Source-linked AI summary
LLM Parkinsonism: Executive-Control Failure, Token-Inefficient Persistence, and an Uncertainty-Aware Global Executive Control Architecture for Autonomous Language-Model Agents
Dongsheng Xiao, Zeyuan Wang, Xuzhe Xia, Bo Zhao, Yankai Cao
arXiv:2609.30662v1cs.AI
TL;DR
Agents can continue with plausible actions even when project-level scope and stopping decisions are unclear. GEC v0.2 separates action generation from project-level governance; in matched-candidate simulations, it achieved comparable hard-goal success to candidate-set local control while improving efficiency, scope discipline, and complexity control.
Problem
Project-level executive control remains a challenge when agents can produce locally coherent work without reassessing whether it still serves the governed objective.
Source: §1 · p. 1, §4.4 · p. 4Method
GEC v0.2 separates model-proposed actions from Goal Contract objectives, independent scope adjudication, verification, candidate selection, state-versioned evidence, and state-level stopping.
Source: §13 · p. 18Results
96.53% candidate-set local control versus 96.57% GEC hard-goal success; relative to that control, GEC reduced mean total tokens 36.4% and RMTTC₄₀ₖ 18.7%.
Source: §9.1 · p. 12, §13 · p. 18Takeaways & Limitations
LPB v0.2 indicates that GEC's distinct contribution is efficiency, scope discipline, and complexity control at comparable success, rather than additional hard-goal success.
Source: §13 · p. 18The independent scope adjudicator is synthetic and oracle-like; the robustness test covers random error, not systematic bias, correlated errors, or strategic self-justification.
Source: §11.4 · p. 17
Abstract
from arXiv · showhide
Large language models (LLMs) can plan, use tools, write code, and execute long-horizon workflows, yet strong local competence does not guarantee project-level executive control. Agents may continue acting after the original objective is satisfied, producing low-value refinements, repeated verification, and repairs to self-created complexity. We use LLM Parkinsonism as a narrowly defined, non-clinical metaphor for this pattern of persistent action despite diminishing task-level value. We argue that the problem is not explained by autoregressive next-token prediction alone, but more directly by concentrating proposal generation, scope interpretation, progress assessment, and stopping authority within the same self-conditioned loop. We therefore introduce Global Executive Control (GEC) v0.2, an uncertainty-aware governance architecture that separates action generation from project-level control. In a 24,000-episode matched-candidate benchmark under a common 40,000-token ceiling, a first-candidate baseline achieved 67.42% hard-goal success, a candidate-set local control achieved 96.53%, and GEC achieved 96.57%. The candidate-set control shows that access to multiple candidate actions explains most of the success gain; relative to that control, GEC preserved success while reducing mean token use from 19,782 to 12,574 (36.4%) and restricted mean tokens to completion at the 40,000-token ceiling from 16,136 to 13,114 (18.7%), while eliminating measured pre-completion drift and sharply reducing gross complexity. Governance-overhead sensitivity remained favorable through an additional 500 synthetic governance tokens per cycle. These mechanistic simulations support explicit governance of scope, evidence, resource use, and stopping, while live-model validation remains necessary.
1 Introduction
The paper frames the central challenge as project-level control: an agent can perform coherent operations yet continue beyond the governed objective. It proposes GEC and LPB v0.2 to separate scope, evidence, resource use, and stopping from local action generation.
1 Introduction
- Coherent local engineering can still become globally misdirected when agents continue adding optional layers and repairing complexity absent from the original goal.The paper distinguishes deciding how to perform an operation from deciding whether it remains part of the governed objective.Source: §1 · pp. 1–2
- The six contributions define LLM Parkinsonism, propose collapsed executive authority as a systems hypothesis, and introduce GEC v0.2 and the LPB v0.2 benchmark.They also revise scope rules, evidence and stopping mechanisms, and evaluation metrics.Source: §1 · p. 2
2.1 Agent loops and self-improvement
- Agent loops support adaptation, but locally plausible next actions do not necessarily serve the global objective, making continuation a control problem.The paper situates this issue in iterative agent designs such as ReAct and Reflexion [1].Source: §2.1 · p. 2
2.2 Cost, budgets, and bounded computation
- Cost-aware reasoning research examines mismatches between capability, computation cost, and efficient allocation; GEC instead targets project-level scope, evidence, and stopping.The paper presents GEC as complementary to token compression alone.Source: §2.2 · p. 2
2.3 Verification and termination
- GEC treats evidence as state-versioned, so evidence valid before a later state change can become stale rather than remain a completion certificate.This extends verification and termination approaches that connect planning to criteria and typed support [14] [18].Source: §2.3 · p. 3
3 The Clinical Metaphor and Its Limits
The paper uses “LLM Parkinsonism” only as a non-clinical analogy for persistent action despite diminishing task-level value. Its operational definition concerns multi-step project trajectories, not verbosity alone.
