PaperlayerView exact revision on arXiv

Source-linked AI summary

Emergi-PersonaOS: A Persona Agent Operating System for Situational Adaptation and Controllable Evolution

Haoluan Fu, Keni Chen, Xinyu Jia, Jinpeng Wang, Yuyu Yin, Yubiao Hu

arXiv:2609.27417v2cs.AI

TL;DR

Long-term human–machine relationships call for persona agents that connect enduring identity with situation-specific expression and development from experience. Emergi-PersonaOS uses a shared three-layer persona object for situational adaptation and reviewed, evidence-traceable belief revision. Its implementation demonstrates long-horizon operation with television-character dialogue, including memory accumulation across episodes.

  • Problem

    Persona research needs an integrated account connecting enduring persona content, current states, and event evidence across lifecycle stages.

    Source: §1 · p. 1, §2.3 · p. 4

  • Method

    Emergi-PersonaOS uses a three-layer persona object to infer current states from situational evidence and retrieved experience, then manage belief revisions through review, evidence, and version records.

    Source: §1 · p. 2, §3.3.2 · p. 10

  • Results

    The television-character implementation accumulated 119 memory entries across six seasons; earlier-season entries in retrieval rose from 43.6% in Season 3 to 80.6% in Season 4 and reached 90% in Season 6.

    Source: §1 · p. 2, §5.2.2 · p. 17

  • Takeaways & Limitations

    Separating current states from enduring beliefs lets the system represent situational variation while subjecting longer-term changes to review and supporting evidence.

    Source: §6 · pp. 20–21

    The findings concern system operation for one fictional character and one longitudinal narrative; extension to real-user personalization requires separate investigation.

    Source: §6 · p. 21

Abstract

from arXiv · show

Symbiosis between humans and digital beings offers a vision for the future of human--machine interaction. In enduring human--machine relationships, personality provides a foundation for continuity of identity, individuality in interaction, and development through experience. We investigate this capacity through persona agents as computational implementations and introduce Emergi-PersonaOS, a psychology-grounded operating system for managing persona objects throughout their lifecycle. The system organizes dispositional traits, characteristic adaptations, and narrative identity into a three-layer persona representation, distinguishing relatively enduring persona beliefs from their activation in the current persona state. During situational adaptation, it integrates the current interlocutor, relationship, event, and retrieved memories to infer a persona state and generate actions and replies; during long-term development, it records experiences and outcomes, and develops and evaluates revision candidates through change attribution, meaning-making, and behavioral testing. Belief updates are managed through explicit review, traceable evidence and version records, and the ability to reject candidates, making persona evolution controllable. Using television-character dialogue as longitudinal material, we demonstrate long-horizon system operation and examine its principal mechanisms in a concrete implementation. This work provides a computational framework for persona agents to maintain individual continuity, produce situation-specific expression, and develop through experience over sustained interaction.

1 Introduction

Long-term human–machine relationships make personality relevant to continuity, individualized interaction, and development through experience. Emergi-PersonaOS addresses the need to connect situational expression with evidence-based long-term persona change.

1 Introduction

  • Persona agents need an explicit current state linking enduring persona content to behavior shaped by interlocutors, relationships, and events.Fixed-profile access offers only a partial account of how one individual responds across situations.Source: §1 · p. 2
  • Long-term persona updating must establish an event’s developmental significance, interpretation, scope, and evidence before revising enduring content.The introduction notes concerns about compressed individual differences and insufficiently supported change directions in reflection-driven evolution.Source: §1 · p. 2
  • Emergi-PersonaOS separates enduring beliefs from current states in a shared three-layer persona object spanning traits, characteristic adaptations, and narrative identity.The same object supports construction, situational expression, and long-term updating.Source: §1 · p. 2
  • Belief revisions require review, traceable evidence and version records, and rejectable candidates; the system demonstrates its operation with longitudinal television-character dialogue.Revision candidates are developed and evaluated through change attribution, meaning-making, and behavioral testing.Source: §1 · p. 2

2 Related Work

Related work addresses persona content, situational expression, and experience-based updating, but these lines often use distinct representations and execution mechanisms. The section frames lifecycle integration as linking enduring content, current states, and event evidence.

