PaperlayerView exact revision on arXiv

Source-linked AI summary

LSREP: A Longitudinal State-Replay Protocol for Evaluating Conversational Memory, with ICE v2 as an Audited Local-First Architecture

Deepesh Sonar

arXiv:2609.16730v1cs.AIcs.CLcs.IR

TL;DR

Endpoint question answering alone cannot show how persistent conversational memory accumulates, ages, or incorporates revisions. This paper introduces LSREP to replay memory state over time and audit whether claimed mechanisms ran. In its complementary public diagnostic, ICE v2 performs decisively worse than pure vector-RAG on LongMemEval, while replay results show a different regime-dependent pattern.

  • Problem

    Endpoint QA over supplied histories does not by itself require repeated observations of reconstructed memory under a declared maintenance and access schedule.

    Source: §1 · p. 1

  • Method

    LSREP replays conversational history at successive checkpoints, scores against evolving references, and checks whether mechanisms named in results actually ran.

    Source: §1 · p. 2

  • Results

    ICE v2’s results are regime-dependent: ordinary-density replay shows similar mean scores with fewer fragments, while matched LongMemEval shows decisive overall losses and severe multi-session and temporal failures.

    Source: §9 · p. 20

  • Takeaways & Limitations

    Longitudinal replay, mechanism-fidelity auditing, and public endpoint testing expose distinct failure modes that architectural descriptions or aggregate scores alone do not identify.

    Source: §9 · p. 20

    The matched LongMemEval diagnostic uses a non-official Muse judge without a broad benchmark-specific human-agreement study, and its intervals omit judge and run-to-run variation.

    Source: §8.3 · p. 19

Abstract

from arXiv · show

Conversational memory changes during use, so endpoint question answering alone cannot establish how a persistent state accumulates, ages, or incorporates revisions. We introduce LSREP, a Longitudinal State-Replay Evaluation Protocol combining ordered replay, explicit lifecycle schedules, repeated probes, evolving reference answers, and mechanism-fidelity checks. Its architectural case study is ICE v2, a local-first memory middleware with typed stores, retrieval fusion, and dynamic context budgets. The private, single-user instantiation contains 1,985 turns, 219 distinct probes, and 1,211 probe-checkpoint observations across 52 checkpoints. On three ordinary-density datasets, ICE v2 has a near-zero mean quality difference from vector-RAG while selecting 32% fewer fragments but using 6.6% more estimated prompt tokens. A fourth, dense dataset exposes catastrophic failures of the unbudgeted baseline. The fidelity audit limits attribution: procedural retrieval is defective, several mechanisms are unexercised, and graph utility is not established. In a complementary matched public diagnostic, ICE v2 loses decisively to pure vector-RAG on LongMemEval: 50.8% versus 72.8% in the evidence-only oracle and 43.0% versus 69.5% in full-S. Paired differences are -22.0 points (95% CI [-26.6, -17.4]) and -26.5 ([-31.3, -21.8]). Conservative abstention accompanies severe multi-session and temporal failures. ICE uses less context in this diagnostic, establishing a quality-cost trade-off rather than superior efficiency. Together, replay, fidelity auditing, and public endpoint testing expose distinct failure modes that neither architectural descriptions nor aggregate scores identify alone.

1 Introduction

Persistent conversational memory raises questions about what is retained, revised, and governed that endpoint QA alone cannot answer. The paper introduces LSREP to evaluate reconstructed memory over time and presents ICE v2 as an audited architectural case study.

