Knowledge graphs make it easy for agents to turn documents into linked facts, merged entities and higher-level summaries. Correction is harder because those products depend on one another. Removing a source does not automatically remove an inference or summary that inherited its content. Rebuilding the graph may also introduce unrelated changes if the constructor is stochastic.
PolyGraph examines maintenance as a first-class part of memory construction. Its approach asks for evidence receipts when each artifact is created, so later deletion can follow recorded dependencies. The key question is whether those receipts actually describe what supports the claim, rather than merely offering a plausible citation list after the fact.
The receipt should be created with the artifact
The October 13 PolyGraph paper asks constructors to attach evidence bundles when they create derived artifacts. Cleanup can then follow those dependencies. Its audit also finds that summary-level receipts can be unreliable when several sentence-level claims are collapsed into one evidence list. Read the research paper.
For each artifact, the constructor emits alternative bundles of source evidence, with each bundle claimed sufficient on its own. Cleanup can then use set operations over the receipts without making new model calls. This separates expensive construction from a more mechanical dependency check when a source is withdrawn.
The paper’s corpus illustrates why direct deletion is insufficient. Removing only directly extracted facts leaves much of the contamination in downstream queries. It also reports substantial variation between two builds of the same corpus, making wholesale reconstruction a noisy way to isolate the effect of one retraction.
Rebuilding can obscure the effect of a retraction
The paper reports that two builds of the same 300-document corpus differ in 29,475 artifacts, compared with 6,248 changes associated with removing 61 retracted documents. The constructor’s variability can therefore be larger than the intervention whose consequences the maintainer wants to trace.
This is why a fresh rebuild is not automatically a clean correction baseline. Differences may reflect sampled merges or summaries rather than the withdrawn evidence. Recording dependencies during construction provides a more targeted way to ask which artifacts have lost support. The re-derivation audit remains necessary because the receipts can be wrong, particularly when a summary combines several claims. Together, the two observations explain the design: targeted arithmetic can be stable and inexpensive, but only if the original dependency records have the right granularity and have been checked against actual reconstruction behaviour.
Provenance itself needs an audit
Receipts are the constructor’s self-report, so the authors test them by removing cited evidence and checking whether the artifact is re-derived. Narrow merge claims show better agreement than flat lists attached to multi-sentence summaries. Moving to a receipt per sentence substantially improves faithfulness in the reported experiments.
This identifies a concrete representational problem. A flat summary citation list cannot express which sentence depends on a deleted source and which survives independently. More citations do not solve that ambiguity. The dependency object must have a granularity compatible with the claim it is meant to preserve or invalidate.
Deletion should preserve independent knowledge
A maintainer has two errors to avoid: retaining unsupported descendants and deleting material that still has another valid basis. Alternative evidence bundles let cleanup distinguish those cases, provided the receipts are faithful. The paper reports strong contamination reductions under its tested constructors, domains and audits, not a guarantee of perfect provenance in arbitrary generated knowledge.
The wider implication is that inexpensive memory creation can produce an expensive liability unless maintenance information is recorded immediately. Source withdrawal, correction and changed predicates belong in the design from the beginning. A system that remembers quickly also needs a precise account of what would make each remembered conclusion cease to be justified.
Questions, news tips or corrections? Contact [email protected].