2.1 Personality Structure and Computational Representation

  • Character-modeling work uses profiles, experiences, and behavioral examples, while a common psychological account of their levels and roles remains an open representational need.RoleLLM uses character profiles, knowledge instructions, and style examples (Wang et al., 2024a); Character-LLM models experiences, knowledge, emotions, and behavior (Shao et al., 2023).Source: §2.1 · pp. 2–3

2.2 Situational Adaptation of Personas

  • Situational methods incorporate context and relevant experience, while evaluation research motivates jointly recording inferred states, behavior, and evaluation conditions.Examples include memory retrieval conditioned on task applicability (Zhang et al., 2026b) and multidimensional dialogue evaluation (Tu et al., 2024).Source: §2.2 · p. 3

2.3 Controllable Persona Evolution

  • Memory and reflection preserve experience for later action, while persona evolution additionally requires determining which events justify long-term change and how character information should be revised.Prior work also examines explicit update procedures, evidence management, developmental evaluation, and belief revision (Li et al., 2026; Wang et al., 2026a; Tang et al., 2026; Zhang et al., 2026c).Source: §2.3 · pp. 3–4
  • Lifecycle integration calls for shared links among enduring persona content, current states, and event evidence, recorded continuously across stages.The related lines often address different time scales and information objects through separate representations and execution mechanisms.Source: §2.3 · p. 4

3 Overall Design of Emergi-PersonaOS

Emergi-PersonaOS manages a shared persona object across construction, situational expression, and long-term development, separating enduring beliefs from context-sensitive states. It governs behavior and revision through evidence-based attribution, testing, and review, with approved updates committed at turn boundaries and recorded in traceable histories.

3 Overall Design of Emergi-PersonaOS

  • The design separates enduring persona beliefs from current states so behavior can adapt to situations while persistent revisions remain traceable to experience.Six functional modules manage the persona object across construction, expression, and development.Source: §3 · p. 4

3.1.1 Theoretical Foundations

  • The theoretical foundation combines a three-layer personality representation with mechanisms for interpreting situational states, linking memories to persona entries, and reviewing developmental changes.The layers are dispositional traits, characteristic adaptations, and narrative identity; state monitoring preserves time-indexed records of situations, behavior, and outcomes.Source: §3.1.1 · pp. 4–5

3.1.2 Design Principles

  • Three design principles guide the system: lifecycle management, separation of enduring content from situational expression, and evidence-based, controllable evolution.Insufficient or conflicting evidence leads to continued observation rather than an unsupported update.Source: §3.1.2 · pp. 6–7

3.2.1 Overall Architecture

  • The architecture organizes six functional modules around a shared persona object, connecting construction, state inference and behavior generation, trajectory recording, and long-term revision.The modules are persona construction, execution engine, state monitoring, memory, reflection, and evolution.Source: §3.2.1 · p. 7, Figure 3.1 · p. 7

3.2.2 The Persona Object

  • The persona object stores beliefs and current states in the same three-layer structure, distinguishing enduring content from its activation, interpretation, and expression in a situation.Each entry records source, evidence location, confidence, applicability, time, version, and modifiability, enabling traceability across beliefs, states, experiences, and revisions.Source: §3.2.2 · p. 8, Table 3.3 · p. 8

3.2.3 Functional Modules

  • The six modules initialize persona beliefs, generate situation-specific behavior, monitor and store execution evidence, interpret changes, and review candidates for long-term updates.Approved candidates create a new belief version at a turn boundary; others may be rejected, deferred, or left pending verification.Source: §3.2.3 · pp. 8–9

