randlee’s plan-anchor tackles a familiar agent failure: an implementation constraint survives the conversation as a vague summary, then disappears from the work. It stores the user’s exact directive and checks both plans and delivery reports against it.

I inspected the implementation and question sets and ran its 73 offline tests successfully. Those use fixtures and stubbed transport, not live Jev. The README explicitly says the author has not yet made a live Jev run; this is a tested harness, not evidence that it prevents drift.

Preserve the constraint, not just its summary

  • Confirmed directives become numbered anchors with original wording, provenance, scope and history.
  • Changing the directive requires a new anchor; requirements and architectural decisions can join the same record.
  • For each anchor and work item, Jev independently judges relevance, compliance and conflict; delivery review adds an evidence question.
  • Ordinary code combines the answers, counts coverage and renders reports quoting the original directive.

A useful separation—with a reproduced hole

  • A development report that claims compliance without checkable artifacts can receive asserted_only, blocking the gate.
  • But unclear evidence does not necessarily block: my synthetic local probe returned satisfied when relevance/compliance/conflict were confident and the evidence answer was unclear at low confidence.
  • The reported minimum confidence excludes the evidence answer. This is a code-path reproduction, not a measured Jev mistake.
  • A paired deterministic check is stored and shown in reports; the script does not execute it. “Satisfied” must not become proof that the code meets the constraint.

Bound the data and the claim

  • Live requests require both a key and explicit remote-judging opt-in; dry runs reveal the request without sending it.
  • Context overflow keeps a shortened prefix of the work text. That can remove the very evidence needed to judge a later section.
  • The implementation’s default confidence threshold is 0.6, a configuration value—not a validated reliability guarantee.
  • Typed answers and passing fixture tests establish plumbing, not correct judgments or calibrated confidence.

The reusable idea is the durable, verbatim constraint register. Before treating this as an acceptance gate, tighten uncertain-evidence handling, run the paired checks, and evaluate real plans and delivery reports.