PaperlayerView exact revision on arXiv

Source-linked AI summary

The Last Human Gate: Forward Deployed Engineering for Governance Automation

Jeremy Canale

arXiv:2609.29345v1cs.AIcs.SE

TL;DR

Governance review must preserve decisions, evidence, and accountable authority, but need not preserve every review task's current human implementation. The paper develops a contract-based substitution framework and controlled evaluations, finding that specified governance reviews can be executed by agents and software while the workforce effect remains a deployment hypothesis.

  • Problem

    Governance needs decisions, supporting evidence, and accountable authority, but those functions do not require every review task to retain its current human implementation.

    Source: §1 · p. 1

  • Method

    The paper treats review gates as executable contracts and links information, decision, authority, and execution requirements to a task-substitution framework and controlled benchmark.

    Source: §1 · p. 2, §1.1 · p. 2, §2.2 · p. 4

  • Results

    20.90 percentage points is Gemini’s gate advantage over DeepSeek in the controlled benchmark, while the paper demonstrates agent execution of specified reviews and complete project routes.

    Source: §6.3 · p. 19, §8 · p. 26

  • Takeaways & Limitations

    Governance obligations can remain in force while software takes over accepted review tasks, with substitution assessed by complete human work at comparable output and quality.

    Source: §8 · p. 26

    No finite benchmark here establishes full elimination of human DGF execution, and the workforce-displacement target remains a deployment hypothesis.

    Source: §7.1 · p. 23

Abstract

from arXiv · show

Enterprise governance requires decisions, evidence, and accountable authority; it does not require every review task to retain its current human implementation. We develop a task-substitution framework for Digital Governance Frameworks (DGF), treating each gate as an executable contract. Substitution requires sufficient accessible information, valid decision and authority checks, and a reduction in total human work after exceptions, verification, correction, and maintenance are counted. We derive a residual-work threshold and show why automating most cases can still increase labor. Forward deployed engineering connects these conditions to an architecture for agents, rule engines, evidence services, and escalation. DGF-Bench supplies controlled evidence from 300 synthetic projects and 899 evaluable model-project runs. Gemini 3.8 Flash, GPT-5.6 Luna, and DeepSeek v4.1 Flash achieve strict gate success of 94.98%, 83.29%, and 74.18%; complete-route success is 76.92%, 42.33%, and 24.67%. A deterministic control passes all 1,700 gates given the supplied rules and structured facts, locating the comparison in execution of a supplied decision kernel. Evidence audits and 135 repeated runs distinguish correct decisions from reliable execution. A document counterexample establishes an information-sufficiency obstruction. These results support the technical feasibility of replacing human execution of specified governance-review tasks with agents and software. The framework specifies a workforce test based on the complete human effort required at fixed output and quality; the present measurements concern review performance. Sources, dossiers, traces, and analyses are public.

1 Introduction: replacing execution, preserving the function

Governance review functions must continue, but their current human implementation is not technically required for every task. The paper frames substitution around executable review contracts and separates gate execution, route reliability, and total human work as distinct questions.

1 Introduction: replacing execution, preserving the function

  • Governance outputs such as decisions and supporting work persist even when software performs tasks previously assigned to analysts, architects, and committee staff.Professional titles allocate tasks organizationally; they do not specify how those tasks must be executed.Source: §1 · p. 1
  • A DGF connects review gates, where agents may complete accepted reviews while humans retain other decisions and software combines language models, deterministic tools, and escalation.A system that produces a draft for a person to redo has assisted rather than performed the review task.Source: §1 · pp. 1–2
  • Delegating review execution requires accessible facts, manageable exception work, faithful evidence, and authorized commitments; the relative priority of DGF adoption remains a hypothesis for deployment settings.Engineering can implement these requirements as tasks with acceptance criteria.Source: §1 · p. 2

1.1 Research questions and units of analysis

  • The paper distinguishes gate execution, complete-route reliability, and labor required across an operating population, which cannot be inferred from an API trace alone.A successful generated-dossier run is not a measured number of hours saved, and refusals can be successful reviews.Source: §1.1 · pp. 2–3
  • A contract does not itself make evidence available or resolve policy disagreement; the proposed engineering work targets adequately informed, authorized reviews and measures the remaining boundary.The manuscript synthesizes an earlier paper and shares its experimental record, with no additional model calls or enterprise observations.Source: §1.1 · p. 3

2 Related work and the evaluation boundary

Prior work frames automation as task displacement within changing demand and persistent intervention work, while process research supplies governance's stage-gate and compliance foundations. This paper evaluates a specified review contract and distinguishes its evidence and acceptance boundary from neighboring enterprise-agent tasks.

