Skip to content

Governance

Source

Mirrors the repo-root GOVERNANCE.md.


title: "EpiHack Arizona 2026 — Governance"

Governance

This document defines how the EpiHack Arizona 2026 stack is governed post-event. It encodes the cross-cutting governance track from plan/05-roadmap.md and is binding on maintainers, agency-affiliated committers, and contributing agents (human or LLM).

The intent is not bureaucracy. It is a small, durable structure that keeps the four founding agencies in the loop on anything that could change how surveillance data is collected, suppressed, or routed — and that gives tribal partners a hard veto on anything touching tribal data.

Standing review board

The standing review board has four seats. Membership is by organization, not by individual, so transitions inside an agency do not require a repo change.

Seat Organization Scope of authority
ADHS Arizona Department of Health Services — Vector-Borne & Zoonotic Diseases program and the Heat Preparedness Network Statewide arbovirus + heat-mortality surveillance; reportable-condition definitions; triage-class thresholds
AZGFD Arizona Game and Fish Department — Wildlife Health Program Wildlife observation flows; WHISPers + iNaturalist enrichment behaviour; species-attribution rules
ITCA-TEC Inter Tribal Council of Arizona — Tribal Epidemiology Center All decisions touching tribal data; veto authority below
Maricopa Vector Control Maricopa County Environmental Services Department, Vector Control Division Mosquito / tick pool data flows; cluster-detection threshold calibration for VBD; Phoenix-metro cooling-center integrations

The board operates by lazy consensus on the GitHub PR thread (any member can request changes; absence is consent after a 5-business-day window). Two domains escape lazy consensus and require explicit signoff:

  • Tribal-data changes require ITCA-TEC explicit approval (see Tribal-partner veto below).
  • Cluster-detection threshold changes require explicit approval from at least one of ADHS or Maricopa Vector Control, depending on vertical.

Proposal-then-merge cadence

Changes fall into three buckets with different review depths.

Schema changes — proposal first

Anything in schema/ — node types, edge predicates, property keys, slug renames, view definitions.

  • Open an issue first describing the proposed change, the affected seeds, and the downstream consumers (mcp/knowledge-graph-mcp/, dashboard/, today/, agents/).
  • Wait at least 5 business days for review-board comment.
  • Then open the PR implementing the change, linking the issue.
  • Slug renames must include a deprecation note in the PR description; the old slug stays valid (as an alias edge) for one full release cycle.

Runtime / orchestrator changes — PR first, fast review

Changes in agents/, app/, dashboard/, today/.

  • Open the PR. The board reviews on the standard lazy-consensus cadence.
  • Cluster-detection threshold edits (agents/src/onehealth_agents/cluster.py, plan/CLUSTER-CALIBRATION.md) trigger the explicit-approval rule above.
  • Consent-enforcement edits (agents/src/onehealth_agents/validation.py) trigger ITCA-TEC explicit approval if the change weakens any tribal-data check, removes any cell-suppression rule, or expands what enters agent_run digests.

MCP server changes — PR first, prefix-scoped review

Changes in mcp/<server>/.

  • New MCP server: follow the recipe in mcp/README.md and CONTRIBUTING.md.
  • Existing MCP server: the seat whose agency owns the upstream data source is the de-facto reviewer (ADHS for adhs-mcp, AZGFD for whispers-mcp + inaturalist-mcp wildlife-context use, Maricopa Vector Control for mag-hrn-mcp and vectorsurv pool tools, etc.).
  • knowledge-graph-mcp SQL-escape-hatch edits require approval from any two seats (the escape hatch is the highest-blast-radius surface in the repo).

Tribal-partner veto

ITCA-TEC and any participating tribal nation hold an unconditional veto on every change that touches tribal data, including but not limited to:

  • New MCP servers that proxy tribal data sources (e.g. the optional navajo-ec-mcp proxy described in plan/05-roadmap.md Phase 2).
  • Changes to the Validation Agent's tribal-data suppression rules.
  • New dashboard views, public or private, that aggregate at a geography below the county level if any constituent ZCTA intersects tribal land.
  • New triage classes, cluster-detection thresholds, or notification routes that change how tribal observations are handled.
  • Changes to consent_profile semantics in schema/deep/application.sql.
  • Changes to what tribal-jurisdiction observations contribute to agent_run digests or to any public-facing artifact.

A veto is recorded on the PR. There is no override mechanism.

Tribal-partner opt-in posture

The default for every new tribal partner is no data flows in either direction until an MOU is signed and a corresponding consent_profile row is added with default = opt_in. There is no opt-out posture and no default that shares tribal data; tribes opt in explicitly per data source.

Concretely:

  • New tribe.* nodes are added to schema/deep/tribes.sql for reference but ship with consent_profile.default = suppress until an MOU attaches an opt-in.
  • MCP servers that proxy tribal data sources do not deploy until the MOU is recorded in the repo and the sunset clause below is encoded in the server's pyproject.toml description and README.
  • Public dashboards apply tribal-suppression before aggregation, not after, so cell-count thresholds cannot be reverse-engineered from neighbouring cells.

Conflict-of-interest disclosure

Committers affiliated with one of the standing-board agencies (ADHS, AZGFD, ITCA-TEC, Maricopa Vector Control) or any state, county, or tribal health authority must disclose that affiliation:

  • In their GitHub profile (the Company field is sufficient), or
  • In the PR description of any PR they author or review when the PR materially affects their agency's data flow or threshold decisions, or
  • In a comment on this file if a structural conflict exists (e.g. serving simultaneously as a committer and a vendor whose product this repo integrates with).

A disclosed conflict does not disqualify the contributor; it informs reviewers, who may ask for an independent second review. An undisclosed conflict that is later discovered is grounds for revoking commit access pending board review.

Sunset clauses

Any MCP server that acts as a proxy for a tribal data source operates only as long as the underlying MOU is live. Specifically:

  • The MOU expiry date is captured in the server's pyproject.toml description field and in the first line of the server's README.md.
  • On MOU expiry, the server's __main__.py entry point refuses to start unless MOU_RENEWED_THROUGH=<ISO-date> is set in the environment to a date later than today. This is a hard runtime gate, not just a comment.
  • Renewal is recorded as a PR updating both files plus a corresponding consent_profile row in the kg.
  • The board reviews proxy-server status quarterly; any proxy whose MOU has lapsed for more than 90 days is removed from the repo (not just disabled) at the next maintenance cycle.

The same posture applies, less formally, to MCP servers proxying non-tribal partner data (ADHS, MAG, 211 Arizona, USGS WHISPers, iNaturalist): if the data-sharing agreement lapses, the server is disabled in deployment configs and the README header is updated.

Amending this document

This file is amended via PR, with the same lazy-consensus rules as runtime changes, plus: amendments to the Standing review board section, the Tribal-partner veto section, or the Sunset clauses section require explicit approval from all four seats. Other sections may be amended by lazy consensus.