3.3.1 System Execution Flow

  • Execution proceeds from persona construction through interaction, event, and developmental timescales, with reflection and evolution assessing accumulated evidence before committed beliefs change.During review, the engine continues using the active belief version and switches only after a new version is committed.Source: §3.3.1 · p. 9, Figure 3.2 · p. 10

3.3.2 Situational Adaptation

  • Situational adaptation retrieves context-relevant experiences, infers a jointly checked three-layer state from beliefs, prior state, and current evidence, then generates and checks actions and replies.Inference identifies implicated identities and meanings, current needs, values, goals and strategies, and salient trait expressions; unsupported activations are removed while uncertain content can remain unknown or pending confirmation.Source: §3.3.2 · pp. 9–11

3.3.3 Controllable Evolution

  • A revision is eligible only when it has sufficient evidence, follows modification rules, passes behavioral testing, and resolves substantive conflicts.Decisions are reviewable, evidence and version histories are traceable, and candidates can be rejected.Source: §3.3.3 · p. 11
  • The system retains beliefs when they and the situation explain a change, defers unresolved explanations, and proposes revisions only for evidence-supported long-term change.Candidates specify a target layer and scope; evidence standards differ across enduring self-understanding, needs and strategies, and persistent trait expression.Source: §3.3.3 · p. 11
  • Evidence review checks sources, attribution, duplicates, counterevidence, field modifiability, scope, and base version before candidates undergo paired frozen-copy tests.Tests hold state and memory snapshots, model, prompts, settings, and situations fixed; target, similar, and out-of-scope situations assess expression, applicability, and unsupported effects. Candidate text and behavior shaped by execution recommendations require corroboration from source evidence.Source: §3.3.3 · p. 12
  • Approved revisions apply only at turn boundaries, updating specified belief fields and reconciling affected state fields while retaining unrelated state content.The current turn remains tied to its original belief version and state, and the reconciled state starts the next turn; jointly committed candidates are checked for compatibility and combined effects.Source: §3.3.3 · pp. 12–13
  • Revision records preserve base versions, edits, evidence, review conclusions, and activation time; new evidence, counterevidence, or scheduled checks can trigger reassessment.Corrections and revocations create new versions without erasing history, and decisions include revise, retain, defer, and statuses such as rejected or pending verification.Source: §3.3.3 · p. 13

4 Experimental Setup

The evaluation follows a longitudinal run of persona inference and response generation across seven seasons, with separate tests for situational adaptation, cross-episode memory, and long-term updating. Shared inputs and controlled comparisons support assessment of the system’s components and outputs.

4.1 Materials and Data Preparation

  • The run used dialogue from 141 episodes across seven seasons; Season 1 built initial persona beliefs, and Seasons 2–7 were processed in broadcast order with six season-end reviews.Each episode produced a narrative, key events and dialogue, information gaps, and provenance records.Source: §4.1 · p. 13

4.2 Evaluation Tasks and Metrics

  • Situational adaptation was tested at 48 fixed nodes across six seasons, while a separate 12-item assessment covered five cross-episode memory categories.The nodes were equally divided between family situations and school or peer situations.Source: §4.2 · p. 13
  • Season-end updating tests evaluated candidate decisions, revision content, and behavior, requiring traceable evidence and narrowing limited-evidence candidates before submission.Behavioral tests checked both expression in target situations and inactivity when the revision did not apply.Source: §4.2 · pp. 13–14

4.3 Comparison Systems and Ablation Conditions

  • Situational-adaptation comparisons included dialogue-only, static-persona, L1-only, no-cross-episode-memory, and full-system conditions; long-term updating compared the full system with no episode-level accumulation.Conditions shared the source scope, initial persona, model version, and generation settings.Source: §4.3 · p. 14, Table 4.1 · p. 14

4.4 Procedure and Implementation

  • Evaluation used repeated response ratings, blinded condition labels, evaluator admission checks, programmatic compliance checks, and paired analyses at nodes or season-end tests.Three samples per node were rated by three evaluator instances; long-term behavioral effects used pre–post differences and directional paired judgments.Source: §4.4 · pp. 14–15

5 Results

Across the evaluations, the full system outperformed the dialogue-only baseline on behavioral ratings and showed strong cross-episode recall, while memory-use precision was imperfect. Season-end revisions were evidence-reviewed and produced measurable behavioral changes, with the detailed case illustrating a narrowed, conditional update.

