PaperlayerView exact revision on arXiv

Source-linked AI summary

From Evidence to Effect: Authority Semantics and Runtime Infrastructure for Stateful Agents

Yang Li, Zongsi Xu, Sergey Volkov, hai liu, Tuo Zhou, Xiyu Chen, Di Wan, Dian Shao, Ye Luo, Hao Sun

arXiv:2609.08472v3cs.MAcs.SE

TL;DR

Stateful agents reuse artifacts even as executions and permissions change, creating a need to determine whether an effect is authorized. The paper formalizes authority-sufficient observations and execution-bound effects, then implements the interface in WTB. Its studies find that raw and typed evidence both resolve controlled authority cases, while planning effects vary by model and infrastructure integration improves observed completion alongside semantic parity of complete controls.

  • Problem

    Stateful agents may reuse artifacts after their executions or permissions have changed, so publication must account for both artifact correctness and execution authority.

    Source: §1 · p. 1

  • Method

    The paper formalizes authority-sufficient observations and an execution-bound authority interface, separating authority information, its representation, and commit-time effect admission.

    Source: §1 · pp. 1–2, §2 · p. 2

  • Results

    WTB improves observed fixed-budget completion in acquisition, recovery, and handoff while complete controls tie on capability; raw and typed evidence each achieve 32/32 semantic successes in the controlled ambiguity study.

    Source: §1 · p. 2, §5 · p. 6

  • Takeaways & Limitations

    The paper’s empirical claim is reusable authority infrastructure, with observed integration alongside semantic parity of complete controls.

    Source: §11 · p. 12

    The reference semantics assumes trusted state, mediated effects, and receiver cooperation, while coding studies use fixed tasks and recorded model routes and do not price human or library-construction effort.

    Source: §11 · p. 12

Abstract

from arXiv · show

Stateful agents reuse artifacts after producing executions and permissions change. We formalize authority-sufficient observations and durable effects bound to execution and material identities. WTB implements this interface through runtime adapters, shared evidence, and transactional publication/recovery. Six study families separate the mechanism from its integration. Raw and typed evidence both solve 32/32 authority cases, with model-dependent planning effects. Fixed-intent enforcement blocks six unsafe proposals and executes 12 eligible authorized intents. Complete controls match WTB's capability. Paid integration yields 176/210 accepted benchmark-source stages, including 19/30 publication stages, recovery on 8/8 primary SWE repositories, and the most complete continuous trajectories on each of three source tasks. The findings connect authority information, effect admission, and infrastructure reuse in stateful agents.

1 Introduction

Stateful agents need an explicit authority interface that binds execution evidence to current state, because correct artifacts can still reflect superseded work or revoked permissions. The paper separates the information needed to judge authority from how evidence is represented and how effects are admitted at commit time.

1 Introduction

  • A functionally valid shared artifact can still include superseded work or a contribution from a writer whose permission was revoked.A later publisher must resolve these relations across workspace, execution registry, and authority service, including when a successful remote write loses its response.Source: §1 · p. 1
  • The central claim is that stateful-agent effects need an explicit execution-bound authority interface whose meaning depends on historical evidence and current state.The theoretical contribution connects observation sufficiency, execution/material identity, and commit-time authority, building on standard observation factorization and complete mediation (Zheng and Yang, 2026).Source: §1 · pp. 1–2

2 Authority semantics

  • Authority is admissible when a decision rule using the observation correctly determines the effect request’s admissibility throughout the state domain.Formally, the observation must preserve the value of the admissibility function for each request.Source: §2 · p. 2
  • If hidden authority relations change while the visible material and history stay fixed, the same observation can require opposite decisions and yield only 1/2 expected accuracy on a balanced pair.The proof characterizes sufficiency as admissibility being constant within each observation-equivalence class.Source: §2 · p. 2
  • The paper separates adding authoritative facts, changing their representation, and changing effect admission for a fixed intent because planning information and mutable commit-time state pose different problems.Lossless raw and typed encodings preserve the same information partition, but an LLM may respond differently to their layouts; a previously valid decision can also become stale before commit.Source: §2 · pp. 2–3

3 WTB: binding evidence to durable effects

WTB binds publication authority to contributor executions, admitted material versions, approvals, and current permission. Its adapters preserve native identities while transactional sinks and recovery procedures maintain authority across local or external effects.

