Nectar

Inspiration

An agent can make a correct decision from the catalog and still become wrong five minutes later when a classification, lineage edge, quality failure, or tag arrives late. Ordinary provenance shows what the agent read; it does not identify downstream claims that silently relied on something not being there.

What it does

Nectar adds negative space to the DataHub context graph. It records bounded, fingerprinted facts proven absent at decision time and attaches claim-level dependencies across agents. When DataHub later contradicts one of those absences, Nectar traces a deterministic trust shockwave, recalls only contaminated claims, preserves independent work, and emits a minimal repair plan.

The 90-second demo starts with two absence proofs and five claims across four agent outputs. One click adds a late PII classification. Nectar recalls exactly three dependent claims across three outputs, retains two independent claims, and tells the workflow what to recompute. No LLM decides what survives.

How we built it

The live adapter uses the official DataHub MCP Server to read entity schema, classifications, ownership, lineage, authenticated identity, versioned policy documents, and prior decision evidence. Nectar turns those graph reads into positive evidence and bounded negative dependencies. Verified decisions can be written back as DataHub documents and recovered by a second agent. Every mutation requires a visible user action plus confirmWrite: true; the recall API itself is read-only.

The implementation includes strict NegativeDependencyV1, TrustClaimV1, TrustGraphV1, MetadataDeltaV1, and RecallResultV1 contracts; recursive rejection of unknown fields, dangling references, cycles, forged query fingerprints, and tampered graph fingerprints; canonical deterministic transitive closure; and selective repair that separates recompute and preserve sets.

Nectar also retains a deterministic preflight substrate that can allow, safely rewrite, review, or block execution using a restricted DataHub-managed PolicyBundleV1. It supports explicit approval, constrained execution, output verification, reusable evidence, and automatic stale-context rejection.

Challenges

Proving absence safely was harder than recording presence. Nectar refuses to create a classification-absence dependency unless the field was actually observed, fingerprints the exact bounded negative query, and rejects a proof when the supposedly absent fact is already present. Cross-agent claim graphs also needed cycle, dangling-reference, and tamper protection before selective recall could be trusted.

DataHub document search is index-backed, so reliable receipt replay required complete document retrieval and bounded indexing retries. Live approval fails closed when identity or ownership cannot be verified.

Accomplishments

• 65 named policy cases plus 5 negative-space cases and 5 safety invariants.
• Deterministic recall of 3 contaminated claims while preserving 2 independent claims.
• Tamper, cycle, expiry, identity, malformed input, MCP failure, and stale-context rejection.
• Machine-readable cold, warm, trust-shockwave, and live-MCP benchmarks.
• Sanitized live DataHub trace and clean-room installation evidence.
• Public credential-free synthetic demo with zero browser console warnings or errors.
• No LLM in the decision or recall path and no real patient data.

Originality

Nectar does not claim policies, provenance, or receipts are new. Its differentiating composition is:

bounded absence proof → exact late-context contradiction → cross-agent claim-level recall → independent-claim preservation → minimal repair

A source-backed comparison against DataHub, Faramesh, Cedar Analysis, PROV-AGENT, and TRACER is committed in the repository. A transparent bounded information-rarity audit scores the declared feature set at 87.7/100. That score is positioning evidence, not proof of universal novelty or a prediction of judging outcomes.

What we learned

For platform teams, the alternative is usually binary and expensive: keep trusting stale agent work or throw away the whole workflow. Nectar localizes the blast radius. Consumers are protected from outputs invalidated by late governance context, while developers keep verified claims that remain safe and recompute only the contaminated path.

What's next

The polished vertical is synthetic healthcare metadata. Next, we would connect the deterministic delta analyzer to production DataHub metadata change events, expand bounded negative-query types, and prepare the generic decision-receipt pattern as a focused upstream contribution. Nectar is not HIPAA certification or a replacement for DataHub authorization.

Built With

• datahub
• datahub-mcp-server
• docker
• node.js
• react
• sha-256
• typescript
• vinext