Architectural truth
How Hyperoru represents what exists, why each relationship exists, confidence, contradictions, and historical change.
Canonical model
Architectural truth is Hyperoru’s evidence-backed belief about a system at a point in time. It is deliberately separate from diagrams, scanner output, model transcripts, and third-party graph schemas.
Core records
Entity
A stable architectural thing: service, API, datastore, queue, identity, cloud resource, deployment unit, pipeline, system, or trust boundary. Entities preserve source-specific identifiers as aliases rather than making any collector’s identifier canonical.
Relationship
A typed edge such as CALLS, READS_FROM, DEPLOYS, OWNS, CONTAINS, or CAN_REACH. Every relationship stores direction, validity window, confidence, contradictions, supersession state, and supporting evidence IDs.
Evidence
An immutable observation that explains why an entity or relationship is believed: file and line, dependency manifest, IaC resource, CI job, cloud asset, API definition, or runtime observation. Evidence records its source type, collection method, source hash, timestamp, and bounded preview.
Confidence is explainable
Confidence is not a model’s feeling. It is derived from source reliability, agreement across independent evidence, freshness, extraction certainty, and unresolved contradictions. The score and every contributing evidence ID remain visible.
Example interpretation:
| Confidence | Meaning | UI behavior |
|---|---|---|
| 0.90–1.00 | independently corroborated or directly declared and observed | solid edge |
| 0.70–0.89 | strong single source or partially corroborated | solid edge with source badge |
| 0.40–0.69 | inferred, stale, or contradicted | dashed edge and review state |
| below 0.40 | hypothesis, not canonical truth | hidden by default; analyst review |
Reconciliation rules
- Normalize source-specific facts before comparison.
- Prefer exact identifiers, then deterministic aliases, then bounded heuristics.
- Preserve both observations when sources disagree.
- Never erase the former relationship; mark its validity window and superseding record.
- Never promote an LLM-only observation to a vulnerability or high-confidence edge.
- Render only from the canonical model, not directly from agent prose.
Temporal behavior
Each audit creates a new observation horizon. Unobserved does not immediately mean deleted: source coverage and freshness determine whether an entity is absent, unknown, or superseded. This lets Hyperoru answer both “what exists now?” and “what changed?” without rewriting history.