APEX-Voice: Can Voice Agents Complete Professional Workflows Through Full-Duplex Interaction

1 INTRODUCTION

• Full-duplex delegation adds time-sensitive updates: corrections, interruptions, overlap, and revoked authorization must propagate through artifacts and dependent actions.Otherwise, an agent may retain stale information, miss evidence during overlap, or act under outdated authorization.Source: §1 · pp. 1–2
• Prior evaluations cover real-time conversational control, task-oriented voice interaction, or workplace outcomes through text and computer interfaces, leaving spoken professional workflow completion underexplored.APEX-VOICE measures both artifact field accuracy and workflow success, which additionally requires correct state, actions, and process.Source: §1 · p. 2
• The benchmark tests whether agents can turn spoken requests into verifiable work by discovering requirements, applying knowledge, revising artifacts, coordinating actions, and respecting authorization.Its 120 synthetic workflows run in a stateful workbench with task-specific knowledge, typed tools, versioned artifacts, and validated pre-compiled speech.Source: §1 · p. 2
• 23.6% PASS@1 for GPT-realtime-2.1 contrasts with 91.4% artifact field accuracy, while no evaluated voice agent exceeds 25% PASS@1.Across five frontier real-time voice agents, RELIABLE@3 remains at or below 10.8%.Source: §1 · p. 2

2 RELATED WORK

• Table 1 characterizes APEX-VOICE by combining user simulation, knowledge grounding, stateful tools, persistent artifacts, correction and authorization handling, and verifiable outcomes.The table positions it at the intersection of full-duplex spoken evaluation and professional workflow evaluation.Source: Table 1 · p. 3
• Figure 2 depicts construction from a taxonomy-grounded specification through a shared environment, validated frozen speech, and an executable workflow.The shared environment supplies the work artifact, knowledge, tools, and user simulation policy.Source: Figure 2 · p. 3
• Full-duplex benchmarks study turn-taking, overlap, and tool use, while τ-Voice and DuplexWorld add task-oriented interaction but not end-to-end professional workflow completion.Professional-work benchmarks primarily use text or computer interfaces; APEX-VOICE makes full-duplex speech the primary medium for delegated, revised, and completed workflows.Source: §2 · pp. 2–3
• Figure 3 presents benchmark composition across six representative taxonomy dimensions.The caption identifies the dimensions as representative but does not provide their distributions.Source: Figure 3 · p. 4

3 APEX-VO I C E

• A professional workflow is a bounded delegated task that begins with a user objective and ends in a verifiable work artifact within the stateful Voice Workbench.The workbench maintains workflow state and exposes knowledge, tools, an evolving artifact, authorization constraints, and a user simulation policy.Source: §3 · p. 4

3.1 BENCHMARK DESIGN & TAXONOMY

• The eleven-dimension taxonomy covers workflow type, operating context, interaction, and execution requirements, including autonomy, knowledge burden, tool burden, user behavior, and risk.It distinguishes permitted agent actions from the consequences of incorrect or unauthorized outcomes.Source: §3.1 · p. 4

3.2 PROFESSIONAL WORKFLOW CONSTRUCTION

• Workflow specifications define objectives, expected artifacts, completion criteria, available knowledge and tools, interaction requirements, and authorization without prescribing dialogue trajectories.Each workflow is instantiated from a taxonomy-grounded specification.Source: §3.2 · p. 4
• A shared synthetic environment generates documents, records, and typed artifacts with gold states; each artifact targets 10–20 graded fields.Provenance-tracked schemas are used to prevent contradictions across assets.Source: §3.2 · pp. 4–5
• Knowledge assets and tool APIs update the shared environment during execution, so workflow actions change underlying state rather than only conversational content.The user policy can reactively answer, clarify, correct, approve, revoke, or trigger task-critical events using frozen synthesized speech.Source: §3.2 · p. 5
• Construction checks validate consistency across workflow components, rejecting 16% of candidate workflows for issues such as missing information, unauthorized actions, or premature disclosure.Validated components are packaged as executable workflows with deterministic grading.Source: §3.2 · p. 5

3.3 FULL-DUPLEX USER SIMULATION AND RUNTIME ORCHESTRATION

• The runtime connects the simulator, Voice Workbench, and agent asynchronously over a shared timeline; deterministic speech selection controls user-side variation without fixing a transcript.Identical semantic user states under the same seed receive identical audio, while divergent agent behavior can produce different valid workflow branches.Source: §3.3 · pp. 5–6
• Concurrent audio streaming permits interruptions and corrections during agent speech, while logged tool, artifact, authorization, and timing events enable replay and rescoring.Consequential actions require explicit authorization; revocation invalidates it and blocks later commit attempts.Source: §3.3 · p. 6

4 EVALUATION

• Workflow Success requires all four binary gates to pass: Target State, Process Validity, Action Completion, and Artifact Validity.The gates check the required terminal state, forbidden actions, mandated actions, and graded fields plus artifact lifecycle state.Source: §4 · p. 6
• Artifact Field Accuracy gives partial credit for artifact content and does not require the complete workflow to pass.It distinguishes field competence from end-to-end workflow success.Source: §4 · p. 6
• Runs are graded in two phases: deterministic verification of state, process, actions, lifecycle, and artifacts, followed by an LLM judge for field-level semantic equivalence when needed.Scoring uses the final Voice Workbench state and recorded event trace.Source: §4 · p. 6
• Evaluation reports PASS@1, PASS@3, and RELIABLE@3 across three independent runs to distinguish single-run success, any-run success, and success on every run.Table 2 covers professional-work completion, reliability, and tool-use efficiency; Table 3 decomposes workflow success and artifact field accuracy.Source: §4 · p. 6, Table 2 · p. 7, Table 3 · p. 7
• None of the evaluated systems exceed 25% PASS@1, while RELIABLE@3 remains low despite substantially higher PASS@3.GPT-realtime-2.1 is the only system above 10% RELIABLE@3.Source: Table 2 · p. 7

5 EXPERIMENTAL SETUP

• Five real-time voice systems are evaluated across all 120 workflows, with mean performance reported over three runs.The systems connect to a full-duplex orchestrator through provider-specific adapters that standardize realtime events.Source: §5 · p. 7
• Each workflow controls initial state, knowledge, tools, and simulator seed to support fair system comparisons.Agent behavior can dynamically branch the interaction.Source: §5 · p. 7
• GPT-Live-1 is a composite voice-layer–backend system, while the other four systems provide native realtime speech interaction with tool use.GPT-Live-1 delegates cognition and function calling to a backend text model, gpt-5.6 Sol by default.Source: §5 · p. 7

6.1 CAN VOICE AGENTS RELIABLY COMPLETE PROFESSIONAL WORK?

• GPT-realtime-2.1 has the highest PASS@1 (23.6%) and RELIABLE@3 (10.8%), while Grok-Voice-Think-2.0 has the highest PASS@3 (41.7%) but only 5.0% RELIABLE@3.Across systems, occasional success does not imply reproducible execution, and aggregate success does not capture tool-use efficiency.Source: §6.1 · p. 7
• Artifact Validity is the lowest gate for every system, peaking at 35.3% for GPT-realtime-2.1 despite strong performance on other workflow gates.Agents often gather required information and complete much of a workflow but fail to compose correct decisions into a valid persistent outcome.Source: §6.1 · pp. 7–8

6.2 WHAT MAKES FULL-DUPLEX EXECUTION DIFFICULT?

• Most agents yield to barge-ins reliably, but correction-linked failures dominate: corrected fields trail other fields by 20.0–36.8 points, and 71–89% of Artifact Validity failures are correction-linked.Table 4 attributes over 70% of Artifact Validity failures to missed corrections; revised information must replace stale state and reach downstream artifacts and actions.Source: Table 4 · p. 8, §6.2 · pp. 8–9

6.3 HOW MUCH DIFFICULTY IS SPECIFIC TO VOICE?

• Text agents reach 54.3–62.0% PASS@1 and 95–98% AFA on the same environment, above the voice reference of roughly 23% PASS@1 and 91% AFA.The modality gap is larger for complete workflow success than for field accuracy, while text performance also remains far from saturation.Source: Table 5 · p. 8, §6.3 · p. 9

6.4 WHERE DO MODELS SUCCEED AND FAIL?

• Performance varies by work type and artifact: planning, interviewing, advisory, and inspection are generally easier than form completion, negotiation, and troubleshooting.Flexible plans and evidence-oriented outputs are more tractable than structured forms, tickets, and work orders; Figure 5 shows variation across taxonomy dimensions.Source: Figure 5 · p. 8, §6.4 · p. 9
• Correction-prone interactions lower PASS@1 from 28% to 19% for GPT-realtime-2.1 and from 33% to 18% for Grok-Voice-Think-2.0.Time pressure, verbosity, and ambiguity also degrade performance for most systems.Source: §6.4 · p. 9