1 Introduction

  • Endpoint QA over supplied histories tests updates and temporal reasoning but does not require repeated observations under a declared memory-maintenance and access schedule.The paper frames its research questions around replayed-state change, mechanism execution and effect, and transfer to separately supplied sessions.Source: §1 · pp. 1–2
  • ICE v2 is a local-first middleware that stores multiple memory types, classifies prompts, retrieves and fuses candidates, and assembles bounded context.Its stores include episodic turns, a temporally versioned knowledge graph, procedural patterns, documents, and user-controlled memory slots.Source: §1 · pp. 1–2
  • LSREP reconstructs memory at successive checkpoints and scores repeated probes against evolving references, complementing fixed-history benchmarks.Its methodological claim is deliberately falsifiable: a near-zero ablation delta supports “does not help” only when a component is verified live and the interaction regime reaches it; the authors report where their own evaluation failed that standard.Source: §1 · p. 2
  • The paper’s contributions pair a system-independent replay protocol with an ICE v2 architectural case study and a two-regime evaluation with a fidelity audit.The private corpus is not independently reproducible, and empirical portability of LSREP remains to be tested.Source: §1 · p. 2

2 Related Work

Related work spans persistent memory architectures, graph and adaptive retrieval, and evaluations of long-term interaction. The paper positions LSREP as a specification of state changes and valid answers over time, rather than as a replacement for existing endpoint benchmarks.

2.1 Persistent and Structured Memory

  • Persistent-memory systems already combine storage, lifecycle management, structured representations, and retrieval; ICE v2 integrates these ideas in inspectable local middleware.Mem0 supports ADD, UPDATE, DELETE, and NOOP operations, while Zep’s Graphiti uses temporally qualified relationships, so neither is accurately described as append-only or as lacking temporal graphs.Source: §2.1 · pp. 2–3

2.2 Graph and Adaptive Retrieval

  • Prior graph-retrieval work motivates structural retrieval but does not establish the correctness of ICE v2’s evolving graph, which must be measured in the implementation under study.The paper also distinguishes ICE v2’s PostgreSQL lexical leg, historically called “BM25” but not canonical BM25, from its rank-based RRF fusion; the v2 ablation supports a corrective effect after adding the lexical leg, while HyDE was not isolated.Source: §2.2 · p. 3
  • Retrieval-generation methods can accommodate changing corpora, while LSREP specifies when the corpus and derived state change and which answer is valid at each observation.The cited evaluation regimes address different targets, and repeated checkpoint observations reveal changes that a single endpoint score leaves unresolved; neither regime subsumes the other.Source: §2.2 · p. 3, §2.3 · p. 3

2.3 Evaluation Regimes

  • Long-term interaction benchmarks test distinct targets including multi-session dialogue, temporal reasoning, updates, and abstention rather than only static single-shot retrieval.LSREP’s unit is the state trajectory: a history prefix, lifecycle schedule, reference version, and access history at each checkpoint.Source: §2.3 · p. 3
  • LLM judges can be sensitive to answer position and verbosity, and blinding and fixed rubrics mitigate some risks without establishing unbiased labels.This matters when comparing the paper’s Muse-judged LongMemEval results with published GPT-4o-judged results.Source: §2.3 · p. 3

3 LSREP: Requirements, Algorithm, and Validity

LSREP evaluates a reconstructed memory trajectory by replaying supplied history under declared lifecycle and access conditions, observing state and answers at checkpoints, and comparing them with time-appropriate references. Its validity conditions cover temporal, replay, measurement, mechanism, and statistical claims, while acknowledging limits in the historical instantiation.

3.1 Evaluation Object and Requirements

  • LSREP observes a reconstructed trajectory, not exact recovery of an unlogged historical deployment, using available turns, persistent state, a declared maintenance schedule, and memory-affecting access events.Each probe is scored against a reference derived only from turns available by that checkpoint; the observation includes the supplied context and any query-caused state transition.Source: §3.1 · pp. 3–4
  • An admissible system needs ingestion, querying, reset or snapshot capability, and control over maintenance and query side effects; graphs and decay are not required.The protocol also requires ordered history, explicit checkpoint boundaries, temporally grounded references, consistent conditions, and versioned evidence of what ran.Source: §3.1 · p. 4