3.1 Why "Parkinsonism" is used only as a phenomenological analogy
- The Parkinsonism analogy concerns a surface trajectory of diminishing-yield actions and persistence, not a biological mechanism; action granularity is not a validated LPB v0.2 endpoint.The evaluated variables include governed scope, externally evaluated utility, token cost, complexity, evidence validity, and stopping.Source: §3.1 · p. 3, Figure 1 · p. 3
3.2 Operational definition
- LLM Parkinsonism includes goal drift from governed causal links and complexity accretion through added states, components, dependencies, interfaces, or abstractions.The construct requires a substantial combination of its defined properties.Source: §3.2 · p. 3
- The construct also includes termination failure and token-inefficient persistence, and applies to multi-step, tool-using project trajectories rather than a single verbose answer.Termination failure includes continuing after verified hard-goal completion or when the best governed continuation has non-positive marginal value.Source: §3.2 · pp. 3–4
4 Why Autoregression Is Relevant but Insufficient
The paper argues that autoregressive response termination does not settle whether a project should stop. Self-conditioning, action bias, and local coherence can sustain work without global objective review, motivating explicit marginal-value and project-level control.
4.1 EOS is not PROJECT DONE
- A clean response ending is distinct from project completion, and a plausible next action is easier to find than a worthwhile one.An orchestration layer can request another step after response termination.Source: §4.1 · p. 4
4.2 Self-conditioning can promote suggestions into requirements
- Self-generated requirement promotion occurs when an optional suggestion is later treated as committed, prompting tests, monitoring, and repair work.The pathway can turn a suggestion into an assumed project fact and then a hard gate.Source: §4.2 · p. 4
4.3 Helpfulness can create action bias
- In an open-ended loop, helpfulness and repeated next-step solicitation can make concrete action seem preferable to recognizing that nothing further is required.The paper describes this as a systems-level interaction, not a deterministic effect of one training algorithm.Source: §4.3 · p. 4
4.4 Local coherence can defeat global objective review
- Locally coherent subproblem work can continue without answering whether that subproblem belongs to the governed objective.The paper uses successive repairs to site-correlation logic to illustrate the missing return to a higher abstraction level.Source: §4.4 · p. 4
4.5 Self-evaluation can confuse activity with progress
- Observable activity is not progress unless it reflects an independently grounded change in completion, prerequisite, risk, or uncertainty state.Self-evaluation can accept cycles with zero or negative measured change in external world state [14].Source: §4.5 · p. 4
4.6 Missing marginal-value computation
- GEC scores candidate actions by expected verified utility gain minus token, complexity, operational-risk, and verification costs; rejecting one action does not itself justify stopping the project.Project-level economic stopping requires evaluating a candidate set or a justified continuation upper bound.Source: §4.6 · p. 5
5 Formalization and Metrics
The framework formalizes project goals, action scope, evidence validity, stopping, and resource-use metrics. Its contract distinguishes mandatory hard requirements from authorized soft objectives and non-goals.
