Evidence and grounding workflows¶
Knowledge work begins with contextual evidence and ends with a reviewable state, not a single confidence score. Sources, claims, contradictions, resolutions, and biological mappings remain separately inspectable.
Ingest evidence¶
For each observation or source:
- assign a stable evidence identifier and evidence kind;
- record source type, locator, origin, extraction method, and curator;
- capture biological and experimental context;
- attach quantitative support and proteomics artifact flags;
- record confidence, strength, observation time, expiry, and derivation;
- validate the typed record before adding it to a versioned bundle.
flowchart LR
source["publication · database · run · assay"]
normalize["contextual normalization"]
record["EvidenceRecord"]
bundle["versioned EvidenceBundle"]
evidence_graph["evidence graph"]
source --> normalize --> record --> bundle --> evidence_graph
Retain imported, observed, inferred, and synthetic origins. Do not convert missing context into a default that appears observed.
Build and audit claims¶
Create claims as references to evidence records, not copies of their prose. Record supporting and contradicting evidence separately along with assumptions, resolution assays, polarity, evidence state, confidence, and decision impact.
Validate that evidence identifiers resolve, target context agrees, and decision lineage reaches the source records. Use evidence-graph integrity checks to find orphans, duplicate identifiers, invalid edges, and cycles.
Reconcile conflict¶
Conflict resolution starts only after disagreement is explicit:
- classify the contradiction and identify affected claims;
- compare source quality, context, freshness, quantitative support, and independence;
- apply a declared resolution policy when the basis is sufficient;
- emit a new resolution record that cites all inputs;
- place unresolved cases in the escalation queue.
Never edit the losing evidence into agreement. Historical evidence remains available for replay, calibration, and later reinterpretation.
Ground biological entities¶
Use the owned resolution reports for protein IDs, features, pathways, complexes, kinase-substrate edges, drugs, disease terms, and orthologs. Review every ambiguous and unresolved row before using the mapped output downstream.
Coverage is a mapping result. It does not establish pathway activity, complex assembly, kinase causality, therapeutic efficacy, disease mechanism, or cross-species equivalence. Those conclusions require analytical evidence and a separate interpretation contract.
Prepare a decision brief¶
A knowledge decision brief should include:
- the bounded decision question and target;
- cited supporting and contradicting evidence;
- context and freshness posture;
- unresolved identity or biological mappings;
- claim status, assumptions, and important trust gaps;
- conflict resolutions and outstanding escalation items;
- questions that intelligence or laboratory work must answer next.
Package canonical JSON with reviewer-facing TSV or narrative views. Flat views must retain identifiers that resolve to the typed record whenever they omit nested context.
Evidence handoff contract¶
Every downstream handoff preserves the evidence state rather than exporting only a conclusion:
| Handoff field | Required meaning |
|---|---|
| bundle identity and schema | exact version of the evidence contract |
| source and evidence identifiers | resolvable path back to origin and extraction |
| context | species, tissue, assay, perturbation, population, and time |
| claim polarity | support, contradiction, qualification, or unresolved relevance |
| derivation | observed, imported, inferred, or synthetic origin |
| freshness | observation time, expiry, and current-use assessment |
| reconciliation | duplicate, superseded, context-specific, or unresolved relationship |
| decision impact | why the record may strengthen, weaken, or block an action |
sequenceDiagram
participant R as Runtime or assay record
participant K as Knowledge bundle
participant I as Intelligence review
participant L as Lab observation
R->>K: artifact identity + scientific context
K->>K: resolve, reconcile, and audit claims
K->>I: versioned support + contradiction state
I->>L: advisory question + evidence references
L-->>K: new observation + QC + deviations
K->>K: append evidence; preserve prior state
Runtime output can be ingested as evidence, but execution success does not establish claim truth. Intelligence can consume a Knowledge bundle, but its ranking policy cannot rewrite evidence history. Lab observations return as new records rather than edits to the recommendation that requested them.
Maintain evidence over time¶
Expiry marks a record stale for current decisions without removing it from history. New observations create new records and can produce a new claim or resolution state. Governed bundles link evidence to runtime outputs, scientific summaries, and review artifacts by schema and digest.
A workflow is complete when source context is recoverable, contradictions are visible, graph integrity passes, ambiguous mappings are retained, freshness is assessed, and downstream consumers receive a versioned artifact rather than an unattributed summary.
Stop conditions¶
Do not produce a decision-ready bundle when:
- a cited evidence identifier does not resolve;
- source context is missing for a claim whose meaning depends on species, tissue, assay, population, or perturbation;
- an inferred or synthetic record is presented as observed;
- an important contradiction is hidden by aggregation;
- entity resolution remains ambiguous but the exported view selects one match;
- an expired record is used without a current-use assessment;
- the evidence graph contains orphaned claims, invalid edges, or cycles;
- a downstream summary cannot be traced to the versioned bundle.
Return an explicit refusal or unresolved state with the failed invariant and next evidence requirement. An incomplete bundle that looks decisive is more dangerous than a visible stop.
See evidence and grounding contracts and knowledge artifact contracts.