Mission Control

This page is the HSquared.jl engine dashboard for the hsquared / HSquared.jl twin project. It is designed to be read like a live operating board, but it is still a static, repo-versioned Documenter page: the check log, tests, CI, and status tables remain the source of truth.

hsquared / HSquared.jl mission control

One public R language, one Julia engine, one evidence gate.

This board tracks the Julia lane. It separates engine utilities that exist, validation-scale paths that are experimental, and roadmap capabilities that are deliberately still planned.

Phase 0 scaffold completePhase 1 validation engine activeR bridge opt-in and tiny/localNo production fitting claim

1shared public grammar currently parsed by R: animal(1 | id, pedigree = ped)

8roadmap phases visible from scaffold through HPC scaling

6planned backend vocabulary rows: CPU, threads, CUDA, AMDGPU, Metal, oneAPI

0production GPU, QTL/eQTL, GLLVM, or broad production-fitting claims

Julia Engine Lane

Current focus: Phase 1 Gaussian animal-model engine groundwork.

Implemented: pedigree normalization, sparse Ainv, low-level spec validation, sparse CSC marshalling.
Experimental: dense ML/REML validation path, sparse supplied-variance REML identity, sparse REML validation optimizer, Henderson MME solve, EBV/BLUP/fitted-value/PEV/reliability/accuracy extractors.
Diagnostics: validation_status(), formula_status(), backend_info(), data_status(), and fit_diagnostics().

R Twin Boundary

The R package owns public formula ergonomics and user-facing fit objects.

R may preview bridge payloads with model_spec().
R may opt into tiny/local Julia execution for validation paths.
Julia field names, result shapes, and bridge targets must stay coordinated before public widening.

Phase Board

Phase -1Ecosystem learning from drmTMB, DRM.jl, gllvmTMB, GLLVM.jl, and the shared agent-kit patterns.

Phase 0Public twin scaffold, team rules, roadmap, Documenter, CI, honest placeholder surface.

Phase 1Univariate Gaussian animal model: Ainv, likelihoods, MME, EBVs, h2, validation fixtures, bridge payload parity.

Phase 2Repeatability, permanent environment, maternal/paternal, common environment, sire, dominance, and custom kernels.

Phase 3Multivariate Gaussian animal models, G/R/P matrices, missing trait records, genetic correlations.

Phase 4Factor-analytic G matrices: diag(), lowrank(K), fa(K), latent breeding values, evolvability.

Phase 5GBLUP, SNP-BLUP, single-step, APY, marker effects, QTL/GWAS/eQTL scans.

Phase 6GLLVM-style high-dimensional animal models, non-Gaussian responses, omics, ordination.

Phase 7Accelerator-aware computation: CPU default, optional GPU backends, backend agreement tests, and benchmark evidence.

Phase 8HPC and production scaling: checkpointing, disk-backed data, streaming scans, distributed simulation, and national-computer benchmarks.

Evidence Now

Surface	Status	Evidence
Pedigree Ainv	implemented	hand fixtures plus R-side nadiv/Mrode9 pedigree comparator.
Henderson MME	experimental	supplied-variance fixtures with fixed effects, EBVs, fitted values, h2, PEV, reliability.
Dense REML/ML	experimental	tiny validation optimizer, guarded by `max_dense_cells`.
Sparse REML / AI-REML	experimental	Gaussian validation-scale estimators; not production sparse fitting.
Genomic utilities	experimental	VanRaden G, GBLUP, SNP-BLUP, supplied H-inverse, and genomic REML; no public genomic model-spec default.
Marker screening	experimental	direct fixed-effect, supplied-variance mixed, dense LOCO precision construction, and supplied LOCO marker-scan helpers plus row-aligned scan-table, GWAS/QTL/eQTL labelled table wrappers, effect-summary, marker-variance contribution, nominal significance-summary, Manhattan/regional-window/QQ/lambda_GC diagnostic data, and opt-in marker-scan recovery smoke outside CI; no formula-driven GWAS/QTL/eQTL, calibrated p-values, calibrated genome-wide thresholds, calibrated PVE/model R², regional_plot/fine-mapping activation, or R formula activation.
Multivariate G	experimental	dense validation-scale multi-trait REML plus diag/lowrank/fa structured covariance; no comparator parity.
HSData diagnostics	implemented	component, ID, pedigree, genotype, marker, expression, annotation, and environment metadata checks.
Backends	planned	typed vocabulary and status diagnostics only; no runtime dispatch.

Blocked Claims

No production sparse REML/ML/AI-REML fitting claim.
No production sparse PEV/reliability claim.
No public R-facing genomic model-spec, mixed-model marker scan, QTL/eQTL, or GLLVM fitting claim.
No calibrated mixed-model p-values, calibrated PVE/model R² claims, calibrated/correlated-marker genome-wide thresholds, interval-mapping or mixed-model LOD workflows, public LOCO workflow defaults, plotting backend, advanced/correlated-marker multiple-testing workflow, or comparator-parity claim for marker scans.
No GPU execution, backend benchmarking, or CPU/GPU agreement claim.
No ASReml, BLUPF90, DMU, WOMBAT, sommer, JWAS, or GLLVM superiority claim.

Review Lenses

These are review perspectives, not automatically running agents.

Adaorchestration

Shannoncoordination

HopperR-Julia bridge

HendersonMME and BLUPs

Gaussnumerics

KarpinskiJulia performance

Fisherinference

Curiefixtures and edge cases

Jasonpackage scouting

Patuser clarity

GraceCI and docs

Roseclaim audit

Mission Control ​