"Policy Discovery: Closing the Loop"

The layer that turns institutional memory back into an input — so the next gate run knows what the last hundred did.

---

The corpus remembers; the pulse analyzes. Policy Discovery is where the loop closes: on the next gate run, the PolicyDiscoveryAuditor compares the current failures against the most recent pulse and asks whether what’s happening now is consistent with what’s been happening.

Three questions

1. Cluster match. Does this failure match a known violation cluster? If the same rule has been failing and getting overridden for weeks across projects, this isn’t a one-off — it’s institutional drift, and the person overriding it right now should know they’re the fifth to do so.
2. Anomaly pattern. Does this checker’s failure rate deviate from the statistical baseline? A sudden spike in safety violations might be a systemic issue — a bad dependency bump, a misunderstood API — not an isolated mistake.
3. Unaddressed policy. Was a policy change proposed in response to this pattern, and never implemented? An unaddressed proposal is institutional debt: the organization noticed the problem, agreed on a fix, and then didn’t do it. Policy Discovery surfaces that debt every time the pattern recurs.

The score

Each match produces a ConsistencyFinding with a risk weight, and the ConsistencyScorer rolls them into a single number from 1.0 (fully consistent with institutional history) downward. The discounting is validity-aware — a finding built on 3 data points deducts a quarter of what one built on 30+ does, so the score inherits the pulse’s honesty about sparse data.

The result is the thing that makes this more than a linter: every gate run now ends with an institutional consistency score. 0.85 means “you’re mostly consistent with what the organization has been doing.” 0.4 means “this looks like drift — here are the patterns you should know about before you ship.”

The gate answered “does the code pass?” Policy Discovery answers “is shipping it consistent with everything we’ve learned?” — and hands that judgment back to the person at the keyboard. (The seam where this reaches into a live gate run is the ConsistencyChecker.)

---

Source: github.com/jpurnell/org-judgement-system