Reproduce provenance atom deduplication for graph retrieval

Finding

GraphEcho, arXiv:2609.17695, submitted September 15, 2026, separates graph path multiplicity from evidential origin multiplicity. Its controlled synthetic benchmark reports that redundant supporting paths increase repeated walks across all four frozen agents. Provenance aware post training raises synthetic accuracy by 41.5 percentage points over frozen Qwen3 4B, but reduces source recall and SciFact accuracy. The negative transfer result matters: reducing revisits is not equivalent to acquiring better independent evidence.

Evidence class: originating team measured result with a controlled benchmark. This is not independently reproduced by RuV.

Opportunity score

4.32 out of 5 under the RuV daily SOTA weighting model.

RuV hypothesis

RuVector and WorldGraph should make an explicit provenance atom available alongside graph path identity. Retrieval, reranking, and graph walk policies can then report both structural multiplicity and independent evidence multiplicity rather than silently treating five paths to one observation as five corroborating observations.

The primitive should remain descriptive. Source count must not become authority, truth, or confidence by itself.

Frozen experiment

Compare:

1. Current graph retrieval and walk behavior.
2. Exact content deduplication only.
3. Canonical provenance atom deduplication with unchanged retrieval scores.
4. Provenance diversity reranking under a fixed walk or token budget.
5. Retrospective oracle over distinct evidence origins.

Use synthetic graph interventions matching path count while varying source count, then transfer to a held out scientific or technical corpus with real source provenance.

Acceptance

Advance only if provenance atoms reduce redundant evidence visits by at least 30 percent under a fixed budget while retaining or improving held out task accuracy and source recall. A result that reduces repeated walks but lowers held out accuracy or distinct source recall is a negative result and must not be promoted.

Benchmark contract

Report exact commits, model and embedding versions, corpus, provenance construction method, seeds, sample size, walk budget, token budget, accuracy, source recall, repeated visit rate, distinct provenance coverage, latency, memory, index overhead, failures, distribution shift, and reproduction steps.

Cross stack mapping

RuVector owns provenance atom representation and retrieval diagnostics. RuVector WASM needs a compact representation for browser and edge graph traversal. WorldGraph consumes independent origin counts for spatial and evidence graphs. MetaHarness owns independent task evaluation. RuFlo can use provenance diversity as a search heuristic only. RVF can bind source identity and evidence digests. Autogenous and Dream Machine may optimize exploration but cannot rewrite provenance. MidStream carries origin lineage events. RuView and RuField can distinguish multiple derived features from independent sensor observations. LatentMesh can carry provenance ids without granting trust. Cognitum can expose evidence diversity diagnostics for enterprise research. MCP resources retain source identity across tool boundaries. Distributed infrastructure must avoid duplicating one physical observation into apparent independent evidence.

Governance

This issue authorizes a reversible benchmark and interface experiment only. No authority changes, no confidence inflation from source count, no autonomous merge, and no production deployment.

Start with RuVector's provenance atom representation and retrieval diagnostics, then map the five frozen experiment comparisons; the issue names no files or tests. Run synthetic graph interventions with matched path counts and varied source counts, recording the listed benchmark contract. Done means reducing redundant visits by at least 30% while retaining or improving held-out accuracy and source recall, or documenting a negative result.