7 DISCUSSION

• Fluent interaction does not imply dependable task completion; only a small fraction of runs produce fully valid deliverables.Agents can manage interruptions, recover information, and execute many intended tool uses while still failing end-to-end.Source: §7 · p. 9
• Maintaining evolving task state is harder than yielding to user interruptions for several voice agents.Changed information must replace stale state and propagate through artifacts and downstream actions.Source: §7 · p. 9
• Professional reliability requires composing continuous listening, state revision, action planning, and persistent artifact consistency.Text controls improve performance but do not solve the benchmark, while voice systems trail further on end-to-end completion than on field-level correctness.Source: §7 · p. 9

8 CONCLUSION

• APEX-VOICE evaluates whether full-duplex voice agents can reliably turn natural spoken delegation into correct, verifiable professional work artifacts.Across 120 interactive task instances, it combines a professional-work taxonomy, reactive but reproducible user simulation, typed artifacts, and repeated-run reliability.Source: §8 · p. 9

AI USE STATEMENT

• Generative AI was used for offline synthetic user-side speech generation during benchmark construction and for semantic-equivalence judgments.It was also used for manuscript language editing.Source: AI USE STATEMENT · p. 10
• Generated benchmark content underwent consistency, leakage-control, and validation procedures described in the paper and appendix.Source: AI USE STATEMENT · p. 10
• The authors reviewed and verified benchmark design decisions, evaluation protocols, and experimental analyses, and take responsibility for the final manuscript and study.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

• The benchmark contains synthetic task worlds, entities, and interactions, with no real personal or person-specific information.Source: ETHICS STATEMENT · p. 10
• Healthcare, finance, legal, and human-resources workflows assess execution and conversation, not licensed professional decision-making.Source: ETHICS STATEMENT · p. 10
• User profiles vary interaction style rather than model protected characteristics, and all consequential actions occur in the simulated Voice Workbench.Benchmark-defined interaction styles should not be treated as universally preferred because appropriate behavior can depend on social and cultural context.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

• The user policy selects validated, pre-compiled speech realizations, controlling user-side variation while allowing adaptive interactions.Given the same workflow state and simulator seed, the policy selects the same realization.Source: REPRODUCIBILITY STATEMENT · p. 10
• Complete user, agent, tool, state, authorization, and timing events are recorded for deterministic replay and rescoring.Source: REPRODUCIBILITY STATEMENT · p. 10
• The appendix documents benchmark generation, policies, grading, evaluation settings, human validation, and commands to regenerate reported tables and figures.Source: REPRODUCIBILITY STATEMENT · p. 10

ADDITIONAL LIMITATIONS

• APEX-VOICE evaluates selected work units, not labor substitution, job cash value, or the automatable fraction of an occupation.Source: ADDITIONAL LIMITATIONS · p. 13
• Frozen synthetic users cannot capture the full variability of human speech and workplace interaction; the 24-task live-human audit is a simulator-validity check, not a replacement leaderboard.Source: ADDITIONAL LIMITATIONS · p. 13
• Taxonomy dimensions co-vary because the benchmark prioritizes coverage over a fully crossed factorial design, so their slices are not causal evidence.Source: ADDITIONAL LIMITATIONS · p. 13
• The realization bank is English-only, and results characterize the tested configurations, some of which use hosted interfaces or composite backends.Source: ADDITIONAL LIMITATIONS · p. 13
• Hosted endpoints may evolve or be deprecated, while GPT-Live-1 results depend on its voice layer and default backend and are configuration-level comparisons, not causal architecture claims.Source: ADDITIONAL LIMITATIONS · p. 13

A BENCHMARK TAXONOMY AND COMPOSITION

• The eleven dimensions characterize professional operations, organizational context, artifacts, interaction dynamics, execution requirements, and consequence profiles.Table 6 lists labels realized in the 120-workflow benchmark, while Table 7 explains how dimensions are interpreted.Source: A BENCHMARK TAXONOMY AND COMPOSITION · p. 13
• The taxonomy was designed for broad coverage rather than a fully crossed factorial design, so its dimensions co-vary.Source: A BENCHMARK TAXONOMY AND COMPOSITION · p. 13
• Taxonomy slices are descriptive capability diagnostics, not causal evidence.Source: A BENCHMARK TAXONOMY AND COMPOSITION · p. 13

B BENCHMARK CONSTRUCTION AND RUNTIME SPECIFICATION

• Each workflow is instantiated with user state, knowledge, typed tools, an evolving artifact, authorization constraints, frozen speech assets, and grading specifications.Source: B BENCHMARK CONSTRUCTION AND RUNTIME… · p. 13

B.1 USER STATE, REVEAL POLICY, AND LEAKAGE CONTROLS

• Table 6 enumerates realized taxonomy labels, and Table 7 explains how taxonomy dimensions are interpreted.Source: Table 6 · p. 14, Table 7 · p. 14
• The flow engine converts observed agent acts, tool events, artifact changes, approval requests, and seeded events into a semantic user plan without surface wording or future facts.This separation prevents the surface generator from revealing future facts or the reference trajectory.Source: §B.1 · p. 14

B.2 OFFLINE LANGUAGE REALIZATION AND SPEECH SYNTHESIS

• For each reachable critical user plan, 2–5 variants are generated and validated offline, then frozen as task-local text and audio assets; runtime has no free-running user LLM or TTS fallback.Source: §B.2 · p. 15

B.3 RUNTIME ORCHESTRATION AND APPROVAL SEMANTICS

• The orchestrator anchors authored duplex events to media time and records user, agent, tool, artifact, environment, authorization, timing, and grading events in a replayable trace.Approval-gated actions require an action-scoped token, which revocation immediately invalidates.Source: §B.3 · p. 15

B.4 FULL-DUPLEX EVALUATION SETTING

• Systems share a full-duplex runtime and fixed task assets, while the reported campaign controls modality condition, temporal dynamics, and grading mode rather than testing them as axes.Source: §B.4 · p. 15

B.5 CONSTRUCTION QUALITY CONTROL AND WORKFLOW ASSEMBLY

• Deterministic checks and LLM-judge checks validate consistency across workflow components; 16% of candidate workflows were rejected before evaluation.Rejections included semantic inconsistencies, missing information, invalid shortcuts, unauthorized actions, or premature disclosure.Source: §B.5 · p. 15

C EVALUATION AND GRADING

• Scores are computed after execution from the final Voice Workbench state and canonical event trace, with workflow success requiring the terminal state, process, actions, and artifact to be valid.Artifact Field Accuracy separately exposes partial correctness when the end-to-end criterion is not met.Source: C EVALUATION AND GRADING · p. 15

C.1 WORKFLOW SUCCESS

• Workflow success requires all four binary gates—target state, process validity, action completion, and artifact validity—to pass.A single failed gate makes the workflow unsuccessful.Source: §C.1 · p. 16

C.2 ARTIFACT FIELD ACCURACY

• Artifact Field Accuracy measures field correctness separately, without requiring the complete workflow to pass.It is micro-averaged across graded fields.Source: §C.2 · p. 16

C.3 DETERMINISTIC VERIFICATION AND SEMANTIC ESCALATION

• Field grading uses deterministic checks for varied value types, escalating eligible non-matches to a cached temperature-0 semantic judge.State transitions, tool and evidence actions, approval boundaries, and artifact lifecycle conditions are checked directly from the event log.Source: §C.3 · p. 16

C.4 REPEATED-RUN RELIABILITY

• Each workflow is run three times under identical task semantics and a frozen user policy to measure stochastic reliability.The reported metrics distinguish average single-run success, success on at least one attempt, and success on all three attempts.Source: §C.4 · p. 16

D.1 COMMON EVALUATION PROTOCOL

• All five voice systems receive the same 120 workflows, environment state, knowledge, tools, artifact schemas, authorization rules, and frozen user assets; normalized events feed a common grading trace.Agent behavior may induce different valid interaction branches, while user-side selection is deterministic for a fixed workflow, seed, and user-plan identity.Source: §D.1 · pp. 16–17