5.1 Governed Goal Contract
- The Goal Contract represents a primary goal, hard requirements, authorized soft objectives, non-goals, safety and feasibility constraints, and an amendment policy.LPB v0.2 does not exercise a dedicated general-purpose constraint predicate engine.Source: §5.1 · p. 5
- A proposal is not a requirement: contract changes require explicit external authorization, and unmet hard requirements cannot be abandoned for convenience.When no feasible path remains before hard completion, the terminal state is BLOCKED; economic stopping applies after hard completion.Source: §5.1 · p. 5
5.2 Governed action linkage
- GEC assesses actions as direct, prerequisite, verification, risk-mitigation, soft, none, or forbidden, with an independent adjudicator deciding scope rather than the generator.Prerequisites can be causally necessary despite producing no immediate hard-goal utility change.Source: §5.2 · p. 5
5.3 Evidence validity and invalidation
- Criterion evidence records status, verifier, state version, confidence, and dependencies; completion requires valid passing evidence above a confidence threshold.Later relevant state changes can invalidate earlier evidence.Source: §5.3 · p. 6
5.4 Primary outcomes and restricted completion cost
- Hard-goal success measures external satisfaction of all hard criteria, while RMTTCB counts each episode up to the budget ceiling, including non-completions.Conditional tokens to first completion among successful trajectories remain a secondary descriptive measure because conditioning on success can induce selection bias.Source: §5.4 · p. 6
5.5 Token Efficiency; 5.6 Direct and process useful-token ratios
- Token efficiency is secondary, and direct versus process useful-token ratios distinguish immediate hard-goal closure from validated prerequisites and other process progress.This separation avoids labeling a legitimate prerequisite chain as waste solely because intermediate steps do not change hard-goal utility.Source: §5.5 · p. 6, §5.6 · p. 6
5.7 Termination Overrun Ratio; 5.8 Pre-completion Goal Drift Rate
- TOR measures the post-completion tail, whereas pre-completion GDR measures the fraction of actions without governed scope before first completion.The two metrics distinguish failure to stop after finishing from drifting before finishing.Source: §5.7 · p. 6, §5.8 · pp. 6–7
5.9 Gross and net complexity; 5.10 Exploratory Executive Persistence Index
- Gross complexity captures added complexity even when later deletion hides it in net change, while EPI combines persistence-related measures for descriptive visualization only.LPB v0.2 does not model systematic delete-after-add cycles, so gross and net values are almost identical; EPI is exploratory, not a primary endpoint.Source: §5.9 · p. 7, §5.10 · p. 7
6 Hypotheses
The revised framework states six hypotheses about persistence, drift, budget constraints, candidate search, component effects, and robustness. It also marks a proposed monotonic relation between unmet criteria and unscoped actions as unestablished.
H1 — Tail persistence
- Uncontrolled local loops are hypothesized to continue after verified completion, producing non-zero termination overrun.Source: §6 · p. 7
H2 — Pre-completion drift
- Uncontrolled loops are hypothesized to execute unscoped actions before completion, while GEC's independent scope authority should reduce their fraction.Source: §6 · p. 7
H3 — Budget insufficiency
- A token ceiling alone is hypothesized not to reliably select high-value work under a constrained budget.Source: §6 · p. 7
H4 — Governance beyond candidate search
- Against best-of-three candidate access, GEC is hypothesized to preserve hard-goal success while reducing restricted completion cost, token use, drift, and complexity.Source: §6 · p. 7
H5 — Component specificity
- Removing independent scope authority or the marginal-value gate is hypothesized to degrade distinct trajectory properties rather than affect all components identically.Source: §6 · p. 7
H6 — Robustness
- Moderate noise in scope linkage and external verification is hypothesized to degrade performance gracefully rather than catastrophically.Source: §6 · p. 7
Scope boundary
- LPB v0.2 does not directly establish that unscoped-action probability rises monotonically as unmet criteria approach zero; live trajectory analysis is reserved for that claim.Source: §6 · p. 7
7 Global Executive Control v0.2
GEC v0.2 separates action proposals from project-level authority, governing scope, evidence, value, and stopping. Its controls distinguish rejecting one action from stopping the project and preserve hard requirements while allowing justified prerequisites and verification.
