Anti-Goodhart Auditor — an epistemic linter for AI/data systems

goodhart_auditor.py

What it does

It flags one precise, common, expensive anti-pattern: a field or metric whose name claims a verified property — reviewed, approved, verified, audited — while its value is backed by nothing that actually verifies it (a default, a constant, an accepted parameter). The name promises a guarantee the data does not carry. Under Goodhart's law, once such a field becomes a target ("ship only reviewed rows"), the gap between the claim and the check is where failures hide.

The auditor takes a list of declared fields and their backing (how each value is produced) and returns findings ranked by severity:

• high — a verification name backed by something weak (default, constant, parameter, assumed).
• medium — a verification name with unknown backing (go find out what sets it), or an unrecognized backing.
• low — a verification name with plausibly-strong backing (human_action, derived, test_run, …) — noted so a reviewer can spot-check it.

Why it's built the way it is

Three deliberate properties, because a governance tool that isn't itself trustworthy is worse than none:

1. Deterministic. Pure function of its input — no randomness, no hidden state, no I/O. Same input → byte-identical output (there's a fingerprint() for CI reproducibility checks).
2. Honest about its limits. It reports suspicions to inspect, never verdicts. Its self-test asserts what it actually catches and demonstrates a real miss (greenlit means "approved" but contains no verification word, so it evades the linter). The blind spot is in the test suite, not hidden.
3. Self-testing. python goodhart_auditor.py runs the self-test before the demo. If the tool's behavior drifts, the test fails loudly.

Usage

from goodhart_auditor import audit, Field

findings = audit([
    Field("reviewed", backing="default"),          # high
    Field("approved", backing="human_action"),      # low
    Field("independence_group", backing="parameter"),# high
])
for f in findings:
    print(f.render())

CLI: python goodhart_auditor.py (self-test + demo).

Limitations (stated plainly)

• Heuristic, not proof. It matches names, not semantics. A name that carries a verification meaning without a verification word (greenlit, qc_pass) is missed. A legitimately-backed field can trip a low-severity note.
• Backing is declared, not inspected. The tool trusts your backing label; it does not read the code that sets the field. Garbage-in applies.
• Scope is field/metric names. It does not audit reasoning, prompts, or model outputs directly — though the same principle (name-claims-vs-what-checks) generalizes.

Use it as a fast first pass that surfaces candidates for human inspection — not as a gate that decides anything on its own. That posture is the point.