2.1 From task automation to the organization of work

  • Automation can displace existing work while demand and new tasks change; exception handling may consume disproportionate labor, so net-work claims require a complete labor account.This task-based distinction follows Acemoglu and Restrepo (2019); monitoring and intervention concerns follow Bainbridge (1983), while handoffs and bottlenecks draw on Demirer et al. (2026) and Gans and Goldfarb (2026).Source: §2.1 · p. 3

2.2 Governance gates and executable process obligations

  • Stage-gate and compliance traditions provide process foundations, while executable governance reviews can combine deterministic policy kernels with agents that interpret evidence and assemble records.In the experiment, formalized policies and structured facts isolate the decision kernel; less-structured fact recovery is a separate comparison.Source: §2.2 · p. 4

2.3 Enterprise-agent benchmarks

  • DGF-Bench measures a review contract that asks for recovery tests, not operational repair, and exposes structured facts rather than requiring browser-based retrieval.These task boundaries prevent interpreting a higher percentage as superiority over agents evaluated on broader or less-prepared tasks.Source: §2.3 · p. 4

2.4 Evidence quality and useful agent comparisons

  • Auditable review separates correct dispositions from supported evidence; the study reports strict evidence conformity, structural audits, deterministic baselines, costs, and repeated trajectories.The authors identify matched-information evaluation and a complete labor account as further tests of incremental agent value and net work transfer.Source: §2.4 · p. 5

3 A gate as a substitutable contract

A Digital Governance Framework organizes project reviews as bounded, recorded tasks whose outputs and authority can be specified. Replacing human execution requires accessible information, valid authorized decisions, and less total human work while preserving the review’s scope and handoffs.

3 A gate as a substitutable contract

  • A DGF composes review checkpoints around a project dossier, with specialist opinions consolidated into a governance record.Organizations may merge, split, or reorder gates while preserving their obligations and information dependencies.Source: §3 · p. 5, §3.1 · p. 5

3.1 The work behind the checkpoints

  • Substitution distinguishes dossier preparation, review, effective authorization, and remediation; completing a review does not complete the actions it requests.A readiness review can issue a conditional decision while leaving the runbook completion condition open.Source: §3.1 · pp. 5–6

3.2 The review contract

  • A review contract fixes permitted evidence and tools, policy version, authorization mandate, acceptance conditions, and the required decision record.A refusal can be a completed review, while an approval fails if it omits a mandatory defect or exceeds the mandate.Source: §3.2 · p. 6

3.3 What is held fixed when execution changes

  • Comparisons must hold the declared population and review boundaries fixed, including difficult cases, missing evidence, legitimate refusals, and work outside the review.A narrower deployment can be useful if its scope and excluded work are reported; document or policy changes also require explicit version-handling rules.Source: §3.3 · pp. 6–7

3.4 Three conditions for a replacement

  • Replacement requires accessible information, valid outputs, authorized action, and reduced total human effort, while quality, coverage, abstention work, enforcement, and correction remain accounted for.Schema validity alone does not establish a substantively correct decision, and escalating every case may preserve authority while performing little review.Source: §3.4 · p. 7

3.5 An information obstruction

  • When two cases look identical under permitted observations but require disjoint accepted outputs, no system can guarantee valid decisions on both.In the counterexample, 26 Word documents have identical extracted text while CSV due-diligence status changes the reference decision from GO to REWORK; this is an information-interface obstruction for human and machine reviewers alike.Source: §3.5 · pp. 7–8

3.6 Why the DGF is a candidate

  • DGF workflows are candidates when their obligations can be named and inspected, but clean rules and accessible evidence are not guaranteed in every enterprise.Table 4 frames these as properties to check in an organization, not universal conditions already met.Source: §3.6 · p. 8, Table 4 · p. 8
  • A route must preserve conditional handoffs and required checks; gate scores alone cannot establish whole-route success without assumptions about dependence.Rule engines, agents, and authorization services can divide execution, but the combined system must satisfy accepted work, quality, and resource requirements.Source: §3.6 · p. 8

4 When agents replace human gate work

The section treats substitution as a complete human-hours accounting problem, not a count of automated cases. It derives a residual-work threshold and shows that route volume and recurring support can reverse labor outcomes.

