nishi code wiki / research
Research briefs
Durable external-research censuses. Every substantive claim carries a source URL and a publication date, and is labelled SHIPPING / CLINICAL / RESEARCH / ANNOUNCED / DEPRECATED. Gaps are declared UNVERIFIED rather than guessed.
Why this area exists. A research finding that lives in one mutable backend is not banked. These briefs are published here so they survive loss of the memory corpus, and so their claims can be checked against their sources by anyone, at any time.
This page is GENERATED by
This page is GENERATED by
nx_research_river from the resbrief- registry — the same registry that renders the research ocean. It is not hand-edited, so this index and the library it indexes cannot drift apart. Publishing a brief adds its row; the row is what puts it on this page and in the ocean at once.PUBLISHED
| Brief | Compiled | Forks off | Scope |
|---|---|---|---|
| Real-time rendering and GPU graphics architecture | 2026-07-31 | - | 6 axes ~40 sourced claims 8 declared gaps; neural rendering, micropolygon geometry, ReSTIR GI, 3DGS, upscaling |
| Sovereign SIMT GPGPU core with Vortex as the oracle | 2026-08-01 | - | warp/thread speedup within 2.4% of oracle geomean; 4 method findings 1 retraction 5 gaps |
| Fixing a defect class at the root: the sovereign scratch arena | 2026-08-01 | - | 455 leak instances across ~210 files fixed by building the missing language primitive; 5 gaps |
| Generational SOTA analysis: method | 2026-08-01 | - | core-periphery DSM economics, dependency-traced ATD, CNCF maturity, topology clustering; FIRST lineage brief; fed /code/sota |
| Real-time rendering on a constrained stack | 2026-08-01 | code-rtrender | voxels WebGL2 no-compute; NAADF traversal, radiance cascades, noiseless GI |
| Character body physics, animation and secondary dynamics | 2026-08-01 | - | DeepMimic-to-SONIC lineage, motion matching as shipped SOTA, XPBD/AVBD, fixed-point physics |
| NPC intelligence, behaviour and social simulation | 2026-08-01 | - | smart-object advertisements, thought ledgers, sourced gossip records, honest state of LLM NPCs |
| World simulation: weather soil plants animals and the farming-life loop | 2026-08-01 | - | Hillaire sky, Minecraft integer crop stencil, Vintage Story soil, Stardew husbandry numbers |
| The style lens: one simulation many art styles | 2026-08-01 | - | anime western-toon painterly pixel from one semantic G-buffer plus swappable style bundle |
| Autonomous AI scientists: the bar and how to beat it | 2026-08-01 | - | Co-Scientist Kosmos and peers; purpose-built scaffolding does NOT beat a well-prompted frontier model end-to-end |
| The brief standard: one document three audiences | 2026-08-01 | - | authored-once-derived-twice rule, claims-table skeleton, one-call publish sequence; a spec that is an example of itself |
| Agent memory / Graph-RAG SOTA | 2026-07-30 | - | knowledge-graph memory does NOT beat a well-tuned embedder+reranker on independent eval; graphs earn cost on 5 named query classes only; corpus source: reference-agent-memory-graphrag-sota-2026-07-30 |
| Measuring code quality qualitatively | 2026-07-29 | - | METR ~half of test-passing agent PRs would be rejected by human maintainers; mutation score as the ungameable gate; corpus source: reference-code-quality-eval-sota-2026-07-29 |
| Systems perf math and model-tier economics | 2026-07-29 | - | do NOT build true-WAND (BMP wins 58x); RTX 5080 is 0.64x a 4090 on prefill not 0.98x; corpus source: reference-sota-perf-math-and-tier-numbers-2026-07-29 |
| Model-tier transcript census | 2026-07-29 | - | 239 sessions mined per-model then corrected by operator - tool-error rate is NOT quality; corpus source: project-nishi-model-tier-transcript-census-2026-07-29 |
| Food science chemistry peptides and supplements | 2026-07-23 | - | 67 pre-existing organs; measured M+H-only detection blind spot in the adulterant screen; public companion /stability; corpus source: project-nishi-food-chem-peptide-supplement-2026-07-23 |
| Perceptual realism oracles: what can actually be gated per build | 2026-08-02 | code-gensota | CGVQM >=87.5 imperceptible, CVVDP >=9.68 JOD, -22dB path-traced fidelity; presence and aliveness have NO automated metric; pretty-vs-physical r=-0.46 |
