Knowledge-Maturity Classifier

knowledge_maturity.py

What it does

It answers a question that is not "is this claim true?" but "how much of the evidentiary work behind this claim has actually been done?" That distinction — between truth and process-completeness — is what keeps AI-assisted analysis honest about its own footing. A model can state something confidently at maturity ANECDOTE; the classifier makes that gap explicit and un-ignorable.

You declare auditable properties of the evidence — how many observations, how many distinct methods, whether it's been independently replicated, whether there's an unresolved contradiction, whether it's been adversarially tested — and it returns a level on a fixed ladder:

ANECDOTE → SUPPORTED → CORROBORATED → REPLICATED → ROBUST

The key property: critical gates

Certain failures cap maturity no matter how much supporting evidence piles up:

• An unresolved contradiction caps the claim at SUPPORTED, even with fifty observations across three methods and a replication.
• No independent replication caps at CORROBORATED, no matter the volume.

This is the anti-Goodhart principle applied to evidence: quantity cannot substitute for a missing kind of evidence. You cannot buy your way to "replicated" by collecting more of the same non-independent data. The self-test enforces exactly this (quantity must not buy replication).

Why it's built the way it is

• Deterministic. Thresholds are fixed constants in the file. Changing them is a visible, attributable editorial decision — the classifier never adapts them on its own. Same input → identical output (fingerprint() provided).
• Self-testing. python knowledge_maturity.py runs the ladder, the gates, and the determinism check before the demo.

Usage

from knowledge_maturity import classify, Evidence

a = classify(Evidence(observation_count=40, distinct_methods=1,
                      independently_replicated=False))
print(a.level.name, a.caps_applied)   # CORROBORATED  ('no_independent_replication -> capped ...',)

Limitations

• It classifies declared properties; it does not verify that your declarations are accurate.
• The ladder and thresholds are a reasonable default, not a universal standard — different domains draw the lines differently. Treat them as a starting convention to adapt openly.
• It rates process maturity, not correctness. A ROBUST claim can still be wrong; an ANECDOTE can still be true. The tool tells you how much work stands behind the claim, which is precisely the thing confident prose tends to obscure.