counter-metric-check

counter-metric-check

Whatever the watched loop's own headline number did in a given period plays no part in this skill's decision for that period. The only input that matters is the second, independently owned reading and where it sits relative to its ceiling.

Instructions

You are a Claude Code skill implementing the counter-metric-loop pattern. Work through these steps in order:

1. Load the governance graph before anything else. Open governance-graph.example.json (or the file the user names) and parse nodes and edges.
2. Locate the two loop nodes. One carries role: "primary" (the loop under watch); the other carries role: "counter" (the independent signal). In this kit's shipped graph those are pr-approval-loop and release-pipeline.
3. Check isolation before doing anything that depends on it holding. Walk every edge whose subject is the primary loop and confirm none of them reaches the counter loop, the threshold node, or a period under a predicate that would let the primary loop read the counter reading (reads, consumes, receives, or anything with that meaning). If one turns up, stop here and report it: a counter-metric the watched loop can see has stopped being a counter-metric, and nothing downstream of that fact is worth computing.
4. Read the ceiling. Follow the counter loop's bound-by edge to its threshold node and note the ceiling and metric fields there.
5. Pick the periods to evaluate. Absent a specific period from the user, use every type: "period" node in the graph — four in this kit's shipped scenario, week-14 through week-17.
6. Take both readings straight off the graph's reported edges — the primary loop's edge into a period, and the counter loop's separate edge into that same period. Neither reading is something to estimate or infer; a period lacking either edge is UNVERIFIABLE, full stop.
7. Let the counter reading alone decide the verdict. Above the ceiling means FLAGGED. At or below it means CLEAN. Keep the primary reading in the record for context, but it never tips a verdict either way — a loud week on the headline number is not itself proof of anything, in either direction.
8. Produce a governance edge for each FLAGGED period. Fields: subject (counter loop), predicate: "flags-divergence", object (primary loop), period, primary_reading, counter_reading, metric, ceiling. CLEAN and UNVERIFIABLE periods get no edge.
9. Write divergence-report.json (or wherever the user asked) at this kit's root, holding: isolation_check (the step-3 result and its detail), a periods array (one entry per period —period_id, primary_reading, counter_reading, ceiling, verdict, and citation, which is the counter loop's exact reported edge on a FLAGGED period and null everywhere else), and new_governance_edges holding step 8's output.
10. Summarize on the way out. State the CLEAN/FLAGGED/UNVERIFIABLE counts and, for every FLAGGED period, the reading pair that triggered it.