fresnellll/atlasnav-artifacts

AtlasNav anonymous research artifacts

This companion release is consumed by AtlasNav 0.2.x. Verify it before use:

atlasnav artifacts verify --artifact-root PATH_TO_RELEASE
atlasnav reproduce --suite paper --artifact-root PATH_TO_RELEASE --output reproduced

BrowseComp-Plus

browsecomp_plus/trajectories contains seven checksummed full-830 archives: one adopted trajectory per query for AtlasNav, DR-DCI, and raw DCI with the DeepSeek and ChatGPT backbones, plus the ChatGPT Gold-document Reference. The archives are clean merged units rather than development/remainder shards. The replay command recomputes strict accuracy—including the paper's 801/830 = 96.51% empirical-reference cell—recorded online cost, paper consistency, paired outcome cells, passive checkpoint endpoints, frozen checkpoint-manifest integrity, and Evidence Blindness. Passive replay is not relabeled as counterfactual active Safe Release accuracy.

browsecomp_plus/atlas/unified_multiplex_v1 contains the complete frozen 100,195-file Atlas: four reduced vector matrices and transforms, sparse graphs, 77-parent/443-leaf hierarchy arrays, cards, labels, catalog, and provenance. browsecomp_plus/router/linear_v1 contains the exact deployed Router. The original full-830 query-vector bundle is no longer available. To preserve exact online reconstruction, browsecomp_plus/frozen_ranking/full830 contains the checksummed post-Router weighted-RRF score for every query/document pair, the aligned query-ID catalog, retrieval audit, and byte-exact frozen first viewports with their navigation metadata. Together with the frozen Atlas, upstream corpus, and full-text index, this recreates the exact initial routes without an embedding API call. This is explicitly a later-stage frozen representation, not a claim that the missing raw query vectors were recovered.

candidate_tasks_v1 contains 30,000 corpus-grounded support candidates, not the exact final 7,163 generated training examples; those rows are unavailable and must not be represented as recovered.

The fragment-level Qrel under qrel/v2_4 covers 830 questions, 889 required answer slots, and 1,753 acceptable spans. Benchmark-derived quotations remain subject to final upstream redistribution review.

MiMo and Qwen are summary-only because their final consolidated trajectories are not recoverable from this server.

Frozen auxiliary construction assets

The frozen/ directories for the reported auxiliary benchmarks now contain the exact final Atlas and the full four-view query embeddings. Each Atlas includes four 192-dimensional reduced document-vector matrices, the fitted PCA transforms, four sparse graphs, hierarchy arrays, cards, labels, and the aligned catalog. Full 2,560-dimensional document matrices are deliberately omitted as a reproducible high-volume intermediate; the public pipeline can regenerate them from upstream corpora. Upstream corpora and query text are not redistributed.

Auxiliary records

Compact per-query or aggregate records are provided for PhantomWiki, FanOutQA, TREC-COVID, 2Wiki-Global-400, and fixed SciFact/ArguAna samples. EnterpriseRAG is explicitly an outcome-aware, one-rerun known-error-replacement upper-bound diagnostic. Its 500 adopted rows and official metrics are merged, but it is not a stable single-trajectory leaderboard submission.

Integrity and scope

release_manifest.json declares schema compatibility and known blockers. MANIFEST.sha256.json binds every admitted path, byte size, and SHA-256. Upstream corpora and question datasets are not redistributed here. No account credentials, provider-specific host configuration, author identity, or local machine path is part of this release.