Biblicus

Make your documents usable by your assistant, then decide later how you will search and retrieve them.

Steering proposals

Steering proposals separate machine-derived signals from reviewable recommendations. Biblicus owns the artifact contract; the application owns human decisions.

Layers

Graph signals

Generate topic-informed graph signals:

biblicus steering graph-signals \
  --corpus corpora/AI-ML-research \
  --classifier ai-ml-research \
  --graph-snapshot simple-entities:<snapshot_id> \
  --steering-feedback papyrus-steering-feedback.json \
  --format json

The command emits a proposal bundle with signals and an empty proposals list. A worker can pass those signals to an agent, a language model, or a human-authored process.

Every accepted taxonomy topic can produce a signal when the graph snapshot has no known topic entity with the deterministic node id topic:<topic_uid>. Classifier topic maps can also produce candidate Item -> Topic membership edge signals. When an accepted taxonomy artifact exists, graph signals use every accepted taxonomy node, including child nodes. When no accepted taxonomy artifact exists, graph signals use the classifier seed manifest as the root-topic source.

When the steering application exports review memory with export_kind: "papyrus-steering-feedback", pass it with --steering-feedback. Biblicus uses the normalized suppressions list to avoid re-emitting rejected graph signals under the same classifier and root topic. This is negative steering only; accepted taxonomy and ontology manifests remain the positive source of truth.

Suppression input shape:

{
  "schema_version": 1,
  "export_kind": "papyrus-steering-feedback",
  "generated_at": "2026-05-16T00:00:00+00:00",
  "source": {
    "system": "papyrus",
    "topic_set_id": "curation-topic-set-example",
    "corpus_id": "curation-corpus-example",
    "classifier_id": "example-classifier"
  },
  "topic_set": {
    "topic_set_id": "curation-topic-set-example",
    "corpus_id": "curation-corpus-example",
    "classifier_id": "example-classifier",
    "display_name": "Example Topics",
    "description": "Reviewed topic set."
  },
  "decisions": [],
  "accepted_proposals": [],
  "rejected_proposals": [],
  "suppressions": [
    {
      "suppression_id": "suppression-create-taxonomy-node-example",
      "proposal_id": "create-taxonomy-node:example",
      "proposal_kind": "create-taxonomy-node",
      "steering_domain": "topic",
      "reason": "Rejected during editor review.",
      "decided_at": "2026-05-16T00:00:00+00:00",
      "decided_by": "editor@example.com",
      "scope": {
        "topic_set_id": "curation-topic-set-example",
        "corpus_id": "curation-corpus-example",
        "classifier_id": "example-classifier",
        "root_topic_uid": "root-topic"
      },
      "match": {
        "topic_uid": "root-topic-child",
        "display_name": "Rejected child label",
        "normalized_display_name": "rejected child label",
        "relationship_type": null,
        "graph_entity_id": null
      },
      "evidence_item_ids": []
    }
  ]
}

Taxonomy and ontology proposal kinds

Unified steering proposal bundles can now carry topic, graph, taxonomy, and ontology recommendations. Supported taxonomy and ontology proposal kinds are:

Accepted human decisions still live outside Biblicus. Agents can create proposal bundles with any method, then validate or record them through the steering proposal commands.

Sample export

A compact steering export with taxonomy and ontology proposals is available at docs/examples/steering-export-taxonomy-ontology.json.

The sample shows:

Accepted taxonomy and ontology manifests are not embedded as top-level steering export fields. The export carries artifact references to those accepted overlays. Workers can load the referenced Biblicus artifacts when they need the full accepted taxonomy or ontology assertion manifest.

Payload field guide

All proposals use the same wrapper fields:

recommend, do_not_recommend, and needs_clarification are not human review actions. They are proposal recommendation labels authored before human review. External applications should keep their own human actions, such as accept, reject, edit, or defer, in application-owned decision records.

create-taxonomy-node

Use when discovery suggests a new accepted taxonomy node.

Expected payload fields:

move-taxonomy-node

Use when a taxonomy node appears to belong under a different parent.

Expected payload fields:

merge-taxonomy-nodes

Use when two or more taxonomy nodes appear semantically redundant.

Expected payload fields:

split-taxonomy-node

Use when one taxonomy node repeatedly contains multiple stable child concepts.

Expected payload fields:

archive-taxonomy-node

Use when a taxonomy node no longer has active analytical value.

Expected payload fields:

add-ontology-relationship

Use when an accepted typed relationship assertion should be added.

Expected payload fields:

When accepted, these fields can be rendered directly into an ontology assertion manifest.

add-relationship-type

Use when a new reusable relationship predicate is needed.

Expected payload fields:

Deprecated or suppressing graph structure

Use suppress-entity-or-edge for graph entities or edges that should be hidden, ignored, or excluded from a future graph view or materialization step. This remains a proposal only; Biblicus does not remove graph data automatically.

Proposal bundles

Validate an externally authored proposal bundle:

biblicus steering proposals validate \
  --input proposal-bundle.json

Record a validated bundle as a reproducible Biblicus artifact:

biblicus steering proposals record \
  --corpus corpora/AI-ML-research \
  --input proposal-bundle.json

Recorded bundles are written under:

analysis/steering-proposals/<snapshot_id>/

The directory contains:

Proposal recommendations

Proposals must use one of:

Negative recommendations are first-class. Applications should show them because they explain what the agent considered and rejected.

Human decisions

The external steering application imports proposal artifacts and records human decisions in its own schema. Biblicus proposal artifacts must not contain human decision fields or accepted/rejected review state.