Precision Crop Genomics

Live window into the Zer0pa lab. Pipeline-vertical research-intake stage; the moat is the typed regulatorily-routed dossier, not the model.

What This Is

Precision Crop Genomics is a license-clean, regulatorily-routed, partner-shaped pipeline that produces a typed Crop Improvement Dossier; Zer0pa is platform-only and refuses trait IP.

This is the eighth Zer0pa workstream. It is a pipeline-vertical research project: six layers (L1 ZPE encoding → L2 license-tagged knowledge → L3 candidate generation across cross/MAS/edit modalities → L4 in silico screening with G×E and regulatory routing → L5 Bayesian active-learning optimisation → L6 typed handoff dossier) backboned by LangGraph + BoTorch, with PLINK 2 PGEN as the canonical genotype store, PHGv2 + TileDB-VCF for haplotypes, BrAPI v2.1 over self-hosted Breedbase as the data spine, and a four-column license stamp on every record. The architectural commitment is pangenome-aware modelling first-class for the seven Tier-1 species (maize, rice, wheat, barley, sorghum, soybean, tomato) with graceful SNP-only degradation for orphan crops. The Zer0pa moat is the license-stamping ETL service plus the typed seven-block Crop Improvement Dossier — a partner-ready package no incumbent (Inari closed; Pairwise wet-lab licensable; Living Models foundation-only; Computomics service-based) currently delivers.

Stage signal (as of 2026-05-02): Operator-authored Pre-PRD Research Synthesis Report 1 of 2 committed (~57 KB, 160 lines, full technology landscape with license decomposition, regulatory routing, intersectional science map, SA/SADC strategy, ten research-question answers); synthesis-agent fresh-eyes pass committed (twelve specific gaps surfaced, three operator-engagement open questions); orchestrator handoff committed; PRD outline drafted and surfaced for operator review (engineering-blocker deep-research lookups resolved, sub-agent dispatch table populated, 22 falsifiers previewed, ZERO-on-CPU before Runpod committed as binding Phase-0 acceptance gate).

Honest blocker: No Phase-0 code. PRD outline awaiting operator confirmation of architectural commitments (dossier-first active-inference reframe; SE(3) protein-structure tower promotion from v1.5 to Phase-1; BoTorch qNEHVI six-axis Pareto Phase-0 binding). Report 2 forthcoming and may displace v1 PRD decisions on partner integration sequence and SA go-to-market.

Pipeline Mechanics

Field	Value
Architecture	Six-layer pipeline-vertical: L1 input encoding → L2 license-tagged knowledge → L3 candidate generation → L4 screening + regulatory routing → L5 Bayesian active-learning → L6 typed dossier
Substrate	PLINK 2 PGEN canonical genotype store; PHGv2 hVCF haplotypes (TileDB-VCF backend); BrAPI v2.1 + self-hosted Breedbase data spine; four-column license stamp on every record
Inputs	Five typed: crop species + agroecological zone + target trait set + breeding-or-editing constraints + regulatory route constraints (plus optional germplasm / pangenome / phenotype context)
Output	Crop Improvement Dossier — JSON-LD with seven blocks (Identity / Intervention Strategy / Performance Forecast / Risk / Regulatory / Provenance / Partner Handoff); regenerable from inputs at any time
Mechanics	LangGraph orchestration; BoTorch qNEHVI multi-objective Pareto over (yield, drought resilience, nutrient efficiency, nutritional quality, regulatory friction, deployment ROI); cross-model disagreement aggregator (rrBLUP + BGLR + sommer ± DNNGP); Regulatory Fit Score as first-class scoring axis
Open Gate	Phase-0 ZERO-on-CPU before Runpod cutover; no Phase-0 code yet (PRD outline awaiting operator confirmation)

