// Glossary

Glossary of agent verification terms

Definitions of the terms used in GDF Research papers on AI agent verification: confirmation ladder, frontier gating, doubt signal, consensus drift and more.

Confirmation ladder
Five levels of evidence for an action: intent observed, action attempted, action acknowledged by the platform, resulting state confirmed by an independent reader, and end-to-end completion at the boundary the user cares about. Proof state only moves up. From Monotonic Proof-State Advancement for Distributed Workflow Verification →
Boundary cap
The highest level of the confirmation ladder an observer can reach from where it sits. Evidence beyond the cap is marked boundary-limited, which is different from missing or failed. A refund's arrival on a customer's card statement is a common example. From Monotonic Proof-State Advancement for Distributed Workflow Verification →
Frontier gating
Evaluating a check only when its confidence threshold is met and every type of evidence it reads has arrived. Without it, out-of-order delivery turned a judge that was perfect on complete evidence into a 91% one. From Evidence-Graded Memory with a Doubt Signal →
Doubt signal (uneasiness)
A cheap score kept beside each remembered fact, rising with staleness, the gap to what a claim needs, unreachability from the agent's vantage, novelty, stakes and the signal's track record. It triggers re-checking. In testing it only changed behavior when it could interrupt a claim. From Evidence-Graded Memory with a Doubt Signal →
Proof of intended outcome
Treating an agent's completion report as a claim tested against what the user wanted, with requirements fixed before execution, evidence bound to exact targets, shared sources counted once and an expiry on every conclusion. From Proof of Intended Outcome →
Handoff verification
The common pattern where a reviewer agent sees only the executor's summary of what it did. It can't catch work done on the wrong target, because all the evidence agrees with the summary. From Proof of Intended Outcome →
Correlated evidence
Checks that agree because they read the same source. Treated as independent, they inflate confidence: seven such verifiers reported 0.99999994 on tasks that succeeded 93.9% of the time. From Proof of Intended Outcome →
Consensus drift
The movement of a verifier panel's verdicts away from the evidence when its members are rewarded for agreeing. Agreement-scored panels drifted toward passing bad work or rejecting good work, depending on the early votes. From Consensus Drift →
Gated evidence
An experimental setup where the verifier must choose to open the evidence, at a small cost, before voting. It separates the decision to look from the judgment made after looking. From Consensus Drift →
Provenance label
A note attached to a vote stating what the voter observed and when. A single sentence of provenance was enough for verifiers to discount approvals that rested only on the executor's message. From Consensus Drift →
Synchronous fast path
Deterministic checks run inline, before anything is sent to a model for review. In the kernel-boundary guardrail it cut p99 verdict latency for those events from 3,617 ms to 1 ms. From Toward LLM-Assisted Policy Enforcement at the Kernel Boundary →