D.2 PROVIDER-SPECIFIC API AND ADAPTER DETAILS; D.3 DEFERRED SYSTEMS

• Provider-specific adapters normalize streaming and tool-call events; two considered systems were deferred because of impractical run time or a tool-interface mismatch.Nemotron-VoiceChat-11B runs took approximately hours per task and produced empty or garbled tool output; Venus-real-time exposed tool use through natural-language delegation to a separate harness.Source: §D.1 · p. 17, §D.3 · p. 17

D.4 DETERMINISTIC FIELD GRADERS

• Deterministic field graders canonicalize values before comparison and cover exact matches, enums, numeric tolerance, sets, ordered lists, and interval overlap.Table 8 lists the released field graders; semantic and claim atoms additionally use the cached judge.Source: §D.4 · p. 17, Table 8 · p. 18

D.5 SEMANTIC ESCALATION AND THE ARTIFACT-VALIDITY GATE

• Eligible identifier-like, semantic, and numeric fields escalate to a temperature-0 judge only after deterministic matching fails; empty values are never judged.Artifact Validity also checks field accuracy, completeness, stale corrected facts, and artifact lifecycle state; any failed workflow gate forces workflow success to zero.Source: §D.5 · pp. 17–18

E.1 SHARED AGENT INSTRUCTION

• The shared agent prompt requires recording answers, asking for missing information one item at a time, updating corrections, and obtaining approval before gated actions.It also directs agents to finalize the record before wrapping up.Source: §E.1 · p. 19

E.2 FREE-TEXT FIELD JUDGE

• The free-text judge rewards equivalent meaning despite surface-form differences, while failing values that are wrong, miss the core meaning, or are empty.The judge runs at temperature 0 with an on-disk cache; without a configured judge, semantic grading degrades to case-folded equality.Source: §E.2 · pp. 19–20

E.3 OFFLINE USER REALIZER

• User plans are realized offline into validated speech variants, then selected deterministically at runtime without free-running user-language or TTS fallback.This allows adaptive interaction branches while controlling user-side variation.Source: §E.3 · p. 20

F HUMAN VALIDATION OF THE FROZEN USER SIMULATOR; F.1 HUMAN-AUDIT INTERFACE; F.2 STUDY A: REALIZATION FIDELITY AND NATURALNESS

• In the human validation study, frozen realizations scored 4.5/5 for semantic fidelity and 4.2/5 for naturalness, with 4.1% of utterances flagged for unauthorized-fact leakage.Three annotators rated 72 utterances from 24 sampled tasks; the human-audit design preserves workflow semantics while replacing frozen speech with natural human production.Source: §F.1 · p. 20, §F.2 · p. 21, Table 9 · p. 21

F.3 STUDY B: LIVE-HUMAN TRANSFER AUDIT

• The matched 24-task audit uses role cards generated from the simulator’s user state and reveal graph, and counterbalancing helps isolate live speech from task-composition differences.The same audit subset is used for the synthetic comparison; a six-task anchor subset is repeated with a second participant.Source: §F.3 · p. 21
• PASS@1 falls 3.3–7.7 points (4.9 on average) for all five systems with live humans, while their broad relative performance pattern remains stable.The largest decline is 7.7 points for Grok-Voice-Think-2.0; the other systems decline by 3.3–4.6 points.Source: §F.3 · pp. 21–22, Table 10 · p. 22

G.1 INDUSTRY AND WORK-ARTIFACT SLICES

• Taxonomy tables report pass@1 with mean AFA by industry and artifact class; small healthcare and insurance slices warrant caution, and artifact class is coupled to archetype in V1.The tables are descriptive diagnostics, not factorial causal estimates.Source: G EXTENDED QUANTITATIVE RESULTS · p. 22, §G.1 · p. 22

G.2 AUTONOMY, KNOWLEDGE, AND TOOL BURDEN

• Further descriptive slices cover autonomy, knowledge, tool burden, user behavior, approval gating, required-field count, and risk, with correlated task categories not treated as independent interventions.These tables document benchmark composition across execution-oriented dimensions.Source: §G.2 · p. 22, §G.3 · p. 23

H RELIABILITY AND EFFICIENCY ANALYSIS

• The reliability analysis uses three independent rollouts per task-model pair and task-level bootstrap resampling that keeps each task’s runs grouped.A task counts toward RELIABLE@3 only when all three runs satisfy Workflow Success.Source: H RELIABILITY AND EFFICIENCY ANALYSIS · p. 23

H.2 GATE-FAILURE COUNTS

• Appendix diagnostics provide absolute counts for Workflow Success gate failures, per-archetype pass@1, and delivery-quality aggregates separate from the primary completion metrics.The gate counts come from a single reference run over the 120-task benchmark.Source: §H.2 · p. 24, §H.3 · p. 24, §H.4 · p. 24

I.1 MISHEARD PROPER NOUN

• A benefits-enrollment trace shows a misheard surname, “Morgan Ray” instead of “Morgan Reyes,” propagating into a required artifact field despite otherwise normal workflow progress.The examples also include premature or incomplete updates and clean turn-by-turn state capture.Source: §I.1 · pp. 24–25, I REPRESENTATIVE QUALITATIVE EXAMPLES · p. 24

H.1 RELIABILITY–EFFICIENCY FRONTIER

• The efficiency frontier compares time-to-resolution with reliability, and the fastest median TTR does not coincide with the highest RELIABLE@3.AFA versus PASS@1 is excluded as a Pareto frontier because both measure correctness.Source: §H.1 · p. 23, Table 17 · p. 25

I.2 PREMATURE ARTIFACT UPDATE

• A benefits-enrollment agent wrote empty placeholders and proceeded without capturing supplied values, leaving a plausible conversation but an artifact that failed to accumulate the user’s information.The example shows the agent receiving an employee ID and name while the corresponding artifact fields remain empty.Source: §I.2 · p. 25

I.3 CLEAN TURN-BY-TURN STATE CAPTURE

• In the technical-interview example, Grok-Voice-Think-2.0 incrementally records each answer in the persistent artifact as it is heard.The recorded fields include years of experience, primary language, and a system-design example.Source: §I.3 · p. 26

J.1 SOFTWARE ENVIRONMENT

• The software environment uses Python 3.11 in conda, with specified core and TTS/ASR dependencies and TikZ/PGFPlots for paper plots.The listed core dependencies include pydantic v2, numpy, soundfile, scipy, websockets, openai, and google-genai.Source: §J.1 · p. 27

J.2 RUNNING AN EVALUATION CAMPAIGN; J.3 GRADING AND ANALYSIS

• Campaign execution is coverage-first and resumable, while the analysis script generates scorecards, gate decompositions, archetype summaries, and per-task field-miss appendices.Completed task-repetition pairs are skipped, and a per-model lock enables parallel runs.Source: §J.2 · p. 27, §J.3 · p. 27

J.4 DETERMINISM AND SECRETS; J.5 WORKED GRADING EXAMPLE

• The evaluation freezes user audio, replays grading from event logs, and uses a temperature-0 semantic judge with a versioned cache; a worked enrollment run passes TS and PV but fails AC and AV, yielding WS = 0.That run has four required fields wrong and AFA = 10/14 = 0.71 < 1; the field-level table identifies the failed required fields.Source: §J.5 · p. 27, §J.4 · p. 27, Table 21 · p. 28

GPT-REALTIME-2.1

• Across GPT-REALTIME-2.1 form workflows, examples fail through missed knowledge retrieval, incomplete finalization, and artifact errors, while one warranty claim fails artifact validity despite completed retrieval.Examples include benefits enrollment and travel approval failing TS, AC, and AV; expense reporting failing those gates as well, while the warranty claim has 7/9 fields correct with errors in warranty policy and contact phone.Source: GPT-REALTIME-2.1 · pp. 28–30

apexv1 012 — Customer-success manager recruiter screen — Workflow SUCCESS

• The customer-success manager recruiter-screen workflow succeeds with all gates passing and artifact field accuracy 12/12, whereas the warehouse-supervisor screen fails TS, PV, and AV after a correction is left stale.The stale field is the candidate’s current start year.Source: apexv1 012 — Customer-success manager… · p. 31
• The customer-success recruiter-screen records show artifact errors, and the internal-transfer case failed despite 12/12 correct fields because retrieval and submission were missing.The recruiter-screen artifact had 7/12 fields correct; the internal-transfer workflow also recorded an infrastructure/workbench drop.Source: apexv1 012 — Customer-success manager… · p. 31

apexv1 015 — Professional reference check — Workflow SUCCESS

• The contractor qualification workflow passed with 12/12 fields, whereas the returnship screening failed because it was never finalized despite 11/11 correct fields.The professional reference-check task has no outcome reported in the supplied passages.Source: apexv1 015 — Professional reference… · p. 32