5.1 Structural Compliance

  • The mean structural-compliance rating was 1.97 out of 2, and no registered critical errors occurred; evaluator agreement was ICC(2,1) = 0.81.Reflection-appropriateness deductions occurred in ten episodes, mainly involving generalization from one event or unsupported motive attribution.Source: §5.1 · p. 15

5.2.1 Behavioral Performance across Conditions

  • 4.48 for the full system versus 2.77 for the dialogue-only baseline, with paired advantages of 1.71 points over dialogue-only and 0.72 over L1-only.The full system scored higher than dialogue-only at every evaluation node; the 1.71-point difference was the overall mean across four behavioral dimensions.Source: §5.2.1 · p. 15, Table 5.1 · p. 16

5.2.2 Cross-Episode Memory Assessment

  • The full system scored 0.88 versus 0.28 without cross-episode memory; it scored 1.00 on cross-episode and cross-season recall, while memory hit rate at history-dependent nodes was 0.79.The memory store accumulated 119 entries, and retrieved entries from earlier seasons reached 90% in Season 6; false-event rejection scored 0.63.Source: §5.2.2 · p. 17, Figure 5.3 · p. 17, Figure 5.4 · p. 18

5.2.3 Case Study of Situational Adaptation

  • In the Paige scene, the full-system reply showed character-specific defensive behavior, while the dialogue-only reply was generic and the no-memory reply received ratings close to the full system.The case comparison used the same input across the three conditions; historical memory supplied additional interaction-specific context.Source: §5.2.3 · pp. 17–18, Figure 5.5 · p. 19

5.3.1 Review Outcomes and Persona Revisions

  • All six season-end checkpoints produced committed revisions that matched reference decisions on target layer and direction, and all revisions concerned characteristic adaptations.Dispositional-trait magnitude evaluation was therefore inapplicable; content outside the revisions remained unchanged.Source: §5.3.1 · p. 19

5.3.2 Behavioral Effects of Persona Revisions

  • The full reflection-and-memory procedure exceeded no episode-level accumulation by 0.36 on eligible target tests, while belief-entry updates contributed more directionally than added runtime policies.Excluding already-at-ceiling pairs, the pooled gain across six seasons was 0.34 across 32 pairs.Source: §5.3.2 · pp. 19–20, Table 5.3 · p. 19

5.3.3 Case Study of Controllable Evolution

  • A Season 3 revision narrowed limited evidence to family-authority intervention or interrupted support; target-situation realization rose from 0.43 to 0.96 after revision.The revised response replaced direct confrontation with self-regulation and temporary acceptance while retaining factual statements or questions about alternatives.Source: §5.3.3 · p. 20

6 Discussion

Emergi-PersonaOS uses a shared persona representation to connect enduring beliefs, situational expression, and long-term updating. The discussion also defines implementation boundaries involving evaluation, data scope, model choice, and revision timing.

