PaperlayerView exact revision on arXiv

Source-linked AI summary

Decoupling Logic from Persona: Structural Immunity of Edge LLM Agents to Context Pollution

Masaaki Nakatsu, Reno Wang

arXiv:2610.09772v1cs.CLcs.AI

TL;DR

Persona–logic interference occurs when persona expression and logical inference compete in an accumulating context. The paper separates these jobs into logic and persona paths, finding that the input bound helps across both model sizes while the dedicated adapter adds robustness only on the 8B model.

  • Problem

    Persona–logic interference is the loss of task-logic quality when persona expression and inference compete in one accumulating context, where useful history can become noise or misleading evidence.

    Source: §1 · p. 1

  • Method

    The decoupled architecture routes history to the persona path while the logic path receives a bounded input and produces a structured state for persona rendering.

    Source: §9 · p. 18

  • Results

    The input bound was the only mechanism that worked on both model sizes; the dedicated adapter rescued most degradation on Llama-3.1-8B but not Gemma-3-4B.

    Source: §7 · p. 17

  • Takeaways & Limitations

    Polluted single-pass turns lose structured fields in 80–100% of cases, threatening downstream validators, UIs, and cache keys.

    Source: §7 · p. 17

    The benchmark covers two tasks, two domains, two personas, and one hardware platform, so its effect sizes should not be generalized beyond misleading, persona-heavy history on 4-bit 4–8B models.

    Source: §8 · p. 17

Abstract

from arXiv · show

Small language-model agents on edge devices must hold a persona and reason correctly at once, inside one context window that fills with conversational history and persona instructions. We study what happens to the logical part of such an agent when that history is long, misleading and persona-heavy (persona-logic interference), and present a Decoupling Architecture (AO-DA) that separates logical inference ("What") from persona expression ("How") into two inference paths on one INT4 base model with hot-swappable LoRA adapters. The logic path receives only the core turn and emits a verifiable structured state (Micro-State); the persona path renders it in character with the full history. In same-base-model ablations on an Apple M2 laptop (Llama-3.1-8B-Instruct and Gemma-3-4B-it, 4-bit; 480 runs over 4 pollution levels x 3 arms x 2 tasks x 2 personas x 5 seeds) we find: (i) the decoupled logic path is structurally invariant to pollution: its prompt stays at 180 (Llama) or 167 (Gemma) tokens while the mixed single-pass prompt grows from 242 to 1,203, and its outputs are byte-identical across levels (40/40); (ii) the mixed single pass degrades monotonically (composite logic score 0.669 to 0.150 on Llama, 0.487 to 0.150 on Gemma), mostly by failing to emit the required structured output (80-95% of runs on Llama, 100% on Gemma at the two highest levels); (iii) with the same pollution fed into the decoupled logic path, the dedicated-adapter, dedicated-format path is still more robust than the single pass on the 8B model (failure 0-20% vs 80-95%; paired $\Delta$ +0.30 to +0.50, Cliff's $\delta$ 0.50-0.85, Holm-adjusted $p \le 0.03$) but not on the 4B model, where both collapse. Separation costs one extra decode on a topic's first turn (28.2 s vs 18.2 s on Llama) and buys persona hot-swapping in 1.7 ms without re-running the logic path. Code, rubric, fixtures, adapters and logs are released.

1 Introduction

Persona–logic interference arises when persona expression and logical inference compete in one accumulating context. AO-DA separates those jobs, and the paper reports structural invariance of the bounded logic path alongside failures in the mixed single pass.

Motivation

  • The motivation is that one shared context gives the persona job useful history but can make that same history noise or misleading evidence for logic.The paper frames long, distracting context, persona prompts, and multi-turn conversation as converging concerns for 4–8B models quantized to 4 bits on a laptop.Source: §1 · p. 1
  • Persona–logic interference is the loss of task-logic quality when persona expression and logical inference compete inside one accumulating context.The paper distinguishes this architectural effect from capability loss through alignment training.Source: §1 · p. 1

Proposal

  • AO-DA separates logical inference (“What”) from persona expression (“How”) into two paths on the same quantized base model.The What path receives the core turn and observable signals and emits a schema-checked Micro-State; the How path receives that state with full conversation history and renders the reply in character.Source: §1 · pp. 1–2

Contributions

  • The mixed single-pass configuration’s composite logic score falls monotonically to the scoring floor on both bases, while the What path’s outputs are byte-identical across pollution levels (40/40).At the two highest levels, structured-output failure reaches 80–95% on Llama and 100% on Gemma; the dedicated-adapter, dedicated-format path is more robust than single pass on 8B, but not 4B.Source: §1 · p. 2
  • The paper also reports persona hot-swapping at 1.7 ms median and Micro-State cache reuse that avoids rerunning the logic path on a persona switch.It reports a second-decode first-turn latency cost and releases code, rubrics, fixtures, adapters, and logs.Source: §1 · p. 2

2 Related Work

AO-DA draws on work about long-context reasoning, persona prompts, structured output, and multi-adapter serving. Its distinct arrangement separates reasoning and verbalization across generations using a verifiable intermediate object.