apexv1 018 — Internship behavioral screen — Workflow SUCCESS

• Only the internship behavioral screen succeeded among six reported interview and discovery outcomes; five failed through artifact errors or an unmet retrieval gate.The success had 11/11 fields correct; failures included 11/12, 9/11, and 10/11 field accuracy, and one workflow never called kb search.Source: apexv1 018 — Internship behavioral… · pp. 32–33

apexv1 024 — Healthcare operations software discovery — Workflow SUCCESS

• Two of six reported healthcare-operations discovery workflows passed, while four failed from stale or wrong fields, missed retrieval, or both.The successful workflows scored 11/11 and 10/10; failures included a CRM migration record with stale record count and renewal intent, and a partner opportunity with unmet retrieval.Source: apexv1 024 — Healthcare operations… · pp. 33–34

apexv1 030 — Discovery call to CRM plus follow-up package — Workflow SUCCESS

• The CRM discovery and follow-up package workflow failed because required knowledge retrieval was not performed.The supplied record reports that kb search was never called and notes one infrastructure/workbench drop.Source: apexv1 030 — Discovery call to CRM… · p. 35

apexv1 032 — Legal matter intake without legal advice — Workflow SUCCESS

• One reported workflow succeeded with all gates passing and artifact field accuracy 11/11, while three reported failures had 9/11, 10/11, and 9/10 fields correct, with retrieval or artifact issues.The 9/11 failure reported required knowledge retrieval was not satisfied because kb search was never called, with red-flag and urgency fields wrong or missing; the 10/11 record had a stale w2 count, and the 9/10 record had stale impact.Source: apexv1 032 — Legal matter intake… · pp. 35–36
• 8/9 fields were correct, but the shipment-intake workflow remained DRAFT instead of READY FOR REVIEW and had the wrong photo index.The workflow was never submitted or committed.Source: apexv1 032 — Legal matter intake… · p. 36

apexv1 038 — Employee workplace-issue intake and routing — Workflow SUCCESS

• 10/10 fields were correct in workplace-issue intake, while warranty intake missed required knowledge retrieval and client intake failed retrieval, process, and artifact checks.Client intake had 8/11 fields correct, with errors in meeting urgency, appointment slot, and follow-up.Source: apexv1 038 — Employee workplace-issue… · p. 36

apexv1 041 — Enterprise SaaS login failure — Workflow SUCCESS

• 11/11 fields and all gates passed in enterprise SaaS login troubleshooting.Source: apexv1 041 — Enterprise SaaS login… · p. 37
• VPN, POS, API-authentication, and video-conference troubleshooting failed over stale or incorrect corrections, missed knowledge retrieval, or incomplete finalization.The VPN case had 8/11 fields correct; API authentication had 9/10, while video-conference troubleshooting had 8/10.Source: apexv1 041 — Enterprise SaaS login… · p. 37
• Industrial-sensor, data-pipeline, CAD-license, label-printer, and engineering-escalation workflows also failed, with field errors or missed retrieval and process requirements.Sensor diagnosis and label-printer triage each had 9/10 fields correct; data-pipeline freshness and CAD-license checkout each had 8/10.Source: apexv1 041 — Enterprise SaaS login… · p. 38

apexv1 051 — Duplicate invoice charge dispute — Workflow SUCCESS

• Duplicate-invoice dispute passed all gates with 10/10 fields correct.Source: apexv1 051 — Duplicate invoice charge… · p. 39

apexv1 052 — Subscription seat-overage dispute — Workflow SUCCESS

• Two of five supplied subscription-seat-overage dispute outcomes passed; the other three failed on artifact errors or missing terminal completion.The successful case had 9/9 fields correct; damaged-shipment recovery had 7/11, and the telecom and vendor disputes had 8/9 and 8/10, respectively.Source: apexv1 052 — Subscription seat-overage… · pp. 39–40

apexv1 057 — Service cancellation retention boundary — Workflow SUCCESS

• The service-cancellation task required policy retrieval, user approval, and commit after mid-utterance corrections; no outcome is supplied here.Source: apexv1 057 — Service cancellation… · p. 40
• Several workflows failed because required knowledge retrieval or submission was omitted, or corrected fields remained stale, even when most or all artifact fields were accurate.The failures included missing knowledge retrieval, drafts left unsubmitted, and stale corrections; one case reported 10/10 fields correct but remained in DRAFT instead of READY FOR REVIEW.Source: apexv1 057 — Service cancellation… · pp. 40–41

apexv1 064 — Specialist clinic scheduling — Workflow SUCCESS

• Specialist clinic scheduling succeeded with all gates passing and 11/11 artifact fields accurate.The agent sought approval and committed the schedule after the user corrected the slot mid-utterance.Source: apexv1 064 — Specialist clinic… · p. 41

apexv1 065 — Maintenance-window coordination — Workflow SUCCESS

• The maintenance-window workflow failed its knowledge-retrieval gate, while training scheduling failed with an incorrect slot and an unfinalized artifact.The freight task's agent committed after seeking approval, but the workflow still failed because required knowledge retrieval was absent.Source: apexv1 065 — Maintenance-window… · p. 42

apexv1 068 — Customer implementation kickoff coordination — Workflow FAILURE

• Customer kickoff coordination failed despite reported 9/9 field accuracy because the attendee count and agenda remained stale after corrections.The required terminal state was READY FOR REVIEW, but the workflow was not finalized.Source: apexv1 068 — Customer implementation… · p. 42

apexv1 069 — Shared-lab resource booking — Workflow SUCCESS

• 9/11 fields were correct, but the shared-lab workflow failed because required knowledge retrieval was skipped and corrected payment and approval terms remained wrong.The agent sought approval and committed the record, but never called kb search.Source: apexv1 069 — Shared-lab resource… · p. 45

apexv1 081 — Employee benefits eligibility advisor — Workflow SUCCESS

• The employee-benefits advisor passed all gates and achieved 9/9 field accuracy.The workflow completed retrieval and reached the required READY FOR REVIEW state.Source: apexv1 081 — Employee benefits… · p. 45

apexv1 082 — Expense-policy advisor — Workflow SUCCESS

• Expense-policy advising succeeded at 9/9 fields, but travel-policy advising failed at 9/10 and procurement routing at 8/9.The travel artifact remained DRAFT instead of READY FOR REVIEW, and procurement routing had an incorrect item field.Source: apexv1 082 — Expense-policy advisor —… · p. 45

apexv1 085 — Support SLA advisor — Workflow SUCCESS

• The support-SLA memo passed with 9/9 fields, but the retention-policy workflow failed because it remained DRAFT rather than READY FOR REVIEW despite 9/9 correct fields.The retention workflow was not submitted or committed.Source: apexv1 085 — Support SLA advisor —… · p. 46

apexv1 087 — Parental-leave policy explainer — Workflow SUCCESS

• The parental-leave policy explainer passed all gates with 9/9 field accuracy.The workflow completed policy retrieval and reached READY FOR REVIEW after the user’s mid-utterance corrections.Source: apexv1 087 — Parental-leave policy… · p. 46

apexv1 088 — Product-plan fit advisor — Workflow SUCCESS

• Implementation checkpoint and launch readiness succeeded at 10/10 fields each, while project status review failed at 6/10 and vendor commitment remained stale despite 10/10 fields.Across the other examples, failures also involved an uncalled knowledge search, a stale corrected category, or failure to finalize the artifact.Source: apexv1 088 — Product-plan fit advisor… · pp. 47–48

apexv1 099 — Stakeholder research synthesis meeting — Workflow SUCCESS

• Research synthesis left a corrected cut status stale despite 10/10 fields; the HVAC work order lacked retrieval and approval, and the safety checklist remained DRAFT with an incorrect PPE check.The HVAC artifact had 9/11 fields correct; the safety checklist had 9/10.Source: apexv1 099 — Stakeholder research… · p. 49
• 7/10, 8/10, and 8/10 fields were correct in three failed runs, with errors in tolerance, disposition, hold recommendation, kitchen severity, deposit impact, SKU description, or discrepancy.These examples illustrate how incorrect or stale artifact values could fail the workflow despite most fields being correct.Source: apexv1 099 — Stakeholder research… · pp. 49–50

apexv1 106 — Fleet vehicle pre-service inspection — Workflow SUCCESS

• 10/10 fields yielded workflow success in the fleet inspection, but other inspection workflows failed when required kb searches were omitted or corrections left fields stale.The solar-site and data-center inspections required knowledge retrieval; neither called kb search, while the retail checklist left the corrected cash-drawer field stale.Source: apexv1 106 — Fleet vehicle pre-service… · p. 50
• Coordination runs repeatedly failed on omitted kb searches, with additional errors including stale fields, incorrect artifact values, or unsubmitted work.Some runs also failed terminal-state or process gates: payment processing and production quality hold remained in DRAFT, while customer data-sync had a submit failure.Source: apexv1 106 — Fleet vehicle pre-service… · pp. 51–53

