Evidence quality¶
Knowledge quality is the preservation of source identity, context, independent lineage, contradiction, uncertainty, and history across ingestion, grounding, reconciliation, and review. A tidy conclusion is not a quality signal if the path to it cannot be inspected.
flowchart LR
S["source quality"] --> L["lineage integrity"]
L --> G["grounding accuracy"]
G --> C["contradiction retention"]
C --> F["freshness and coverage"]
F --> U["intended-use sufficiency"]
U --> P{"review posture"}
P -->|adequate| B["bounded review bundle"]
P -->|inadequate| D["deficit or refusal"]Quality dimensions¶
| Dimension | Evidence | Blocking failure |
|---|---|---|
| source custody | source/version/license/retrieval fixtures | untraceable or impermissible evidence |
| normalization | round-trip and provenance tests | normalized record cannot reach its source |
| graph integrity | node, edge, orphan, cycle, and lineage checks | claims reference missing or incompatible evidence |
| grounding | known, ambiguous, unresolved, species, alias, and coordinate cases | ambiguous identity presented as unique |
| reconciliation | duplicate, conflicting, contextual, and hold cases | contradiction disappears without a recorded action |
| independence | derivation and duplicate-source analysis | copied assertions counted as separate support |
| freshness and coverage | dated source and deficit reports | stale or sparse evidence presented as comprehensive |
| review fidelity | bundle-to-memory comparison | review omits material conflict or uncertainty |
Evidence is fit for an intended use¶
Quality does not reduce to one universal score. Descriptive context, scientific interpretation, candidate prioritization, and experimental action have different evidence burdens. The review artifact identifies the intended use and reports whether its coverage, directness, contradiction, and freshness meet that burden.
An unresolved result is high-quality when the available evidence genuinely cannot resolve the question and the deficit is precise. A confident scalar answer can be low-quality when its source and context are missing.
Proof by change type¶
| Change | Minimum proof |
|---|---|
| source connector | provenance, malformed input, license metadata, duplicate handling |
| identifier resolver | known, ambiguous, unresolved, alias, organism, and source-version cases |
| biological relationship | positive, negative, contextual, coverage, and stale-source cases |
| reconciliation policy | every action, competing contexts, hold, history, before-and-after bundle |
| evidence or claim model | schema compatibility, graph integrity, serialization, consumer impact |
| review workflow | fixed memory revision, complete support and contradiction, deterministic assembly |
Test strategy and change validation map these proof obligations to package and repository checks.
Invariants¶
- evidence history is append-only in meaning;
- normalized records retain original provenance;
- claims do not outlive or lose their evidence edges;
- ambiguity, contradiction, staleness, and missing coverage remain explicit;
- evidence duplication is not mistaken for independent corroboration;
- review bundles name their memory revision and assembly policy;
- downstream recommendations cannot mutate Knowledge truth.
The complete contract is in invariants.
Active limitations and risk¶
External sources may be incomplete, stale, biased, unavailable, or differently licensed. Identifier and biological relationship coverage varies by species and context. Automated reconciliation cannot settle every conflict. These limits belong in known limitations and unresolved ownership or review risks belong in the risk register, not in hidden fallback behavior.
Review route¶
Use dependency governance for reference and parsing dependencies, documentation standards for public evidence language, and review checklist before handoff. Definition of done requires both passing evidence and an explicit record of remaining gaps.