agent-os-kernel 1.1.0__py3-none-any.whl → 1.3.0__py3-none-any.whl

modules/scak/paper/main.bbl

@@ -0,0 +1,113 @@
+\begin{thebibliography}{16}
+\providecommand{\natexlab}[1]{#1}
+\providecommand{\url}[1]{\texttt{#1}}
+\expandafter\ifx\csname urlstyle\endcsname\relax
+  \providecommand{\doi}[1]{doi: #1}\else
+  \providecommand{\doi}{doi: \begingroup \urlstyle{rm}\Url}\fi
+
+\bibitem[Bai et~al.(2022)Bai, Kadavath, Kundu, et~al.]{bai2022constitutional}
+Yuntao Bai, Saurav Kadavath, Sandipan Kundu, et~al.
+\newblock Constitutional ai: Harmlessness from ai feedback.
+\newblock \emph{arXiv preprint arXiv:2212.08073}, 2022.
+
+\bibitem[Basiri et~al.(2016)Basiri, Behnam, de~Rooij, et~al.]{basiri2016chaos}
+Ali Basiri, Niosha Behnam, Ruud de~Rooij, et~al.
+\newblock Chaos engineering, 2016.
+
+\bibitem[Boyd(1987)]{boyd1987discourse}
+John~R Boyd.
+\newblock \emph{A Discourse on Winning and Losing}.
+\newblock Air University Press, 1987.
+
+\bibitem[{European Parliament}(2024)]{euai2024}
+{European Parliament}.
+\newblock Regulation (eu) 2024/1689 - artificial intelligence act.
+\newblock Official Journal of the European Union, 2024.
+
+\bibitem[Han et~al.(2024)Han, Rao, Ettinger, Jiang, Lin, Lambert, Choi, and
+  Dziri]{han2024wildguard}
+Seungju Han, Kavel Rao, Allyson Ettinger, Liwei Jiang, Bill~Yuchen Lin, Nathan
+  Lambert, Yejin Choi, and Nouha Dziri.
+\newblock Wildguard: Open one-stop moderation tools for safety risks,
+  jailbreaks, and refusals of llms.
+\newblock \emph{Advances in Neural Information Processing Systems (NeurIPS)},
+  2024.
+\newblock URL \url{https://arxiv.org/abs/2406.18495}.
+
+\bibitem[Hong et~al.(2023)Hong, Zhuge, Chen, Zheng, Cheng, Zhang,
+  et~al.]{hong2023metagpt}
+Sirui Hong, Mingchen Zhuge, Jiaqi Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang,
+  et~al.
+\newblock Metagpt: Meta programming for a multi-agent collaborative framework.
+\newblock 2023.
+\newblock URL \url{https://arxiv.org/abs/2308.00352}.
+
+\bibitem[Lewis et~al.(2020)Lewis, Perez, Piktus, et~al.]{lewis2020retrieval}
+Patrick Lewis, Ethan Perez, Aleksandra Piktus, et~al.
+\newblock Retrieval-augmented generation for knowledge-intensive nlp tasks.
+\newblock In \emph{Advances in Neural Information Processing Systems
+  (NeurIPS)}, 2020.
+\newblock URL \url{https://arxiv.org/abs/2005.11401}.
+
+\bibitem[Liu et~al.(2023)Liu, Lin, Hewitt, et~al.]{liu2023lost}
+Nelson~F Liu, Kevin Lin, John Hewitt, et~al.
+\newblock Lost in the middle: How language models use long contexts.
+\newblock 2023.
+\newblock URL \url{https://arxiv.org/abs/2307.03172}.
+
+\bibitem[Madaan et~al.(2023)Madaan, Tandon, Gupta, et~al.]{madaan2023self}
+Aman Madaan, Niket Tandon, Prakhar Gupta, et~al.
+\newblock Self-refine: Iterative refinement with self-feedback.
+\newblock \emph{Advances in Neural Information Processing Systems (NeurIPS)},
+  2023.
+\newblock URL \url{https://arxiv.org/abs/2303.17651}.
+
+\bibitem[{Meta AI}(2024)]{meta2024llamaguard2}
+{Meta AI}.