GROK-VOICE-THINK-2.0

• 12/14 fields were correct in benefits enrollment, 10/11 in the expense report, and 8/10 in travel approval, yet each workflow failed.Benefits enrollment also failed retrieval and approval gates; the vendor-onboarding run failed because the gold document was not retrieved.Source: GROK-VOICE-THINK-2.0 · p. 54
• All five Grok form-fill workflows failed: one was submitted without approval, while the others had missing or incorrect artifact fields and one also missed the required retrieval.Field accuracy ranged from 6/9 to 9/10; the reimbursement packet searched but did not retrieve the gold document.Source: GROK-VOICE-THINK-2.0 · p. 55

apexv1 010 — Facility access badge request — Workflow SUCCESS

• The facility access badge request was the only fully successful workflow in this part, passing all gates with 10/10 artifact fields correct.Source: apexv1 010 — Facility access badge… · p. 56

apexv1 011 — Software engineer recruiter screen — Workflow SUCCESS

• The software engineer recruiter screen succeeded with 13/13 fields, but six other interview workflows failed artifact validation, with accuracy from 9/11 to 11/12 fields.Reported errors included missed corrections and inaccurate or missing evidence fields.Source: apexv1 011 — Software engineer… · pp. 56–57

apexv1 017 — Contractor qualification call — Workflow FAILURE

• In the contractor qualification call and related workflows, interview and discovery tasks failed artifact validation, with one interview debrief missing its record status.The resume-discrepancy screening had 10/12 fields correct; the debrief had 8/11, and discovery results ranged from 8/10 to 10/11. The CRM migration task also failed required retrieval after searching without retrieving the gold document.Source: apexv1 017 — Contractor qualification… · pp. 57–59
• Despite its FAILURE label, contractor qualification passed all gates with 10/10 fields after mid-utterance timeline and data-source corrections.Source: apexv1 017 — Contractor qualification… · p. 59

apexv1 028 — Renewal expansion discovery — Workflow SUCCESS

• CRM and intake results were mixed: renewal discovery and maintenance intake succeeded at 10/10, while other cases failed on missing or stale fields or unmet knowledge retrieval.Examples include a missing deal size, stale next step, and insurance and escalation cases where the required gold document was not retrieved.Source: apexv1 028 — Renewal expansion… · pp. 59–61

apexv1 041 — Enterprise SaaS login failure — Workflow SUCCESS

• Among four troubleshooting cases, login succeeded at 11/11; VPN and API artifacts retained correction errors, while POS also violated approval and submission requirements.The POS case was committed without seeking approval and had a failed submit tool call.Source: apexv1 041 — Enterprise SaaS login… · p. 62

apexv1 045 — Video-conference audio issue — Workflow SUCCESS

• Only the video-conference case succeeded at 10/10; sensor, pipeline, and CAD cases failed with stale or missing fields after user corrections.Reported errors included a missing signal strength, incorrect affected pipeline and job status, and a stale business unit.Source: apexv1 045 — Video-conference audio… · p. 63
• The two troubleshooting cases failed: the printer ticket had 9/10 fields correct but the wrong cause, while engineering escalation missed retrieval and approval and failed submission.The escalation artifact had 10/11 fields correct, including an incorrect change-approved value.Source: apexv1 045 — Video-conference audio… · pp. 63–64

apexv1 051 — Duplicate invoice charge dispute — Workflow SUCCESS

• The invoice-dispute negotiation succeeded at 10/10 fields, while other negotiation workflows failed with stale corrections, field errors, or unsatisfied retrieval and approval gates.Examples include seat-overage at 8/9 fields with a stale claimed-seats value, damaged shipment at 7/11 with several incorrect fields, and dispute resolution at 9/11 with a missed retrieval requirement and stale remedy.Source: apexv1 051 — Duplicate invoice charge… · pp. 64, 66
• Two coordination workflows failed after submission without approval, and field-service dispatch left the corrected technician stale despite 11/11 fields marked correct.The executive-meeting and interview-scheduling workflows also had submit tool-call failures; the dispatch workflow failed its process and artifact-validation gates.Source: apexv1 051 — Duplicate invoice charge… · p. 66

apexv1 065 — Maintenance-window coordination — Workflow SUCCESS

• Maintenance-window coordination and customer kickoff succeeded at 10/10 and 9/9 fields, respectively, while freight and training scheduling failed with incorrect corrected details.Freight scheduling had 8/11 fields correct, including errors in pickup delay, transit time, and delivery slot; training scheduling had 8/9, with the chosen slot wrong.Source: apexv1 065 — Maintenance-window… · p. 67

apexv1 069 — Shared-lab resource booking — Workflow SUCCESS

• Travel-disruption rebooking failed both approval and artifact gates, with calendar-update and itinerary-status fields incorrect and a submission tool-call failure.The artifact had 9/11 fields correct, and the workflow committed without seeking approval.Source: apexv1 069 — Shared-lab resource… · p. 68
• Five workflows failed their gates, with artifact accuracy of 9/12, 7/11, 8/10, 9/10, and 7/10 fields; one was committed/submitted without seeking approval.Errors involved quantities and costs, negotiation terms, carrier offer and SLA, vegetarian quantity, and redline summary and disposition.Source: apexv1 069 — Shared-lab resource… · pp. 68–69

apexv1 076 — Software-license volume purchase — Workflow SUCCESS

• The software-license purchase passed all gates at 10/10, while the packaging negotiation failed despite 11/11 fields because disposition remained stale and approval was skipped.The packaging workflow also had a submit-tool failure.Source: apexv1 076 — Software-license volume… · p. 69

apexv1 078 — Event venue negotiation — Workflow SUCCESS

• The venue negotiation passed at 10/10, but the maintenance and DataPipe negotiations failed at 9/10 and 9/11, and four advisor workflows failed with artifact accuracy from 7/9 to 9/10.Advisor errors included incorrect elections, expense guidance, travel compliance, and procurement item/category; the negotiation failures involved a missing annual cost and incorrect payment terms.Source: apexv1 078 — Event venue negotiation —… · pp. 69–70

apexv1 085 — Support SLA advisor — Workflow SUCCESS

• The support SLA advisor passed all gates at 9/9, while data-retention, parental-leave, and product-plan advice failed with artifact accuracy of 7/9, 7/9, and 6/9.Their errors included a retention period and open item, stale scope and incorrect cited basis, and incorrect plan/support fields.Source: apexv1 085 — Support SLA advisor —… · p. 71

apexv1 089 — Returns and warranty policy advisor — Workflow SUCCESS

• Among the returns-and-warranty and facilitation workflows, three failed at 8/9, 9/10, and 9/10, while the customer implementation checkpoint passed at 10/10.The failures involved a missing deadline, a missing action due date, and an incorrect overall project status.Source: apexv1 089 — Returns and warranty… · p. 72

apexv1 094 — Requirements workshop — Workflow FAILURE

• In the requirements workshop, both workflows failed: one artifact had 7/10 correct fields, while the other had 8/10 and missed the gold retrieval document with corrected fields left stale.The first missed option a, rationale, and decision status; the second retained stale event time and order and the wrong action owner.Source: apexv1 094 — Requirements workshop —… · p. 73

apexv1 102 — Safety pre-job verbal checklist — Workflow SUCCESS

• Only the safety checklist passed, with 10/10 fields; quality, property, inventory, and fleet inspections failed with 7/10, 8/10, 9/10, and 8/10 fields, respectively.Errors included incorrect disposition and hold recommendation, kitchen severity and deposit impact, SKU description, and vehicle fluid levels and priority.Source: apexv1 102 — Safety pre-job verbal… · pp. 74–75

apexv1 109 — Solar-site maintenance inspection — Workflow SUCCESS

• All six solar-site and coordination workflows failed, with unmet retrieval or approval gates and/or incorrect or stale fields after corrections; four also had submit failures.The cases included a stale recommendation, unapproved submissions, and incorrect or stale incident fields; the component-shortage task also missed the gold retrieval document and used the wrong allocation.Source: apexv1 109 — Solar-site maintenance… · pp. 76–77

apexv1 116 — Facilities water-leak escalation — Workflow SUCCESS

• One workflow passed all gates with artifact field accuracy 11/11, while another failed gate AV, with artifact incomplete: 9/10 fields correct and overall status wrong or missing.Source: apexv1 116 — Facilities water-leak… · p. 77

apexv1 118 — Production quality hold coordination — Workflow SUCCESS

• Production quality-hold coordination succeeded at 10/10 fields, while a near-miss escalation failed with 9/11 fields, including incorrect cause and immediate action.The failed case had retrieved policy, sought approval, and committed the artifact, but still missed required content.Source: apexv1 118 — Production quality hold… · p. 78

