Observation Kernels

Turns VerificationRegistry data into a real evidence-completeness measurement — the layer that feeds the Research State Kernel with numbers instead of hand-typed test values.

PurposeCompute evidence_fraction and verification_status for real, from actual registry data, instead of narrating a judgment.

StatusVerified — executed directly against a real VerificationRegistry instance, output shown below

Built fromverification_registry.py

Depends onverification_registry.py

Superseded by—

EvidenceReal execution output, this page, run 2026-07-15.

Source

def evidence_completeness_observer(registry, claim_id):
    total_invariants = len(registry._invariants)
    if total_invariants == 0:
        return {"evidence_fraction": 0.0, "reason": "No invariants registered yet -- nothing to measure against."}

    matching_records = [r for r in registry._records if r.claim_id == claim_id]
    if not matching_records:
        return {"evidence_fraction": 0.0, "reason": "No verification records found for '%s'." % claim_id}

    most_recent = matching_records[-1]
    if not most_recent.passed:
        return {"evidence_fraction": 0.0, "reason": "Most recent verification for '%s' did not pass." % claim_id}

    checked = len(most_recent.invariant_names_checked)
    fraction = checked / total_invariants

    return {
        "evidence_fraction": round(fraction, 3),
        "checked": checked,
        "total_invariants": total_invariants,
        "is_stale": registry.is_stale(claim_id),
        "reason": "%d of %d currently-registered invariants checked and passed." % (checked, total_invariants),
    }


def verification_status_observer(registry, claim_id):
    result = evidence_completeness_observer(registry, claim_id)
    if result["evidence_fraction"] == 0.0:
        return "none"
    if result.get("is_stale", True):
        return "partial"
    if result["evidence_fraction"] >= 0.99:
        return "full"
    return "partial"

Real execution output

Run directly against a fresh VerificationRegistry, three invariants registered in sequence — not a mock, the real class.

=== Real test: claim checked against ALL current invariants ===
{'evidence_fraction': 1.0, 'checked': 2, 'total_invariants': 2, 'is_stale': False,
 'reason': '2 of 2 currently-registered invariants checked and passed.'}
Status: full

=== Real test: claim checked against only SOME current invariants (should be lower fraction) ===
{'evidence_fraction': 0.5, 'checked': 1, 'total_invariants': 2, 'is_stale': True,
 'reason': '1 of 2 currently-registered invariants checked and passed.'}
Status: partial

=== Real test: a NEW invariant gets added -- claim_A should now be genuinely stale/partial ===
{'evidence_fraction': 0.667, 'checked': 2, 'total_invariants': 3, 'is_stale': True,
 'reason': '2 of 3 currently-registered invariants checked and passed.'}
Status: partial

=== Real test: claim with no records at all ===
{'evidence_fraction': 0.0, 'reason': "No verification records found for 'never_checked_claim'."}

The third test is the one that actually proves something: adding a third invariant correctly dropped a previously-full claim to partial, without that claim's own history changing at all. This is the staleness mechanism working end to end, not just in isolation.