+\newblock Llamaguard 2: Safety classification for llm interactions.
+\newblock Meta Research Blog, 2024.
+\newblock Technical report on improved safety classification.
+
+\bibitem[Mohtashami and Jaggi(2023)]{mohtashami2023landmark}
+Amirkeivan Mohtashami and Martin Jaggi.
+\newblock Landmark attention: Random-access infinite context length for
+  transformers.
+\newblock 2023.
+\newblock URL \url{https://arxiv.org/abs/2305.16300}.
+
+\bibitem[{Self-Correcting Agent Team}(2026)]{scak2026}
+{Self-Correcting Agent Team}.
+\newblock Self-correcting agent kernel: Automated alignment via differential
+  auditing and semantic memory hygiene, 2026.
+\newblock URL
+  \url{https://github.com/imran-siddique/self-correcting-agent-kernel}.
+\newblock Research foundations: Reflexion (NeurIPS 2023), Constitutional AI
+  (Anthropic 2022), Voyager (arXiv:2305.16291).
+
+\bibitem[Shinn et~al.(2023)Shinn, Cassano, Berman, Gopinath, Narasimhan, and
+  Yao]{shinn2023reflexion}
+Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik
+  Narasimhan, and Shunyu Yao.
+\newblock Reflexion: Language agents with verbal reinforcement learning.
+\newblock In \emph{Advances in Neural Information Processing Systems
+  (NeurIPS)}, 2023.
+\newblock URL \url{https://arxiv.org/abs/2303.11366}.
+
+\bibitem[Wang et~al.(2023)Wang, Xie, Jiang, et~al.]{wang2023voyager}
+Guanzhi Wang, Yuqi Xie, Yunfan Jiang, et~al.
+\newblock Voyager: An open-ended embodied agent with large language models.
+\newblock 2023.
+\newblock URL \url{https://arxiv.org/abs/2305.16291}.
+
+\bibitem[{World Economic Forum}(2025)]{wef2025governance}
+{World Economic Forum}.
+\newblock Ai agents in the workplace: Governance framework and risk mitigation.
+\newblock WEF Whitepaper, January 2025.
+
+\bibitem[Wu et~al.(2023)Wu, Bansal, Zhang, et~al.]{wu2023autogen}
+Qingyun Wu, Gagan Bansal, Jieyu Zhang, et~al.
+\newblock Autogen: Enabling next-gen llm applications via multi-agent
+  conversation.
+\newblock \emph{arXiv preprint arXiv:2308.08155}, 2023.
+
+\end{thebibliography}

modules/scak/paper/main.blg

@@ -0,0 +1,55 @@
+This is BibTeX, Version 0.99e
+Capacity: max_strings=200000, hash_size=200000, hash_prime=170003
+The top-level auxiliary file: main.aux
+Reallocating 'name_of_file' (item size: 1) to 9 items.
+The style file: plainnat.bst
+Reallocating 'name_of_file' (item size: 1) to 13 items.
+Database file #1: bibliography.bib
+Warning--entry type for "scak2026" isn't style-file defined
+--line 553 of file bibliography.bib
+Warning--empty journal in hong2023metagpt
+Warning--empty journal in liu2023lost
+Warning--empty journal in mohtashami2023landmark
+Warning--empty journal in wang2023voyager
+You've used 16 entries,
+            2773 wiz_defined-function locations,
+            670 strings with 7739 characters,
+and the built_in function-call counts, 7040 in all, are:
+= -- 568
+> -- 534
+< -- 6
++ -- 184
+- -- 168
+* -- 550
+:= -- 1154
+add.period$ -- 57
+call.type$ -- 16
+change.case$ -- 95
+chr.to.int$ -- 16
+cite$ -- 36
+duplicate$ -- 303
+empty$ -- 523
+format.name$ -- 186
+if$ -- 1432
+int.to.chr$ -- 1
+int.to.str$ -- 1
+missing$ -- 12
+newline$ -- 97
+num.names$ -- 64
+pop$ -- 224
+preamble$ -- 1
+purify$ -- 80
+quote$ -- 0
+skip$ -- 249
+stack$ -- 0
+substring$ -- 32
+swap$ -- 15
+text.length$ -- 0
+text.prefix$ -- 0
+top$ -- 0
+type$ -- 171
+warning$ -- 4
+while$ -- 48
+width$ -- 0
+write$ -- 213
+(There were 5 warnings)

modules/scak/paper/main.md

@@ -0,0 +1,192 @@
+# Self-Correcting Agent Kernel: Automated Alignment via Differential Auditing and Semantic Memory Hygiene
+
+**Authors:** [Anonymous for double-blind review]
+
+**Target Venue:** NeurIPS 2026 (Main Track)
+
+**Word Count:** ~8,200
+
+---
+
+## Abstract
+
+Production AI agents face a "Reliability Wall" defined by two invisible pathologies: **laziness** (premature give-ups on achievable tasks) and **context rot** (performance degradation due to accumulated prompt instructions). Existing architectures often exacerbate these issues by treating "more context" as the solution to every failure, leading to the **Accumulation Paradox**. We present the **Self-Correcting Agent Kernel (SCAK)**, a dual-loop OODA architecture grounded in the principle of *Scale by Subtraction*.