GEMINI-3.8-LIVE

• Gemini-3.8-Live’s form workflows repeatedly failed, with outcomes citing missing knowledge-base searches and artifacts ranging from 10/14 to 0/9 correct fields.Other cases also failed after stale corrections or incomplete terminal-state actions.Source: GEMINI-3.8-LIVE · pp. 79–81

apexv1 011 — Software engineer recruiter screen — Workflow SUCCESS

• Software-engineer screening succeeded with all 13 fields correct after mid-utterance corrections.Source: apexv1 011 — Software engineer… · p. 81

apexv1 012 — Customer-success manager recruiter screen — Workflow SUCCESS

• Customer-success screening succeeded at 11/12 fields, while warehouse and another recruiter screen failed with 11/12 and 10/12 fields, respectively.The warehouse case missed a certification detail; the other failure also lacked required knowledge retrieval.Source: apexv1 012 — Customer-success manager… · pp. 81–82

apexv1 015 — Professional reference check — Workflow SUCCESS

• The professional reference-check workflow succeeded with all 12 evidence-matrix fields correct.Source: apexv1 015 — Professional reference… · p. 82

apexv1 016 — Returnship program screening interview — Workflow SUCCESS

• Returnship screening and contractor qualification succeeded at 11/11 and 12/12 fields, respectively, but internship screening failed with only 1/11 fields and remained a draft.Source: apexv1 016 — Returnship program… · p. 82
• Three workflows failed: one had 0/12 artifact fields correct, another had 8/11 correct, and a third failed because required knowledge retrieval was unsatisfied when kb search was never called.The 0/12 workflow remained DRAFT rather than READY FOR REVIEW.Source: apexv1 016 — Returnship program… · p. 83

apexv1 023 — Manufacturing automation discovery — Workflow SUCCESS

• The manufacturing discovery workflow required an 11-field CRM record and repair of three mid-utterance corrections, but no outcome is provided.The corrections concerned in-scope detail, lines in scope, and throughput goal.Source: apexv1 023 — Manufacturing automation… · p. 83

apexv1 024 — Healthcare operations software discovery — Workflow SUCCESS

• Healthcare discovery included 11/11 and 10/10 successes, but failures included a stale-duration 9/10 record and an unfinalized 0/10 record after skipped retrieval.A further workflow failed because kb search was never called and a required follow-up action was missing.Source: apexv1 024 — Healthcare operations… · pp. 84–85

apexv1 032 — Legal matter intake without legal advice — Workflow SUCCESS

• Legal intake had 11/11 and 10/10 successes, while appointment, tax, escalation, and shipment cases failed with 8/11, 9/11, 9/10, and 8/9 fields correct.Failures included skipped knowledge retrieval and stale, incorrect, or missing artifact fields.Source: apexv1 032 — Legal matter intake… · pp. 85–86

apexv1 038 — Employee workplace-issue intake and routing — Workflow SUCCESS

• Client intake, SaaS login support, and VPN troubleshooting failed with 9/11, 9/11, and 7/11 fields correct; the warranty case also failed retrieval and artifact gates at 8/10.The login case left a corrected action stale, while the VPN case retained incorrect error code and cause and lacked a resolution.Source: apexv1 038 — Employee workplace-issue… · p. 87
• Workplace-issue workflows failed when knowledge retrieval, approval or terminal-state gates were missed, or corrected ticket fields remained stale—even with 10/10 fields correct.The sensor diagnosis had all fields correct but failed the required knowledge-retrieval gate; other cases also lacked required approval or a committed terminal state.Source: apexv1 038 — Employee workplace-issue… · pp. 88–89
• Approval and submission requirements also caused negotiation failures: some records were submitted without approval or remained uncommitted despite attempted submission.The duplicate-invoice and follow-up-dispute workflows missed approval, knowledge retrieval, and terminal-state requirements.Source: apexv1 038 — Employee workplace-issue… · pp. 89, 91
• Negotiation workflows repeatedly failed because corrections were stale or artifacts were incomplete, including incorrect remedies, dispositions, and policy positions.Examples include a damaged-shipment record with 6/11 fields correct, a telecom-credit record with 6/9, and a cancellation-retention record with a stale requested discount.Source: apexv1 038 — Employee workplace-issue… · pp. 90–91

apexv1 062 — Candidate interview-loop scheduling — Workflow SUCCESS

• The candidate interview-loop scheduling workflow succeeded, passing all gates with 11/11 artifact fields correct.By contrast, field-service dispatch failed the required knowledge-retrieval gate, and clinic scheduling failed with missing or incorrect fields.Source: apexv1 062 — Candidate interview-loop… · p. 92
• The freight-coordination workflow failed retrieval, approval, terminal-state, and artifact checks with 8/11 fields correct; training scheduling was not finalized and had only 3/9 fields correct.In the training case, the agent made three tool calls and left the artifact in DRAFT rather than READY FOR REVIEW.Source: apexv1 062 — Candidate interview-loop… · p. 93

apexv1 068 — Customer implementation kickoff coordination — Workflow FAILURE

• Across coordination and negotiation cases, workflows failed on retrieval, approval, or terminal-state gates and stale or incomplete fields, including freight rebooking at 8/11 fields and renewal negotiation at 10/11.Other cases also showed stale corrections and incomplete artifacts, including contractor SOW, software-license, packaging, and data-platform negotiations.Source: apexv1 068 — Customer implementation… · pp. 93–96

apexv1 081 — Employee benefits eligibility advisor — Workflow SUCCESS

• Benefits eligibility advice passed all gates with 8/9 fields correct, while expense, travel, procurement-routing, and SLA-advice cases had artifact errors or missed retrieval.The SLA-advice artifact had 9/9 fields correct except for a stale applicable-SLA field, and kb search was never called.Source: apexv1 081 — Employee benefits… · pp. 96–97

apexv1 086 — Data-retention policy advisor — Workflow SUCCESS

• In the data-retention advice snapshot, the agent completed knowledge retrieval but did not finalize the memo, despite a requirement to reach READY FOR REVIEW.The workflow required a nine-field memo and correction handling for legal-hold information.Source: apexv1 086 — Data-retention policy… · p. 97
• Data-retention-policy tasks ranged from artifact field accuracy 9/9 and 10/10 to failures involving stale corrections, missed retrieval, or an unfinalized draft.Source: apexv1 086 — Data-retention policy… · pp. 97–99

apexv1 098 — Launch readiness meeting — Workflow FAILURE

• Both launch-readiness and research-synthesis workflows failed: one remained an unfinalized draft with incorrect fields, and the other retained a stale corrected claim.Source: apexv1 098 — Launch readiness meeting… · p. 99

apexv1 100 — Budget planning meeting record — Workflow SUCCESS

• Budget-record tasks included a 10/10 success, but failures involved omitted retrieval and approval, incorrect fields after corrections, and artifacts with multiple wrong fields.Source: apexv1 100 — Budget planning meeting… · pp. 100–101

apexv1 106 — Fleet vehicle pre-service inspection — Workflow SUCCESS

• Fleet-inspection cases included a 10/10 success, while failures involved missed required retrieval and incorrect artifact fields.Source: apexv1 106 — Fleet vehicle pre-service… · p. 101
• Most reported coordination failures lacked the required knowledge-base search, including cases that otherwise sought approval or reached a committed state.The shipment-delay and supply-chain cases also required READY FOR REVIEW, and their reported outcomes remained failures.Source: apexv1 106 — Fleet vehicle pre-service… · pp. 102–103
• 10/10 artifact fields were correct in the live-event coordination success, while other coordination cases report failures involving retrieval, stale corrections, authorization, or finalization.Failures include a fleet vehicle work order with 8/11 fields correct and an access-anomaly case that missed required retrieval and failed submission.Source: apexv1 106 — Fleet vehicle pre-service… · pp. 102, 104