Key Metrics

Metric	Value	Baseline
BRIEFING_ASSETS_COMMITTED	5/5 (Report 1 + synthesis + RESISTANCE + MODUS-OPERANDI + HANDOFF-TO-ORCHESTRATOR)	repo-resident; sibling-workstream pattern
ENGINEERING_BLOCKER_LOOKUPS	10/10 resolved (EU NGT / US APHIS / PHGv2 / plant DNA FM weights / AlphaFold DB v4 / Living Models BOTANIC / GRF-GIF / Corteva WUS2-BBM / Pairwise Fulcrum / BrAPI v2.1)	2026-05-02 lookup pass
ANTI_DECISIONS_PRE_REGISTERED	18 enumerated as binding falsifiers	sourced from Report 1 §3-5 + synthesis #11
READINESS_STAGE	STAGE-2 RESEARCH-INTAKE; PRD outline awaiting operator confirmation	sibling-workstream stage-3 gate (Synbio / Materials / Energy / Health are stage-3 INTERNAL)

Source: source-briefs/01-precision-crop-genomics-report-1-full-technology-landscape.md, synthesis/01-fresh-eyes-on-precision-crop-genomics-report-1.md, PRD-OUTLINE-FOR-OPERATOR-REVIEW.md, and MODUS-OPERANDI.md § Acceptance gates.

Repo Identity

Field	Value
Identifier	Precision-Crop-Genomics
Repository	https://github.com/Zer0pa/Precision-Crop-Genomics
Portfolio	Crop Genomics
Visibility	Public · research-intake dossier surface
Default Branch	main
Authority Source	PRD-OUTLINE-FOR-OPERATOR-REVIEW.md (pending) → PRD.md (forthcoming)
License	Code license set when first code lands (Zer0pa Source-Available License v7.1); briefing materials are operator-authored

Readiness

Field	Value
Stage	STAGE-2 RESEARCH-INTAKE
Posture	prd_outline_drafted_pending_operator_confirmation
Checks	All five briefing assets committed; engineering-blocker lookups resolved; sub-agent dispatch table populated; 22 falsifiers previewed
Authority metric	phase_0_zero_on_cpu_acceptance_gate not yet attempted (no Phase-0 code)
Authority	MODUS-OPERANDI.md § Acceptance gates; RESISTANCE.md; HANDOFF-TO-ORCHESTRATOR.md

Honest Blocker

PRD outline at PRD-OUTLINE-FOR-OPERATOR-REVIEW.md awaits operator confirmation of ten architectural commitments before the full PRD can be written. No Phase-0 code is in the repo. The synthesis-flagged Report 2 (forthcoming) may displace v1 PRD decisions on partner integration sequencing (Pairwise / Living Models / Syngenta / Corteva) and the South Africa go-to-market motion (ARC-BTP / SACTA / CIMMYT-Zimbabwe). The Living Models BOTANIC fine-tune deliverable is conditional on per-model-card license verification (currently checking on huggingface.co/living-models); a class-C/D verdict triggers anti-decision falsifier f022. Pairwise Fulcrum sub-licensing terms are not publicly disclosed; the v1 technical scaffold is built but the partnership conversation is operator-deferred.