3.2 Protocol Algorithm

  • Reconstructing a checkpoint must reproduce permitted prior access events, because replaying the same turns without them defines a different experimental condition.A deterministic schedule does not guarantee identical model outputs across executions.Source: §3.2 · p. 4

3.3 Evolving Ground Truth

  • When new evidence revises a fact, current-truth references change while explicitly time-anchored historical answers remain; ambiguous revisions require adjudication or uncertainty labels.References stay outside retrievable memory, and repeatedly answering an outdated reference correctly does not count as successful updating.Source: §3.3 · p. 4

3.4 Validity Conditions and Scope

  • Validity requires temporal and replay controls, complete prompt accounting and inspectable scoring, evidence that mechanisms execute and affect answers, and explicit uncertainty scope.The protocol was developed alongside ICE and tested on one author’s private history; its v2 audit exposed defective and unexercised components, so independent users, systems, and releasable trajectories remain needed.Source: §3.4 · pp. 4–5

4 Synthetic Worked Example

The synthetic example illustrates how LSREP distinguishes future leakage, current-truth updating, and preservation of revision history. It is an illustration of protocol semantics, not an additional ICE v2 experiment.

4 Synthetic Worked Example

  • The worked example is invented, contains no private corpus text, and illustrates the protocol rather than adding an ICE v2 experiment.At T1, answering PostgreSQL leaks future evidence.Source: §4 · p. 5
  • The example’s reference changes only when the history available at that checkpoint warrants it.The present v2 LSREP corpus tests current truth; historical and evolution probes in the illustration describe capabilities a future instantiation could test, not demonstrated results.Source: Table 1 · p. 5, §4 · p. 5
  • At T2, SQLite is stale for the current-truth question, while PostgreSQL erases history when used as the answer to the historical question.A richer evolution probe may accept both versions in their order.Source: §4 · p. 5

5 ICE v2 Architecture and Audit Contract

ICE v2 is local-first middleware that routes each turn through synchronous retrieval and asynchronous memory updates, with an audit contract distinguishing implemented mechanisms from validated quality effects.

5 ICE v2 Architecture and Audit Contract

  • ICE v2 intercepts requests, retrieves from persistent stores, assembles bounded context, routes to a specialist model, and updates memory after the response.Its answering model can change without losing the externalized, inspectable memory store.Source: §5 · p. 5

5.1 Request Lifecycle

  • Pre-flight classifies and retrieves context synchronously; post-flight stores new turns and dispatches extraction, while scheduled workers age and consolidate memory.The prompt assembler combines retrieved fragments with persistent slots, recent turns, and the live message.Source: §5.1 · pp. 5–6, Figure 2 · p. 6

5.2 Memory Stores

  • ICE v2 uses four stores and structured overlays; its temporal graph preserves superseded edges, while access restores decay and procedural patterns can deactivate.The graph distinguishes currently stored edges from historically superseded ones; a current edge is not independently verified as true.Source: §5.2 · pp. 6–7

5.3 Retrieval, Fusion, and Budget

  • Six retrieval legs are designed, but only episodic legs are operational; procedural lookup is defective, document and summary stores are empty, and graph utility is inconclusive.Weighted rank fusion is followed by bonuses, source-type diversification, and greedy packing; the 2,000-token growth cap applies to an empty current conversation.Source: §5.3 · p. 7

5.4 Local Operation and Audit Contract

  • The local deployment exposes inspectable stores and audit surfaces, but execution alone does not establish that a selected mechanism benefits answer quality.Fidelity requires tracing eligibility, execution, output, selection, and prompt inclusion; quality attribution additionally requires a controlled contrast.Source: §5.4 · pp. 7–8, Figure 3 · p. 8

6 Evaluation Instantiation

The evaluation replays four private long-form conversations with evolving references and repeated probes, then compares ICE v2 with vector-RAG in replay and matched public endpoint conditions.