7.1 Candidate-set governance
- GEC evaluates multiple candidate actions, allowing it to reject a bad proposal while retaining alternatives and to make economic stopping a state-level decision.The architecture separates proposal generation from project-level authority through a governed Goal Contract, independent scope adjudication, state-versioned evidence, candidate-set value selection, and explicit terminal semantics.Source: §7.1 · p. 8, Figure 2 · p. 8
7.2 Independent scope gate
- Independent scope assessments reject forbidden or unsupported actions while allowing sufficiently confident direct, prerequisite, verification, risk-mitigation, and soft links.The scope gate does not trust the target criterion supplied by the proposing LLM.Source: §7.2 · p. 8
7.3 Contract-aligned expected utility
- Expected hard utility derives from criterion weight and candidate success probability; prerequisite and risk-mitigation links receive discounted value, while soft value is separately weighted by β.Soft-objective value cannot silently become hard utility.Source: §7.3 · pp. 8–9
7.4 Action-level rejection versus state-level economic stopping
- GEC rejects an individual action when its value is at or below τ, but permits economic stopping only after hard requirements are complete and no governed continuation has positive net value.If hard requirements remain unmet and no feasible governed path exists, the terminal state is STOP_BLOCKED; the prototype relies on a planner, environment, or external oracle for that feasibility assessment.Source: §7.4 · p. 9
7.5 Process-progress circuit breaker
- After k cycles without hard-goal progress, validated prerequisite completion, or new valid external evidence, the breaker suspends local repair and requests global replanning.The prototype sets k=3; the current matched exogenous benchmark does not establish an independent efficiency benefit because replanning cannot change the future candidate stream.Source: §7.5 · p. 9
7.6 Evidence-carrying termination
- STOP_SUCCESS requires valid evidence for every hard criterion, and later state changes can invalidate that evidence to prevent stale completion certificates.This makes terminal success depend on current evidence rather than a permanent boolean flag.Source: §7.6 · p. 9
7.7 Governed amendments and non-goals
- Non-goals are executable contract elements, and only an externally authorized ContractAmendment can revise priorities; the executor may propose but not authorize amendments..Source: §7.7 · p. 9
7.8 Deletion-first search
- Deletion-first search prioritizes removing self-created structure, but LPB v0.2 does not quantitatively isolate this principle and future benchmarks should test add-delete-thrash trajectories.The supplied passage introduces the prioritization without specifying its full sequence.Source: §7.8 · p. 9
8 Methods
The methods evaluate GEC v0.2 against local controls in a synthetic mechanism-isolation testbed, using matched candidate opportunities and shared resource ceilings. Separate experiments examine budget sensitivity, component ablations, governance noise and overhead, and whether beneficial soft work is suppressed.
8.1 Implementation
- The dependency-free Python prototype includes contract, scope, evidence, controller, metric, and simulator objects, with 13 unit tests covering core governance and simulation behavior.Tests include evidence-dependent completion, non-goal rejection, matched candidate streams, common budget ceilings, and separated drift and tail metrics.Source: §8.1 · pp. 9–10
8.2 LPB v0.2 task families
- LPB v0.2 tests six synthetic task families with three or four hard criteria, explicit non-goals, and selected prerequisite structures; it is a mechanism-isolation testbed, not a real-model prevalence estimator.The families cover disaster recovery, API release, database migration, CI, backup/restore verification, and reproducible research.Source: §8.2 · p. 10
8.3 Matched exogenous candidate sets
- Policies receive the same pre-generated candidate sets and execution draws, while governance can select or reject candidates but cannot force the generator to produce a required next action.The candidate-set local control uses generator-declared links and a token-cost tiebreaker; GEC adds independent scope adjudication, value scoring, and complexity and risk penalties.Source: §8.3 · p. 10
8.4 Candidate distribution
- The stream contains 28% optional proposals, including misleading self-declared links and a 10% branch of expensive, low-yield in-scope verification candidates; unvalidated prerequisites reduce hard-action success probability.Each cycle costs 120 planning tokens plus the selected action cost; optional work can create maintenance debt or stale evidence, and baseline completion recognition is probabilistic while GEC uses valid-evidence termination.Source: §8.4 · p. 10
8.5 Governance conditions
- Four policies compare first-candidate selection, budget-only skipping, candidate-set local selection, and GEC governance under the same 40,000-token ceiling.The candidate-set control lacks independent scope authority, evidence-aware value gating, and complexity/risk governance; GEC includes these controls and process-progress monitoring.Source: §8.5 · p. 10
8.6 Main experiment
- The main benchmark runs 24,000 episodes across six scenarios and four policies, with a 40,000-token ceiling and 40-cycle horizon.Binary hard-goal success uses Wilson 95% confidence intervals; continuous means are reported as mean ± 1.96 SE conditional on the fixed synthetic scenarios and parameters.Source: §8.6 · pp. 10–11
8.7 Budget sensitivity
- Budget sensitivity compares budget-only and GEC at eight token ceilings across 14,400 episodes and reports normalized success area under the log-budget curve.The curve summarizes efficiency across the full resource range rather than at one threshold.Source: §8.7 · p. 11
8.8 Component ablation
- The 7,200-episode ablation compares full GEC with versions lacking independent scope authority, the marginal-value gate, or the no-progress breaker.Each policy is evaluated on 300 episodes per scenario.Source: §8.8 · p. 11
9 Results
In the matched synthetic benchmark, GEC reached essentially the same hard-goal success as candidate-set local control while using fewer tokens and reducing drift and complexity. Additional experiments report budget sensitivity, component ablations, noise robustness, overhead limits, a soft-work probe, and a watchdog case study.