What We Prove

• Operator-authored Pre-PRD Research Synthesis Report 1 (160 lines, ~57 KB) is the substrate: full technology landscape, four hard-exclusion license traps (PLAZA / TAIR live / WorldClim 2.1 / EOX cloudless mosaics) plus three soft row-level traps (T3 Toronto Statement / CGIAR SMTA / Doench-Fusi sgRNA scoring), Tier-1/2/3 species deployment matrix, base-editing-default-prime-editing-R&D-only empirical commitment, Regulatory Fit Score formula RFS = Σ over target markets [ MarketWeight × JurisdictionScore(0-5) × TraitMultiplier(0.3-1.0) × EditTypeCoefficient(0.2-1.0) ] − ProcessRiskPenalty, eight architecture-changing intersectional disciplines, ten research-question direct answers.
• The synthesis fresh-eyes pass surfaces twelve specific gaps Report 1 does not see when read top-to-bottom as a unified specification, including: dossier-first active-inference architectural reframe; license-stamping ETL service is the moat and underspecified; seven-block dossier JSON-LD schema must be locked before any other layer's interface contract; Regulatory Fit Score master decision matrix needs versioned encoding with citation chain; BoTorch-for-breeding qNEHVI optimiser is highest-leverage build piece and a recurring cross-workstream pattern (fork from Synbio MFMO); SA carveout has architectural implications not just narrative; Benson Hill Chapter 11 / Chapter 7 lesson encoded as binding architectural constraint (no vertical integration into grain supply chains).
• The orchestrator's PRD outline closes all twelve synthesis gaps plus ten orchestrator fresh-eyes additions (closed-loop dossier mode as v1 not v1.5; license-class taint cross-class metric specified; regulatory-text-staleness falsifier; orphan-crop prior-strength differential operationalised in BoTorch acquisition reweighting; per-species pangenome-readiness JSON; six-axis Pareto front Phase-0 acceptance test specified; four SE(3) tower use cases enumerated; 18 anti-decisions encoded as falsifier IDs not just narrative).
• All ten engineering-blocker deep-research lookups are resolved with current-state findings (May 2026): EU NGT Council-adopted 21 April 2026 with mid-2028 application; US APHIS pre-SECURE regime in effect with replacement rule pending; PHGv2 Apache-2.0 last updated 12 March 2026; PlantCaduceus Apache-2.0 / HyenaDNA BSD-3 / PlantRNA-FM MIT / GPN MIT confirmed commercial-licensable; AlphaFold DB v4 CC-BY-4.0 confirmed; Living Models BOTANIC available on HF (license per-model-card check pending); UC Davis GRF-GIF requires partner-side commercial license; Corteva WUS2/BBM no public ag-pool terms (VIGE remains primary IP-arbitrage); Pairwise Fulcrum CIMMYT-NARS sub-license confirmed June 2025; BrAPI v2.1 still current.
• The 18 anti-decisions are pre-registered as binding falsifiers that fail the brain-functionality gate at PRD review: APSIM commercial use; AgroNT / Nucleotide Transformer weights; AlphaFold3 / ESM3 main weights; WorldClim 2.1; PLAZA; TAIR live curated data; Google Earth Engine free tier for commercial use; Ultralytics YOLOv5/v8 default license; OpenDroneMap as hosted SaaS; EOX Sentinel-2 cloudless mosaics; herbicide tolerance traits in v1; PIP-equivalent insect resistance via novel insecticidal protein; pharma traits; trait IP ownership by Zer0pa; SA gene-edited product deployment absent GMO Act amendment; gene drives; cultivation bans; vertical integration into grain supply chains.
• The fork-and-own discipline carries from sibling workstreams without runtime co-dependency: Synbio's UniversalLayerEnvelope + SBOL3+PROV-O audit-trail shape + Unknown Enzyme Generative Sub-Pipeline (RFdiffusion3 + MACE-OFF + ProteinMPNN + AlphaFold DB v4 + OpenFold3 + ESMFold) reusable as the SE(3) protein-structure tower for NLR receptor / Rubisco / seed protein / TF DBD trait engineering; Materials' cross-model ensemble disagreement aggregator code structure; Energy's httpx.MockTransport golden-fixture invariance test for Mac-CPU-vs-Runpod-GPU equivalence; Health's runs/<rid>/tuples.jsonl reasoner-tuple discipline.
• RESISTANCE.md doctrine is binding before any executor begins: forbidden meta-proxies fp-shapematchRE, fp-rushtoend, fp-NULLasout, fp-approvalseek, fp-flatteryasfreedom, fp-shapematch, fp-efficiency-as-corner-cutting, fp-pretty-greenflag, fp-less-work-disguised-as-focus are forbidden behaviours.

