drift_status - Linksee Memory

Reads reality against what you decided and returns every active anchor in one of five states. The colour is a claim; the reality field is the evidence for it.

State	Meaning
🔴 `drift`	The detector has hard evidence against this anchor and nobody has answered it — an open `contradicts` edge, an expired acknowledgement, or an `at_risk` lifecycle.
🟡 `review`	A soft signal awaiting a human: a pending candidate, or “declared but not found in reality” (`absent`).
⚪ `held`	Acknowledged and time-boxed. Reopens as 🔴 when the review date passes.
🔵 `aligned`	Verified. A recorded resolution, or the detector observed reality matching (`implements`).
⚫ `unverified`	Declared, never checked. No edges, no resolution. Not a problem — but not “fine” either. Give it `affects` / `violation_signal` so the detector can look, or leave it as a note.

unverified exists because on a real machine 35 of 45 anchors had never been checked and were all painted 🔵. Different colour, different meaning: the detector has no eyes there.

Parameters

Name	Type	Description
`anchor_id`	number	Deep-dive into one decision instead of the map: state, premises, drift edges, pending candidates. (Absorbs the old `check_decision`.)
`domain`	string	Filter by domain (`strategy`, `product`, `engineering`, `growth`, …).
`decision_mode`	string	Filter: `hypothesis` · `constraint` · `commitment` · `source_of_truth`.
`verbose`	boolean	Return full nodes and all candidates. Default is compact: attention items in full, aligned/unverified one line each per domain, candidates as counts.

Response (compact)

{
  "ok": true,
  "triage": "46 anchors: 🔴 2 drifting · 🟡 1 needs review · 🔵 8 verified · ⚫ 35 unverified",
  "attention": [
    {
      "id": 2,
      "state": "drift",
      "statement": "Never rebuild the memories table; add fields to content JSON instead.",
      "reality": "1 open contradiction — db/migrate.ts hit \"alter table memories drop\" (2026-06-12). resolve_drift: fix if reality is wrong, dismiss if false positive."
    }
  ],
  "aligned": [{ "domain": "engineering", "count": 3, "nodes": [{ "id": 58, "statement": "…", "reality": "Accounted for by recorded resolution" }] }],
  "unverified": [{ "domain": "strategy", "count": 12, "nodes": [{ "id": 27, "statement": "…" }] }],
  "candidates": { "auto": 10, "suppressed": 3 },
  "hint": "verbose:true for full nodes; drift_status({ anchor_id }) for one node."
}

The full form was ~15k tokens for 43 anchors and buried the two that mattered under 41 that did not; the compact form is the default for that reason.

One decision in depth

{ "anchor_id": 2 }

Returns the same node plus affects, detect_terms, violation_signal, driftEdges (each with verdict, confidence, status) and pendingCandidates. drift_status({ anchor_id }) and the map agree by construction — they read the same state machine.

When to call

At session start, if recall()’s brief shows 🔴 or 🟡 items
Before making a decision — is there already an anchor on this topic?
After finishing work — did the state change?
When the user asks what’s broken, stale, or what to revisit