| Biological plausibility as a build gate: the unoccupied niche | 2026-08-02 | code-realism-oracles | CI biological validation confirmed NONEXISTENT; phyllotaxis/allometry/Koppen constants; our forest measured at HALF real stem density (372->643 trees/ha) |
| Cloth and soft tissue: what is real-time, what is validatable | 2026-08-02 | code-bio-oracles | do NOT build CCD (2025 SOTA deletes it); O(1) containment beats substepping; WebGL 60fps only to 4K nodes; gate ratios+kinematics never material params |
| Social & physical NPC interaction: approach, conversation, contact | 2026-08-02 | code-npcmind,code-charphys | F-formations + proxemics = integer geometry for approach rings; ledger-driven dialogue beats LLM NPCs at our constraints (anti-hallucination gateable); paired clips on a shared anchor is what SHIPS for two-character contact; S1-S4 beach engagement rungs mapped to live doors; May-July 2026 tail DECLARED unverified |
| Digital twinning of human anatomy: exact-person capture to surgical precision | 2026-08-03 | code-deformables | COMPLETES the 07-31 brief + S5 the corpus question: photos/videos/GIFs reach the bar on a 3-TIER ladder (T1 class from GIFs; T2 consult-grade 1-3mm from HD video = the Crisalix class surgeons already use; T3 sub-mm approachable on pro photo sets) -- binding constraint is SCALE ANCHOR + per-region measured-vs-inferred confidence, never pixel count; clinical 0.2mm stays scanner territory until validated; 7-layer NXA container = the moat |
| Soft-tissue dynamics: the plant, the drive, and the difference that reorders the work | 2026-08-03 | research_deformables | the physics of breast and glute soft tissue: natural frequency vs drive frequency, damping, per-axis anisotropy, identified stiffness parameters, controllability, and which of it can be gated per build |
| Physics as a per-world regime: what transforms with gravity, what does not, and what must refuse | 2026-08-03 | research_soft_tissue_dynamics | physics as data per world with Earth as the calibrated benchmark: which quantities are gravity-invariant, which scale and by what law, which refuse to transform, the 0.4g validity floor on Froude scaling, and the pendant-vs-inverted stiffness sign flip |
| Ruler-craft: descriptive geometry to modern metrology, applied to precision realism | 2026-08-03 | code-digitaltwin | the freemason ruler answered: Monge descriptive geometry (2 orthogonal views determine a solid) exposes that our 1-view instrument CANNOT SEE depth -- nose/chin projection was inference not measurement; ladder M1 orthogonal pair, M2 Procrustes+EDMA (kills anchor bias, 120 distances vs 10), M3 TPS decomposition = the marble order as MATH, M4 silhouette bounds, M5 feed-forward 2-view reconstruction; law: precision comes from the CRAFT before the instrument |
| The measurement constitution: metrology across every domain, down to pore scale | 2026-08-03 | code-metrology | REFRAME: our worlds look fake because everything is PERFECT -- realism is measured IMPERFECTION and metrology is the science of specifying it; the operator's Class-III mandible paper answered = 4-METHOD BATTERY (cephalometric=medical interface, Procrustes=global verdict, EDMA=localiser, TPS=carve order) with agreement required; Procrustes ALREADY overturned our anchored jaw verdict (16.3mm, anchor was hiding it); the full outline mapped: VIM vocabulary, GUM uncertainty budgets, traceability, GD&T tolerances for the built world, ISO 25178 as ONE texture ruler for skin+bark+plaster+sand, forensic standard for the scoreboard, Allan deviation for frame time, unit discipline (our q8 64x error); micro-detail = a STATISTICS problem: point processes + published distributions for nevi/Montgomery/stubble/pores, ITA+dE00 retires eyeball colour |
| The council of instruments: consensus metrology for verdicts we can trust | 2026-08-03 | code-measurement | the 7-of-10 question answered with real metrology: CIPM key comparisons + degrees of equivalence, NIST-class consensus estimators, BIRGE RATIO = the literal do-we-know-for-sure test (do experts disagree MORE than their claimed uncertainties allow), robust median/MAD, z/zeta/En proficiency scores, Bland-Altman, Dawid-Skene latent truth (learns rater reliability with NO ground truth), conformal intervals; FIRST FINDING: a council of CORRELATED experts is ONE expert -- our 4 shape methods share one detector = 1 seat of 10; independence must be in the SENSING; measured gauge floor U(k=2)=+-1.2mm so any error under that was never real |