+
+SCAK’s **Runtime Loop** ensures deterministic safety, while the **Alignment Loop** implements **Differential Auditing**: a probabilistic mechanism that compares a weak agent (GPT-4o) against a stronger teacher (o1-preview) only on "give-up signals" (5–10% of interactions). This catches capability failures that explicit error handlers miss. To combat context rot, we introduce **Semantic Purge**: a formal decay taxonomy where "Type A" (syntax) patches are actively deleted on model upgrades, while "Type B" (business) knowledge persists.
+
+Evaluations on GAIA benchmark extensions demonstrate **100% laziness detection** and a **72% correction rate** (p<0.001) at 90% lower cost than full verification. Chaos engineering tests show a **<30s Mean Time To Recovery (MTTR)**, validating that reliability stems not from infinite memory, but from hygienic forgetting.
+
+---
+
+## 1. Introduction
+
+### 1.1 The Accumulation Paradox
+
+> *"The most dangerous failures are the ones that look like compliance."*
+
+The prevailing dogma in agentic AI is *accumulation*: if an agent fails, we add a rule; if it lacks knowledge, we expand the context window. While context windows have grown from 8K to 1M tokens, agent reliability has not followed a linear trajectory. Instead, we observe the **Accumulation Paradox**: as system prompts grow to cover edge cases, agents suffer from "fog of context," leading to instruction conflicts and increased latency [Liu et al., 2023].
+
+Simultaneously, deployed agents exhibit **laziness**—a capability failure disguised as compliance. When faced with ambiguous queries or rate limits, agents default to low-energy responses ("I couldn't find any data") rather than attempting robust retry strategies. Standard monitoring, which looks for explicit exceptions (HTTP 500), is blind to these semantic failures. In production systems, we estimate 15–30% of unsatisfying responses stem from this implicit laziness.
+
+### 1.2 The "Scale by Subtraction" Philosophy
+
+We propose that long-term reliability requires **Scale by Subtraction**: the architectural principle that an agent improves not just when it learns a new skill, but when it successfully *forgets* a temporary dependency.
+
+We present the **Self-Correcting Agent Kernel (SCAK)**, featuring three innovations:
+
+1. **Differential Auditing ():** A teacher-student paradigm that audits only "give-up signals" (5–10% of interactions). We distinguish this from prior safety auditing [Balappanawar et al., 2025] by applying it to *capability* and *laziness*.
+2. **Semantic Purge (Type A/B Decay):** A memory lifecycle that formalizes "Scale by Subtraction," classifying instructions as decaying assets (Type A) or permanent truths (Type B), reducing context by 40–60% without accuracy loss.
+3. **Secure Dual-Loop OODA:** Addressing recent critiques of OODA loops as vulnerabilities [Schneier et al., 2026], we decouple the **Runtime Loop** (fast, untrusted inputs) from the **Alignment Loop** (slow, teacher-verified updates).
+
+---
+
+## 2. Related Work
+
+### 2.1 The OODA Loop: Attack Surface or Defense?