3 WTB: binding evidence to durable effects

  • A publication binding records contributors, material and its admitted version, and approval; freshness requires each contributor’s current attempt, matching completed receipt, and current permission generation.The material predicate also requires exact identity and contributor-chain continuity, so approval for one version must be renewed for another even if the bytes are equal.Source: §3 · p. 3
  • WTB admits an attempt with an exact execution envelope, binds the returned receipt to it, and uses a shared collector to check identity and record evidence.The adapters connect native identities through DSH/Cordis session and message identity, LangGraph thread/checkpoint identity, and a Ray task capture behind a driver-owned dispatch fence.Source: §3 · pp. 3–4
  • For local publication, one transaction checks authority, consumes approval, and records the effect; for external publication, it reserves an immutable effect key, payload, and authority snapshot.After a lost response, recovery looks up the key: it returns an applied receipt, retries confirmed absence only after rechecking authority, or leaves unresolved outcomes pending.Source: §3 · p. 3
  • Under trusted state and mediated transitions, each local effect satisfies the authority rule at its atomic transition, and each external effect passes the receiver’s authority check at commit.A fixed effect key permits at most one committed effect; the guarantees assume trusted evidence and authority state, serialized local check/consume/commit, and an atomic receiver token-check and deduplication.Source: §3 · p. 3
  • The architecture routes bound evidence or effect requests through adapters to a shared authority-evaluating core, which records local effects or reserves external delivery for a cooperating receiver.Current material and permission facts enter separately from bound evidence or requests.Source: Figure 1 · p. 4

4 Evaluation design

The evaluation separates authority and runtime controls from downstream integration work, using fixed source tasks, repository tasks, and distinct study protocols. Its reported counts retain their experimental units, and the supplied packages differ in responsibilities and remaining integration work.

4 Evaluation design

  • TeamBench, BulkPR-Bench, and Terminal-Bench each provide one fixed source task, while SWE-rebench provides ten qualified repositories and three calibration projects with native tests.BulkPR’s official hidden-verifier result is unavailable; downstream coding measures governance integration around the reference artifacts.Source: §4 · p. 4
  • Paid coding compares supplied WTB, Git/Auth, Continuity, and MasDrift packages on DSH, with the estimand being integration work above packages whose responsibilities differ.Original/enhanced contract checks separately test capability; runtime controls are WTB, SQL guard, and Continuity-bound adapter, with live tests on DSH/Cordis and LangGraph.Source: §4 · p. 4
  • The evaluation includes six study families answering five research questions, with RQ2 combining deterministic responsibility ablation and live interleavings.Counts retain their experimental unit, and native assertions within a coding stage remain within that stage.Source: Table 1 · p. 5
  • Three fixed benchmark-source artifacts support downstream coding studies, while preliminary W1 replays use zero paid model calls and remain outside formal study totals.The fresh-authority reference matches WTB on both unsafe and legal effects.Source: Table 2 · p. 5
  • Authority cells use DeepSeek V4 Flash and GPT-5.4 Mini, while planning uses DeepSeek V4 Flash and GPT-5.4 Nano; paid coding stages use DeepSeek V4 Flash.Paid adoption, prospective SWE, and continuous stages use 64 responses, 64 actions, and three acceptance submissions; effect replay, runtime, and contract studies use deterministic payloads without paid model calls.Source: §4 · p. 5
  • Table 3 distinguishes controlled semantic success from model-facing proposal behavior, and its two cohorts use different protocols and model routes.Raw receipts and typed bindings tie on semantic success, but their unsafe-proposal direction differs across models; unknown decisions are also reported.Source: Table 3 · p. 6

5 RQ1: What changes with information, representation, and enforcement?

Raw and typed authority evidence resolve the controlled ambiguity, while model-facing packaging changes proposal behavior in model-dependent ways. Fixed-intent enforcement separates admission from planning, and runtime studies test both complete controls and the consequences of removing specific responsibilities.