4 When agents replace human gate work

  • The labor account includes baseline and post-deployment case work, recurring upkeep, rework, and hours across teams and suppliers.Each activity is allocated once; time categories must not overlap.Source: §4 · p. 9, Table 5 · p. 9
  • A target fractional workload reduction d is met exactly when ϕη + (1 − ϕ)µ + r + bA ≤ (1 − d)(1 + bH).Source: §4 · p. 9
  • Case share alone misses selection: 20% of cases can retain 40% of baseline labor, and 50% higher exception effort raises that burden to 60%.Source: §4 · p. 10
  • Repair effort from erroneous accepted outputs contributes to residual work, but low repair cost does not guarantee quality and deployment values of q and hC are unmeasured.Source: §4 · p. 10
  • In the synthetic example, exceptions are 23.43% of cases but retain 46.86% of baseline labor; fixed coverage and partition do not determine labor’s direction.None of the four scenarios reaches an 80% reduction; with ϕ ≃ 0.4686 and no other burden, η must be ≤ 0.427 for that target.Source: §4 · pp. 10–11
  • 43.52, 85.32, 89.07, and 168.92 FTE result under four assumptions versus a 140-FTE baseline; none meets the 80% reduction target.Source: Figure 2 · p. 10, §4 · p. 11

4.1 Route volume and the allocation of recurring support

  • At five Integrate and 95 Build visits, combined labor is 994.40 hours (ratio 0.414), yet the low-volume Integrate route exceeds its baseline; at 95 Integrate visits its ratio is 0.930.The fixed 66 support hours are spread over more visits in the alternative scale calculation.Source: §4.1 · p. 12, Table 6 · p. 12
  • Reusing maintained connectors, policy tests, and evidence interfaces can distribute recurring work across more reviews, while route-level results and support allocation remain necessary.Required hours translate to headcount only under additional staffing assumptions; support remains a floor unless eliminated or automated.Source: §4.1 · p. 12

5 The FDE implementation: engineering the whole workflow

Forward deployed engineering turns governance review into an engineered workflow, with explicit task boundaries, usable evidence, valid authority, and preserved obligations. Rollout is judged by quality and total human work, not by apparent automation alone.

5 The FDE implementation: engineering the whole workflow

  • FDEs build information, policy, evidence, and authorization interfaces; agents, deterministic services, and escalation divide review work, but this proposed architecture was not evaluated by the benchmark.A technical check alone does not certify complete substantive correctness.Source: §5 · pp. 12–13, Figure 3 · p. 13

5.1 Specify the transfer before building the agent

  • A transfer starts with a process-owner-approved task boundary specifying inputs, accessible systems, outputs, service requirements, and decisions that remain outside the mandate.Acceptance cases should include deficient, incomplete, conflicting, and out-of-scope inputs; missing information must trigger the declared request or abstention behavior.Source: §5.1 · p. 13

5.2 Make evidence usable without reconstructing the dossier

  • Evidence interfaces identify authoritative sources, versions, and conflicts, then preserve observed values and locations; confirming an excerpt was observed is distinct from confirming it supports each premise.Manual preparation of project snapshots belongs in the human-work account, and inaccessible or stale sources should be reported rather than replaced with plausible values.Source: §5.2 · pp. 13–14

5.3 Separate policy execution, commitments, and failure handling

  • Executable policies can be evaluated by deterministic services, while owners retain responsibility for policy approval and exceptions; separate mandate checks govern commitments, and failures must produce specified states.The benchmark tests supplied policies, not the correctness of an organization’s policy conversion; retries cannot resolve a missing mandate or make an unsupported premise true.Source: §5.3 · p. 14

5.4 Preserve obligations across gate interfaces

  • Gate handoffs must preserve unresolved findings, actions, evidence, policy version, and authorization; closure follows only when declared evidence is supplied.Proposed integration tests also revalidate earlier records when their source or mandate changes.Source: §5.4 · p. 14

5.5 Roll out against quality and labor criteria

  • Substitution requires fixed quality and service criteria plus lower total human work, counting exceptions, correction, and recurring engineering; more automated cases alone do not establish savings.Rollout begins with frozen contracts and representative cases; shadow operation adds an agent beside existing work, while the proposed DGF-first sequence is not an observed adoption ranking or measured savings.Source: §5.5 · pp. 14–15, Table 7 · p. 15

6 Evidence: executing governance contracts

DGF-Bench tests whether models can execute specified review contracts across synthetic project routes, with strict success requiring correct decisions and complete evidence. Models often passed individual gates but less reliably completed whole routes; audits and repeated runs expose separate limits in evidence fidelity and repeatability.