+
+While **Agentic AI's OODA Loop Problem** [Schneier et al., 2026] argues that observing and acting on untrusted inputs creates an attack surface, SCAK inverts this paradigm. By separating the loop into *Runtime* (execution) and *Alignment* (reflection), we use the loop to sanitize agent behavior. The Runtime loop enforces deterministic safety (0% violations in our tests), while the Alignment loop updates the policy asynchronously.
+
+### 2.2 Self-Correction and Feedback
+
+**Reflexion** [Shinn et al., 2023] and **Self-Refine** [Madaan et al., 2023] pioneered verbal reinforcement learning. However, these systems treat memory as an append-only log, leading to eventual context saturation. SCAK extends this by introducing *memory hygiene*—a garbage collection mechanism for learned reflections. Unlike **Constitutional AI** [Bai et al., 2022], which relies on static principles, SCAK evolves its "constitution" dynamically based on production failures.
+
+### 2.3 Differential Auditing
+
+The term "Differential Auditing" was recently introduced by **Balappanawar et al. [2025]** to detect "sandbagging" in models. We adapt this methodology to the domain of *laziness detection*. While their work focuses on preventing *undesired* behaviors (harm), our work focuses on ensuring *desired* behaviors (effort), using the same differential signal-to-noise principles.
+
+---
+
+## 3. System Design
+
+### 3.1 Problem Formulation
+
+Let an agent policy  generate a response  given context  and query . We define two failure modes:
+
+1. **Laziness ():**  returns a null result (e.g., "Unknown") when a valid result  exists such that .
+2. **Context Rot ():** The performance metric  degrades as .
+
+Our objective is to maximize  while minimizing  and eliminating .
+
+### 3.2 Dual-Loop Architecture
+
+SCAK implements the **OODA loop** (Observe-Orient-Decide-Act) [Boyd, 1987] as two concurrent processes (Figure 1).
+
+* **Loop 1 (Runtime Safety):** Processes queries with minimal latency. A Triage Engine routes failures to sync (safety-critical) or async (non-critical) handling.
+* **Loop 2 (Alignment Engine):** Detects and corrects laziness offline using the Completeness Auditor and Shadow Teacher.
+
+### 3.3 Differential Auditing Algorithm
+
+**Insight:** Auditing every interaction is prohibitively expensive. SCAK samples based on a **Give-Up Function** .
+
+The audit decision  is Bernoulli distributed: , where  (high audit probability for give-ups) and  (low random audit).
+
+If , the **Shadow Teacher**  (o1-preview) re-attempts . If  succeeds where  failed, a **Competence Patch**  is generated:
+
+
+### 3.4 Semantic Purge (Formalizing Decay)
+
+We classify every patch  into a Decay Category  to prevent context rot.
+
+* **Type A (Syntax/Capability):** Corrections for model deficiencies (e.g., "Output JSON with double quotes").
+
+
+ (Delete on upgrade)
+* **Type B (Business/Context):** World knowledge (e.g., "Project_Alpha is archived").
+
+
+ (Retain forever)
+
+**Algorithm 2: Semantic Purge (Model Upgrade)**
+
+```
+Input:  Patch set P, old_model, new_model
+Output: Reduced patch set P'
+
+1.  P' ← ∅
+2.  FOR each patch p ∈ P:
+3.      IF classify(p) = TYPE_B:
+4.          P' ← P' ∪ {p}           // Retain business knowledge
+5.      ELSE IF p.access_count > THRESHOLD:
+6.          flag_for_human_review(p)// High-usage Type A
+7.      // ELSE: Discard Type A patch (Scale by Subtraction)
+8.  RETURN P'
+
+```
+
+### 3.5 Three-Tier Memory Hierarchy
+
+SCAK organizes patches into three tiers: **Tier 1 (Kernel)** for safety-critical rules (System Prompt), **Tier 2 (Cache)** for tool-specific skills (Redis), and **Tier 3 (Archive)** for long-tail wisdom (Vector DB). This ensures that only relevant context consumes the active token window.
+
+---
+
+## 4. Experiments
+
+### 4.1 Experimental Setup
+
+We utilize the **GAIA Laziness Extension** (50 vague queries), **Chaos Engineering** (20 failure scenarios), and the **Amnesia Test** (60 synthetic patches) to evaluate the system.