| A council of experts, done honestly: artifact diversity beats headcount | 2026-08-03 | research_world_physics_regimes | true metrology for physics and morphometrics validation: a seven-member council chosen for disjoint blind spots, the formal V&V and uncertainty standards, why an agreement count is not evidence strength until effective independence is measured, and how to route disagreement as a finding |
| Character surface floor -- measured, not asserted | 2026-08-03 | - | character asset floor vs a 50-package reference corpus; the no-UV finding |
| Rig floor -- their skeleton vs ours, measured from morph data | 2026-08-03 | charfloor2026 | skeleton comparison derived from morph deltas; the lateral pectoral pair finding |
| The Reciprocal Corpus Program -- measure, don't take; give the measurement back | 2026-08-04 | charfloor2026 | roadmap: ingest reference corpora as data for our generator, publish the aggregate back as open research |
| Fixing any car: diagnosis and repair planning at SOTA, and past it | 2026-08-05 | code-gensota | SOTA July 2026 for vehicle diagnosis+repair planning via 2012 Sonata Theta II and 2013 Sienna 2GR-FE head-gasket ladder; campaign-first triage, evidence-gated test plans, parts-quality upgrade registry; exceed rungs R1-R5 each with a falsifiable gate |
31 in this section.
MAP
| Brief | Compiled | Forks off | Scope |
|---|---|---|---|
| The research family tree | 2026-08-01 | - | lineage map over the library; missing relatives drawn as visible gaps rather than left off the page |
1 in this section.
SYNTHESIS
| Brief | Compiled | Forks off | Scope |
|---|---|---|---|
| Nishi Craft: the SOTA roadmap | 2026-08-01 | code-charphys,code-npcmind,code-worldsim,code-stylelens | the synthesis those briefs fed; phases A-E, every rung with a falsifiable exit criterion |
1 in this section.
Coverage of this index
| Measure | Rows | Meaning |
|---|---|---|
| Declared by the registry | 34 | the plane's own row count, all areas |
| Loaded | 34 | rows actually read back |
| In this area | 33 | rows whose area column matches this page |
| Rendered above | 33 | must equal “in this area”, else a row is malformed and is missing |
| Reachable beyond the count | 0 | records past the declared count; non-zero means reconcile |
The process — standing, not optional
| 0 | CHECK IN FIRST, ALWAYS. Before ANY research session: read this index and search the corpus for the topic. If a brief exists, append to it, supersede specific claims with newer sources, or add a sub-brief that forks off it — never start a parallel brief. Re-deriving an existing brief is the redo-work failure this process exists to end. |
| A | ARCHIVE EVERY CITATION — no link rot. At publish time every external URL a brief cites is ingested into the sovereign library, so the CONTENT survives the link. The live URL stays for attribution; the library copy is the durable evidence. VERIFY each ingest landed by searching for it — an ingest acknowledgement is not an index entry. |
| M | FEED THE MINERS. Briefs live here as served pages so the corpus miners index them — publishing here IS registering the work as done. Work that exists only in a session transcript is invisible to every miner and will be redone. |
| R | REGISTER, OR IT IS NOT PUBLISHED. A page nobody can navigate to is not published, it is merely written. The registry row is what places a brief on this index and in the ocean; writing the file is only half the act. Equally: a registry row that disagrees with what is actually served gets rendered as authority — reconcile against the filesystem, which is the oracle for what a reader can reach. |
What counts as a brief here
A brief earns this area when it meets all four:
| 1 | Every substantive claim carries a source URL and a publication date. |
| 2 | Each claim is labelled SHIPPING / CLINICAL / RESEARCH / ANNOUNCED / DEPRECATED — a shipped product and a preprint are not the same evidence. |
| 3 | Measured numbers carry units and their measurement context (hardware, resolution, cohort, dataset). |
| 4 | Everything the author could not verify is listed in a declared UNVERIFIED section rather than omitted or guessed. A brief with no declared gaps is a brief whose author did not look for them. |