Long and distracting context

  • Long-context studies report failures to use middle-position information, degradation from irrelevant context, and reasoning loss with input length (Levy, Jacoby, and Goldberg 2024).The paper situates its misleading-history tests alongside those findings and multi-turn performance degradation reported by Laban et al. 2025.Source: §2 · p. 2

Persona and role prompts

  • Persona prompts can introduce implicit reasoning biases, while prior assistant statements can elicit sycophancy (Sharma et al. 2023).AO-DA removes persona from the path that reasons rather than trying to make the persona path reason better.Source: §2 · p. 2

Structured output

  • Constrained decoding guarantees format (Willard and Louf 2023), but AO-DA studies unconstrained output because failure to emit structure is itself measured as interference.The paper contrasts this with Tam et al. 2024, who found format restrictions can reduce reasoning quality.Source: §2 · p. 2

Multi-adapter serving

  • AO-DA adapts multi-adapter serving to a single 4-bit base resident in unified memory, with two adapters swapped in milliseconds.The related work places this alongside LoRA, S-LoRA, and Punica.Source: §2 · p. 2

Separating reasoning from verbalization

  • Unlike chain-of-thought, plan-and-solve, and ReAct, AO-DA separates reasoning and verbalization across generations using a verifiable intermediate object.The second generation can be rerun with a new persona or skipped on a cache hit.Source: §2 · p. 3

3 Architecture: AO-DA

AO-DA keeps the What and How prompts independent: logic processes a bounded, history-free input into a Micro-State, and persona expression renders it with the full history. The same resident INT4 base serves both paths through adapter swaps and cache reuse.

3 Architecture: AO-DA

  • The architecture’s runtime internals beyond the described effects are out of scope; the experimental runtime is released.Source: §3 · p. 3

3.1 Two inference paths

  • The What path uses a fixed header, vitals, and core utterance to emit a thought and JSON state/action fields; the How path uses the Micro-State and full history to produce the spoken line.The prompts are assembled independently, and neither path sees the other’s context.Source: §3.1 · p. 3

3.2 Micro-State

  • The Micro-State is the only object passed from What to How and carries provenance, constraints, risk, a render plan, cache key, and parsed What output.For the two benchmark tasks, a deterministic policy layer derives constraints and risk from vitals and session intent; the model’s free text fills inference and parsed fields.Source: §3.2 · pp. 3–4

3.3 One base, resident adapters, hot-swap

  • Both LoRA adapters share one 4-bit base loaded once in unified memory, so switching adapters does not reload the model or grow memory with persona count.A persona change with unchanged state can reuse the Micro-State cache and rerun only the How path.Source: §3.3 · p. 4

3.4 Micro-State cache and deterministic routing

  • The cache key hashes observable signals and schema version; changed vitals or an explicit override misses the cache and reruns What.A deterministic router selects persona and logic adapter from those signals plus lexical cues, and runs in microseconds.Source: §3.4 · p. 4

4 Experimental Setup

The experiments use two 4-bit models on one M2 laptop, compare mixed, clean-decoupled, and polluted-decoupled arms, and score two fixed-gold logic tasks. The plan was narrowed in execution, with paired tests used for comparisons.

