SCIENTIA A2A evidence-gathering tasks
Orchestrator / mesh A2A can delegate read-heavy, idempotent jobs that return structured JSON for metadata_json.scientia_evidence or publication_status_events. This document names task kinds for operators and agent authors; routing uses existing RemoteTaskEnvelope types in vox-orchestrator (a2a / envelope modules).
Allowed task families
| Task kind (logical) | Goal | Must not |
|---|---|---|
scientia.gather.benchmark_lineage | Collect baseline/candidate run ids and report paths | Invent benchmark outcomes |
scientia.gather.repo_docs | List ADR/research paths and linked corpus | Summarize novelty |
scientia.gather.repro_artifacts | Find checksum / manifest paths | Claim reproducibility passed |
scientia.gather.venue_requirements | Fetch venue checklist text (cached) | Assert submission eligibility |
scientia.gather.credential_presence | Clavis/env presence bits only | Expose secret values |
Envelope rules
- Payload is JSON with
task_kind,publication_id,repository_id(when known), andidempotency_key. - Result merges into
scientia_evidenceor appends a status event withdetail_jsonpointing at file paths and digests. - Refusal: if grounding artifacts are missing, return
blocked_reasons— never backfill with LLM prose. - Human loop: meaningful advance, novelty, and final abstract remain human-attested per how-to: Scientia publication.
Related
- Discovery ranking:
vox_scientia_publication_discovery_scan/vox scientia publication-discovery-scan - LLM assist (bounded):
vox_scientia_assist_suggestions(use_llm=falsefor heuristic-only)