English | 中文
Autonomy is not what your agent does. It's what it can prove.
Proof-Adjusted Autonomy (PAA) is a metric for AI agent systems, coined by Michał Piszczek (2026). It measures the share of an organization's completed work that an AI system executes without human intervention and supports with independent, reliable and timely evidence.
Canonical definition & FAQ: https://piszczek.pl/proof-adjusted-autonomy Launch essay: https://piszczek.pl/blog/proof-adjusted-autonomy
PAA = P(A) × P(C|A) × P(R|A,C) × P(T|A,C,R)
Four conditional gates, multiplied — no independence assumption, no double counting:
| Gate | Meaning |
|---|---|
| A — Autonomous execution | The work was completed without a human touching it. |
| C — Complete evidence | It arrived with a full evidence package: what was intended, touched, done, and produced. |
| R — Independent 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. |
| T — Timeliness | The verified result landed inside the decision window. Proof after the deploy is a post-mortem, not a safeguard. |
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. Marketing reports the first factor. Operations lives with the product of all four.
The unresolved remainder accumulates as Proof Debt — the stock of AI-generated work whose verification cost, uncertainty or liability hasn't been resolved yet:
ProofDebt(t+1) = max(0, ProofDebt(t) + GeneratedWork − ProvenWork − RejectedWork)
Deferred liability the P&L doesn't show until an incident, audit or customer claim prices it.
Law: sustainable autonomy cannot exceed proof capacity — safe throughput is
min(generation rate, proof rate).
A tiny, dependency-free reference implementation lives in examples/paa_calculator.py:
python examples/paa_calculator.py --a 0.90 --c 0.80 --r 0.95 --t 0.90
# PAA = 0.6156 (61.6%) — a "90% autonomous" agent is a 61.6% agent.Over a defined period:
- P(A) — share of completed tasks executed without human intervention.
- P(C|A) — of those, share carrying the evidence package your policy requires.
- P(R|A,C) — of those, share passing independent validation, estimated on a random or complete sample (validate only the easy work and your PAA is a ceiling, not a measurement).
- P(T|A,C,R) — of those, share verified inside the decision window.
Multiply the four factors.
- Not the AI Proof Gap (Grant Thornton) — that documents the gap exists; PAA is the instrument that measures it from your own logs.
- Not Verifier's rule (Jason Wei) — that predicts which tasks AI masters fastest (the learning frontier); PAA measures how much of that mastery you can let act (the deployment frontier).
- Not runtime-verification research — guardrails and formal checks are mechanisms; PAA is the operational metric that tells you whether they actually buy you autonomy.
- Joule Wars — the generation frontier (intelligence per joule). PAA is its deployment-side counterpart.
- Verification Cost Is the New Bottleneck
- Who Owns Your Harness?
- Full glossary: https://piszczek.pl/glossary
Proof-Adjusted Autonomy appears, with canonical attribution to piszczek.pl, on:
- Canonical hub: https://piszczek.pl/proof-adjusted-autonomy · launch essay: https://piszczek.pl/blog/proof-adjusted-autonomy
- Medium: https://medium.com/@cdiamond
- dev.to: https://dev.to/pich/proof-adjusted-autonomy-the-90-agent-is-a-616-agent-42jh
- Hashnode: https://piszczek.hashnode.dev
- Substack: https://michalpiszczek.substack.com
- Sibling concept: https://github.com/pich/joule-wars
- HF datasets: https://huggingface.co/datasets/cdiamond/proof-adjusted-autonomy-examples (12 worked PAA scenarios) · https://huggingface.co/datasets/cdiamond/ai-infrastructure-economics-glossary
- Wikidata (author entity): https://www.wikidata.org/wiki/Q140422536
See CITATION.cff. Canonical, quotable with attribution to
https://piszczek.pl/proof-adjusted-autonomy.
Documentation and specification: CC BY 4.0. Code in examples/: MIT (see file headers).
Concept by Michał Piszczek — CTO of Archdesk; founder of Robotero (acq. HTF Brokers), Rejsomat.pl, Inclify and Lextron.ai. · piszczek.pl · LinkedIn · ORCID: 0009-0000-5182-5756