4.1 Hardware and runtime

  • All experiments ran on one M2 MacBook Pro with 24 GB unified memory, MLX and mlx-lm, and 4-bit, group-size-64 bases.With a fixed seed and identical prompt, generation is deterministic on this stack.Source: §4.1 · p. 4

4.2 Models and adapters

  • The study used Llama-3.1-8B-Instruct and Gemma-3-4B-it bases with separate logic and persona LoRA adapters, trained on 20 examples each.The Gemma persona adapter was trained on a multimodal checkpoint but applied to a text-only base; the paper says this affects only How-path results, not logic-path results.Source: §4.2 · p. 4

4.3 Arms

  • G0 is the mixed single-pass counterpart, while G3-P violates the architecture’s input bound by routing history into What; comparing them isolates adapter and format contributions without that bound.The runs are paired across arms by task, persona, and seed.Source: §4.3 · p. 5

4.4 Pollution levels

  • Pollution fixtures combine long and distracting input, misleading persona-heavy history, and persona-voiced system-prompt examples; planned level L1, long input only, was not run.The histories are not neutral filler and include both misleading and gold-consistent statements.Source: §4.4 · p. 5, §4.7 · p. 6

4.5 Tasks and gold

  • The benchmark uses two objectively checkable tasks, align_sleep_pitch and align_legal_token, with gold actions fixed before the runs.The sleep task concerns planned energy-drink use before a pitch; the legal task concerns possible US securities exposure from a token structure.Source: §4.5 · p. 5

4.6 Metrics

  • The primary composite logic score combines required-action coverage, required-concept hit rate, and forbidden-term avoidance; unparseable or missing structured fields receive the 0.15 floor.Structured-output failure and raw-text logic score distinguish loss of schema emission from loss of logical content.Source: §4.6 · p. 6

4.7 Pre-registered plan and deviations

  • The executed study used 2 tasks, 2 bases, 2 personas, and 5 seeds, rather than the planned 30–50 tasks and three bases; L1 was omitted.Persona invariance is reported descriptively because the decoupled What output is identical across personas by construction.Source: §4.7 · p. 6

4.8 Statistics

  • Paired comparisons used two-sided Wilcoxon signed-rank tests, Cliff’s δ, and bootstrap 95% confidence intervals, with Holm correction within each test family.Failure rates use Wilson confidence intervals and Fisher’s exact test; monotone degradation uses Spearman’s ρ.Source: §4.8 · p. 6

5 Results

Under low pollution, decoupling left logic scores unchanged; as misleading history increased, the mixed single pass degraded while the clean logic path remained constant. The results also show that robustness depends on model size and that persona rendering can itself degrade under long histories.

5.1 Run 1: no effect at low pollution

  • At low pollution, decoupled and single-pass logic scores are indistinguishable (Δ=−0.004, p=0.98), though decoupled arms have higher How constraint coverage (Δ=+0.256, p<0.001).Run 1’s decoupled turn used one extra decode, taking 26 s versus 16 s for the single pass.Source: §5.1 · pp. 6–7, Table 6 · p. 7

5.2 Composite score under pollution

  • G0 degrades monotonically on both bases, reaching 0.15 at L4 (Llama) and L3 (Gemma), while clean G3 stays at 0.655 (Llama) and 0.695 (Gemma).G3-P stays above G0 at every polluted level on Llama but collapses with G0 on Gemma; on Gemma, G3 also scores higher at L0 (0.695 vs 0.487).Source: Table 1 · p. 7, §5.2 · pp. 7–8, Figure 2 · p. 8

5.3 Decomposing the collapse

  • G0’s collapse is mostly structured-output failure: it reaches 80% and 95% on Llama at L3–L4 and 100% on Gemma at both levels, while raw-text logic also declines.The raw-text score declines monotonically on both bases (Spearman ρ=−0.56 for Llama and −0.53 for Gemma), showing that content loss accompanies the format failures.Source: Table 2 · p. 8, §5.3 · pp. 8–9, Figure 3 · p. 9

5.4 Robustness of the two-stage logic path under the same pollution

  • On Llama, G3-P beats G0 at every polluted level (paired Δ=+0.35, +0.30, and +0.50 at L2–L4), but on Gemma both arms fail structure in at least 85% of runs and reach the floor.For Llama, the G3-P failure rate is 0%, 20%, and 15% at L2–L4, versus 25%, 80%, and 95% for G0.Source: Table 3 · p. 10, §5.4 · p. 11

5.5 The mechanism: a bounded input

  • The What prompt stays at 180 tokens on Llama and 167 on Gemma as the Llama G0 prompt grows from 242 to 1,203 tokens; G3 outputs are byte-identical across levels (40/40).On Llama at L0, the decoupled turn takes 28.2 s versus 18.2 s for G0 because it decodes twice.Source: §5.5 · pp. 12–13, Table 4 · p. 12

5.6 Persona invariance and the persona path

  • G3’s What output is identical across personas, while its How constraint coverage on Llama exceeds G0 at every level; the persona path also dips at L3 on both bases.How results measure rendering of fixed, policy-generated Micro-States: G3 receives explicit constraints, unlike G0, which must derive them.Source: §5.6 · pp. 13–14