9.1 Main matched-candidate benchmark
- 96.57% hard-goal success for GEC versus 96.53% for candidate-set local control, while GEC reduced mean total tokens by 36.4% and RMTTC₄₀ₖ by 18.7%.GEC also increased TE by 1.63×, reduced pre-completion GDR from 0.1290 to 0, and reduced gross complexity accretion from 0.8224 to 0.1228; candidate-set control explains most of the success gain over the 67.42% first-candidate baseline. Figure 3 depicts the operating-point comparison.Source: Figure 3 · p. 12, §9.1 · p. 12
9.2 Budget-sensitivity advantage
- 0.467 for GEC versus 0.167 for budget-only control on normalized log-budget success AUC, a 2.79× ratio.The reported interpretation is allocation of the fixed candidate stream toward governed causal relevance and higher expected value, not creation of capability; Figure 4 presents hard-goal success under token ceilings.Source: §9.2 · pp. 12–13, Figure 4 · p. 13
9.3 Component ablations
- Removing independent scope authority produced pre-completion drift and more than doubled gross complexity, while removing the value gate preserved success but added about 1.77k mean tokens.Removing the no-progress breaker slightly improved synthetic efficiency because replanning cannot change the exogenous future candidate stream; the experiment does not establish an independent benefit for that breaker.Source: §9.3 · p. 13
9.4 Robustness to imperfect governance signals
- 96.33% to 96.00%: GEC success across injected noise levels 0 to 0.10, with TE approximately 0.095-0.096.The results show no collapse under modest random errors in this synthetic setup, not calibration to real-world verifier performance; Figure 5 concerns synthetic scope/verifier noise.Source: §9.4 · p. 13, Figure 5 · p. 14
9.5 Governance-overhead sensitivity
- At 500 additional synthetic governance tokens per cycle, GEC achieved 89.88% success and RMTTC₄₀ₖ = 23,281, versus 67.42% and 29,024 for the first-candidate baseline.At 1,000 tokens per cycle, GEC's advantage largely disappeared; these stress-test costs are not calibrated to provider prices.Source: §9.5 · p. 15
9.6 Beneficial soft work is not automatically suppressed
- The soft-work probe approved the beneficial action at net value 0.2321, rejected the low-yield action at -0.1605, and returned CONTINUE.This tests rejection of one candidate without treating it as a reason to stop the project.Source: §9.6 · p. 15
9.7 Watchdog case study
- In the watchdog case, independent scope adjudication rejects SITE DOWN aggregation as forbidden or unlinked, but can admit a staging or synchronization prerequisite for a hard criterion.The generator's claimed link alone is insufficient to authorize the aggregation proposal.Source: §9.7 · p. 15
10 Discussion
The discussion locates persistent over-action in executive control: agents can generate coherent work while losing reliable authority over scope, value, evidence, and stopping. It argues for uncertainty-aware governance that distinguishes necessary prerequisites from scope drift and treats stopping as a project-level decision.