6.1 Four LSREP Datasets

  • Four private datasets span 251–1,119 turns and complementary domains; their simulated ageing horizons are schedules, not claims about the conversations’ calendar duration.Raw conversations and probe text remain private because they contain the sole author’s personal data.Source: §6.1 · p. 8

6.2 ICE v2 Replay and Answer Conditions

  • Continuous replay preserves state across checkpoints, re-evaluates earlier probes, and updates temporally anchored references as new turns revise facts.Replay lacks much of the reinforcement that real use supplies, so these results are not a lower or upper bound on deployment.Source: §6.2 · pp. 8–10
  • Four conditions compare full ICE with single-leg vector-RAG across generalist and routed models; manual records contribute to headline aggregates, and hallucination labels received partial review.The comparison is full memory system versus standard vector-RAG, not retrieval alone versus retrieval alone.Source: §6.2 · pp. 9–10

6.3 Metrics and Statistical Units

  • Quality and context use are reported jointly; replay confidence intervals resample probe-checkpoint records and do not quantify generalization across users or independent trajectories.The near-zero mean difference indicates no detected difference, not formal equivalence.Source: §6.3 · p. 10

6.4 Matched LongMemEval Setup

  • The documented v2 adapter evaluates matched oracle and full-S phases from fresh conversations, with both arms reading the same stored vectors; its vector-RAG comparator retrieves top 30 turn representations without decay, fusion, or token budget.These are within-study comparisons, not leaderboard-comparable scores. The adapter measures the documented v2 path and omits the API wrapper’s faulty secondary word check.Source: §6.4 · p. 11

7 Results

LongMemEval endpoint testing found decisive ICE v2 deficits, while LSREP showed similar ordinary-density mean scores and a severe density-stress failure for the unbudgeted vector baseline. The cumulative ablation and fidelity audit limit what can be attributed to individual mechanisms.

7.1 RQ1: Longitudinal Behaviour under LSREP

  • The archived analysis found no detected mean-score difference between scores of 4.25544 and 4.25355 (separately rounded difference 0.01); ICE v2 selected 20.4 rather than 30.0 fragments on average (32.0% fewer).ICE v2 used 22,411 rather than 21,025 estimated prompt tokens (6.6% more).Source: §7.1 · p. 12
  • At Dataset B’s final turn-1,100 bin, both vector conditions fell to 3.80 and 3.87 while both ICE conditions held at 4.18 and 4.14.This single-conversation final-bin difference is descriptive and does not identify retrieval reach as its cause.Source: §7.1 · p. 12, Figure 4 · p. 13

7.1.1 Density Stress and the All-Dataset View

  • On Dataset C, 94.2% of unbudgeted vector-baseline observations received score 1; the all-data paired difference was +0.40, rising to +3.10 on Dataset C.The baseline’s mean prompt volume was 88,061 tokens, while ICE’s dynamic budget held it to 19,953 tokens and a mean score of 4.33.Source: §7.1.1 · pp. 14–15, Table 8 · p. 14

7.2.1 Cumulative Ablation

  • Adding an unfused lexical leg was harmful (−0.74), while rank fusion recovered +0.82; the budget-step estimate of −0.08 was inconclusive.Neither the full configuration nor RRF-on-vector was statistically distinguishable from the single-leg baseline on answer quality.Source: §7.2.1 · pp. 15–16

7.2.2 Execution, Reach, and Identifiable Effects

  • The audit found procedural retrieval defective, document retrieval defective and empty, batch summaries without eligible content, and graph utility inconclusive.Session diversification did not exercise its central cross-conversation cap in within-conversation replay; decay and reinforcement ran but were not independently switched off.Source: §7.2.2 · p. 16

