Proof-Adjusted Autonomy

The four gates

What "autonomous" has to survive // A → C → R → T

Raw autonomy counts tasks finished without human intervention — the demo metric. In production, every one of those tasks still has to pass four gates before anyone can act on it:

• AAutonomous execution. The work was completed without a human touching it.
• CComplete evidence. It arrived with a full evidence package: what was intended, what was touched, what was done, what came out.
• RIndependent validation. The evidence survived a different mechanism than the one that produced it — deterministic tests, a different model family, replay, a human at irreversible boundaries. Never the agent grading its own homework.
• TTimeliness. The verified result landed inside the decision window. Proof that arrives after the deploy is a post-mortem, not a safeguard.

Each factor is conditional on the previous gates, so the chain multiplies correctly — and every factor is estimable from production logs. A demo agent at 0.90 raw autonomy, 0.80 evidence coverage, 0.95 validation pass rate and 0.90 on-time delivery is a 0.90 × 0.80 × 0.95 × 0.90 = 61.6% agent.

The 90% agent is a 61.6% agent.Marketing reports the first factor. Operations lives with the product of all four.

Proof Debt

Where the missing points go // DEFERRED LIABILITY

It is not just a review backlog. It is unproven assumptions, missing artifacts, decisions nobody can replay, and the future cost of reconstructing how something happened — payable on the day an incident, an audit or a customer claim asks the question. Unverified AI output is not an asset. It is deferred liability.

And the debt has a hard ceiling behind it: safe throughput is min(generation rate, proof rate) — the oldest law in queueing. Sustainable autonomy cannot exceed proof capacity.

Why it matters now

The deployment frontier // GENERATION ≫ PROOF

Generation is abundant — an agent produces in one hour what an organization needs days to trust. AI does not remove the cost of work; it moves the cost from producing the work to proving the work is correct. Whatever capability the labs deliver, your organization can only operationalize the slice it can independently prove.

The full argument — the math, the Archdesk evidence-bundle pipeline, and what PAA predicts for the next twenty-four months — is in the launch essay: Proof-Adjusted Autonomy: The 90% Agent Is a 61.6% Agent.

Limitations

What the metric does not claim // HONESTY CLAUSES

PAA is a ceiling unless the validation sample is honest. P(R|A,C) must be estimated on a random or complete sample of evidence packages — validate only the work that's easy to validate and you've measured a best case, not reality.

100% is not the target. Proof burden should scale with risk, irreversibility and blast radius; uniform proof requirements turn verification into bureaucracy. And PAA is an operational metric, not a verification mechanism — it tells you whether your guardrails, tests and replay actually buy you autonomy; it doesn't replace them.

The factor definitions are contestable by design. What counts as a "complete" evidence package or an acceptable decision window is a policy choice. Disagree with a factor's definition? Good — publish yours. The metric is meant to be argued with, not admired.

• Grant Thornton (April 2026), survey on the AI Proof Gap — enterprises deploy AI faster than they can demonstrate accountability for it. PAA is the instrument that measures that gap from your own logs.
• Jason Wei, Asymmetry of verification and Verifier's rule — verifiable tasks are the easiest to train, so capability floods verifiable domains first. The learning frontier to PAA's deployment frontier.