behaviour and social simulationSmart-object advertisements, thought ledgers, gossip records with sources, and the honest state of LLM NPCs. STUBDigital twinning of human anatomyIn progress on the research index — anatomical, material and mechanical 1:1. MISSINGMocap acquisition and retargetingresearch_charphys concludes motion matching is the shipped SOTA and that its real cost is animation data. No brief covers how a one-operator team acquires or synthesises that data. MISSINGSpeech and TTS for NPCsresearch_npcmind covers what NPCs think and say as text. Nothing covers how they say it out loud, or the licensing and performer-consent questions that shut down at least one vendor. MISSINGBelievability evaluationresearch_npcmind states plainly that no industry-standard NPC believability benchmark exists. That absence is itself the research question.

gen 1 · world simulationthe world the characters live in

PUBLISHEDWeather, soil, plants, animals and the farming-life loopHillaire sky, the Minecraft integer crop stencil, Vintage Story N/P/K, Stardew husbandry numbers, and the FLIAW mechanic checklist. MISSINGVoxel fluids and the water cycleDeferred explicitly by research_worldsim as correctness-heavy and easy to make quadratic. Deferred is not answered. MISSINGProcedural audio and adaptive musicThe capability benchmark scores audio HAVE with a real artifact, but no brief has ever examined the field. A whole sense with no research parent. MISSINGNetcode and multiplayer authorityA lockstep-plus-rollback part exists and is gate-proven; the SOTA around it has never been surveyed.

gen 1 · substratesilicon, GPU, kernel — the layers under everything

STUBSubstrate program order and costsreference-substrate-program-order-and-costs-2026-08-01 · the measured argument that silicon is the one axis where sovereignty is free, and the resulting build order. STUBSystems perf math and model-tier economicsreference-sota-perf-math-and-tier-numbers-2026-07-29 · includes two corrections to the banked roadmap. MISSINGFPGA and open silicon toolchainsAn entire first RTL design exists in the corpus with a 10/10 behavioural gate and no HDL simulator available. The tooling landscape has never been surveyed. PUBLISHED · gap now filledSovereign SIMT GPGPU core, benchmarked against VortexThis slot was drawn as MISSING on 2026-08-01 and was filled the same day by a concurrent workstream. Recording the fill rather than the gap is the point of the map.

gen 1 · agents and memoryhow the builder itself works

STUBAgent memory and Graph-RAGreference-agent-memory-graphrag-sota-2026-07-30 · verdict: knowledge-graph memory does not beat a well-tuned embedder plus reranker; graphs earn their cost on five named query classes only. STUBAgentic SOTA and what we adoptedThe memory lever is bounded growth, not retrieval. STUBModel-tier transcript census239 sessions mined, then corrected by the operator: tool-error rate is not quality. MISSINGNegative-result ledgerresearch_autosci names this as one of three sovereign differentiators — the anti-publication-bias mechanism nobody ships. No design exists.

gen 1 · life sciencesthe non-software wing

STUBFood science, chemistry, peptides and supplements124,685 bytes in the corpus, 497 assertions across 43 gates, public companion at /stability. The largest untranscribed brief in the library. IN PROGRESSDigital twinning of human anatomyToward an implant digital twin and an accurate cosmetic-surgery before/after. MISSINGWet-lab-free validation strategyEvery validated result in research_autosci required a human wet lab. A sovereign stack with no wet lab must know which claims it can and cannot ever close.

gen 2 — what the map is telling us

Read as a family tree, three things are visible that a list of documents would hide.

Reading	What it says
The roots are the thinnest part	Only one published brief sits at gen 0, and the two most-inherited ideas in the whole ecosystem — the evidence law and the citation contract — are unwritten. Every child brief silently depends on parents that do not exist on paper.
Rendering is the only mature wing	Three published briefs with a real parent-child relationship. Every other domain is one brief deep or entirely stubs, which means their conclusions have never been challenged by a sibling.
Eight stubs is the cheapest work available	These are finished research already sitting in the corpus, some very large, that has never been transcribed to a checkable page. Transcription is not new research; it is banking research already paid for.

How a brief earns a place here

The same four criteria as the research index, plus one more that this map adds.

1	Every substantive claim carries a source URL and a publication date.
2	Each claim is labelled SHIPPING / RESEARCH / ANNOUNCED / DEPRECATED.
3	Measured numbers carry units and measurement context.
4	Everything unverified is declared, not omitted.
5	It names its parents and its gaps. A brief states which brief it forked from, and ends by naming the slots it could not fill — which is how the red nodes above get created. A brief that declares no gaps has not looked for them, and a brief with no parent is either a root or is lying about being one.

Honest limit of this page. It is hand-authored, exactly like the briefs it indexes — the node states below are correct as of 2026-08-01 and will drift the moment a brief is published without editing this file. That drift was demonstrated within hours of first publication: three sibling briefs (gensota, scratch, sovgpu) had been published by a concurrent workstream while this page was being written, and one of them filled a slot this map had drawn as MISSING. The counts above were corrected by hand the same day — which is the argument for the generator, not a defence of the snapshot. The sibling organ tree does not have this problem because nx_eco_graph_pedigree generates it from knowledge/store/ecograph_full on every run. The next rung for this map is to earn the same property: a registry of briefs with parent links and declared gaps, and an organ that emits this page from it, so the tree cannot silently disagree with the library. Until that exists, treat the counts above as a snapshot, not a live gauge.