6 Discussion

  • The three-layer representation gives construction, situational expression, and long-term updating a common basis while separating enduring beliefs from current states.Psychological categories keep content assignments consistent across tasks, preserving representational continuity as control passes between processes.Source: §6 · p. 20
  • Situational evidence shapes a current state under active beliefs, while belief changes require a separate review of eligibility and supporting evidence.State evaluation uses pre-response material so assessment is tied to the state available when the response was generated.Source: §6 · p. 20
  • Long-term belief revisions proceed through change attribution, meaning-making, and behavioral testing, with candidates subject to evidence review, traceability, rejection, and reassessment.Provenance records and source checks keep the basis for each belief inspectable; candidates remain uncommitted until eligible.Source: §6 · pp. 20–21
  • Formal ratings depend on fixed-setting evaluators; human evaluation and its effects on absolute scores and condition rankings have not been quantified.Randomized labels and blinded comparisons address condition-identification effects, but evaluator-model dependence remains.Source: §6 · p. 21
  • The longitudinal material covers one fictional character in one narrative, so real-user personalization requires separate investigation with authorization and ethics review.The authors also plan to test the revision procedure across multiple characters and calibrate change magnitude.Source: §6 · p. 21
  • Every stage uses the same underlying model, leaving the effects of separating generation and evaluation models or replacing that model untested.Future work will examine model-level separation and sensitivity to the model used.Source: §6 · p. 21
  • Persona revisions are reviewed only at season-end boundaries; within-season changes remain monitoring records, and event-level review rules are not implemented.Future work proposes event-level review with evidence requirements and tests.Source: §6 · p. 21

7 Conclusion

The conclusion presents a persistent persona object linking enduring beliefs, current states, and event evidence, demonstrated over 141 episodes. The appendices specify the study’s data-use scope and implementation procedures for comparison conditions and runtime policies.

7 Conclusion

  • Across all 141 episodes of Young Sheldon, the system’s shared persona object and situational and long-term processes operated together over an extended run.The demonstration preserved links among enduring persona content, current states, and event evidence.Source: §7 · p. 21

A Data Use and Research Ethics

  • The study interprets textual material for one fictional character, excludes real-participant psychometrics and private or sensitive data, and limits claims to specified character-dialogue tasks.Persona inference excludes physical and vocal performance; future information and unclear-provenance material are excluded from pre-response inference and season-end evidence.Source: A Data Use and Research Ethics · p. 26

B.1 Implementation of Comparison Conditions and Behavioral-Test Versions

  • The behavioral-test comparison holds revised belief entries fixed while varying whether current-season behavioral policies are added.The three versions are the preceding-season version, revised beliefs with current-season and inherited policies, and revised beliefs with inherited policies only.Source: §B.1 · p. 26

B.2 Compilation and Attachment of Runtime Behavioral Policies

  • Runtime policies encode triggers, selection conditions, rules, scope, exclusions, and fallback behavior; triggers must be observable and scope limited to evidence-supported situations.A trigger depending on interaction history cannot apply at a first encounter, so programmatic checks reject it during compilation.Source: §B.2 · p. 26
  • Required policies are attached to candidate beliefs for frozen-copy testing, and successful beliefs, reconciled states, and applicable policies are committed together at a turn boundary.Programmatic checks block activation when applicable policy generation, compilation, attachment, or record-writing checks fail.Source: §B.2 · pp. 26–27

B.3 Implementation Limitations

The implementation has three stated limitations: prompt wording can affect outputs, sequential modules can propagate errors, and runtime policies cover only part of how persona content influences behavior.

B.3 Implementation Limitations

  • Prompt wording may affect module outputs, although prompt versions were fixed before the formal run to limit this variation.Source: §B.3 · p. 27
  • Sequential modules allow upstream state errors to propagate into response generation and evaluation, while monitoring and programmatic checks intercept only some errors.Source: §B.3 · p. 27
  • Conditional runtime policies cover only part of how persona content influences behavior; when no policy triggers, revisions act through the belief entry itself.Source: §B.3 · p. 27

C.1 Examples of Core Data Objects

The appendix specifies the persona-belief structures, construct categories, and episode materials used to support persona records and longitudinal execution.