10.1 The deeper failure is executive, not merely linguistic
- Executive failure is broader than verbosity: concentrating proposal, scope, progress, world-state judgment, and stopping authority can make continued action locally coherent but globally poor.As capability expands, the larger action set can raise cost without raising utility; being able to act does not make an action desirable.Source: §10.1 · p. 15, §10.2 · p. 15
10.3 Prerequisites matter: scope is causal, not merely lexical
- GEC v0.2 treats scope causally, allowing enabling prerequisites while distinguishing them from self-invented work that lacks a justified causal link.A conservative scope gate should not reject legitimate multi-step chains merely because their intermediate actions do not immediately close a criterion.Source: §10.3 · p. 15
10.4 Independent scope authority matters
- Trusting a generator’s own claim that work is required reintroduced pre-completion drift and complexity in the ablation, motivating a scope adjudicator separate from the proposer.The independent linker need not be perfect, but its authority should be organizationally or computationally separate.Source: §10.4 · p. 16
- Candidate-set access alone raised success to nearly GEC’s level; GEC’s distinct contribution was comparable success with lower token cost, restricted completion cost, drift, and complexity.The discussion cautions against crediting GEC with the full success increase over the first-candidate baseline.Source: §10.4 · p. 16
10.5 Economic stopping is a state-level decision
- Economic stopping is a state-level decision: reject one low-value action without rejecting every alternative, and stop only after hard completion when no governed continuation has positive net value.A token budget limits spending but does not determine which action merits the next token; unmet hard requirements with a feasible positive-value path can justify continuing.Source: §10.5 · p. 16, §10.8 · p. 16
10.6 Evidence has a lifetime
- Evidence can become stale when later actions change relevant world state, so verification should be versioned and invalidated rather than treated as a permanent completion certificate.A previously passing test may need rerunning after a dependency or interface changes.Source: §10.6 · p. 16
10.7 Success and restricted completion cost should dominate TE
- Evaluation should pair externally adjudicated hard-goal success with restricted-mean completion cost over all trials; conditional completion time can mislead when policies have different success rates.When final utilities are equal, TE is an inverse-cost transformation rather than independent evidence beyond token cost.Source: §10.7 · p. 16
10.9 Complexity deletion is a first-class operation, but remains under-tested
- Gross complexity accounting exposes add-then-delete thrashing, but LPB v0.2 lacks a dedicated delete-thrash task family, so deletion-first search remains a design hypothesis.The broader direction is uncertainty-aware governance of scope, progress, evidence, and stopping, not an assumption that the controller knows the correct state.Source: §10.9 · p. 16, §10.10 · p. 16
11 Threats to Validity and Limitations
The limitations constrain LPB v0.2 to synthetic demonstrations with simplified governance, task structure, and evaluation. Several reported properties are partly enforced by controller rules, while key mechanisms—including no-progress breaking and deletion-first search—remain unvalidated or under-tested.