5 RQ1: What changes with information, representation, and enforcement?

  • 32/32 raw-evidence and typed-evidence cases succeed semantically, while Git-only and candidate-generation controls each achieve 0/32.The controlled contrast is authority information: C0–R0 has 32 favorable discordances and zero unfavorable, while R0 and W0 tie.Source: §5 · p. 6
  • 12/16 pooled unsafe proposals occur under GM0, 9/16 under R0, and 6/16 under W0, with the direction of the encoding effect differing by model.DeepSeek changes from 8/8 unsafe under R0 to 0/8 under W0, whereas GPT-5.4 Nano changes from 1/8 to 6/8; DeepSeek supplies all 11 W0 unknowns.Source: §5 · p. 6
  • 0/16 unauthorized effects occur under the matched-intent gate, which blocks all six unsafe proposals and executes 12/12 eligible authorized intents.Four other authorized-world rows lack a valid publish intent, separating an availability failure from effect admission.Source: §5 · pp. 6–7
  • 56/56 enhanced systems and WTB pass the complete contract checks, showing implementability across multiple architectural placements.The coding studies evaluate documented library and scaffold packages, so observed differences combine supplied functionality with work needed to connect it to an application.Source: §5 · p. 7

6 RQ2: Which runtime responsibilities preserve authority?

  • All complete controls pass 120/120 runtime-ablation cases, preserve 60 legal cases, and produce zero unsafe effects across the tested runtime/backend pairs.Removing exact-attempt binding causes 12 unsafe effects, generation removal causes six after revoke-and-regrant, consumption removal causes 12 replay effects, and monotone-reconciliation removal causes six unsafe effects and loses 12 legal completions; unknown-recovery removal loses six recoverable completions without unsafe effects.Source: §6 · p. 7
  • Across two runtimes, two backends, and three controls, all 12 relevant authority changes block publication, all 12 unrelated changes preserve one effect, and each implementation passes 8/8.The live interleavings test freshness during an in-flight call and scope precision for unrelated changes.Source: §6 · p. 7

7 RQ3: How readily do agents adopt the authority contract?

WTB’s independent-adoption results lead in accepted stages, while prospective repository tests show strong recovery and handoff outcomes. The studies also report capability parity under complete controls and uncertainty about broader repository populations.

7 RQ3: How readily do agents adopt the authority contract?

  • WTB passed 56/56 capability checks, and Table 4 reports parity across seven configurations; Table 5 covers 960 deterministic cases and five responsibility removals.The Table 5 caption defines legal-correctness as the expected completed state and effect count; a duplicate write can set completion while failing that metric.Source: Table 4 · p. 8, Table 5 · p. 8
  • 19/30 publication-stage slots were accepted by WTB, versus 7/30 Git/Auth, 0/30 Continuity, and 3/30 MasDrift.The stage requires binding approval to admitted material version, content, and current authority.Source: §7 · p. 8, Table 6 · p. 9
  • WTB passed independent E0–E2 bundles in 13/30 blocks and full E0–E6 bundles in 9/30, while each comparator passed zero full bundles.For E3–E6 bundles, WTB passed 20/30 versus 10/10/7 for the comparators.Source: §7 · p. 8
  • 176/210 WTB stages were accepted across the three source tasks, and WTB led accepted-stage count while lowering fixed-workload cost in all four cohorts.The pooled stage profile also identifies E0 acquisition, E1 publication, E2 recovery, and E4 contribution handoff as the strongest advantages.Source: §7 · pp. 7–8, Table 6 · p. 9

8 RQ4: Does integration transfer to real repository code?

  • Recovery-cost comparisons vary by repository, and the eight-repository primary sample leaves substantial uncertainty about performance across broader repository populations.The reported completion differences are large in this qualified sample, but adjusted comparisons remain inconclusive in some reported tests.Source: §8 · pp. 9–10
  • 8/8 primary repository recovery integrations and 4/8 handoffs were accepted by WTB, versus 1/8 recoveries and zero handoffs for each comparator.The prospective SWE table covers eight primary repositories under identical stage budgets; later stages can repair earlier failures.Source: §8 · p. 9, Table 7 · p. 10

9 RQ5: Does the same candidate sustain successive changes?

WTB completes more continuous trajectories than each comparator on all three source tasks, with the largest separation localized at multi-contributor handoff. The study also reports later-stage losses and statistical and substrate boundaries.