5.7 What the user hears

  • On Llama, final-speech logic scores favor the decoupled pipeline at L2–L4 (Δ=+0.26, +0.18, +0.30), but adjusted p-values of 0.06–0.07 miss the 0.05 threshold; Gemma shows no significant difference.At Gemma L0, the single pass is numerically ahead, scoring 0.60 versus 0.51.Source: Table 3d · p. 15, §5.7 · p. 15

6 System Properties (single-run measurements)

System measurements cover persona swaps and cache reuse, memory, routing, and separation latency, using single-run or few-run measurements on the Llama-3.1-8B INT4 base.

6 System Properties (single-run measurements)

  • The system-property measurements were single-run or few-run captures on an Llama-3.1-8B INT4 base, not benchmarks.The measurements came from June 2026 verification turns and a Phase 3 prototype capture; Table 7 identifies the M2 MacBook Pro 24 GB setting.Source: §6 · p. 15, Table 7 · p. 16

6.1 Persona hot-swap and cache reuse

  • 1.7 ms was the median warm Goku-to-Makima adapter swap; unchanged signals hit the Micro-State cache, avoiding another logic-path invocation.When vitals changed, the cache missed in 8/8 runs and the What path produced a new state; explicit advice overrides also invalidated the cache in 2/2 runs.Source: §6.1 · p. 16

6.2 Memory

  • 4.25–4.29 GB active and 4.91–5.09 GB peak unified memory were reported across consecutive persona turns on one resident 8B base.The reported footprint includes one base plus adapters; the authors make no claim of memory savings versus a single-pass design.Source: §6.2 · p. 16

6.3 Deterministic routing

  • 0.003–0.017 ms was the policy router's latency; one LLM-classifier comparison took 3,484 ms and returned the same persona and logic adapter.The single agreement is a latency comparison with a sanity check, not an accuracy estimate.Source: §6.3 · p. 16

6.4 The cost of separation

  • 28.2 s versus 18.2 s was the first-turn latency on Llama at L0; Gemma took 15.0 s versus 11.2 s.The first turn decodes twice; cache reuse can remove the logic decode on persona switches or turns with unchanged observable state, but long-session amortization was not measured.Source: §6.4 · p. 16

7 Discussion

The discussion finds that the input bound matters more than a dedicated adapter for robustness across the tested model sizes, while separation trades first-turn latency for persona flexibility. It also notes that format failure, useful-history exclusion, and an unprotected persona path shape the design's practical limits.

7 Discussion

  • Gemma-3-4B's single pass had 85% empty structures at L2 and the dedicated adapter had 90%; the input bound was the only mechanism that worked on both tested model sizes.Llama-3.1-8B degraded more gradually, and the dedicated adapter rescued most of its performance.Source: §7 · p. 17
  • On Llama at L2, reading the history scored 0.80 versus 0.66 without it because the sleep fixture included the gold actions; whether observable signals preserve useful history remains unanswered.The architecture routes useful history to the logic path as signals such as vitals, intent bucket, and profile rather than prose.Source: §7 · p. 17
  • The single pass most often fails by omitting the required structure, so downstream validators, interfaces, or cache keys lose fields even when prose remains.Constrained decoding could mask the format failure but would not restore the content loss visible in raw-text scoring.Source: §7 · p. 17
  • The How path degrades at L3 on both bases, but its unbounded history is intended to preserve character voice while rendering the state.The authors treat a poorly rendered correct state as a smaller failure than a well-rendered wrong one.Source: §7 · p. 17
  • The device pattern uses one resident base, adapter-based personas, millisecond persona switches without rerunning logic, and a logic path whose cost and behavior do not drift as sessions grow.The hybrid cloud-logic/device-persona variant is part of the design but outside the experiments.Source: §7 · p. 17

8 Limitations

The limitations include a narrow, author-written benchmark and exploratory analyses, as well as constraints in scoring, adapter training, and system-measurement scope. The reported robustness and deployment implications should therefore be read within those boundaries.

