Context
This is external research feedback after reviewing the public repository at 6328de6, its efficacy ledger, and the current implementation boundaries.
HSWM is clearly connected to active work in Hypergraph RAG, graph memory, temporal knowledge graphs, and verifiable retrieval. The risk is that the broad architecture is already crowded:
- HyperGraphRAG: n-ary hypergraph RAG;
- SiReRAG: recursive similarity/relatedness structures for multihop retrieval;
- HippoRAG 2: graph-based non-parametric continual memory;
- GFM-RAG: learned graph retrieval at much larger scale;
- Zep/Graphiti: temporal graph memory and invalidation of superseded facts;
- AriGraph: graph memory used with planning in interactive environments.
The repository is commendably explicit that it is not yet a reasoner or production runtime. The checked-in results also show narrow positive static retrieval, traversal certified OFF on the real datasets, failed cross-dataset cognitive uplift, and an unsuccessful B2.1 learned router.
Candidate contribution to isolate
The most specific potentially distinctive HSWM hypothesis appears to be:
Can one versioned semantic weight field consistently serve retrieval, selection, and knowledge revision better than separate task-specific retriever/planner/revision heads?
That is more defensible than a generic "new Hypergraph RAG" or "world model" claim.
Currently this hypothesis is not established:
plan() is a compatibility alias for selection_distribution(), not task/cost/risk planning;- prototype
supersede() is multiplicative salience decay;
- the positive additive-j comparison gives HSWM offline LLM judgments that the lightweight retrieval baselines do not receive;
- structural composition tests establish closure/invariants, not reasoning or retrieval efficacy.
Suggested decisive experiment
Run a preregistered shared-field versus separate-head comparison under equal budgets.
Arms:
- Frozen embedding/cosine baseline.
- Strong graph/hypergraph retrieval baseline.
- Separate retrieval, selection, and revision heads with the same total parameters/calls.
- HSWM shared field with the same total parameters/calls.
- Shared-field ablations: no slow weight, no query weight, no topology, no supersession.
Tasks should include both:
- multi-hop/cross-document retrieval, where graph structure has a plausible role;
- evolving-knowledge cases with contradictions, stale facts, as-of queries, and repeated revisions.
Measure:
- retrieval recall/nDCG and downstream answer quality;
- stale-fact suppression and current-fact preservation;
- as-of correctness and revision confluence;
- disagreement between retrieve/select/revise decisions;
- replayability/audit success;
- latency, memory, model calls, and offline judgment cost.
Use LakatoTree only as the experiment-governance layer and attribute the preregistration/receipt/verdict machinery to LakatoTree rather than to HSWM.
Success boundary
A strong HSWM result would show that the shared field either:
- improves quality at equal cost, or
- preserves quality while materially reducing cross-head inconsistency and improving audit/replay guarantees.
If neither happens, HSWM remains a well-engineered evidence-preserving graph retrieval substrate, but the shared-field research claim should be rejected or narrowed.
Context
This is external research feedback after reviewing the public repository at
6328de6, its efficacy ledger, and the current implementation boundaries.HSWM is clearly connected to active work in Hypergraph RAG, graph memory, temporal knowledge graphs, and verifiable retrieval. The risk is that the broad architecture is already crowded:
The repository is commendably explicit that it is not yet a reasoner or production runtime. The checked-in results also show narrow positive static retrieval, traversal certified OFF on the real datasets, failed cross-dataset cognitive uplift, and an unsuccessful B2.1 learned router.
Candidate contribution to isolate
The most specific potentially distinctive HSWM hypothesis appears to be:
That is more defensible than a generic "new Hypergraph RAG" or "world model" claim.
Currently this hypothesis is not established:
plan()is a compatibility alias forselection_distribution(), not task/cost/risk planning;supersede()is multiplicative salience decay;Suggested decisive experiment
Run a preregistered shared-field versus separate-head comparison under equal budgets.
Arms:
Tasks should include both:
Measure:
Use LakatoTree only as the experiment-governance layer and attribute the preregistration/receipt/verdict machinery to LakatoTree rather than to HSWM.
Success boundary
A strong HSWM result would show that the shared field either:
If neither happens, HSWM remains a well-engineered evidence-preserving graph retrieval substrate, but the shared-field research claim should be rejected or narrowed.