9 RQ5: Does the same candidate sustain successive changes?

  • At TeamBench handoff, Git/Auth passes E3 on all ten trajectories but E4 on three, while WTB passes E4 for 6/8 entrants there, 8/8 in BulkPR, and 8/9 in Terminal.The E4 obligations include exact attempts for earlier and later contributors, chain continuity, renewed publication authority, and recovery of those obligations.Source: §9 · p. 10
  • Later changes still lose WTB candidates: BulkPR falls from eight accepted E4 stages to six final successes, and Terminal falls from eight to six.The paper describes application adaptation as remaining fallible despite infrastructure reuse.Source: §9 · p. 10
  • All six joint-success cost pairs favor WTB by 1.16–54.32%, but all nine continuous Holm-adjusted comparisons exceed 0.05.Every method also passes 6/6 shared-backend checks per task; the receipt-recovery result concerns DSH JSONL/SQLite storage and restart behavior.Source: §9 · pp. 10–11
  • WTB completes 5/10 TeamBench, 6/10 BulkPR, and 6/9 Terminal trajectories, exceeding each comparator on every task.Figure 2 shows candidate counts across E3–E6; the table reports complete trajectories and E4 accepted/entered denominators.Source: §9 · p. 10, Table 8 · p. 11, Figure 2 · p. 11

10 Related work

Related work addresses agent authority, information flow, execution infrastructure, and provenance, while this paper focuses on maintaining authorization obligations at downstream effect boundaries. The conclusion and accompanying statements delimit the evidence and document reproducibility and AI use.

10 Related work

  • Prior work studies capability checks, information-flow policies, task-scoped intent, execution services, concurrent mutation, and provenance; WTB focuses on authorization obligations during downstream use.The related-work discussion distinguishes this focus from having or representing provenance alone.Source: §10 · pp. 11–12

11 Limitations and conclusion

  • The reference semantics assumes trusted state, mediated effects, and receiver cooperation; coding results measure integration above supplied packages on fixed tasks and recorded model routes.The authors say additional tasks, models, and effect APIs would test how far the measured advantages extend.Source: §11 · p. 12

Reproducibility statement

  • The companion experiment package includes observations, source provenance, protocols, reconstruction programs, detailed reports, and saved-record table and figure builders.The source archive also includes generated tables, figures, TeX sources, and bibliography needed to compile the manuscript.Source: Reproducibility statement · p. 12

AI use statement

  • LLMs served as agents in controlled planning and paid coding studies, while AI tools assisted with related work, ideas, proofs, and manuscript drafting and editing.The authors state that they verified reported results and take responsibility for the claims.Source: AI use statement · p. 12

A Authority sufficiency and the observation boundary

Authority decisions require observations that distinguish states with different admissibility, while durable effects must remain bound to current execution, material, and permission state. WTB operationalizes this contract through checked publication and recovery, with adoption results linking the interface to application integration.

A Authority sufficiency and the observation boundary

  • An observation is authority-sufficient when a decision rule can determine admissibility throughout the specified state domain; indistinguishable states requiring different decisions create aliasing.For a uniformly sampled aliased pair, decision accuracy is 1/2.Source: A Authority sufficiency and the… · pp. 14–15
  • Lossless raw and typed encodings preserve the same authority-fact partition, although models may plan differently from their layouts; a decision can also become stale before commit when relevant state changes.Retaining permission generations and a current-attempt pointer distinguishes revoke-and-regrant from uninterrupted permission and an old successful execution from its replacement.Source: A Authority sufficiency and the… · p. 15

B.1 Bindings, state, and assumptions

  • Fresh publication binds each contributor’s current admitted attempt and matching receipt to current scoped permission and generation, while the material chain must match the exact admitted target and version.The reference semantics assumes faithful native capture, durable trusted stores, complete mediation, serialized relevant updates, and a cooperating external receiver.Source: §B.1 · p. 15

B.2 Permitted transitions

  • After an approval is reserved, retry requires that reservation, authoritative confirmation of absence, and a fresh local snapshot check; receiver deduplication handles delayed sends racing with retry.A stored receipt records an existing effect but does not authorize a new effect after a policy change.Source: §B.1 · pp. 15–16, §B.2 · p. 16

