BioNexus v0.8.0

[0.8.0] - 2026-08-15

🚀 Added

• Machine-Readable Capability Contracts (bionexus.capabilities):
  • Implemented formal CapabilityContract, SemanticInputType, Precondition, RefusalTrigger, EvidenceRequirement, and CapabilityEvaluationResult.
  • Registered 8 canonical capabilities: scrna.pseudobulk_de, scrna.exploratory_clustering, spatial.morans_svg, survival.kaplan_meier, scvi.probabilistic_vae, allotrope.format_conversion, nextflow.pipeline_launch, variant.acmg_classification.
  • Added CLI subcommands: bionexus capability [list|show|check].
• 6-Stage Scientific Intent & Invariant Router (bionexus.intent_router):
  • Implemented 6-stage routing pipeline: Intent Extraction $\to$ Data Semantics $\to$ Preconditions $\to$ Capability Matching $\to$ Backend Probe $\to$ Decision.
  • Added authoritative routing statuses: PERMITTED, NEEDS_DATA, ABSTAIN, DEGRADED_ADVISORY.
  • Added CLI command: bionexus route "<query>" [--data <path>] [--min-replicates <N>].
• BioNexus Eval: Agent Behavior & Scientific Reliability Benchmark (evals/):
  • Implemented 8 Core Reliability Metrics: Routing Accuracy, Unsafe Invocation Rate, Abstention Precision/Recall, Capability Hallucination Rate, Backend Fidelity, Scientific Semantic Error Rate, Evidence Calibration, and Composite Reliability Index (CRI).
  • Created 6 evaluation datasets across 29 structured prompts: routing.yaml, refusal.yaml, capability_claim.yaml, scientific_semantics.yaml, backend_failure.yaml, adversarial.yaml.
  • Added CLI command: bionexus eval [--suite <name>] [--report <path>] [--json].
• Single Source of Truth (SSOT) Multi-Platform Manifest Compiler (bionexus.registry):
  • Canonical bionexus.registry.yaml compiling into Agent Plugins 1.0 (plugin.json, mcp.json), Claude Plugin (.claude-plugin/plugin.json), and OpenAI Codex (.codex/config.json).
  • Added CLI command: bionexus registry [--generate|--check|--validate-endpoints].
• Skill Scaffolding Tool (bionexus.scaffold):
  • Added CLI generator bionexus create-plugin <name> generating Gold Reference skill directories, pipelines, and offline unit test fixtures.
• Ecosystem Governance Documentation:
  • Added docs/versioning-policy.md, docs/compatibility-matrix.md, docs/migration-guide.md, docs/deprecation-policy.md.
  • Added automated GitHub release workflow .github/workflows/release.yml.

🔄 Changed

• EvidenceCard 2.0 (Three-Layer Epistemic Model):
  • Decoupled into ExecutionState (EXECUTED, DEGRADED, REFUSED, FAILED), DimensionGrade (A, B, C, UNTESTED, NOT_APPLICABLE, INSUFFICIENT, CONFLICTED), and ConclusionMaturity (ABSTAIN, FRAGILE, CONFLICTED, PRELIMINARY, SUPPORTED, ROBUST, REPLICATED).
  • Added non-breaking backward compatibility aliases ConclusionStatus and EvidenceGrade.
• Parameter Robustness Auditing:
  • Added audit_parameter_stability() in bionexus.integrity computing Adjusted Rand Index (ARI) and Jaccard similarity across parameter perturbation sweeps.

🛡️ Fixed

• Fixed single-cell condition differential expression to require biological replicates ($n \ge 2$) and refuse single-sample pseudoreplication.
• Fixed negative binomial GLM inputs to refuse continuous normalized floats and enforce discrete integer counts.
• Fixed cell-type naming invariants to prohibit unverified hallucinated biological labels.

---