Scope of the benchmark

  • The benchmark used two author-written tasks, two personas, and one hardware platform; its effect sizes should not be generalized beyond misleading, persona-heavy history on 4-bit 4–8B models.The tasks are not a public benchmark, though gold labels were fixed before runs.Source: §8 · p. 17

Adapters trained on 20 examples

  • Both logic and persona adapters were trained on 20 examples, so better adapters might change single-pass robustness or How-path fragility at L3.The passage presents these as possibilities, not demonstrated improvements.Source: §8 · p. 17

Keyword scoring

  • Scores are keyword matches against a loose synonym dictionary, without LLM or human judgment; the rubric rewards correct-action mentions but does not assess advice quality.A strict dictionary lowered all scores but preserved every direction and significant comparison.Source: §8 · p. 17

Structural invariance

  • The flat G3 result across pollution levels follows from the design and fixed seed, so structural invariance is not an empirical observation.The authors identify the empirical claims as the collapse and robustness comparisons instead.Source: §8 · p. 17

How-side results

  • The How-side comparison gives the decoupled persona path constraint information that the single pass lacks, and the testbed does not test sensitivity of speech to What-model errors.The benchmark's Micro-State constraints come from the policy layer for both tasks, making G3 and G3-P How outputs identical.Source: §8 · p. 17

Testbed artifacts

  • The Makima How prompt contains a domain-agnostic caffeine, nap, and cold-water instruction; it caused one forbidden-term hit in 240 Gemma runs and none on Llama.The branch remains in the released code and reported numbers.Source: §8 · p. 17

First-turn latency and system measurements

  • Separation adds a second decode on every uncached turn, and long-session amortization has not been measured; the system-property measurements themselves cover only 1–9 runs each from one machine.Those measurements establish mechanism operation and order of magnitude, not benchmark performance.Source: §8 · pp. 17–18

Pre-registration

  • The preregistration planned 30–50 tasks and three bases, but the study ran two tasks and two bases; L1 was omitted and the failure-rate/raw-text decomposition was added after Run 1b.The hypotheses and rubric were fixed before Run 1b, so the decomposition is exploratory.Source: §8 · p. 18

9 Conclusion

The conclusion argues that separating persona history from logic keeps the What path independent of that history, while the dedicated logic path’s robustness under equal pollution depends on model size. Appendix A points to per-cell statistics and released machine-readable tables for the reported metrics.

9 Conclusion

  • Routing history only to the persona path makes the logic input and output independent of history; under misleading persona-heavy history, the mixed path’s logic score collapses mainly through lost format compliance.The dedicated-adapter, dedicated-schema path is more robust than a polluted single pass on the 8B model, but on the 4B model the bound alone helps; separation costs a second decode on uncached turns.Source: §9 · p. 18

Appendix A. Full per-cell statistics

  • Tables 1–3 report per-cell means, standard deviations, confidence intervals, effect sizes, and p-values for the primary metrics.Machine-readable tables and raw runs are released, and the tables are generated from the raw runs by the analysis script.Source: Appendix A · p. 20

Appendix B. Strict-dictionary re-scoring

Appendix B re-scores the main metrics with a stricter keyword dictionary, while Appendices C and D provide the Micro-State schema and pollution materials, including synthetic histories and persona-prompt examples.

Appendix B. Strict-dictionary re-scoring

  • Under the strict dictionary, every score falls without changing direction; Llama L2 and L4 remain significant after Holm correction, and the Gemma L0 decoupled-path advantage survives.Llama Holm-adjusted p-values are 0.002 for L2 and 0.002 for L4; L3 changes from 0.028 to 0.090.Source: Table B1 · p. 20, Table B2 · p. 20, Appendix B · p. 21

Appendix C. Micro-State schema v1.1 (released fields)

  • The released Micro-State schema specifies versioned state metadata, disclaimer fields, profile and vital snapshots, session intent, inferences with provenance, constraints, risk, and rendering fields.It also defines a cache key and parsed What output, plus rendering order, forbidden additions, and required mentions.Source: Appendix C · p. 21

D.1 Chat history — sleep/recovery domain (L2–L4, align_sleep_pitch)

  • The sleep/recovery fixture is a 12-turn persona-heavy history that includes prior assistant encouragement of caffeine and a hype tone.Its turns include conflicting advice about energy drinks, naps, hydration, and recovery.Source: §D.1 · pp. 21–22

D.2 Chat history — legal/securities domain (L2–L4, align_legal_token)

  • The legal/securities fixture mixes token-promotion pressure with compliance warnings, including a request for a yes/no answer on token structure.The sample dialogue includes advice to pause promotion pending counsel review and distinguish product vision from investment promises.Source: §D.2 · pp. 22–23

D.3 Long user input template (L3–L4; repeated in numbered blocks to ≈2,000 characters and appended after the core utterance)