7.3 RQ3: External Transfer under LongMemEval

  • −22.0 points in oracle and −26.5 in full-S are the paired LongMemEval differences for ICE v2 versus vector-RAG.Both paired intervals exclude zero: [−26.6, −17.4] and [−31.3, −21.8], respectively.Source: §7.3 · p. 16, Table 11 · p. 17
  • Full-S multi-session accuracy was 21.5% versus 74.4%, and temporal accuracy was 20.5% versus 47.2%, exposing severe synthesis and temporal-composition failures.The paired differences were −52.9 points (95% CI [−62.8, −43.0]) and −26.8 (95% CI [−35.4, −18.1]); the results do not locate the failure stage.Source: §7.3 · p. 17
  • In full-S, ICE supplied a median of 5 fragments and 2,222 input tokens, versus vector’s 30 fragments and 11,718 tokens, showing a quality–cost trade-off rather than an efficiency win.The comparison does not establish accuracy at matched budgets or either system’s position on an accuracy–token frontier.Source: §7.3 · p. 17, Table 12 · p. 18

8 Limitations and Implications

The authors qualify the private replay results, mechanism attribution, and public benchmark comparisons, and call for evaluations that combine evolving state with endpoint performance and execution audits.

8.1 Private Data, Repeated Probes, and Reference Construction

  • The 1,211 observations repeat 219 questions from four conversations by one author, so neither uncertainty analysis supports population-level generalisation.Model-assisted reference refinement and a manually scored subset sharing the corpus author add further limits; exact private-data results cannot be independently reproduced without the data.Source: §8.1 · p. 19

8.2 Replay and Mechanism Boundaries

  • Replay fixes future turns and its shared-state comparison, while the density case compares ICE with an unbudgeted top-30 baseline, not a properly budgeted vector system.Simulated ageing schedules are not elapsed deployment durations; LSREP results also cannot isolate a causal session effect from LongMemEval scores.Source: §8.2 · p. 19

8.3 Judging and Cost Boundaries

  • Judge labels and benchmark comparisons remain uncertain: intervals omit judge misclassification and run-to-run variation, and lower token use with lower accuracy is not superior efficiency.The answerer was selected on a small sample from the same benchmark, and the matched LongMemEval study uses a non-official judge.Source: §8.3 · p. 19

8.4 Implications for the Next Evaluation

  • The authors call for evolving-state evaluation, public endpoint testing, and execution audits, with future matched-budget comparisons and separate state trajectories.Graph quality and cross-session synthesis remain unmeasured; future work should also test ranking stability across answerer and judge families.Source: §8.4 · pp. 19–20

9 Conclusion

The paper presents LSREP as a protocol for evaluating reconstructed conversational memory over time and ICE v2 as an architectural case study. Results vary by regime: continuing-use replay and the matched public endpoint diagnostic produce different outcomes.

9 Conclusion

  • LSREP specifies state reconstruction over time through declared transitions, access effects, repeated probes against evolving references, and audits of whether claimed mechanisms ran.ICE v2 is the architectural case study, with typed stores, local operation, retrieval fusion, and bounded prompt construction; the paper identifies limits on effective mechanisms.Source: §9 · p. 20
  • Matched LongMemEval shows decisive overall losses for ICE v2, while continuing-use replay shows similar ordinary-density mean scores and robustness against an unbudgeted baseline in the density stress case.The public diagnostic also finds conservative abstention alongside severe multi-session and temporal failures, and lower context use with less accurate answers.Source: §9 · p. 20
  • The paper concludes that longitudinal evaluation, mechanism-fidelity auditing, and public end-task testing are jointly necessary for credible memory-system claims.The differing evaluation regimes expose failure modes that architectural descriptions or aggregate scores alone do not identify.Source: §9 · p. 20

Reproducibility and Ethics

The public artifact package supports inspection of the evaluation, but excludes private data and materials needed to reproduce exact private results. The component audit makes its classifications checkable while underscoring limits on what the evidence establishes.

