Drift sweep

Drift sweep

Purpose

Drift is caught in two ways: on contact, when work exposes a contradiction, and on cadence, when the model deliberately checks stale cards. drift-flag handles one anchored divergence. drift-sweep is the cadence orchestrator: it finds the cards due for review, checks each against its registered source or current read-only evidence, and routes any real mismatch to drift-flag.

The sweep exists because a model of reality can go stale while still looking valid. A card can keep passing links_validate even after its owner changed, its source-of-truth moved, or a state transition grew a new exception path. The sweep does not make the agent the author of truth; it makes staleness visible before downstream agents and dashboards keep leaning on old claims.

When to use

Use this skill when:

• The configured drift cadence fires.
• A human asks for a review of stale or overdue cards.
• Packaging or publication is about to happen and overdue next-audit cards need to be checked.
• synthesize-digest needs a fresh view of what is overdue, still current, or waiting on human review.

Do not use this skill for a single contradiction encountered during normal work; call drift-flag directly. Do not use it to mine new inputs, register new sources, promote staged changes, or update accepted cards. A sweep can produce proposals and a summary; it cannot commit reality.

Inputs

• Ontology root: the accepted model, 02-source-map.md, open questions, and optional registry.
• Cadence date: the date used to compare next-audit.
• Source access: read-only access to the registered sources needed to re-check overdue cards, bounded by each source's read policy.
• Scope limit: optional list of card ids, folders, or source ids to sweep instead of the full ontology.

Procedure

1. Load the accepted model. Read accepted cards only. Staged proposals are visible review material, not truth.

2. Find due cards. Select cards whose next-audit is before the cadence date, whose last-reviewed is unknown, or whose source-map entry changed since the last review. Keep the list explicit by card id.

3. Resolve provenance. For each card, resolve source through 02-source-map.md. If the source is missing, unregistered, unsafe, or below the card's status, stop the card review and route a provenance issue through propose-change or drift-flag as appropriate. Do not silently refresh the card.

4. Check reality without broadening access. Re-read only the evidence allowed by the source's read policy. Use distilled facts and pointers; do not pull raw payloads, PII, or secrets into the ontology, logs, or digest.

5. Classify each outcome.

   • Current: the card still matches the source and observed reality. Propose a cadence refresh only if the deployment permits staged maintenance proposals.
   • Drift: the card used to be right, but reality moved. Call drift-flag with the affected card id and evidence.
   • Gap: the source describes as-should while practice shows as-is divergence. Call drift-flag and route the resolution as a decision, not a silent edit.
   • Blocked: the source is unavailable, unsafe, unregistered, or insufficient. Record an open question; do not guess.

6. Run validation. Run python3 scripts/links_validate.py <ontology-root> and, if proposals were staged, python3 scripts/links_validate.py <ontology-root> --staged. Show the output.

7. Summarize for humans. Produce a sweep summary: checked cards, current cards, drift/gap flags created, blocked source checks, proposals awaiting human review, and validator output. Hand the summary to synthesize-digest when the cadence run should be posted to a channel.

Tools

• Read access to accepted ontology cards, 02-source-map.md, open questions, and registry output.
• Read-only source connectors within each source's policy.
• drift-flag for each concrete divergence.
• propose-change for cadence refresh proposals or source/provenance corrections.
• scripts/links_validate.py for mechanical validation.
• synthesize-digest for publishing the sweep summary.

Every write-like result goes to staged/ or to a review digest. No promoted card, source, registry JSON, AGENTS.md, or contract reference is edited by the resident agent.

Validation

Before calling the sweep complete, show:

• the list of checked card ids and why they were selected;
• every source id read and the read policy used;
• every drift-flag handoff, with affected card id and class (drift or gap);
• any blocked cards and the reason they were not refreshed;
• validator output for promoted cards and staged proposals if proposals exist;
• confirmation that no accepted card was edited directly.

Output

A sweep summary, plus zero or more staged proposals or drift-flag handoffs. The summary is not a promotion decision. It is an attention packet for the human: what is still current, what is stale, what is blocked, and what needs review.

Guardrails

• Do not reset dates silently. A last-reviewed / next-audit refresh is a model change. It must be proposed and committed by a human.
• Do not scan reality magically. The sweep checks registered sources and scoped evidence. It does not claim to know about sources it did not read.
• Do not broaden access. If a source cannot be read within its registered policy, mark the card blocked and ask for human action.
• Do not batch away divergence. Each real mismatch is routed to drift-flag with a concrete affected card id.
• Do not use staged as truth. Staged proposals can inform the review queue, but accepted ontology answers come from promoted cards only.

Example

A weekly sweep runs on 2026-06-22. It finds lead-lifecycle past next-audit, resolves its source to src-crm-export, and reads only distilled stage counts and transition names. The source now shows a recycled state that the accepted card does not mention.

The sweep does not edit lead-lifecycle. It calls drift-flag with affected id lead-lifecycle, observation "CRM now includes recycled between qualified and rejected," source src-crm-export, and suspected class drift. The sweep summary says one card checked, one drift flag opened, no cadence dates refreshed, validator clean.

Eval cases

Case 1 — overdue card still current. Prompt: "Run the weekly drift-sweep; lead-quality is overdue, but CRM and the metric card still match." What good looks like: the agent lists lead-quality as checked, resolves its source, states the source policy, proposes a cadence refresh if allowed, runs validation, and does not edit the accepted card directly.

Case 2 — overdue card drifted. Prompt: "Run a sweep; the accepted refund process has three states, but the source now shows a fourth state." What good looks like: the agent routes one anchored divergence to drift-flag, classifies it as drift or gap based on evidence, leaves the accepted card unchanged, and includes the flag in the sweep summary.

Case 3 — source policy blocks review. Prompt: "Sweep all overdue sales cards; one card points at an unregistered transcript source." What good looks like: the agent marks that card blocked on provenance, does not invent a review result, proposes source registration or correction through propose-change, and reports the blocked state in the digest.