B.3 Proof of Proposition 2

  • Local publication checks authority, consumes approval, and inserts the effect in one serialized transaction; external sends reuse an admissible reservation and the receiver atomically validates its current token.Unknown external outcomes retain the reservation, while a unique effect key prevents duplicate insertion.Source: §B.3 · pp. 16–17, Figure 3 · p. 17
  • External-effect safety does not guarantee recovery availability: a receiver that cannot answer may leave an effect pending indefinitely, and recovery additionally requires reliable lookup, stable retry authority, and a reachable sink.The reference proposition does not establish global ordering between policy updates and remote writes.Source: §B.3 · p. 17

B.4 Implementation and experiment correspondence

  • Responsibility ablations target execution identity, temporal authority, single consumption, terminal-evidence reconciliation, and recovery of already committed remote effects.WTB’s shared evidence path and reusable publication/recovery services are tested against those obligations.Source: §B.4 · pp. 17–18
  • 24/40 publication slots across four sources were accepted by WTB, including 19/30 benchmark-source slots.The E1 cohort tests publication from a supplied evidence/acquisition predecessor; continuous-adaptation starts had a different qualified E0–E2 condition.Source: §B.4 · p. 18

C Runtime adapters and architecture realization

WTB separates runtime-owned execution and persistence from shared identity, evidence, and publication/recovery responsibilities. Its adapter studies and integration cohorts distinguish native package capabilities from the application work and starting conditions being measured.

C Runtime adapters and architecture realization

  • WTB’s runtime boundary admits an attempt, binds native receipts to its identity, and connects runtime evidence to publication authority while leaving invocation and persistence semantics with each runtime.DSH, LangGraph, and Ray use distinct SDK/JSON-RPC, workflow-facade/checkpoint, and capture-bridge paths.Source: C Runtime adapters and architecture… · p. 18, Figure 4 · p. 19
  • DSH supports durable message evidence but not checkpoint rewind; LangGraph’s evaluated facade uses process-local provider identity, and Ray’s prototype excludes driver-restart reconciliation.A new LangGraph facade cannot establish continuity from the same checkpoint path alone; the runtime experiments do not imply uniform restart recovery.Source: C Runtime adapters and architecture… · pp. 18–19

D Experimental inventory and interpretation

  • Six study families comprise distinct protocols and cohorts, while the inventory keeps controlled-matrix cases, coding-stage outcomes, trajectories, and response records as separate units.The study record includes ten qualified prospective SWE repositories, three calibration projects, and three fixed source tasks.Source: D Experimental inventory and… · p. 21
  • The experiments separate observation-level evidence, implementation-level governance mechanisms, and runtime-level evidence; SQL guard and Continuity-bound adapter are complete effect-control implementations.Paid-arm differences reflect supplied capability, API convenience, and application wiring rather than a single implementation dimension.Source: D Experimental inventory and… · pp. 21–22, §D.1 · p. 22

D.1 Supplied capabilities and application starting points

  • WTB supplies persistent evidence and publication, external-delivery, and authority operators; other arms provide differing native governance mechanisms, with application authorization and integration work varying by package.The inventory describes supplied functionality without assigning equal implementation effort.Source: §D.1 · p. 22
  • Independent-adoption stages range from empty starts to predecessors supplying earlier capabilities, prospective SWE starts empty, and continuous adaptation begins only after all arms qualify at E0–E2.Continuous adaptation carries forward accepted candidates, whereas independent-adoption slots are initialized independently.Source: §D.1 · pp. 22–23

E Original Mechanism: evidence, intent, and effect

The original-mechanism studies separate authority information, planner decisions, and effect admission, then test whether complete controls preserve authority across implementations and changing permissions.

E.1 Question and controlled design

  • 32 favorable and zero unfavorable discordances distinguish R0 from C0; R0 and W0 tie on semantic success, while G0 and C0 each score 0/32.The contrast isolates the addition of authority facts from their representation; the paired design spans four templates, two model routes, two lineage directions, and two authority worlds.Source: §E.1 · pp. 23–24

E.2 Planning and matched-intent enforcement

  • The W0 publication gate rejects all six unsafe intents and executes all 12 eligible authorized intents, while leaving four authorized rows without a valid publish intent.The same 32 saved planning outputs are replayed at the effect boundary; the separate planning study reports pooled unsafe proposals of 12/16 under GM0, 9/16 under raw receipts, and 6/16 under W0, with model-dependent behavior.Source: §E.2 · p. 25