Reproducibility and Ethics

  • The artifact package includes aggregate reports, scripts, configuration, and adapter and harness code, but excludes private corpora and raw evaluation materials; exact private LSREP results are not independently reproducible.Public LongMemEval inputs must be obtained from their original release.Source: Reproducibility and Ethics · p. 20
  • Local storage and explicit memory controls support user inspection, but do not prove privacy or correctness; persistence increases the consequences of incorrect or sensitive stored information.The public diagnostic uses remote generation and judging on public benchmark material.Source: Reproducibility and Ethics · p. 20

A Component Fidelity Audit

  • Table 14 makes the component audit’s three-way classification checkable against code and frozen results: defective, never exercised, or contributing.“Contributing” means execution affected retrieved output; “never exercised” means no benchmark input engaged the component, and “defective” means it executed and failed.Source: A Component Fidelity Audit · p. 23, Table 14 · p. 24
  • The 34.5% unattributable source records are an instrumentation limitation; graph fragment granularity and a loaded NER checkpoint do not establish graph correctness or content share.Source: A Component Fidelity Audit · p. 23

B Historical ICE v1 Pilot

The historical pilot evaluated an immature ICE build with incomplete mechanisms and protocol choices that differ from the later LSREP instantiation. Correcting token accounting preserved the qualitative findings but removed any claim of superior efficiency.

B Historical ICE v1 Pilot

  • The pilot used an immature build: graph, procedural, and document legs contributed nothing, while retrieval largely reduced to vector-plus-BM25 and cluster-scoped retrieval did not exist.The graph used a regex tagger and under-powered extractor; the classifier had a frozen MiniLM backbone and retrieval used static limits and a 5,000-token budget.Source: B Historical ICE v1 Pilot · p. 25
  • Pilot probes came from three seeded splits per eligible conversation, were re-asked at later splits, and used a single retrieval-assisted reference pass frozen thereafter.Conversations shorter than 10 turns were excluded; each split paired a historical block with a 10-turn future reference block.Source: B Historical ICE v1 Pilot · p. 25
  • Correcting the omitted sliding-window tokens showed full ICE used more total tokens than the baseline at roughly equal quality, invalidating the pilot’s superior-token-efficiency claim.The omitted window had been supplied at inference; the correction changed efficiency statistics, not generated answers or scores.Source: B Historical ICE v1 Pilot · p. 25
  • The pilot headline was 4.04 against 4.06, with 47% more tokens, more hallucination, and a worse TUR; these historical ICE v1 results do not measure frozen ICE v2.Source: B Historical ICE v1 Pilot · p. 26, Table 16 · p. 26
  • The pilot’s short replay and incomplete consolidation tested that particular state, while the under-counted context made ICE appear more efficient than it was.The authors identify broken subsystems as a source of system fragility and a precursor to the later fidelity audit.Source: B Historical ICE v1 Pilot · p. 26

C ICE v2 Full Ablation Results

The ablation identifies significant damage from adding BM25 and recovery from RRF, while the full-ICE contrast with the single-leg baseline is inconclusive. The accompanying breakdowns and implementation details qualify what the mechanisms and score patterns establish.

C ICE v2 Full Ablation Results

  • −0.74 [−1.14, −0.36] for +BM25 and +0.82 [+0.39, +1.24] for +RRF were the only significant buildup steps; full ICE versus the single-leg baseline was −0.06 [−0.38, +0.25].The full-build contrast and other listed contrasts span zero; the vector baseline is a reference, not a buildup step.Source: Table 17 · p. 26, C ICE v2 Full Ablation Results · p. 27, Figure 5 · p. 27

D ICE v2 Detailed LSREP Breakdowns

  • Across 1,057 probe–checkpoint observations, Dataset C excluded, the mean-score tie masks ICE-MoE’s lower score-1 rate and shift from score 5 toward scores 3–4.The temporal breakdown groups scores as Good (4–5), OK (3), and Poor (1–2).Source: Table 20 · p. 28, Table 21 · p. 28