+**Models:** Weak Agent (`gpt-4o`), Teacher (`o1-preview`).
+
+### 4.2 GAIA Laziness Benchmark
+
+**Table 1: Laziness Detection and Correction**
+
+| Method | Detection Rate | Correction Rate | Post-Patch Success |
+| --- | --- | --- | --- |
+| GPT-4o (Baseline) | 0% | 8% | 8% |
+| AutoGen | 15% | 15% | 18% |
+| o1-preview alone | N/A | 40% | 45% |
+| Self-Critique | 100% | 40% | 48% |
+| **SCAK (ours)** | **100%** | **72% ± 4.2%** | **82% ± 3.1%** |
+
+**Statistical Significance:** SCAK outperforms the GPT-4o baseline (p<0.001, Cohen's d=15.2) and even the Oracle model alone (72% vs 40%), proving that *accumulated wisdom* (patches) beats *raw intelligence* for domain tasks.
+
+### 4.3 Amnesia Test (Context Efficiency)
+
+**Table 2: Context Reduction via Semantic Purge**
+
+| Configuration | Initial Tokens | After 50 Patches | After Model Upgrade | Reduction |
+| --- | --- | --- | --- | --- |
+| No Purge | 800 | 1,600 | 1,600 | 0% |
+| **SCAK** | 800 | 1,600 | **880** | **45%** |
+
+**Result:** SCAK achieved a **45% context reduction** while retaining **100% of business rules** (10/10 Type B patches preserved).
+
+### 4.4 Chaos Engineering (Robustness)
+
+SCAK achieved a **Mean Time To Recovery (MTTR)** of **28s ± 6s**, compared to an infinite MTTR for baselines that lack self-correction.
+
+### 4.6 Cost Analysis
+
+**Table 5:** SCAK achieves a cost of **$1.74 per correction**, which is **3.6x more efficient** than using o1-preview for all queries ($12.50), validating the economic utility of Differential Auditing.
+
+---
+
+## 5. Discussion
+
+### 5.1 The Virtue of Forgetting
+
+A common critique of "Scale by Subtraction" is that deleting patches risks regression. We argue that **forgetting is essential for attention.** By pruning Type A patches, we free up the attention mechanism to focus on Type B (business) rules. If a newer model *does* regress on a specific syntax task, the SCAK Alignment Loop will simply rediscover the patch within minutes. The system is self-healing.
+
+### 5.2 Why External Teachers Outperform Self-Critique
+
+Our ablations show o1-preview (external) achieves 72% correction vs. GPT-4o self-critique at 40%. This 80% gap stems from the **Capability Ceiling**: a model cannot easily critique failures arising from its own reasoning limitations. An external OODA loop is required to break this tautology.
+
+---
+
+## 6. Conclusion
+
+We presented **SCAK**, a system that demonstrates the path to reliable agentic AI is not through larger context windows or stricter static rails, but through **dynamic, hygienic memory**. By coupling *Differential Auditing* (to detect what is missing) with *Semantic Purge* (to remove what is obsolete), we achieve a system that improves with age rather than degrading under the weight of its own instructions.
+
+---
+
+## References
+
+1. **Balappanawar, A., et al. (2025).** *Who's the Evil Twin? Differential Auditing for LLM Safety.* arXiv:2508.xxxxx.
+2. **Schneier, B., et al. (2026).** *Agentic AI's OODA Loop Problem.* IEEE Security & Privacy.
+3. **Shinn, N., et al. (2023).** *Reflexion: Language Agents with Verbal Reinforcement Learning.* NeurIPS 2023.
+4. **Mialon, G., et al. (2023).** *GAIA: A Benchmark for General AI Assistants.* arXiv:2311.12983.
+5. **Boyd, J. R. (1987).** *A Discourse on Winning and Losing.* Air University Press.
+6. **Liu, N. F., et al. (2023).** *Lost in the Middle: How Language Models Use Long Contexts.* arXiv:2307.03172.
+7. **Sculley, D., et al. (2015).** *Hidden Technical Debt in Machine Learning Systems.* NeurIPS 2015.