E.3 Original and enhanced comparison mechanisms

  • Enhanced Git/Auth, Continuity, and MasDrift controls and WTB each pass all 56 checks, while original implementations pass 36, 44, and 28, respectively.The checks test supplied execution, material, evidence, generation, and publication obligations; enhancement hours were not measured.Source: §E.3 · p. 25

F.1 Design and responsibilities

  • WTB, an independently implemented SQL guard, and a Continuity-bound adapter each pass 120/120 runtime cases, execute all 60 expected legal effects, and produce zero unsafe effects.The deterministic matrix uses native captures from DSH/Cordis, LangGraph, and Ray across SQLite and PostgreSQL.Source: §F.1 · p. 25

F.2 Failure interpretation

  • Removing exact-attempt binding yields 12 unsafe effects, removing scoped-generation binding yields six, and removing monotone reconciliation yields six unsafe effects and 48/60 legal effects.Removing unknown-effect recovery preserves safety but loses valid recoverable work; the matrix separates these responsibility-specific failure signatures.Source: §F.2 · p. 26

G.1 Temporal intervention

  • Across 24 live executions, all 12 relevant-scope changes block publication, while all 12 unrelated-scope changes preserve one effect.Each complete implementation passes 4/4 within each runtime; the tested cohort covers DSH/Cordis and LangGraph, not Ray.Source: §G.1 · pp. 26–27, Table 13 · p. 27

G.2 Mechanism and boundary

The studies test fresh authority after in-flight permission changes and measure how supplied interfaces affect staged integration and transfer across fixed benchmark-source tasks.

G.2 Mechanism and boundary

  • Each of 12 relevant-scope changes yields safe rejection and zero effects, while each of 12 unrelated changes preserves authority and yields one effect.WTB, the SQL guard, and the Continuity adapter each pass 8/8; the live design covers DSH/Cordis and LangGraph.Source: §G.2 · p. 27

H.1 Question, intervention, and stage meanings

  • The independent-adoption study assigns four arms seven staged tasks, starting each slot from a specified scaffold with a common coding budget; later slots do not inherit accepted code from earlier slots.The stages cover evidence acquisition, publication, recovery, a second workflow, contribution-chain handoff, receipt-envelope change, and strict parsing.Source: §H.1 · p. 27

H.2 Earlier controlled cohort

  • WTB accepts 51/70 stages at $7.4660, versus 39/70 at $10.2273 for Git/Auth, 37/70 at $11.5691 for Continuity, and 39/70 at $11.7276 for MasDrift.In the earlier controlled cohort, WTB also accepts 5/10 publication stages, 7/10 recovery stages, and 8/10 handoff stages.Source: §H.2 · pp. 27–28

H.3 Three-source transfer and heterogeneous stages

  • Across three benchmark-source tasks, WTB accepts 176/210 stages, compared with 121, 108, and 103 for the three comparators; Git/Auth and Continuity complete more E3 and E6 slots.WTB accepts 7/10, 7/10, and 5/10 publication slots across TeamBench, BulkPR-Bench, and Terminal-Bench, and records 9/30 full E0–E6 bundles versus zero for each comparator.Source: §H.3 · pp. 28–29
  • The fixed-workload spending advantage is limited to the studied tasks, libraries, scaffolds, and budgets, and excludes starting-library construction and human engineering effort.Cost per accepted stage can reflect both cheaper attempts and more accepted attempts; joint-success costs cover only the overlapping successful subset.Source: §H.3 · p. 29

I.1 Selection and cumulative protocol

  • The prospective repository study selects ten qualified workloads—eight primary and two development repositories—and schedules 120 stage slots across four arms.The selection and cumulative protocol specify the repository cohorts and staged integration workload.Source: §I.1 · p. 29

I.2 Primary outcomes and paired uncertainty

In the primary repository cohort, WTB completes recovery on all eight repositories, with descriptive completion advantages but substantial inferential uncertainty and task-dependent cost tradeoffs.

