Skip to content

SWARMS_AND_EVIDENCE.md

Latest commit

 

History

History
62 lines (46 loc) · 2.34 KB

SWARMS_AND_EVIDENCE.md

File metadata and controls

62 lines (46 loc) · 2.34 KB
Raw

Swarms And Evidence Contracts

Scientific swarms are optional public-safe planning lanes. They break evidence collection, assumption review, factor-universe construction, observability, and handoff writing into bounded work items while keeping the Python package local and deterministic.

Evidence Table Contract

Evidence workers write rows that match:

  • templates/evidence-table.template.csv
  • templates/evidence-executor-agent-brief.md

The compact public header is:

evidence_id,source_type,source_ref,source_url,claim,confidence,review_status,decision_impact,reuse_notes,transformation_notes

Each row should contain one paraphrased decision-relevant claim, a stable source reference, source type, confidence, review status, and transformation notes. Public or reference evidence must not be represented as target-specific experimental proof.

Issue Packs

Public issue packs live under packs/issue-packs/. They use pack-local ids, relative paths, and validation commands that run with stdlib Python. They do not carry private tracker ids, provider launch fields, or private artifact paths.

Current public packs:

  • adaptive-wave2-assay-power-v0
  • campaign-arms-v1
  • doe-parity-v0
  • doe-parity-v1
  • evidence-executor-v0
  • fermentation-readiness-v0
  • scientific-swarm-v0

The evidence executor pack is for source scoping, public literature or protocol extraction, and evidence-table handoff. The scientific swarm pack is for using those evidence rows to create planning artifacts such as factor universe, assumption attack, observability, control strategy, and adjudication notes. The DoE parity packs cover reference-software import / export and the high-ROI parity upgrades. The fermentation readiness pack covers the readiness scoring work that gates campaign launch. The campaign-arms pack covers first-class multi-arm support for coupled campaigns.

Activation Pattern

  • Keep contract/setup work serial.
  • Run independent evidence lanes in parallel only when touched areas are disjoint.
  • Merge evidence rows serially before they influence factor-space or tournament decisions.
  • Keep final dossier and claim-boundary review serial.

Non-Claims

Evidence and swarm artifacts are planning inputs. They do not validate physical execution, process performance, assay power, scale transfer, production readiness, or regulatory compliance.