MEMORY SUBSTRATE / PUBLIC RESEARCH PROTOTYPE
HSWM은 출처가 고정된 증거를 불변 world artifact로 컴파일하고, 가중 하이퍼그래프 field와 인증된 readout 경계로 읽는 메모리 substrate입니다. 답을 스스로 만들어내는 reasoner가 아닙니다.
TECHNICAL CONTRACT
각 단계는 다음 단계가 사용할 수 있는 명시적 artifact 또는 typed refusal을 만듭니다.
Frozen sources and recorded observations compile into immutable, content-addressed WorldArtifactV1 records or a typed rejection.
World, embeddings, topology, revision cut, policy, and candidates are byte-addressed as one immutable read boundary.
Evidence-bearing nodes and n-ary relations receive bounded retrieval weights; the measured positive path is currently static additive-j.
Only an exact certified tuple reaches the scoring kernel; scope mismatch returns a payload-free typed refusal.
Revision can suppress stale evidence while retaining it for audit, but durable replay, compensation, signatures, and concurrent publication remain open.
PUBLIC CLAIM LEDGER
아래 수치는 공개 저장소에 추적된 영수증과 효능 장부만 사용합니다. 현재 작업트리의 미공개 또는 미추적 결과는 포함하지 않습니다.
The checked-in conformance run admitted 40/40 valid controls, refused 400/400 scope-fault calls before scoring, and refused 9/9 injected mutants without returning a payload.
This confirms local conformance for the tested cut. It does not establish retrieval quality, cryptographic authenticity, exhaustive security, or production durability.Across three checked-in MuSiQue/2Wiki runs, HSWM measured +0.0364 support recall@3, +0.0259 nDCG@10, and +0.0729 downstream answer F1 over cosine.
HSWM used 100 offline LLM judgments per run. The ladder omitted strong late-interaction and production graph retrievers, while cosine still led hit@3 and MRR; no state-of-the-art claim follows.The preregistered cross-dataset criterion failed: pooled HSWM minus direct-LLM answer F1 was -0.1489; the positive 2Wiki delta was not significant at the stored paired-bootstrap threshold.
The supported statement is retrieval improvement over the listed lightweight baselines on this ladder, not a smarter reasoner.MuSiQue and 2Wiki certificates selected mu=0, and none of nine tested traversal settings beat the static field on hop-drop.
The implementation falls back to the same snapshot's static field. This is not evidence of successful graph reasoning.A full-dose non-destructive supersede write removed the maximally confusable stale fact from top-10 in the two checked-in datasets while preserving it for audit.
An external graded revision arm was bit-exact, and a wrong write reduced current recall by 12.69 points on MuSiQue and 31.0 points on 2Wiki. Durable correction and replay are still required.Crash-safe event replay, compensation, concurrent publication, signatures, and external trust distribution are not present-tense capabilities.
The public implementation is a compiler boundary and measurement prototype, not a production memory service.A synthetic experiment establishes mechanism sufficiency when judge-readable aboutness survives vector dilution.
No real NoCha, QASPER, NarrativeQA, or book-scale result has landed, so the website makes no long-document efficacy claim.EXPLICIT NON-CLAIMS
REPRODUCE BEFORE BELIEF
소스 트리 검증은 NumPy와 pytest 개발 extra로 실행됩니다. GPU/LLM production receipt는 별도 계층입니다.
uv sync --extra dev uv run python verify_efficacy_claims.py --pretty PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 uv run pytest -q -p no:cacheprovider uv run python certified_cut_compare.py
OPEN RESEARCH, INCLUDING NEGATIVE RESULTS