What We Don't Claim

• This is not a deployable medical product, not a regulatory certification, not a clinical or human-subject system. Outputs are research artifacts.
• Zer0pa does not take trait IP. Zer0pa is platform-only. Any sub-brief proposing Zer0pa take trait IP fails the brain-functionality gate at PRD review.
• Zer0pa does not perform environmental release of GMOs or gene-edited material directly. The platform produces dossiers; physical wet-lab work is performed by partner organisations under their own biosafety regimes.
• South Africa is not a near-term ag deployment market for gene-edited products. The Section-19 ministerial decision (January 2024) keeps gene-edited crops under the full GMO Act; the SA strategy defaults to MAS-and-prediction pathways with editing performed in Kenya / Nigeria / Mauritius and germplasm imported via conventional crossing under SACTA-recognized cultivar registration.
• The v1 build does not cover herbicide tolerance traits (automatic NGT-2 in EU regardless of edit size), PIP-equivalent insect resistance via novel insecticidal protein (EPA PIP registration trigger plus EU NGT-1 exclusion), or pharma traits (TraitMultiplier 0.30 in the Regulatory Fit Score).
• Zer0pa does not vertically integrate into grain supply chains. The Benson Hill Chapter 11 / Chapter 7 lesson is encoded as a binding architectural constraint.
• This is not an MVP, not a first-paying-customer build, not a toy. The target is the most overdesigned, best-in-class, science-and-evidence-anchored platform the technology landscape supports — a partner-ready Crop Improvement Dossier that no incumbent currently delivers.
• Pangenome-aware modelling claims do not extend to species without public graph pangenomes. Orphan crops fall back to SNP-only mode with explicit orphan_crop_pangenome_unavailable flag in the Risk Block.
• Foundation-model performance claims await Phase-1 fine-tunes on Runpod. PlantCaduceus / GPN / HyenaDNA / Evo-2 / PlantRNA-FM / Living Models BOTANIC are consumed-and-fine-tuned, not pre-trained from scratch.
• The v1 PRD does not commit to direct partnership outreach (Pairwise Fulcrum sub-licensing; ARC-BTP / SACTA / CIMMYT-Zimbabwe SA go-to-market). Technical scaffolds are built; partnership conversations are operator-deferred and may be sharpened by Report 2.

Verification Status

Code	Check	Verdict
V_01	Canonical README spine present with parser-safe headings and field shapes	PASS
V_02	Boundary block carried verbatim across all five briefing assets	PASS
V_03	RESISTANCE.md doctrine read order locked first; forbidden meta-proxies named	PASS
V_04	18 anti-decisions enumerated as binding falsifiers at PRD review	PASS
V_05	Engineering-blocker deep-research lookups (10 items) resolved with current-state findings	PASS
V_06	Sub-agent dispatch table populated with Sonnet 4.6 / Opus 4.7 only	PASS
V_07	ZERO-on-CPU before Runpod committed as binding Phase-0 acceptance gate	PASS
V_08	Operator review of PRD outline	PENDING
V_09	Phase-0 layer code (license-stamping ETL service through L6 dossier validators)	UNTESTED
V_10	Runpod cutover golden-fixture invariance test	UNTESTED

Proof Anchors

Path	State
source-briefs/01-precision-crop-genomics-report-1-full-technology-landscape.md	VERIFIED
synthesis/01-fresh-eyes-on-precision-crop-genomics-report-1.md	VERIFIED
PRD-OUTLINE-FOR-OPERATOR-REVIEW.md	DRAFT-PENDING-OPERATOR
RESISTANCE.md	VERIFIED
MODUS-OPERANDI.md	VERIFIED
HANDOFF-TO-ORCHESTRATOR.md	VERIFIED