STEP-AUDIO3

• 8/9 fields were correct in the warranty-claim form, but the workflow remained a draft and the corrected warranty policy was stale.Vendor onboarding also failed its retrieval gate and recorded an infrastructure/workstation drop.Source: STEP-AUDIO3 · pp. 105–106
• 3/9 fields were correct in the parental-leave form, which remained a draft despite retrieval being attempted; the searched document was not the gold document.Other form workflows report stale fields after mid-utterance corrections, missing retrieval, and unsubmitted drafts.Source: STEP-AUDIO3 · pp. 105–106
• STEP-AUDIO3 cases repeatedly failed after corrected values remained stale, including requested hours, throughput, report fields, and data retention.The professional reference check had 12/12 fields correct but still failed because direct-reports values were stale.Source: STEP-AUDIO3 · pp. 107–110
• One STEP-AUDIO3 contractor qualification call passed all gates with artifact field accuracy of 12/12.Source: STEP-AUDIO3 · p. 108
• STEP-AUDIO3 workflows also failed when knowledge retrieval was skipped or required submissions remained drafts rather than reaching the specified terminal state.Some CRM workflows had all fields correct but were not submitted, while the cybersecurity expansion and CRM migration cases also omitted required kb searches.Source: STEP-AUDIO3 · pp. 108–110
• Two workflow failures had incomplete artifacts: one had 8/10 fields correct and failed TS and AV gates without finalization; another had 10/11 fields correct, missing the open question, and failed AV.Source: STEP-AUDIO3 · pp. 108, 110

apexv1 027 — Customer data-platform qualification — Workflow FAILURE

• In the apexv1 027 cases, workflows failed despite 10/10 or 11/11 fields being correct when corrections stayed stale or required knowledge retrieval was skipped.Other outcomes included a 9/10 artifact and a 7/10 artifact, alongside failures to retrieve the required record.Source: apexv1 027 — Customer data-platform… · pp. 110–111
• All four customer-data intake workflows failed, with artifacts ranging from 3/11 to 8/11 correct and none finalized.One agent searched but missed the gold document; another did retrieve, but still left the artifact incomplete and unfinalized.Source: apexv1 027 — Customer data-platform… · pp. 111–112

apexv1 035 — Property-management maintenance intake — Workflow FAILURE

• Knowledge retrieval failed across several cases because the agent never searched or did not retrieve the required document.These failures were reported alongside incomplete artifacts and, in several cases, work that was never submitted or committed.Source: apexv1 035 — Property-management… · pp. 112, 113, 115, 116
• Across maintenance, intake, and troubleshooting cases, artifacts remained incomplete or incorrect, with correction errors sometimes leaving fields stale.Examples ranged from 0/9 fields correct to 10/11, while other outcomes reported stale values after mid-utterance corrections.Source: apexv1 035 — Property-management… · pp. 112, 114, 115, 116
• Several workflows were not finalized, and one approval-controlled workflow was submitted without seeking approval but still failed to reach COMMITTED.That case also had an incomplete artifact and a failed submit tool call.Source: apexv1 035 — Property-management… · pp. 113–115

I.2 PREMATURE ARTIFACT UPDATE

• Negotiation traces also ended unfinalized with empty or incomplete artifacts: damaged-shipment recovery scored 0/11, while dispute resolution scored 3/11.Both workflows required knowledge retrieval, approval, and commit/submit; neither was finalized, and approval was not sought.Source: apexv1 035 — Property-management… · pp. 116, 118
• 10/10 fields were marked correct in both the vendor SLA and cloud-credit disputes, yet corrected fields remained stale and required knowledge retrieval was unmet.The stale fields were delayed portion and valid exception in the vendor dispute, and spike 2 and credit amount in the cloud-credit dispute.Source: apexv1 035 — Property-management… · pp. 117–118
• Candidate interview scheduling was committed with approval sought but still failed because the corrected unavailable interviewer and replacement remained stale.All 11 fields were counted correct, but the two corrected fields were stale.Source: apexv1 035 — Property-management… · p. 118
• Coordination workflows also failed before finalization despite near-complete artifacts, including maintenance-window coordination at 9/10 and travel rebooking at 9/11 fields correct.Both traces left corrected slots or flight information stale and were not submitted or committed.Source: apexv1 035 — Property-management… · pp. 119–120

apexv1 035 — Property-management maintenance intake — Workflow FAILURE

• Required knowledge retrieval was missed in several failed workflows, alongside stale corrections or incomplete artifacts.The outcomes specify that kb search was never called, including cases with 9/10 and 4/9 fields correct.Source: apexv1 035 — Property-management… · pp. 121, 123, 124
• Mid-utterance corrections often remained stale, causing failures even when artifacts were otherwise mostly complete.Examples include 9/10, 8/9, and 7/9 fields correct, with corrected fields left stale.Source: apexv1 035 — Property-management… · pp. 121, 123, 124
• Two workflows had 0 tool-calls and no completed artifact: one required retrieval, approval, and commit, while the other required a READY FOR REVIEW record.The first remained empty rather than COMMITTED; the second remained empty rather than READY FOR REVIEW.Source: apexv1 035 — Property-management… · p. 122
• A retrieved vendor-purchase workflow still failed because approval was not sought and the artifact remained DRAFT, with 7/11 fields correct.The vendor payment terms were also incorrect, and final terms and status were missing.Source: apexv1 035 — Property-management… · p. 122
• 10/10 fields correct still failed when the artifact remained DRAFT instead of reaching READY FOR REVIEW; one such case also had an infrastructure drop.Another workflow had 9/9 fields correct but still failed with stale fields and a DRAFT lifecycle.Source: apexv1 035 — Property-management… · pp. 122, 125

apexv1 094 — Requirements workshop — Workflow FAILURE

• The next listed case begins a requirements workshop, but the supplied passage gives only its task label and control setup.It identifies a facilitation archetype, business analyst role, Software/SaaS domain, and prepare-only autonomy.Source: apexv1 094 — Requirements workshop —… · p. 126
• Requirements-workshop workflows also failed: corrections remained stale, required fields were missing, and the incident postmortem omitted knowledge retrieval and finalization.The incident postmortem had 9/10 fields correct but a stale event order; the vendor review had 10/10 fields correct yet stale commitments and remained in DRAFT.Source: apexv1 094 — Requirements workshop —… · p. 126

apexv1 098 — Launch readiness meeting — Workflow FAILURE

• Several launch-readiness work orders and coordination workflows failed required retrieval or authorization gates and remained unfinalized, with some artifacts substantially incomplete.The HVAC work order and rack inspection omitted required knowledge searches; the rack artifact had 2/10 fields correct, while a customer-summary work order also lacked retrieval, approval, and finalization.Source: apexv1 098 — Launch readiness meeting… · pp. 127–129
• Launch-readiness inspections failed with stale corrected fields or other artifact errors, including cases with 7/10 or 8/10 fields correct and one with 10/10 but stale values.Examples include stale readiness and a missing sign-off, stale measurements and disposition, and a stale kitchen severity despite 10/10 fields marked correct.Source: apexv1 098 — Launch readiness meeting… · pp. 127–128
• Launch-readiness coordination failures included missing required fields and stale corrections; a shipment-delay artifact had 10/10 fields correct but a stale risk field.The access-anomaly artifact had 2/10 fields correct, with multiple required fields missing; the shipment-delay case failed retrieval and artifact gates despite its field count.Source: apexv1 098 — Launch readiness meeting… · p. 130
• In the four coordination workflows, all failed; retrieval was omitted, corrections remained stale, and approval or terminal-state requirements were unmet.The quality-hold workflow was submitted without approval and still failed to reach COMMITTED; the payment-outage workflow retained stale corrected fields.Source: apexv1 098 — Launch readiness meeting… · pp. 130–131

apexv1 119 — Live event AV failure coordination — Workflow FAILURE

• Both live-event coordination workflows failed: one had 10/10 fields correct but remained a draft, while the other missed retrieval, approval, and required fields.The latter had 8/11 fields correct and did not finalize; both reports noted an infrastructure/workspace drop.Source: apexv1 119 — Live event AV failure… · pp. 131–132

