Skip to content

Graph explanation artifacts ​

Persistent knowledge workspace ​

Open View > Exploration > Knowledge on a map to inspect persistent sources, concepts, and evidence-bearing relationships. Build from sources processes the corpus in resumable pages; Import selected points handles the current selection, up to 100 points. Subsequent imports update activated workspaces through the existing background worker. Connect generated graph promotes the latest GraphFusion output into individually addressable proposed assertions.

Selecting an object shows its relationships, exact evidence quotations, producer, and status. Select source on map selects the original loaded point. Objects outside loaded geometry remain inspectable as knowledge, but their positions are never fabricated. Filters and Show more help navigate large results.

Source-backed keyword mentions are supported only when an exact quotation exists. Generated relationships remain proposed; a model confidence score is not a proof. Quotation matching validates provenance, not entailment. Private spaces retain their existing access policy; viewers can inspect but cannot import or edit knowledge.

The panel reports full object counts and explicitly discloses bounded views. Stop cancels the current request; completed import pages remain saved. Refresh before retrying a timeout or inspecting partial work. Graph revisions protect concurrent edits from overwriting each other.

The workspace requires the linked MantisAPI native-knowledge migrations and Celery worker update. If that API is unavailable, the panel surfaces the server error and keeps the existing graph visible.

Switch between Objects, Semantic, Relationships, and Evidence inside Knowledge. Object filters and selection stay shared across these views. The semantic view uses original document coordinates and marks evidence-centroid concepts with hollow circles. Concepts without source geometry stay in the list. The directed relationship layout uses arrows and dashed proposals; its distances have no semantic meaning. The evidence ledger exposes grounded quotations and ungrounded assertions together, with clear status labels. Dense relationship views preserve the selected neighborhood and disclose omitted objects.

Temporary Composer explanations ​

Composer can attach a small, interactive subgraph to a response by writing a JSON file with the .mgraph extension and citing it as an agent file:

text
@cite/agent_files/user@example.com/ran-hcc1937.mgraph

Selecting the citation hydrates any missing points, focuses their map, reveals the edges in rank order on the default WebGL map, and attaches the explanation to the focal point's flashcard. Selecting it again or opening another citation clears the temporary overlay. It does not create or modify persisted map links.

Version 1 ​

json
{
  "schema": "mantis.graphical-explanation/v1",
  "map_id": "map-uuid",
  "title": "RAN essentiality in HCC1937",
  "focal_point_id": "ran-idea-uuid",
  "links": [
    {
      "id": "ran-nemp2-attribution",
      "subject_idea_id": "ran-idea-uuid",
      "object_idea_id": "nemp2-idea-uuid",
      "predicate_name": "protein_protein_interaction",
      "description": "Top integrated-gradient attribution",
      "properties": {
        "integrated_gradient": 0.42,
        "pruning_delta": -0.18
      },
      "produced_by": "HIDRA",
      "display": {
        "rank": 1,
        "weight": 0.42
      }
    }
  ],
  "card": {
    "title": "Why RAN is essential in HCC1937",
    "summary": "The model relies on a nuclear-transport neighborhood.",
    "metrics": {
      "predicted_fitness": -3.21,
      "observed_fitness": -4.56
    },
    "caveats": [
      "This is a model attribution, not a causal mechanism."
    ]
  },
  "provenance": {
    "checkpoint": "abl_ESS_only_hero_v2/i1-s1-l1-a1-m1"
  }
}

Each entry in links uses the ordinary Mantis Link fields. The optional display object is artifact-only: rank controls reveal order and weight controls relative line width. point_ids may include unlinked contextual points. The client accepts at most 100 links, 200 total points, and a 2 MiB file.

The same projected SVG layer renders ordinary persisted Links when they are loaded for a hovered or selected point, bounded to 1,000 visible links. Dense points, picking, and LOD remain inside the custom WebGL renderer; sparse links and cards use its public world-to-screen projection.

Search current knowledge ​

Open Semantic and relationship search in the native Knowledge panel. Choose the original embedding model and field, enter an intent, or use a selected source as the semantic anchor. Source matches show cosine similarity in that model space. A typed relationship filter follows only recorded assertions, with supported-only filtering enabled by default.

Each result path preserves its source quotations, producer, and review status. Similarity alone creates no relationship and does not verify a claim. Results disclose invalid vectors, sources missing native graph objects, path limits, and changes since the search revision. Source selection uses the loaded map region; an unavailable point produces an explicit notice.

Editors can name and save a dynamic query, then rerun it as imports and relationships change. Viewers may run shared saved queries. Query edits and deletions require the current workspace revision and enter graph history.

Keep Composer reasoning in native knowledge ​

After opening a cited .mgraph explanation, choose Save to native knowledge. Mantis retains its descriptions as proposed claims and its graphical relationships as proposed native assertions in that exact map and Space. Saving requires edit access and a current workspace revision; the original overlay remains available for inspection. A generated explanation is not treated as verified evidence.