Repo Shape

Field	Value
Proof Anchors	6 display anchors
Portfolio	Crop Genomics
Stage	STAGE-2 RESEARCH-INTAKE
Source Material	Operator-authored Pre-PRD Research Synthesis Report 1 of 2 + synthesis fresh-eyes pass
Authority Source	PRD-OUTLINE-FOR-OPERATOR-REVIEW.md (pending) → PRD.md (forthcoming) → HANDOFF-TO-OVERNIGHT-EXECUTOR.md (pending)
Support Sections	Read order; Licensing; Upcoming Workstreams; Anti-decisions enumerated; Cross-workstream fork-and-own

Read order (incoming agents, Sonnet / Opus only per operator directive 2026-05-01)

1. RESISTANCE.md — anti-corruption doctrine. Read first. Headspace.
2. README.md — this file.
3. MODUS-OPERANDI.md — multi-agent pattern; § Operator refinements binding all workstreams 2026-05-01 + sharpened 2026-05-02 (ZERO-on-CPU before Runpod, no false-flag reward hacking, no interim reporting); § Sub-agent topology Sonnet/Opus only.
4. HANDOFF-TO-ORCHESTRATOR.md — orchestrator role, inheritance, expected output, ten engineering-blocker lookups, three operator-engagement open questions, eighteen anti-decisions.
5. source-briefs/01-precision-crop-genomics-report-1-full-technology-landscape.md — operator-authored Pre-PRD Research Synthesis Report 1 of 2 (Sections 1-15: v1 architecture; L1 input encoding; L2 knowledge layer + license traps; L3 candidate generation; L4 screening; L5 BoTorch; L6 dossier; phenomics + remote sensing; Regulatory Fit Score; intersectional science map; commercial landscape; SA / SADC / orphan crops; ten research questions; frontier opportunities; build order). Report 2 forthcoming.
6. synthesis/01-fresh-eyes-on-precision-crop-genomics-report-1.md — synthesis fresh-eyes pass: dossier-first active-inference reframe; cross-model + license-class + Tier-1-vs-orphan-crop disagreement as three first-class falsification primitives; twelve specific gaps; pressure-test points; three operator-engagement open questions; heavy fork-and-own from Synbio.
7. PRD-OUTLINE-FOR-OPERATOR-REVIEW.md — orchestrator's PRD outline closing the synthesis gaps and adding ten fresh-eyes additions; sub-agent dispatch table; ten engineering-blocker resolutions; three operator-engagement open questions resolved with carry-forwards; ZERO-on-CPU binding gate; 22-falsifier registry preview.
8. ORCHESTRATOR-STARTUP-PROMPT.md — paste-ready orchestrator startup prompt for cross-machine handoff.

Anti-decisions (18 enumerated, binding falsifiers)

APSIM commercial use; AgroNT / Nucleotide Transformer weights (CC-BY-NC-SA); AlphaFold3 / ESM3 main weights (non-commercial); WorldClim 2.1 (CC-BY-NC-SA); PLAZA (CC-BY-NC-SA); TAIR live curated data (paid Phoenix Bioinformatics); Google Earth Engine free tier (not commercially usable for any private company); Ultralytics YOLOv5/v8 default license (AGPL-3.0 with required Enterprise license); OpenDroneMap as hosted SaaS (AGPL-3.0 network-use trigger); EOX Sentinel-2 cloudless mosaics (CC-BY-NC-SA); herbicide tolerance traits in v1; PIP-equivalent insect resistance via novel insecticidal protein; pharma traits; trait IP ownership by Zer0pa; SA gene-edited product deployment absent GMO Act amendment; gene drives; cultivation bans; vertical integration into grain supply chains.

