What the programme asks, and how it keeps its evidence apart

The manuscript’s contributions, in its own framing: a bounded QuestionArtifact, five epistemic functions for continual discovery, a proposed (not validated) sparse question-priority mechanism, and four distinct evidence classes that test retrieval, repair, acquisition and distribution under failure.34 The programme does not claim graph memory, persistent memory, reconstruction or hierarchy as standalone novelties; the contribution is the combination: first-class travelling questions over a lineage-typed fabric, with explicit separation of acquisition, availability and independent support.5

The programme studies three things together: bounded questions that travel between memory holders as first-class objects, a fabric in which every claim carries its lineage, and an evaluation discipline that separates what was acquired, what stayed available and what remained independently supported.61 Graph memory, persistent memory, reconstruction and hierarchy are prior art and are not claimed as inventions.

1. Every figure on the site has a source note with the document and section, and a row in the claims register; the register wins if a page disagrees. Claims register →
2. Two different system ladders exist. The manuscript’s Class-D ladder (B0 to B7, B3Q, B6C) and the organisational benchmark’s ladder (B0 to B9 plus ORACLE) are always prefixed “manuscript” or “bench” so that “B7” is never ambiguous.2930
3. LoCoMo category labels are shown “as labelled” because the inherited category constants swap single-hop and multi-hop; aggregates are unaffected.31
4. Superseded numbers appear only in the corrections log.32 Corrections log →
5. Wording rules: “behaviour is stable across four simulated scales”, never “scales to 100,000 agents”; “designed so that raw data does not leave the device”, never “guaranteed”.3334
6. Manuscript-derived content (Class C and D tables, verbatim quotations, formal definitions) is not in the deployed pages. In the committed source it sits between gate markers; tools/build.py, which the deploy workflow runs on every push to main, strips it and fails the build if any of it, or a citation to it, survives. It returns only when the build is run with --manuscript-cleared, after the authors confirm that publication does not breach the conditions of review.3550 Where an ungated sentence rests on the manuscript, the site says it in its own words and keeps the citation. A source note that cites the manuscript by section and margin line is a reference, not content: it stays in the deployed page whenever an ungated sentence points to it, and is held out only when every marker that points to it is held out with its copy.

One manuscript describing the lineage-first fabric, the QuestionArtifact, continual discovery and the four evidence classes is under review.35 The venue is not named and the manuscript is not distributed from this site while review is in progress; the site describes its content in the lab’s own words and marks every simulation figure as “reported in the manuscript”. After the review decision this section will link the paper or its preprint and the corrections log will record the date.

Note first: source paths on this site cite the NeuralGraph repository, which is public to read at github.com/anovruzov/NeuralGraph but has no licence, so all rights are reserved and nothing in it is offered for reuse.3648 “An auditor can check every sentence” therefore holds for any reader: the rows below link to the commit or branch each artifact is recorded at, and a request to the lab is needed only for what is not in the repository.

1. Class B figures. The repository’s benchmark record gives random path diversification 0.704 survival and 5,327 bytes.40 The manuscript gives 6/9 and 5,206 bytes for the same strategy.41 The manuscript reports a different figure for the same strategy; the comparison is held back while review is in progress. This site uses the repository record (bitwise reproducible) throughout and says so.
2. LoCoMo labels. Inherited category constants swap single-hop and multi-hop; all tables say “as labelled”; a corrected mapping is pending in the repository.31
3. Two ladders. The manuscript’s B0 to B7/B3Q and the benchmark’s B0 to B9 plus ORACLE share ids; the site always prefixes.30
4. Two Tesseracts. The repository contains an older single-node retrieval module and the cross-node coordinator, both called Tesseract; this site means the coordinator unless it says otherwise.42
5. Test counts. The coordination suite was 107 tests before a real-storage restart test added two; the site cites 109.43
6. Default model. The README lists an Ollama default of qwen2.5:7b-instruct while the code’s default is an OpenAI-compatible endpoint at 127.0.0.1:1234 with google/gemma-4-e4b; the site therefore states no default model and treats model names as configuration.4445
7. Gossip baseline. Table 3 gives manuscript-B4 knowledge survival .749 at 48.437 replicas per claim under the 50% white-box attack;46 the prose (§5.6 ll.206-207) gives B4 .669 at the same storage in the sentence that follows the 90%-churn comparison.47 We read .669 as the 90%-churn value; the manuscript does not label it explicitly, so the site shows neither as a headline and records the reading here.

Every future discrepancy is added here and dated in the corrections log. Corrections log →