Composer can retain structured answer graphs when asked, with individual claim text, exact source citations, and acyclic premise dependencies. Select an answer or claim in Knowledge to inspect its retained reasoning. Editors can record a claim review with an explicit explanation. Supporting or refuting a claim requires current citations; supporting a conclusion also requires reviewed premises. Changed premises and corrected sources flag dependent conclusions for review.

Contextual source embeddings ​

Open Contextual embeddings in Knowledge to train topic, method, mechanism, or outcome retrieval contexts. Select an original model and field, state the intent, and provide distinct relevant/irrelevant training examples. Keep validation examples in the separate held-out partitions. Mantis learns a retrieval direction without changing document vectors or map geometry.

Each feedback publication creates a new immutable version. Inspect contextual and intent cosine scores, exact source excerpts, and stored held-out relevance AUC. Missing validation classes make AUC unavailable. Held-out examples are marked and cannot enter training through result feedback. Source and embedding changes are disclosed; historical metrics describe their original version. Viewers can inspect versions and retrieve sources, while editors can train and publish feedback versions.

Discover testable connections ​

Select a native knowledge object and open Discover connections. Choose the relationship to investigate and, optionally, a contextual embedding version. Cards show recorded graph paths, source quotations, and review states; contextual scores express relevance. Mantis excludes relationships already recorded and discloses incomplete graph/candidate coverage. A missing edge in this workspace does not establish a gap in published research.

Editors can refine a falsifiable test plan and retain the lead as a proposed hypothesis and relationship. The proposal stays unverified until its evidence is tested and reviewed. Viewers can inspect leads and original sources without changing native knowledge.

Investigate a selected subgraph ​

Select an object and open Subgraph investigations. Editors can capture a question and a bounded neighborhood as an immutable investigation basis. Its source text, assertions, and objects remain available through separate pages, even if the live sources later change. Coverage limits are disclosed.

Prepare in Composer stages the investigation instructions in the Composer for this Space and map. Review the draft and press Send. Composer reads the retained basis and submits a versioned proposal; it cannot accept its own proposal. Preview the complete before/after patch, provide a review explanation, and accept or reject it. Changed source evidence or a newer graph revision blocks stale acceptance. Superseded proposals remain inspectable.

Accepting a proposal applies its graph changes while keeping inferred claims proposed. Reviewing the evidence for a claim is a separate action. Viewers may inspect investigation evidence and proposals without applying changes.

Follow knowledge through time ​

Open Knowledge history and impact to browse attributed graph changes and inspect complete retained patches. Select an object and compare two revisions to reconstruct its state and incident assertions, including original quotations and review states. This comparison covers that selected slice, not the entire historical map. Budget errors require a narrower or more recent comparison; Mantis never substitutes incomplete history for an exact result.

Dependency impact traces recorded source citations and premise dependencies to conclusions that may need review. Coverage limits are explicit, and the paths do not establish causal effects. History inspection is read-only and does not restore old claims or overwrite current evidence.

Explore a permission-aware knowledge atlas ​

Open Cross-Space knowledge atlas and explicitly select up to eight existing knowledge workspaces. The catalog includes your current workspace and Spaces you belong to. Every contribution is checked for current read access and map containment; hidden members, labels, quotations, and review records are omitted. Inspect an entity to read its own workspace's source neighborhood.

Matching labels are identity candidates, not automatic equivalence. Select entities, explain their equivalence, and choose Link selected identities. The action requires edit access and current revisions for every admitted workspace, keeps all original objects and edges, and changes only the selected identities. Recent fully visible review records offer Reverse identity review with an explicit explanation. Later conflicting identity decisions must be resolved before reversal. Atlas coverage and recent-history limits are disclosed.

Test hypotheses with reproducible experiments ​

Select a hypothesis and open Hypothesis experiments. Retain the question, protocol, controls, original source inputs, metric/unit, and explicit target criterion. Review the stated prior, sensitivity, specificity, and cost assumptions. Next-test rankings calculate expected binary information gain and compare information per cost only within the same cost unit; they do not predict guaranteed scientific discovery.

Create and start an attributed run when carrying out the real test. Mantis records the workflow and executes no external experiment. Record the actual measured value, adherence, observed controls, and an exact result source quotation. Completed/failed runs are terminal; corrections belong in a new run. Original inputs and result evidence remain inspectable with their hashes and attribution.

Review scientific meaning separately with an explicit explanation. Meeting the retained target does not automatically support the hypothesis. Corrected inputs/citations block stale certification, and new results require renewed review of hypotheses and dependent conclusions. Viewers can inspect protocols, runs, evidence, and planning assumptions.

Semantic Triggers can turn a future explicit human refutation into one bounded investigation review draft with exact source receipts, cooldown and a private no-action dry-run history.