Cross-workstream fork-and-own (no runtime co-dependency)

Sister workstream Zer0pa/Synthetic-Biology (highest-value source — bio-domain pipeline; SBOL3-attested audit-trail with sha256 hash chain → JSON-LD-attested dossier; four-column license decomposition; Unknown Enzyme Generative Sub-Pipeline tools as SE(3) protein-structure tower; MFMO multi-fidelity BoTorch shape as qNEHVI breeding-cycle optimiser); Zer0pa/Health (TxGemma reasoner-tuple discipline → dossier-tuple discipline; audit-log shape with hash chain); Zer0pa/Materials (falsifier-registry-first build discipline; cross-model ensemble disagreement aggregator code structure most mature implementation); Zer0pa/Energy (same-shape Runpod cutover with httpx.MockTransport golden-fixture invariance — Mac-CPU-vs-Runpod-GPU equivalence test); Zer0pa/Polymath-AI and Zer0pa/Polymath-Diffusion-AI (Sonnet/Opus-only sub-agent dispatch; RESISTANCE.md augmentation with fp-pretty-greenflag and fp-less-work-disguised-as-focus; HF Pro streaming-corpus regime); Zer0pa/RCE (surgical sub-brief shape for schema-locking tasks).

Licensing

This repository is part of the Zer0pa Precision Crop Genomics workstream and is publicly visible as a research-intake dossier surface; no Phase-0 code has landed yet (PRD outline awaiting operator confirmation).

Code license is set when the first code lands (Apache-2.0 default per the Zer0pa workstream pattern; see Zer0pa/Synthetic-Biology and Zer0pa/Glyph-Engine for the canonical instantiation).

Briefing materials in this repository — RESISTANCE.md, MODUS-OPERANDI.md, HANDOFF-TO-ORCHESTRATOR.md, source-briefs/, synthesis/, PRD-OUTLINE-FOR-OPERATOR-REVIEW.md — are operator-authored or agent-authored under the operator's authority. Re-use outside Zer0pa is by operator grant only.

Trademarks. "Zer0pa" and "Precision Crop Genomics" used in workstream identity are trademarks of Zer0pa. See TRADEMARKS.md.

Data and fixtures. No genomic data, phenotypic data, satellite imagery, or model weights are vendored in this repository. Bulk artifacts go to private Hugging Face under the Architect-Prime user (NOT the Zer0pa org), with manifests + metadata only on disk. See DATA_POLICY.md. The 18 anti-decisions enumerated above bind every dataset and model artifact considered for ingestion.

Public audit limits. This repository is publicly visible as a research-intake dossier surface. The public audit surface is bounded by this repository's structural files plus the briefing assets named in Proof Anchors. No partner data, no field trial data, no off-the-shelf licensed regulatory text bodies, no model fine-tune weights are part of the public audit surface. See PUBLIC_AUDIT_LIMITS.md.

Quick Start

The repository is at STAGE-2 RESEARCH-INTAKE. There is no Phase-0 code yet. The "quick start" is the read order in ## Repo Shape. Incoming agents (Sonnet / Opus only per operator directive 2026-05-01) read the eight files in order, then proceed by role:

# Synthesis-agent role — already discharged
# Inputs:    operator-authored Report 1
# Output:    synthesis/01-fresh-eyes-on-precision-crop-genomics-report-1.md  ✓ committed

# Orchestrator role — in progress
# Inputs:    Report 1 + synthesis + RESISTANCE + MODUS-OPERANDI + HANDOFF-TO-ORCHESTRATOR
# Output:    PRD-OUTLINE-FOR-OPERATOR-REVIEW.md  ✓ drafted (pending operator confirmation)
#            PRD.md                              ☐ pending operator confirmation
#            HANDOFF-TO-OVERNIGHT-EXECUTOR.md    ☐ pending PRD.md

# Overnight-executor role — pending
# Inputs:    PRD.md + HANDOFF-TO-OVERNIGHT-EXECUTOR.md
# Output:    Phase-0 ZERO-on-CPU implementation across 11 waves; falsifier registry; license-stamping ETL service; dossier JSON-LD all-seven-blocks-locked; RFS master decision matrix; rrBLUP+BGLR+sommer prediction stack; AlphaSimR+PyBrOpS breeding sim; CROPSR+Cas-OFFinder edit design; off-target enumeration; BoTorch qNEHVI six-axis Pareto on 100×5 AlphaSimR test corpus; SE(3) protein-structure tower CPU-side stub adapters; PCG validation triple end-to-end (seeds 1-3)

