did-we-drift

Project-drift audit for agent-built software, plus /did-we-drift init to start projects audit-ready. Grades two axes separately: the WORK (ON TRACK / DRIFTED (MINOR|MATERIAL|CAPTURED) / INCONCLUSIVE / BLOCKED) and the BASIS it was measured against (RATIFIED / PROVISIONAL / NONE), so a real-but-scattered plan gets a usable verdict instead of a refusal. Init converts a fresh plan into one ratified, provenance-checkable baseline document BEFORE the build starts — the user's finish-line sentence lands in its own commit with their verbatim words, so authorship is provable, not asserted. Classifies every candidate goal USER-RATIFIED / UNKNOWN / AGENT-DRAFTED on mechanical git evidence. Basis recovery for projects that never wrote a plan; a ratified sentence governs FORWARD from its own date. Refuses to certify on-track without an enumerable work map. Grade-gated drift interrupts, snooze-with-memory, transactional repair, run-to-plan binding for /loop and /goal sprints. QUICK single-agent audit by default; contamination-scaled DEEP fan-out when triggers fire. Generates and maintains a self-contained plain-language dashboard (docs/drift-dashboard.html) showing planned-vs-built with a computed completion meter, per-step real-world explanations with technical detail behind a disclosure, build history, what-changed-since-last-look, and waiting-on-you decisions — a derived VIEW that never becomes a second authority. Ships a sandboxed checker that verifies the page renders, escapes its inputs, keeps jargon out of the default view, and (with --ssot) that its quoted goal and steps exist verbatim in the plan. /did-we-drift update maintains a mature baseline at milestones: freshness-gated (refuses to refresh over unresolved MATERIAL/CAPTURED drift), appends dated verified-facts blocks (append-only history), forces a named disposition of every carried/open item (RESOLVED/CARRY/ESCALATE — nothing dropped silently), records user-attested facts distinctly from machine-verified ones, and