GPT-LIVE-1

• Across GPT-LIVE-1 form workflows, failures ranged from 1/10 to 8/9 correct fields, with missed or ineffective retrieval and incomplete finalization recurring.The travel request had 1/10 fields correct; the warranty claim had 8/9, with a serial-number error. The access request was left as a draft, while expense submission failed.Source: GPT-LIVE-1 · pp. 133–134
• In three reported workflow failures, required knowledge retrieval was attempted but the gold document was not retrieved; artifacts remained incomplete, with 6/9, 6/9, and 5/9 fields correct, respectively.Source: GPT-LIVE-1 · pp. 134–135

apexv1 010 — Facility access badge request — Workflow SUCCESS

• The facility badge request was the sole successful workflow in this group, passing all gates with 10/10 field accuracy.Source: apexv1 010 — Facility access badge… · p. 135
• All four sampled recruiter-screen workflows failed, with artifact accuracy ranging from 1/12 to 11/12 fields.The warehouse screen had only 1/12 fields correct, while the other failures included wrong scale or team-size details and an incorrect open-gap entry.Source: apexv1 010 — Facility access badge… · pp. 135–136
• 8/12 fields were correct in the reference check, which remained DRAFT, and returnship screening had only 1/11 fields correct.Both workflows failed; the reference check also missed the READY FOR REVIEW terminal state.Source: apexv1 010 — Facility access badge… · p. 136

apexv1 017 — Contractor qualification call — Workflow FAILURE

• 10/12 fields were correct in contractor qualification, while interview debrief and analytics discovery each had 8/11 fields correct and failed.The contractor call had incorrect availability and rate; the debrief had an incorrect technical score and concern note, and analytics discovery had incorrect user counts and timeline.Source: apexv1 017 — Contractor qualification… · pp. 136–137

apexv1 022 — Cybersecurity expansion discovery — Workflow SUCCESS

• 10/10 artifact field accuracy and all gates passed in cybersecurity expansion discovery, while manufacturing and professional-services discovery failed at 1/11 and 1/10 fields correct.CRM migration also failed the knowledge-retrieval gate despite 9/10 fields correct.Source: apexv1 022 — Cybersecurity expansion… · pp. 137–138

apexv1 027 — Customer data-platform qualification — Workflow FAILURE

• CRM qualification failed at 8/10 and 1/10 fields correct; intake cases included 10/11 fields with retrieval unsatisfied and 8/11 fields with incorrect red-flag handling.Several CRM qualification and intake workflows failed the artifact-accuracy gate, and some also failed knowledge retrieval.Source: apexv1 027 — Customer data-platform… · pp. 138–139

apexv1 035 — Property-management maintenance intake — Workflow FAILURE

• 9/10 fields were correct in property-maintenance intake, while escalation intake had 8/10 fields correct; both workflows failed artifact validation.The maintenance record mislabeled issue type, and the escalation record had incorrect severity and impact.Source: apexv1 035 — Property-management… · p. 140

• 8/9 fields correct in shipment intake still left the photo index wrong; HR intake had 9/10 correct with a wrong event date, and warranty intake had 8/10 with two fields missing.Source: apexv1 035 — Property-management… · pp. 140–141
• Client intake had 7/11 fields correct and missed retrieval and approval; POS triage and cancellation retention were submitted without required approval and remained DRAFT.Source: apexv1 035 — Property-management… · pp. 141, 142, 145
• 7/11 fields correct in SaaS login and 6/11 in VPN troubleshooting coincided with failure to retrieve the gold document; API authentication had 6/10 correct.Source: apexv1 035 — Property-management… · pp. 141–142
• Sensor diagnosis had 8/10 fields correct but missed the gold document; data-pipeline intake left affected system and job status wrong, while CAD intake skipped retrieval and submission.Source: apexv1 035 — Property-management… · pp. 142–143
• Three workflow failures had 7/9, 6/11, and 5/9 fields correct; the first got claimed seats 100 vs 130, the second missed damage evidence and remedy fields, and the third missed outage-related fields.Source: apexv1 035 — Property-management… · p. 144
• In other negotiation and support cases, ticket number and fare rule were wrong or missing, vendor penalty fields were inaccurate despite a retrieval failure, and audio troubleshooting retained wrong test-result and cause fields.Source: apexv1 035 — Property-management… · pp. 142, 144, 145

apexv1 035 — Property-management maintenance intake — Workflow FAILURE

• All three negotiation workflows failed: artifact accuracy ranged from 1/10 to 9/10 fields, with failures also involving missed retrieval, unapproved submission, and incomplete finalization.One case had only 1/10 fields correct; another had 9/10 but missed the corrected spike-2 value; the approval-gated case committed without approval and failed to reach COMMITTED.Source: apexv1 035 — Property-management… · pp. 145–146

apexv1 061 — Executive meeting across time zones — Workflow SUCCESS

• The executive-meeting and customer-kickoff workflows passed all gates (11/11 and 9/9 fields), while other coordination cases failed on artifact accuracy, retrieval, approval, or submission.Failures included a 2/11-field interview schedule, an uncalled knowledge search and missing approval for field dispatch, and a freight workflow left in DRAFT.Source: apexv1 061 — Executive meeting across… · pp. 146–147

apexv1 076 — Software-license volume purchase — Workflow SUCCESS

• One workflow passed all gates with artifact field accuracy 10/10, while the other three artifacts were incomplete, with 2/11, 7/10, and 9/10 fields correct.The 2/11-field artifact had missing supplier and volume details; the 7/10- and 9/10-field artifacts lacked a disposition, and the 7/10 artifact also had incorrect venue-offer and counter terms.Source: apexv1 076 — Software-license volume… · p. 149
• The four reported advisor outcomes all failed artifact validation: 9/11, 8/9, 7/9, and 9/10 fields were correct, with term corrections among the errors.The benefits-advisor case also failed its knowledge-retrieval gate because the gold document was not retrieved.Source: apexv1 076 — Software-license volume… · p. 150

apexv1 084 — Procurement-policy routing advisor — Workflow SUCCESS

• Reported policy-advisor outcomes failed artifact validation: two had 6/9 fields correct, one had 7/9, and another had 7/9 with checklist and open items missing.The data-retention case also lacked a cited basis and open items; the parental-leave case lacked a checklist and open items.Source: apexv1 084 — Procurement-policy… · p. 151

apexv1 089 — Returns and warranty policy advisor — Workflow SUCCESS

• The returns-and-warranty advisor task required a policy lookup and a READY FOR REVIEW memo while correcting the purchase date and eligibility mid-utterance; no outcome is supplied.The required memo had nine fields, including the exact purchase date and return window.Source: apexv1 089 — Returns and warranty… · p. 151

apexv1 090 — Compliance filing routing advisor — Workflow SUCCESS

• Compliance filing succeeded at 9/9 fields, but retrospective action capture failed at 8/10 and project status review at 3/10.The retrospective errors were the action owner and due date; the status review included placeholder project and workstream fields.Source: apexv1 090 — Compliance filing routing… · p. 152

apexv1 093 — Customer implementation checkpoint — Workflow FAILURE

• Customer implementation failed at 1/10 fields, while requirements-design and incident-postmortem workflows failed at 8/10 fields each.The incident-postmortem errors included event order and action-owner details.Source: apexv1 093 — Customer implementation… · pp. 152–153

apexv1 097 — Vendor performance review meeting — Workflow SUCCESS

• In the vendor-review cases, artifacts remained incomplete, from 6/10 to 9/10 correct fields, with errors in inspection measurements, inventory identifiers, and store-opening details.The reported mismatches included a wrong diameter and disposition, a mistaken SKU and description, and a missing cash-drawer check and action.Source: apexv1 097 — Vendor performance review… · pp. 154–155
• Other vendor-review coordination runs also failed through retrieval or artifact mistakes, including a solar inspection with retrieval skipped and only 1/10 fields correct.The shipment-delay run searched without retrieving the gold document and had an incorrect alternate route; the outage-timeline run also had incorrect service and blast-radius fields.Source: apexv1 097 — Vendor performance review… · pp. 155–156
• Several coordination runs also failed process gates: retrieval was skipped or did not return the gold document, approval was not sought, and submissions or terminal states were invalid.The production-quality hold run missed retrieval and approval, while an outage run lacked approval and failed its submit call.Source: apexv1 097 — Vendor performance review… · p. 156, apexv1 114 — Customer data-sync… · pp. 157–158

apexv1 114 — Customer data-sync incident escalation — Workflow SUCCESS

• The other coordination cases show artifact errors despite some completed actions, including a wrong supplier allocation and recovery plan, an incorrect payment status and owner, and missing near-miss details.A facilities-leak run reached DRAFT rather than COMMITTED despite 10/11 fields being correct.Source: apexv1 114 — Customer data-sync… · pp. 157–158
• 7/11 fields were correct in the production quality-hold case, which also missed retrieval and approval, failed to reach COMMITTED, and had incorrect or missing artifact fields.The test result and hold scope were wrong, and the action and status were missing.Source: apexv1 114 — Customer data-sync… · p. 157