SOTA: epistemic ancestry tracking so more agents do not count as more evidence

Finding

Epistemic Sybil Resistance (arXiv:2609.01873, submitted 2026-09-01) formalizes a failure mode that directly affects distributed RuV intelligence: another agent report is not another independent observation.

The originating team reports more than 20,000 controlled LLM report and extraction calls. Holding one evidence root fixed while report multiplicity grows from 1 to 32 collapses naive posterior coverage from 0.940 to 0.263. Increasing true evidence-root multiplicity closes the gap. They also measure correlated extraction error among repeated agent calls and show that report-space deduplication tracks surface similarity much more strongly than actual evidential ancestry.

The public reproducibility repository is MIT licensed for code and includes frozen outputs, tests, a standalone benchmark, pinned dependencies, fixed seeds, and a no-network reproduction path.

RuV implication

RuVector, Ruflo, Autogenous, LatentMesh, Core Memory, and MetaHarness should reason over evidence roots and dependence, not agent count, report count, vote count, or embedding similarity.

This complements earlier correlation-aware memory work but gives a sharper invariant:

N reports from one evidence root never become N independent observations.

At the same time, common ancestry does not imply complete redundancy because independent extraction passes can still add bounded information. The model must track both ancestry and extraction dependence.

Proposed primitive

Add an experimental EvidenceAncestry graph above normal RuVector retrieval.

Each evidence-bearing artifact should be able to carry:

• immutable evidence root identifiers
• parent artifact identifiers
• observation or source class
• extractor or model family identity
• extraction attempt identity
• known shared context roots
• tenant and resource scope
• confidence and calibration metadata
• explicit authority: none

Aggregation should estimate effective independent evidence from roots plus dependence, while retaining raw report multiplicity as descriptive telemetry only.

Benchmark

First reproduce the public ESB benchmark without integrating RuVector. Then implement a RuVector adapter and compare:

1. naive report voting
2. embedding or lexical deduplication
3. evidence-root clustering only
4. evidence-root clustering plus extraction-dependence correction

Add RuV-specific cases using multiple agents reading the same document, different documents with similar wording, the same source through multiple retrieval paths, and independent sensors producing superficially similar observations.

Metrics

Calibration coverage, Brier score, log loss, effective evidence count, false corroboration rate, missed corroboration rate, aggregation latency, index bytes, provenance traversal cost, and task utility.

Falsification

Reject the extra graph machinery if simple source-root counting performs within variance, if ancestry metadata is unavailable often enough that estimates become mostly unknown, or if dependence estimation is too unstable to improve held-out calibration.

Security and governance

Evidence ancestry is not identity authority and not execution authority. A signed agent cannot turn a duplicated source into independent evidence merely by signing it. A trusted source can still be wrong. RVM remains the privileged action boundary.

Test forged ancestry, omitted parents, cross-tenant roots, cyclic ancestry, source aliasing, identical reports from independent roots, distinct reports from one root, and correlation shifts across model upgrades.

Acceptance

Advance only if held-out calibration error improves materially over both naive voting and similarity deduplication, false corroboration falls by at least 50%, task utility does not regress by more than 2 absolute points, provenance traversal remains bounded, and zero cross-scope or authority regressions occur.

Coordinate independent reproduction through MetaHarness and record promotion state in Core Memory federation. No autonomous merge.

Start by reproducing the public ESB benchmark without integrating RuVector, using its frozen outputs, tests, standalone benchmark, pinned dependencies, fixed seeds, and no-network path. Then compare the four listed aggregation approaches and define the RuVector adapter and RuV-specific cases. Done means held-out calibration improves materially, false corroboration falls by at least 50%, task utility drops no more than 2 points, traversal remains bounded, and no cross-scope or authority regressions occur.