C.1 Examples of Core Data Objects

  • Versioned belief entries carry evidence identifiers, applicability scopes, and runtime-policy slots, with examples of characteristic adaptations and a triggered behavioral policy.The policy example links ritual cues, requested physical participation, and the absence of a verifiable factual or logical basis to its trigger conditions.Source: §C.1 · pp. 27–28

C.2 Predefined Construct Taxonomy

  • The fixed taxonomy assigns general construct categories to persona layers; character-specific entries are generated from Season 1, and unsupported constructs are marked unknown.L1 uses the five Big Five domains; categories outside the predefined taxonomy are excluded.Source: §C.2 · p. 28, Table C.1 · p. 28

D.1 Material Construction

  • Episode materials preserve dialogue and provenance while organizing each episode into a narrative, key events, information gaps, and pre- and post-response evidence.Only records passing structural checks and LLM content review enter the formal material set; source files, prompt and schema versions, and fingerprints are retained.Source: §D.1 · p. 28

D.2 Data Splits

The appendix describes season-based data splits, evaluation materials, and rating procedures for situational adaptation, structural compliance, responses, persona updates, and behavioral tests.

D.2 Data Splits

  • Post-commit behavioral tests are derived by a separate designer after a season’s revision is committed, while evaluation nodes and memory-check situations are fixed before outputs.Initialization and longitudinal-execution sets are divided by season.Source: §D.2 · p. 29

D.3 Selection of Situational-Adaptation Evaluation Nodes

  • Situational-adaptation nodes are screened for observable responses and outcomes, then registered with fixed dialogue spans, relationship categories, historical evidence, and exclusions.The final selection covers relationship categories and seasons; evaluators rate the registered spans without expanding context.Source: §D.3 · p. 29

D.4 Season-End Evidence Sets and Cross-Episode Memory Assessment

  • The 12-item memory assessment tests recall, partner and similar-experience discrimination, and rejection of false events; natural-node checks separately report memory hits and misuse.The 12-item score difference between the full system and the no-cross-episode-memory condition estimates memory’s contribution.Source: §D.4 · p. 29

E.1 Structural Compliance Ratings

  • Structural compliance scores six episode processes across eight dimensions from 0 to 2, while any recorded critical error causes the unit to fail.Full structural compliance requires all units to pass.Source: §E.1 · pp. 29–30

E.2 Character-Response Ratings

  • Character responses receive four 0-to-5 ratings covering persona fidelity, situational coherence, state–behavior consistency, and dialogue continuity.Ratings record reasons and evidence identifiers; paired comparisons use the same node and metric.Source: §E.2 · p. 30

E.3 Long-Term Persona Update Ratings

  • Persona-update ratings assess revision decisions, evidence scope, preservation, and fidelity, with semantic change and applicability scope reported separately for L2 and L3 entries.Candidates supported by a single episode or partly consistent evidence must first be narrowed in scope and meet the minimum evidence standard to be committed.Source: §E.3 · p. 30

E.4 Behavioral Tests and Directional Paired Judgments

  • Behavioral tests assess target-situation realization and out-of-scope inapplicability; directional paired judgments compare anonymized outputs with and without current-season policies and belief revisions.The formal run used eight target situations and four to six out-of-scope situations per season; target realization averages four equally weighted ratings and is scaled from 0 to 1.Source: §E.4 · pp. 30–31

E.5 Statistical Analysis

The analysis uses paired and reliability statistics to summarize condition comparisons while following predefined analysis rules.

E.5 Statistical Analysis

  • Situational-adaptation comparisons pair conditions at the same node, report ICC(2,1) and Cronbach’s α separately, and summarize paired preferences with ties weighted 0.5.The analysis introduces no unplanned weighted composite metrics or post hoc subgroup analyses.Source: §E.5 · p. 31

F Generation and Evaluation Settings

The appendix specifies generation and evaluation procedures, then reports grouped situational-adaptation summaries and season-level revision decisions and content.