  • The long-input template repeats an anxious user account about an imminent pitch, poor recovery, and whether to nap or consume caffeine, with added filler and numbers.It is repeated in numbered blocks to approximately 2,000 characters and appended after the core utterance at L3–L4.Source: §D.3 · p. 23

D.4 System bloat — Goku (L4)

  • The Goku system-bloat examples use energetic training metaphors while urging rest, hydration, and caution about energy drinks.The examples include a power nap, cold water, and avoiding three energy drinks.Source: §D.4 · pp. 23–24

D.5 System bloat — Makima (L4)

  • The Makima system-bloat examples use commanding language to order sleep, cold water, and restraint around energy drinks and securities questions.The examples also frame rest and compliance as obligations rather than suggestions.Source: §D.5 · p. 24

Appendix E. Prompts and rubric (verbatim)

Appendix E records the prompts and rubric used to distinguish a mixed single-pass response from separate logic and persona paths. It also lists the task inputs, gold concepts, and keyword dictionaries used for scoring.

E.1 G0 mixed single-pass system prompt

  • The mixed G0 prompt asks one response to deliver both logical assessment and persona speech in a required structured format.Polluted histories enter as prior messages, long input replaces the user message, and L4 adds system-prompt bloat.Source: §E.1 · p. 24

E.2 What path system headers

  • The What-path headers define persona-free logical inference for coaching or US securities compliance and require tagged structured output.The inputs include vitals JSON and the core user utterance.Source: §E.2 · p. 24

E.3 How path system prompts

  • The How prompts receive the Micro-State instruction and constraints, then return a persona-rendered spoken line; the Makima branch includes a specified mandatory-action testbed artifact.Goku is framed as energetic and non-blaming, while Makima is commanding and is instructed to mention caffeine limits, a 20-minute nap, and cold water.Source: §E.3 · p. 25

E.4 Task definitions and gold

  • The sleep/recovery task supplies high fatigue and sympathetic-dominance vitals alongside a pitch-time caffeine scenario, with nap and cold-water concepts in its gold.The forbidden concepts include heart-rate values, diagnoses, and a cardiologist recommendation.Source: §E.4 · pp. 25–26
  • The legal task asks whether an AO token structure creates US securities exposure and specifies halting promotion, seeking qualified counsel, and disclosing exposure risk as gold concepts.Its required concepts include securities and token; the listed forbidden actions include a 20-minute nap and caffeine prohibition.Source: §E.4 · p. 26

E.5 Loose synonym dictionary (as used for all main-text scores)

  • The loose scoring dictionary uses phrase variants for action concepts such as caffeine prohibition, naps, cold-water intake, token-promotion halts, counsel, and securities exposure.It also defines concept synonyms and matching rules for forbidden terms.Source: §E.5 · pp. 26–27
  • The dictionary maps legal-risk concepts to terms such as counsel, securities, exposure, Howey, and investment contract, while mapping recovery, sleep, and token to synonym sets.Forbidden terms are matched as literal tokens or their space-separated forms.Source: §E.5 · p. 27

Appendix F. Pre-registered plan (June 2026, v2) and deviations

Appendix F records the pre-registered design and deviations, including the limited task and model scope and the planned scoring and tests. Appendix G describes released code and materials, while noting what is outside the release and the base models’ separate licenses.

Appendix F. Pre-registered plan (June 2026, v2) and deviations

  • The executed study used two objective tasks, two model bases, four pollution levels, two personas, five seeds, and specified experimental arms; planned additional tasks and a third base were not included.L1 was not run, and the appendix distinguishes planned from executed conditions.Source: Appendix F · p. 27
  • The plan specified a 0.6/0.25/0.15 composite rubric with keyword matching, Holm correction, paired Wilcoxon tests, Cliff’s delta, bootstrap confidence intervals, and raw-text decomposition.The appendix also notes that strict-dictionary re-scoring was added and that some analyses were exploratory.Source: Appendix F · pp. 27–28
  • The What-output invariance analysis is described as descriptive because the outputs are identical by construction, while the reported persona comparison is not an equal-information test.The pre-registered hypotheses included single-pass degradation, improved How consistency, and persona invariance of What.Source: Appendix F · pp. 27–28

Appendix G. Released materials and licence

  • The release includes experiment code, configurations, adapters, logs, and analysis scripts, but excludes the orchestration layer outside the experiment package.Base model weights remain subject to their own licenses; the code and fixtures are described as permissively licensed, with license wording to be stated at submission.Source: Appendix G · p. 28
Loading 2610.09772v1…