Upcoming Workstreams

Active — operator review pending

• Operator review of PRD outline — PRD-OUTLINE-FOR-OPERATOR-REVIEW.md surfaces ten architectural commitments for confirmation: dossier-first active-inference reframe; SE(3) protein-structure tower promotion from synthesis-v1.5 to Phase-1; BoTorch qNEHVI six-axis Pareto Phase-0 binding; closed-loop dossier mode Phase-1; 22 falsifiers committed Wave-0; Pairwise partnership outreach deferred to v2; SA direct partnership outreach deferred to Report 2; 18 anti-decisions encoded as falsifier IDs; Phase-0 ZERO-on-CPU acceptance gate binding; sub-agent dispatch Sonnet 4.6 / Opus 4.7 only.

Active — engineering once PRD is locked

• Full PRD writing — orchestrator writes PRD.md (~1,200-1,500 lines mirroring Synbio template) once operator confirms PRD outline; commits and pushes.
• HANDOFF-TO-OVERNIGHT-EXECUTOR — orchestrator writes role-specific handoff describing inheritance, expected output, authorities, and constraints.
• Phase-0 ZERO-on-CPU overnight execution — overnight executor agents on a separate Mac dev machine read PRD + HANDOFF + RESISTANCE; decompose into 11 waves; complete every CPU-tractable component to 110% before any Runpod GPU is requested.

Operations / External Dependency

• Report 2 reconciliation sub-brief — Report 2 forthcoming will sharpen architectural specifications, partner integration sequences, and the SA go-to-market motion through ARC-BTP / SACTA / CIMMYT-Zimbabwe; orchestrator runs reconciliation pass once Report 2 lands; surfaces displacement against v1 PRD.
• Living Models BOTANIC license verification — per-model-card check on huggingface.co/living-models/Botanic0-{S,M,L}; class-A/B verdict triggers Phase-1 fine-tune deliverable in Zer0pa-secure HF Pro environment per Living Models partner pattern; class-C/D verdict triggers anti-decision falsifier f022.

Research-Deferred — Investigation Underway

• Plant single-cell foundation model — scPlantLLM exists with unclear commercial licensing; commercial path is fine-tuning scGPT (MIT) or Geneformer (Apache-2.0) on Plant Cell Atlas; if commercial-licensed plant scFM emerges in 2026, cell-type-resolved trait modelling warrants a dedicated model branch.
• ImageNet-for-crop-GSxE benchmark — license-clean cross-crop benchmark with held-out test sets and standardized evaluation protocols; currently unowned; Zer0pa community contribution candidate; Phase-2 frontier.
• BoTorch-native multi-objective breeding-cycle optimizer with closed-loop active-inference — promoted from frontier-deferred to Phase-1 in PRD outline; six-axis Pareto front demonstration on AlphaSimR 100-genotype × 5-environment test corpus is Phase-0 acceptance test.

Zero-Base Scientific Thinking — GPD Research and Planning Pending

• Causal inference for trait discovery — Phase-2 first frontier; PC-algorithm graph induction plus Mendelian-randomization effect-size pruning over multi-trait biobank-scale panels; CI-GWAS demonstrates multi-trait causal graphs from 17 traits in <10 minutes on GPU.
• G×E-VAE (climate physics + active inference joint) — Phase-2 second frontier; learned environment encoder branch meeting genome encoder in joint latent space; forward projection onto CMIP6-2050 scenarios for climate-resilient cultivar selection.
• GNN over heterogeneous knowledge graph — Phase-2 third frontier; KG substrate built Phase-0; GNN message-passing with eigenvector and betweenness centrality as priors on marker effect.