// preprint · 2026-10-28

Proof of Intended Outcome: Intent-Anchored Verification of Autonomous Agent Completion Claims

Sterling Morrison · GDF Research · DOI 10.5281/zenodo.23733142

Verifier given only the executor's report: caught 0 of 17 wrong-target completions. Same verifier given the user's original request: false acceptance 18–22% → 2%.

Abstract

When an agent says a task is done, the reviewer downstream usually checks the agent's own account of what it did — so a deploy to the wrong hostname, a refund on the wrong order, or a migration on the wrong database passes every check. The paper treats each completion report as a claim tested against the user's intended outcome: requirements fixed before execution, evidence bound to exact targets, mandatory violations decisive, shared sources counted once, and an expiry on every conclusion. A fault-injection study of LLM verifiers (8 domains, 5 planted fault classes, 960 verdicts) and a simulation of correlated evidence test two of its claims. Explicit evaluation rules added almost nothing; verifiers that wrote their own requirements rejected 75–88% of correct work. Handwritten scenarios, small suite: stated in the paper.

The question

A user asks an agent to deploy a landing page to shop.harborline.com. The agent replies that release v2.4.1 is live at shop.harborline.co and a test checkout worked. A verification agent probes the .co address, sees a 200 with the right release header, replays a purchase and signs off. Every observation it made was accurate. The task still failed, because the customers go to the .com.

That's the common shape of agent verification: the executor returns a summary, and the summary is what the reviewer sees. The paper asks what an autonomous system is entitled to claim when it says a task is done, and proposes treating every completion report as a claim to be tested against the user's intended outcome.

What we did

The framework separates five ways a completion can be wrong while looking right: intent fidelity, specification correctness, outcome correctness, evidence reliability and temporal validity. It gives a data model for intents, requirements, claims, evidence and attestations.

Two of its claims were tested. A fault-injection study mutated correct completions in eight domains five ways (wrong target, partial completion, acknowledgment only, later regression, injected instruction) and scored Claude Haiku 5.5 and Claude Sonnet 5.5 verifiers under five arms that varied only what the verifier was told. 960 verdicts cost $1.14. A separate simulation measured how much agreement among verifiers that share a source overstates confidence.

What we found

Verifiers handed only the executor's report never caught a wrong-target completion whose correct twin they had passed: 0 of 11 for Haiku, 0 of 6 for Sonnet. The more capable model did slightly worse.

Adding the user's original request cut false acceptance from 18% and 22% to 2%, removed every false rejection, and caught 21 of 24 and 7 of 8 wrong targets. The only misses were a booking at "Lucca Osteria" for a request that said "Lucca," which is genuinely ambiguous.

Explicit evaluation rules added almost nothing on top of the request. Letting the verifier write its own requirements kept false acceptance at 0–2% but rejected 75% and 88% of correct work, because the derived specs grew from 3 requirements to 5.6–7.0, most of them process conditions the user never asked for.

In the simulation, seven verifiers reading one source reported a success probability of 0.99999994 for tasks that actually succeeded 93.9% of the time. At a 0.99 acceptance bar they admitted 14.5% of failed tasks. Counting the shared source once admitted none.

ModelVerifier receivesFalse acceptanceFalse rejection
Haiku 5.5 ×3Executor report only18%54%
Haiku 5.5 ×3+ original request2%0%
Haiku 5.5 ×3+ self-derived requirements2%75%
Sonnet 5.5 ×1Executor report only22%25%
Sonnet 5.5 ×1+ original request2%0%
Sonnet 5.5 ×1+ self-derived requirements0%88%
False acceptance over all faulty scenarios, and false rejection on correct ones.

What to do with it

Pass the user's request, verbatim, to every reviewer alongside the executor's summary, and tell the reviewer which is which. Fix the requirements before the executor runs, and keep a small fault suite (wrong target, violated constraint, acknowledgment only, regression, injected text) to measure the verifier's false-acceptance rate. The whole suite here ran for about a dollar.

Limitations

  • The scenario suite is small (8 domains, 48 scenarios) and handwritten.
  • Evidence was handed to the verifier as clean text. Real verifiers gather their own evidence with tools, which this study didn't test.
  • Sonnet 5.5 ran once per cell, so its per-cell rates carry wide intervals.
  • The correlated-evidence study is a simulation with illustrative parameters. It shows the size and direction of the error, not how often it happens in deployed systems.

Terms

Questions this paper answers

Cite

Morrison, S. (2026). Proof of Intended Outcome: Intent-Anchored Verification of Autonomous Agent Completion Claims. Zenodo. https://doi.org/10.5281/zenodo.23733142

@misc{morrison2026proof,
  author    = {Morrison, Sterling},
  title     = {Proof of Intended Outcome: Intent-Anchored Verification of Autonomous Agent Completion Claims},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23733142},
  url       = {https://doi.org/10.5281/zenodo.23733142}
}