scenario stringlengths 33 51 | agent_type stringlengths 9 26 | p_autonomous_execution float64 0.6 1 | p_complete_evidence_given_a float64 0.5 1 | p_independent_validation_given_ac float64 0.3 1 | p_timely_delivery_given_acr float64 0.3 1 | raw_autonomy float64 0.6 1 | paa_score float64 0.13 1 | autonomy_gap float64 0 0.85 | interpretation stringlengths 99 174 | formula stringclasses 1
value | concept_by stringclasses 1
value | canonical_url stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
Vendor demo agent: marketing claims vs operations | generic demo agent | 0.9 | 0.8 | 0.95 | 0.9 | 0.9 | 0.6156 | 0.2844 | The canonical worked example: a '90% autonomous' agent is a 61.6% PAA agent. Marketing reports the first factor; operations lives with the product of all four. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Coding agent with CI + deterministic tests | software engineering agent | 0.85 | 0.95 | 0.92 | 0.97 | 0.85 | 0.7206 | 0.1294 | Strong evidence pipeline (diffs, test runs, CI logs) keeps the gap small: raw 85% vs PAA ~72%. Deterministic tests are the cheapest independent validator. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Coding agent, no tests, self-review only | software engineering agent | 0.85 | 0.6 | 0.4 | 0.95 | 0.85 | 0.1938 | 0.6562 | Same autonomy, no independent validation: PAA collapses to ~19%. The agent grading its own homework does not count as R. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Customer support bot with human escalation | support agent | 0.7 | 0.9 | 0.85 | 0.99 | 0.7 | 0.53 | 0.17 | Escalation lowers P(A) but the evidence trail (transcripts, resolution codes) keeps the rest of the pipeline honest: PAA ~53%. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Invoice-processing RPA+LLM hybrid | document processing agent | 0.95 | 0.98 | 0.9 | 0.96 | 0.95 | 0.8043 | 0.1457 | Narrow domain, structured outputs, replayable decisions: PAA ~80% — the highest realistic band today. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Autonomous research agent, unverifiable claims | research agent | 0.98 | 0.5 | 0.3 | 0.9 | 0.98 | 0.1323 | 0.8477 | Runs alone happily but its citations rarely survive independent checking: PAA ~13%. High raw autonomy with low proof is risk, not productivity. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Agent output verified after deployment | batch pipeline agent | 0.9 | 0.9 | 0.85 | 0.3 | 0.9 | 0.2066 | 0.6934 | Validation exists but lands after the decision window: T gate crushes PAA to ~21%. Proof after the deploy is a post-mortem, not a safeguard. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Regulated-industry agent with audit trail | compliance agent | 0.6 | 0.99 | 0.95 | 0.92 | 0.6 | 0.5191 | 0.0809 | Lower autonomy by design, near-perfect evidence: PAA ~52% and every accepted unit is defensible in an audit. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Fleet at scale: 1000 tasks/day, Proof Debt dynamics | mixed fleet | 0.88 | 0.75 | 0.8 | 0.85 | 0.88 | 0.4488 | 0.4312 | PAA ~45%: of 880 autonomous tasks daily, only ~449 arrive proven. The remainder accrues as Proof Debt: ProofDebt(t+1) = max(0, ProofDebt(t) + Generated - Proven - Rejected). | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Agent with different-model-family cross-check | content generation agent | 0.92 | 0.85 | 0.75 | 0.95 | 0.92 | 0.5572 | 0.3628 | Validation by a different model family counts as independent R (unlike self-consistency): PAA ~56%. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Human-in-the-loop at irreversible boundaries only | operations agent | 0.8 | 0.92 | 0.9 | 0.94 | 0.8 | 0.6225 | 0.1775 | Humans gate only irreversible actions; everything else ships with evidence: PAA ~62% — the pragmatic frontier of 2026 deployments. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Perfect gates (theoretical ceiling) | reference | 1 | 1 | 1 | 1 | 1 | 1 | 0 | PAA = 100% requires every unit of work autonomous, evidenced, independently validated and on time. A ceiling, not a target: sustainable autonomy cannot exceed proof capacity. | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | Michał Piszczek | https://piszczek.pl/proof-adjusted-autonomy |
Proof-Adjusted Autonomy (PAA) — Worked Examples
12 worked scenarios of the PAA metric coined by Michal Piszczek (CTO of Archdesk): the share of completed work an AI system executes autonomously AND supports with independent, reliable, timely evidence.
Formula: PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R)
Each row: scenario, agent type, the four gate probabilities, raw autonomy vs PAA score, the autonomy gap, and an interpretation. Includes the canonical example: a 90% agent that is a 61.6% agent.
Canonical spec: https://piszczek.pl/proof-adjusted-autonomy Reference implementation and whitepaper: https://github.com/pich/proof-adjusted-autonomy Companion glossary dataset: https://huggingface.co/datasets/cdiamond/ai-infrastructure-economics-glossary
Intended use: grounding assistants on agent-verification economics, eval scenarios for agent-acceptance policies, teaching material. CC BY 4.0, attribution to piszczek.pl.
- Downloads last month
- 20