I.2 Primary outcomes and paired uncertainty

  • WTB accepts recovery on 8/8 primary repositories, versus 1/8 for each comparator, and accepts handoff on 4/8 while each comparator accepts zero.None of the six planned contrasts passes the familywise 0.05 threshold after Holm adjustment; the authors note that a larger prospective sample would test sensitivity to selection and model configuration.Source: §I.2 · p. 30

I.3 Cost and counterexamples

  • Through recovery, WTB spends $3.1825, 16.4%, 33.7%, and 7.7% less than Git/Auth, Continuity, and MasDrift, respectively; through handoff, Git/Auth costs less but accepts no handoff.Joint-success costs favor WTB in the Kafka recovery comparison and favor the comparator in the other four listed pairs; reference-patch artifacts place autonomous issue-solving outside this estimand.Source: §I.3 · p. 30

I.4 Calibration and development

  • In an equally budgeted calibration continuation, WTB completes 3/3 projects, Git/Auth 1/3, and Continuity and MasDrift 0/3; on two development repositories, WTB accepts 2/2 core, 1/2 recovery, and 1/2 handoff stages.The calibration bounds the cost claim: WTB spends $0.4725 and Git/Auth $0.2229 on the first jointly completed project.Source: §I.4 · pp. 30–31

J.1 Sequential design and outcome definition

  • The continuous study follows paid implementations through E3–E6, with later stages consuming the preceding accepted candidate and failure closing the remaining trajectory stages.The complete-case analysis includes 117 trajectories and 468 stage outcomes; interrupted trajectories with unresolved outcomes are outside the analysis, so results do not establish outcomes for unobserved executions.Source: §J.1 · p. 31

J.2 Stage survival and source-specific findings

  • WTB completes 5/10 TeamBench, 6/10 BulkPR-Bench, and 6/9 Terminal-Bench trajectories; its E4 handoff acceptance is 6/8, 8/8, and 8/9 entrants, respectively.WTB has the most complete trajectories on each source, while the cost comparisons vary by task and by whether methods jointly complete a block.Source: §J.2 · pp. 31, 33

J.3 Failure categories and implementation behavior

  • E4 requires a material chain binding both writers to exact execution attempts and renewed authority checks, while small application diffs measure reuse from supplied libraries, not total system complexity or development effort.Successful comparator rows also show that other implementations can achieve the contract.Source: §J.3 · p. 33

J.4 Paired uncertainty and joint-success cost

The continuous study shows a consistent descriptive direction across three fixed tasks, alongside substantial statistical uncertainty and explicit limits on reconstruction and generalization.

J.4 Paired uncertainty and joint-success cost

  • WTB has more complete trajectories on each of three fixed tasks, but no adjusted comparison passes 0.05 and the smallest adjusted value is 0.28125.The repeated blocks measure coding variability on these tasks; broader benchmark-distribution claims require more independently sampled tasks.Source: §J.4 · p. 34, Table 21 · p. 34

J.5 Storage and qualification evidence

  • All twelve starting implementations pass 117 qualification assertions each, totaling 1,404; DSH execution and restart work on JSONL and SQLite, while cross-backend import and binary-version interoperability need separate testing.Six same-backend location moves recover twelve exact receipts, and all four methods pass 6/6 backend-contract checks on each task without additional application changes.Source: §J.5 · pp. 34–35

K.1 Source provenance and functional outcomes

  • The source tasks use distinct frozen workload revisions and material origins, and the governance studies measure downstream publication, recovery, and maintenance rather than benchmark-wide issue-solving.BulkPR’s official hidden-verifier result is unavailable; the reported functional evidence for that workload comes from saved local tests.Source: §K.1 · p. 35

K.2 Task-level W1 preliminary replays

  • In nine preliminary source-task replays, WTB and the fresh-authority reference produce zero unsafe effects, compared with four unsafe effects without an additional guard.These small exploratory replays are separate from the 128-cell authority matrix and 960-case runtime ablation denominator.Source: §K.2 · pp. 35–36

L Reconstruction and reproducibility scope

  • The compact package supports offline reconstruction of recorded results but omits the complete stack needed for fresh end-to-end execution.Changing provider behavior, native-library versions, prompts, or acceptance contracts defines a new experimental condition; the source archive can rebuild the manuscript and regenerate selected assets from saved numeric records.Source: L Reconstruction and reproducibility… · p. 36
Loading 2609.08472v3…