This guide introduces the main Paper Harness views using a captured, read-only workspace. The navigation and availability labels are part of the product contract: incomplete enrichment remains visible and does not hide independently usable source metadata.
The screenshots contain public research-paper metadata and short evidence excerpts. That content remains owned by its authors and upstream sources; its appearance here does not relicense it under Apache-2.0. AI-generated summaries, claims, comparisons, and inferred relations are unverified unless the interface explicitly states otherwise.
Use the topic selector in the upper-right corner to switch between Broad LLM Agents, Brain-Computer Interfaces, and World Models. Topic selection scopes the Dashboard, Reports, Papers, Graph, Trends, and Runs views. A deep link to a paper, comparison, lineage, or run preserves the relevant identity; selecting a different topic returns to the corresponding topic-level view.
The screenshots show one captured state, including its dates, paper counts, and availability. They demonstrate the interface rather than promising current or exhaustive research coverage.
The Dashboard answers "what changed in this research domain?" It summarizes tracked papers, active topics, and the latest completed work, then surfaces the current Daily publication and report. Status badges distinguish the overall pipeline from the published product outcome.
Start here to:
- confirm the active topic;
- see whether the latest day is
COMPLETE,PARTIAL, orNO_UPDATE; - open the current report or report history; and
- move into papers, evidence, graph, trends, or run diagnostics.
Reports group one topic's publishable results by logical date. The date picker opens a historical Daily publication, while "Return to latest" restores the current result. The publication summary exposes discovered, selected, completed, and failed counts before listing paper cards and optional enrichment.
A PARTIAL report remains usable: paper cards with valid source metadata stay
visible, and failed stages and stable error codes explain what is missing.
NO_UPDATE means the pipeline completed normally but found no newly eligible
paper for that topic and date.
The Papers view lists canonical papers and explicit arXiv versions. Opening a paper shows bibliographic metadata, its source link, analysis scope, structured summary, research problem, method, results, limitations, claims, related work, comparisons, evidence, and lineage links when available.
The analysis-scope banner is important. FULL TEXT means the selected arXiv PDF
was parsed by GROBID before analysis. The product never silently turns a failed
full-text operation into abstract-only analysis. Model identity, prompt version,
generation time, and verification state are retained with generated records.
The Evidence viewer connects concise source excerpts to the claims they support or contradict. Each record identifies the exact paper version, section or passage, page or coordinates when available, extraction source, and generation provenance.
Use evidence links to inspect why a claim or relation appears. An excerpt is traceability material, not independent validation of the paper's finding or the model's interpretation.
The Graph connects papers, research problems, methods, tasks, datasets, and benchmarks. Search by name, paper title, or arXiv ID across all published nodes in the active topic, including nodes outside the displayed overview. Results are paginated. Choosing a result opens that node's published neighborhood.
The overview displays at most 60 nodes and 120 relations and admits complete edge endpoints before filling remaining slots. Node-type filters include connected nodes of other types; relation and provenance filters constrain the displayed relations. Selecting a visible node opens its details without resetting the canvas. Use Explore connections to open a new neighborhood, Focus selected to zoom to a node, Fit graph to fit the visible graph, and Return to overview to clear the neighborhood and filters. Labels appear as zoom permits, with full names available in the detail panel and node index.
Visual style preserves provenance. Explicit or deterministically derived edges are distinct from AI-inferred edges, and inference never implies human verification. Follow a node's lineage link to inspect the bounded research path behind it.
Trends compare persisted structured records across equal 7-, 30-, or 90-day windows. The page shows current and preceding counts, absolute change, growth status, paper activity, and top-entity activity.
Top-entity activity uses one row per entity with a shared scale for current and preceding counts. Names occupy at most two lines; hover over a name or focus its link with the keyboard to read the complete label. Each name links to its Graph neighborhood. On narrow screens, counts appear below the label.
Trend values are deterministic calculations, not model-generated statistics. Zero denominators, small samples, and insufficient preceding windows stay explicit instead of becoming unsupported growth claims.
The remaining product routes provide deeper inspection:
- Runs shows pipeline executions, terminal state, duration and accounting,
plus item-level stages and stable failures. It is the first place to inspect
an unexpected
PARTIALorFAILEDresult. - Comparisons presents structured dimensions for a new paper and compatible historical work, along with comparability and evidence links. Missing or incompatible historical analysis remains an availability state.
- Lineages traces a bounded, provenance-aware path through papers and graph entities. It does not claim an exhaustive history of a field.
COMPLETE: usable Daily source metadata was published; optional enrichment may still be unavailable.PARTIAL: usable metadata was published, but one or more selected papers failed core metadata or source-analysis processing.NO_UPDATE: a normal complete publication with no newly eligible papers.FAILED: a run-level configuration, authentication, source, persistence, or publication boundary prevented a usable publication.UNVERIFIEDor AI-inferred provenance: generated content has not been confirmed by a human.INSUFFICIENT_DATAor another unavailable label: the product intentionally declines to fill a missing evidence, comparison, graph, trend, or lineage input.
For the precise rules, see the Failure policy and Product and trust boundaries. Return to the README for local setup, API examples, and pipeline operation.