11.1 Synthetic construction
- LPB v0.2 uses synthetic stress-test parameters rather than fitted estimates of a named LLM, so its numerical results are demonstrations under that construction, not forecasts of commercial-agent behavior.Zero post-completion TOR and zero pre-completion GDR under perfect independent scope adjudication are partly controller invariants; improvement must instead be assessed under matched action opportunities.Source: §11.1 · p. 17, §11.2 · p. 17
11.3 Candidate-set abstraction
- Candidate sets are matched rather than adaptively generated, and synthetic governance overhead is not calibrated to model-equivalent provider usage.Live studies should include the full inference cost of the generator, adjudicator, verifier, and controller.Source: §11.3 · p. 17
11.4 Scope and evidence remain simplified
- The scope adjudicator is oracle-like, while robustness tests random errors rather than systematic bias, correlated errors, or strategic self-justification.Real scope links can be ambiguous, and evidence can be incomplete or adversarial.Source: §11.4 · p. 17
11.5 Utility, prerequisites, constraints, and feasibility are simplified
- LPB uses short synthetic prerequisite chains and simplified criterion weights, lacks a dedicated general-purpose constraint gate, and relies on an externally supplied feasibility assessment for STOP_BLOCKED.Real projects may have partially observable goals, changing priorities, interacting constraints, and long prerequisite DAGs.Source: §11.5 · p. 17
11.6 No-progress breaker not supported by the current ablation
- The no-progress mechanism showed no efficiency benefit in the matched exogenous benchmark; removing it modestly improved TE, so its independent benefit is not validated here.It is retained for live-agent testing.Source: §11.6 · p. 17
11.7 Deletion-first search not isolated
- Gross and net complexity metrics are implemented, but systematic add-delete thrashing is absent from the benchmark, leaving deletion-first search unisolated.A dedicated complexity-reversal benchmark is needed.Source: §11.7 · p. 17
11.8 Action granularity is conceptual
- LPB v0.2 does not validate a cross-domain action-granularity metric; its operational construct instead rests on scope, evidence, utility, cost, and stopping variables.The sequence-like reduction in action amplitude motivates the metaphor but is not a validated endpoint.Source: §11.8 · p. 17
11.9 The medical metaphor can be misunderstood
- The authors limit “LLM Parkinsonism” to a non-clinical analogy and do not propose biological homology, but note that the metaphor risks clinical misunderstanding.They suggest future work test whether it aids conceptual clarity without avoidable misunderstanding.Source: §11.9 · p. 18
12 Prespecified Live-Model Validation
The proposed next step is a paired live-model experiment comparing baseline, budget-only, and GEC governance under matched task conditions. It prioritizes externally adjudicated success and restricted completion cost, with auditing, diverse task types, and full cost accounting.
12 Prespecified Live-Model Validation
- Use paired tasks with the same base model, settings, tools, Goal Contract, and acceptance criteria across baseline, budget-only, and GEC conditions.Share exogenous task states and common randomness where technically feasible.Source: §12 · p. 18
- Primary endpoints should be externally adjudicated hard-goal success and restricted mean tokens-to-completion over all paired trials up to a prespecified ceiling.The restricted endpoint avoids silently excluding non-completion; success should use a prespecified non-inferiority or superiority hypothesis.Source: §12 · p. 18
- Secondary measures should include conditional completion tokens, TE, drift, complexity, tool calls, wall time, amendment and invalidation events, and heavy-tail token percentiles.Record exact input/output tokens and provider-billed totals, and include the full inference cost of adjudicators and verifiers.Source: §12 · p. 18
- Independently audit scope-link decisions and avoid adjudicating completion solely with the LLM that proposed and executed the work when external checks are available.A sample of candidate links should be double-coded by humans or a blinded secondary evaluator.Source: §12 · p. 18
- Include prerequisite chains, beneficial soft improvements, authorized goal changes, evidence-invalidating changes, and cases where the correct outcome is BLOCKED.Test at least one frontier reasoning model and one lower-cost model, use paired inference for cost, and report P90/P95 tails.Source: §12 · p. 18
13 Conclusion
The conclusion frames the problem as high capability under weak executive control and presents GEC v0.2 as a separate project-governance layer. It emphasizes that candidate-set access accounts for most of the success gain, while independent governance primarily improves efficiency and control at comparable success.
13 Conclusion
- The central weakness is not lack of capability but unreliable judgment about whether to act next.An agent may answer “What can I do next?” better than “Should I do anything next?”Source: §13 · p. 18
- GEC v0.2 separates proposal generation from project governance through a Goal Contract, independent scope linker, verifier, state-versioned evidence, candidate selection, and state-level stopping.Evidence expires when relevant state changes.Source: §13 · p. 18
- Candidate-set access alone explains most of the hard-goal success gain over a first-candidate loop, while GEC governance primarily improves efficiency, scope discipline, and complexity control at comparable success..Source: §13 · p. 18
- The compact rules distinguish mandatory scope, rejection of a low-value action, economic stopping after hard completion, and blocked status when hard requirements lack a feasible path.A low-value candidate is rejected without rejecting the project.Source: §13 · p. 18
- Long-horizon evaluation should assess objective preservation, prerequisite discrimination, evidence validity, economic computation, recognition of sufficiency, and stopping for the right reason.Source: §13 · p. 19