6.1 Design, dossiers, and information conditions

  • The benchmark uses 300 synthetic projects across Buy, Integrate, and Build routes, with structured facts, executable policies, and balanced decision outcomes.The generated cases include missing, stale, or inconsistent evidence; their diversity statistics do not establish equivalent diversity of organizational contexts or difficulty.Source: §6.1 · pp. 15–16

6.2 Execution and scoring protocol

  • Strict success requires the correct disposition, exact finding and action sets, observed evidence references and excerpts, and the correct authorization indicator.A complete route passes every gate, so a case that passes four of five gates has zero complete-route success.Source: §6.2 · p. 17

6.3 Deterministic control and aggregate results

  • The deterministic control passes 1,700 gates and 300 complete routes using the public policy and structured-fact interface, showing the formalized review kernel is software-executable.Its zero model API expenditure excludes programming and local computation and does not measure the cost of producing the supplied inputs.Source: §6.3 · p. 18
  • 94.98% of Gemini’s gates and 76.92% of its complete routes pass strictly; the gate–route gap also appears for Luna and DeepSeek.Luna and DeepSeek also exhibit substantive errors, while Gemini’s 85 strict failures concern evidence conformity.Source: Table 8 · p. 18, §6.3 · p. 18
  • Gate-family performance varies: Gemini passes all IT and Compliance reviews, while Luna exceeds it in Architecture and Tech Readiness; sequential routes also expose upstream errors to General.The single handoff condition does not isolate a causal propagation effect.Source: §6.3 · p. 19, Table 9 · p. 19

6.4 Authorized conditional decisions

  • Conditional approvals are bounded by policy eligibility and an active mandate, while the finding and required action remain in the final review.In one readiness trace, Gemini’s authorized approval leaves the runbook finding unresolved and retains its action.Source: §6.4 · p. 19, §6.5 · pp. 20–21

6.5 Evidence failures and two inspectable traces

  • A structural evidence audit recovers 69 Gemini, 38 Luna, and 52 DeepSeek gates, but does not establish full premise coverage or independently validated semantic support.In Procurement, authentic observed excerpts can still be contract-excluded, incomplete, or structurally nonliteral.Source: §6.5 · p. 20

6.6 Repeated trajectories, expenditure, and interpretation

  • Across 135 repeated runs, Gemini, Luna, and DeepSeek pass 35, 19, and 11 of 45 routes, respectively; provider routing is not fixed, and only 15 distinct synthetic projects are included.The paper separates decision production, evidence preservation, and repeated whole-route delivery; recorded costs exclude development, local computation, human review, and deployment.Source: §6.6 · pp. 22–23

7 From task substitution to a workforce hypothesis

The section distinguishes successful review execution from reliable project-wide performance and from reduced workforce requirements. It proposes a falsifiable deployment test that measures complete human effort alongside output, quality, and service.

7.1 The 2033 hypothesis and a falsifiable deployment test

  • The 2033 hypothesis targets at least 80% fewer workload-equivalent FTE at comparable output, quality, and service, using the 2026 reference year.The author-selected forecast is tested against a complete labor boundary; it is not extrapolated from benchmark scores.Source: §7.1 · p. 23
  • Full elimination of human DGF execution remains an undated conjecture requiring case-level effort and recurring support to disappear.No finite benchmark in the paper establishes that endpoint.Source: §7.1 · p. 23

7.2 Observing substitution without losing the work boundary

  • A credible substitution study measures accepted output and complete human effort, while distinguishing actual hours from standardized hours for comparable output.The account includes preparation, exceptions, correction, handoffs, support, and transition costs; quality and service conditions must also be met.Source: §7.2 · pp. 23–24
  • An upper workload-ratio bound at or below 0.20 supports the target only when quality and service conditions are met; an interval crossing it is inconclusive.Missing material work cannot be assigned zero, and uncertainty includes case variation, organizational clustering, and reconstructed baseline hours.Source: §7.2 · p. 24
  • The author’s 80% reduction forecast cannot be attributed to agents without a suitable comparison, and no enterprise observations are claimed.Staged or randomized adoption can help separate agent effects from other changes; ecosystem-wide inference also requires a declared sampling frame.Source: §7.2 · p. 24, §7.4 · p. 25