F.1 Model and Parameters

  • The formal run used deepseek-v4-pro-0813 through an OpenAI-compatible interface, with temperature 0.2 and task-dependent maximum output lengths generally set to 6,000 tokens.Model identifiers, prompt versions, and interface configurations were fixed before execution, and every retry was recorded.Source: §F.1 · p. 31

F.2 LLM Reference Generation and Evaluation

  • Reference generators produce states, acceptable ranges, or revision decisions from defined evidence; evaluators then return dimension scores, reasons, evidence identifiers, and inapplicability markers.References can record multiple acceptable directions when the material supports multiple interpretations.Source: §F.2 · p. 32

F.3 Condition Randomization and Blinded Evaluation

  • Condition outputs were anonymized with randomized labels, and each situational node had three response samples rated by three separately randomized evaluator instances.Units with missing or duplicated labels were excluded from effect calculations and recorded as evaluation failures.Source: §F.3 · p. 32

F.4 Evaluator Admission and Precalibration

  • Evaluators entered formal rating only after meeting a predefined error-identification threshold on ten control pairs containing deliberately inserted errors.Admission checks occurred in three batches.Source: §F.4 · p. 32

G.1 Node-Level Paired Differences in Situational Adaptation

  • The 48 situational-adaptation nodes comprise 24 family and 24 school-or-peer situations; pairing within nodes accounts for differences in node difficulty.Table G.1 presents condition means and paired differences by relationship category.Source: §G.1 · p. 32, Table G.1 · p. 32

G.2 Grouped Summaries

  • Table G.1 reports grouped condition means and paired differences; memory comparisons use a restricted set of 21 nodes identified as requiring memory.Other comparisons use all 24 nodes in each relationship category.Source: Table G.1 · p. 32, §G.2 · p. 32

G.3 Season-Level Decisions and Revision Content

  • Both the system and reference procedure committed revisions at all six season-end checkpoints; evidence was multi-episode in two seasons and single-episode or partly consistent in four.The revisions addressed factual priority, family coping, autonomy, and ritual-related expression across Seasons 2–7.Source: §G.3 · p. 33

G.4 Run Failures and Missing Evaluation Items

All formal adaptation nodes, cross-episode memories, and season-end checkpoints entered evaluation, with run failures and missing evaluation not causing exclusions.

G.4 Run Failures and Missing Evaluation Items

  • All formal situational-adaptation nodes, cross-episode memory items, and season-end checkpoints with test situations entered evaluation.Source: §G.4 · p. 33
  • No evaluation unit was excluded because of a run failure or missing evaluation.Source: §G.4 · p. 33
  • Evaluation-unit exclusions, failure categories, and their handling were reported separately from effect scores.Source: §G.4 · p. 33

H Run Records, Error Categories, and Replication Information

The appendix records run and call provenance, reports formal-run call and retry counts, classifies errors and their handling, and describes the replication materials and access boundary.

H.1 Run Records

  • Run records capture scope, settings, identifiers, timing, call-level inputs and outputs, hashes, reuse status, and reasons for replacing failed or superseded results.Matching calls receive the same request fingerprint; structurally and content-checked results may be reused.Source: §H.1 · p. 33

H.2 Model-Call Statistics

  • The formal run recorded 2,327 model calls, with 2,325 accepted, two abandoned, and 165 retries following failures.Source: §H.2 · p. 33

H.3 Error Categories

  • Errors fall into six categories; temporal, evidence, version, and evaluation errors exclude affected units from effect calculations, while material errors are checked against source records.Reported categories are material, temporal, persona-representation, evidence, version, and evaluation errors.Source: §H.3 · p. 33

H.4 Replication Materials

  • Replication materials include manifests, evaluation and evidence materials, condition specifications, prompts, criteria, checks, and scripts, but exclude copyrighted continuous scripts and complete episode dialogue.Episode identifiers and evidence locations let researchers consult legally available sources to verify original text.Source: §H.4 · p. 33

Related papers

Loading 2609.27417v2…