Recurring checks to run on a cadence (by a higher-reasoning model), not items parked in TODO.md. The point is to surface drift — places where the system has quietly fallen out of line with its axioms or its own accuracy — and open tickets for what's found, rather than narrating it inline forever.

How to use: walk the sections, note findings, file/triage tickets. Most checks have a tool that does the heavy lifting.

Axiom adherence

Against the three anti-failure axioms (defined in README.md; summarized in architecture.md):

• Clarity — does each part still have a stated purpose that matches what it actually does? Unstated or stale purposes are drift; name/restate or subdivide them. (The accuracy checks below feed this.)
• Interchangeability — run agent-tool interop-points; review the leaky list (covered ops called raw → route them) and uncovered list (no native interface → warranted?). Confirm the Model/Provider/AgentRunner seam still substitutes cleanly and no unnecessary entanglement has crept in.
• Accountability — does every contribution carry a claim + evidence (report / verify / specs), and is each judged against its stated purpose? Does reported status reflect the real execution path, not a proxy — provider status must probe the path, not just key presence (ticket b12463b8); spot-check that a deliberately-broken provider shows red. Watch for any new AGENT_*-style constant re-conflating product and mechanism.

Accuracy & hygiene

• Docs vs reality — docs/architecture.md, TODO.md, and section acceptance criteria still match the built system (one binary + interfaces; ext/docs present; symlink deploy). Fix stale statements at the source.
• Completed work — anything in TODO.md that is actually done moves to PROGRESS.md and out of TODO.
• Secret handling — review references to secrets: specs/UI should reference secret uses, not bake specific secret names/values; confirm no secret reaches HTML, JSON, the binary, or logs.
• Ticket health — are open tickets grouped and ordered into prioritized milestones (the gap that keeps TODO alive)? Re-tag/re-order; close stale ones.

Pipeline & UX

• Gate depth — run the full crystal spec and crystal tool format before claiming green/merging. The worker's verify_crystal_change and Accept-and-merge's verify_live do not run integration/cross-file checks, so regressions can slip to trunk.
• Accept / revert UX — can an operator review a task's change, accept it, and revert cleanly to avoid regressions? Is the Agent Console navigation actually pleasant to use? Note UX rough edges as tickets.