7.3 The next discriminating experiment

  • The next experiment should vary information preparation and policy representation separately while controlling access, budgets, tools, and evidence requirements.The proposed matched-information design is unexecuted; removing ready-made snapshots alone would not isolate source-reading ability because runtime tools can return them.Source: §7.3 · pp. 24–25, Table 11 · p. 25

7.4 Scope of the observations

  • Review-performance measurements do not establish enterprise operating effort, staffing changes, or the relative adoption speed of business functions.The paper specifies a deployment protocol for measuring labor and evaluating the dated workforce hypothesis, but claims no enterprise observations of this kind.Source: §7.4 · p. 25

8 Conclusion

The conclusion argues that specified governance-review tasks can transfer to software while governance obligations remain. It distinguishes this task-level feasibility from the workforce consequences, which require deployment measurement.

8 Conclusion

  • Agents and a deterministic control execute specified governance-review operations in a controlled environment, including specialist reviews and complete project routes.The control executes the formalized policy kernel; agents complete many of the same operations.Source: §8 · p. 26
  • The proposed DGF architecture connects agent investigation with evidence services, policy execution, authority enforcement, and exception handling.Recurring dossiers, named responsibilities, explicit policies, and recorded handoffs provide identifiable work for implementation.Source: §8 · p. 26
  • At comparable output, quality, and service, satisfying the substitution frontier means fewer human hours are required for governed work.The complete account includes those who operate, maintain, and resolve exceptions around the system.Source: §8 · p. 26
  • The paper’s 80% fewer-FTE target by 2033 is a deployment hypothesis, while the reported measurements concern review performance.Complete replacement would also require remaining exception and support work to be executed in software.Source: §8 · p. 26
  • The proposed path to task substitution is to define contracts, establish dependable execution, measure remaining work, and expand where accepted output requires less human effort.The conclusion frames human review as an implementation of governance that can change as these conditions are met.Source: §8 · p. 26

Data, code, and attribution

The paper identifies its author and directs readers to a public repository containing the benchmark materials and analyses.

Data, code, and attribution

  • The public repository contains the dgf-bench-300-20260923 release with generated dossiers, model traces, failed attempts, scores, and cost ledgers.The release also includes Word documents and architecture materials.Source: Data, code, and attribution · p. 26
  • The paper locates original-run materials, follow-up controls, evidence audits, repetitions, and source checks in separate research directories.The listed repository paths identify where results and reproduction materials are kept.Source: Data, code, and attribution · p. 26
  • The author is Jeremy Canale, who states responsibility for the claims and interpretation.The passage provides author contact details and repository locations for the manuscript and verifier.Source: Data, code, and attribution · p. 26

A Reproducing the claims from the released artifacts

The appendix describes how to reproduce benchmark and labor-account claims from the released artifacts. It distinguishes aggregate checks from rescoring and from validation of the policy’s professional validity.

A Reproducing the claims from the released artifacts

  • The original offline reproduction uses the frozen benchmark source and expects 899 evaluable trajectories, 5,094 gates, and 299 dossiers in the common comparison.The run archive’s analysis_20260923_final is the specified source for reproduction; no new API calls are needed for the summaries.Source: A Reproducing the claims from the… · p. 27
  • Recomputing saved-score aggregates checks arithmetic and exclusions, while rescoring saved submissions separately checks evaluator implementation.Neither operation independently establishes the professional validity of the policy.Source: A Reproducing the claims from the… · p. 27
  • The repeated-run reproduction pools successful gates over 255 gates per model and complete routes over 45, while grouping trajectories by dossier for the all-three endpoint.Resampling must preserve the grouping of the 15 dossiers rather than treating 45 trajectories as distinct projects.Source: A Reproducing the claims from the… · p. 27
  • The labor computation recovers 43.52, 85.32, 89.07, and 168.92 FTE from specified assumptions, not estimates fitted to the benchmark.The calculation uses a declared 120-hour monthly FTE convention.Source: A Reproducing the claims from the… · p. 27

Related papers

Loading 2609.29345v1…