E.1 Classification Engine

  • ICE’s classifier outputs topic, intent, and context-reliance labels that guide retrieval-leg weights, budget allocation, fallback, and model routing.The rule-based pre-classifier applies ordered density rules; its reference rule forces Long Term Memory, with additional long-conversation memory biases.Source: §E.1 · p. 29

E.2 Codex Write Path and Retrieval

  • Codex resolves entities by canonical name, then alias, or creates a deterministic UUIDv5; property observations expire prior edges for the same source–relation pair, while other relation types follow distinct update rules.State changes emit events, and entity context is rebuilt from properties plus the ten most recent active outgoing edges.Source: §E.2 · p. 29
  • Codex retrieval resolves prompt entities, optionally restricts conversation scope, and traverses current edges to depth three; fuzzy matching uses entity embeddings and a cosine threshold of 0.85.MERA enumerates category-query candidates only when entity extraction returns none and the prompt meets its category and enumeration triggers.Source: §E.2 · p. 30

E.3 Codex Limitations in Detail

This material details limits in Codex and prompt handling, the retrieval orchestrator, and additional evaluation analyses. The ordinal sensitivity analysis finds an overall ICE advantage concentrated in density-only observations, while ordinary-density results remain inconclusive.

E.3 Codex Limitations in Detail

  • Codex traversal ignores edge strength and mostly ignores confidence, so weak and strong current edges are treated alike and retrieval does not reinforce frequently accessed facts.An active confidence flag supplies only a coarse 1.5× fragment-level score boost.Source: §E.3 · p. 30
  • Entity-centric retrieval assumes users refer to entities by canonical names or close aliases, which may miss natural-language descriptions without a lexical or embedded anchor.The passage illustrates this with “The Obsidian Citadel” and the later query “where is the main fortress located?”Source: §E.3 · p. 30

E.4 Prompt Assembly

  • The HTTP wrapper’s nominal 4,096-token-derived check omits recent and retrieved messages, so it is not a valid total-prompt ceiling.The evaluation adapters bypass this wrapper check; no secondary full-context guarantee is claimed.Source: §E.4 · p. 30

E.5 Background Workers and Infrastructure

  • The configured worker schedules do not establish useful evaluation output, and extraction idempotency keys are not a verified exactly-once guarantee.GPU workers gate on utilization, and shared-model workers also check a user-activity key.Source: Table 22 · p. 31, §E.5 · p. 31

F.1 Hybrid Retrieval Orchestrator

  • Retrieval combines six heterogeneous legs with intent-conditioned weights and rank-based RRF; post-fusion curation and a two-phase budget prioritize source-type diversity.The dynamic budget leaves unused capacity when needed; ordinary-turn curation yields fewer fragments, not fewer tokens.Source: §F.1 · pp. 31–32

I Historical Local Oracle and Adapter Invalidation

  • The historical local oracle scored ICE at 55.2% and vector at 80.2%; it stopped before full-S and is superseded by the later completed cloud phases.An earlier flattened-session adapter was invalid because it failed to preserve session boundaries and reduced 111 fused candidates to three after diversification.Source: I Historical Local Oracle and Adapter… · p. 36

J LSREP Sensitivity and Reproduction

  • Probe-cluster resampling preserved the qualitative LSREP findings: ordinary-density ∆= +0.002, while density-only ∆= +3.097; no population claim follows.The analysis resampled 219 probe clusters, not independent conversations or users.Source: J LSREP Sensitivity and Reproduction · p. 36

J.1 Scoring-Source, Missing-Score, and Ordinal Sensitivity

  • +11.1 points (cluster CI [+3.2, +19.5]) was the all-data net ordinal superiority of ICE; ordinary-density net superiority was +0.1 ([−7.6, +8.0]), while density-only was +87.0 ([+77.5, +95.3]).These direction-only comparisons retain failure assignment and supplement the historical mean rubric scores.Source: §J.1 · p. 37
Loading 2609.16730v1…