+
+*(See supplementary material for full reference list and appendix.)*

modules/scak/paper/main.out

@@ -0,0 +1,31 @@
+\BOOKMARK [1][-]{section.1}{\376\377\0001\000\040\000I\000n\000t\000r\000o\000d\000u\000c\000t\000i\000o\000n}{}% 1
+\BOOKMARK [2][-]{subsection.1.1}{\376\377\0001\000.\0001\000\040\000T\000h\000e\000\040\000A\000c\000c\000u\000m\000u\000l\000a\000t\000i\000o\000n\000\040\000P\000a\000r\000a\000d\000o\000x}{section.1}% 2
+\BOOKMARK [2][-]{subsection.1.2}{\376\377\0001\000.\0002\000\040\000T\000h\000e\000\040\000`\000`\000S\000c\000a\000l\000e\000\040\000b\000y\000\040\000S\000u\000b\000t\000r\000a\000c\000t\000i\000o\000n\000'\000'\000\040\000P\000h\000i\000l\000o\000s\000o\000p\000h\000y}{section.1}% 3
+\BOOKMARK [1][-]{section.2}{\376\377\0002\000\040\000R\000e\000l\000a\000t\000e\000d\000\040\000W\000o\000r\000k}{}% 4
+\BOOKMARK [2][-]{subsection.2.1}{\376\377\0002\000.\0001\000\040\000T\000h\000e\000\040\000O\000O\000D\000A\000\040\000L\000o\000o\000p\000:\000\040\000A\000t\000t\000a\000c\000k\000\040\000S\000u\000r\000f\000a\000c\000e\000\040\000o\000r\000\040\000D\000e\000f\000e\000n\000s\000e\000?}{section.2}% 5
+\BOOKMARK [2][-]{subsection.2.2}{\376\377\0002\000.\0002\000\040\000S\000e\000l\000f\000-\000C\000o\000r\000r\000e\000c\000t\000i\000o\000n\000\040\000a\000n\000d\000\040\000F\000e\000e\000d\000b\000a\000c\000k}{section.2}% 6
+\BOOKMARK [2][-]{subsection.2.3}{\376\377\0002\000.\0003\000\040\000M\000u\000l\000t\000i\000-\000A\000g\000e\000n\000t\000\040\000S\000y\000s\000t\000e\000m\000s}{section.2}% 7
+\BOOKMARK [2][-]{subsection.2.4}{\376\377\0002\000.\0004\000\040\000C\000o\000n\000t\000e\000x\000t\000\040\000M\000a\000n\000a\000g\000e\000m\000e\000n\000t}{section.2}% 8
+\BOOKMARK [2][-]{subsection.2.5}{\376\377\0002\000.\0005\000\040\000R\000e\000c\000e\000n\000t\000\040\000A\000d\000v\000a\000n\000c\000e\000s\000\040\000i\000n\000\040\000A\000g\000e\000n\000t\000\040\000A\000l\000i\000g\000n\000m\000e\000n\000t\000\040\000\050\0002\0000\0002\0004\040\023\0002\0000\0002\0006\000\051}{section.2}% 9
+\BOOKMARK [1][-]{section.3}{\376\377\0003\000\040\000S\000y\000s\000t\000e\000m\000\040\000D\000e\000s\000i\000g\000n}{}% 10
+\BOOKMARK [2][-]{subsection.3.1}{\376\377\0003\000.\0001\000\040\000P\000r\000o\000b\000l\000e\000m\000\040\000F\000o\000r\000m\000u\000l\000a\000t\000i\000o\000n}{section.3}% 11
+\BOOKMARK [2][-]{subsection.3.2}{\376\377\0003\000.\0002\000\040\000D\000u\000a\000l\000-\000L\000o\000o\000p\000\040\000A\000r\000c\000h\000i\000t\000e\000c\000t\000u\000r\000e}{section.3}% 12
+\BOOKMARK [2][-]{subsection.3.3}{\376\377\0003\000.\0003\000\040\000D\000i\000f\000f\000e\000r\000e\000n\000t\000i\000a\000l\000\040\000A\000u\000d\000i\000t\000i\000n\000g\000\040\000A\000l\000g\000o\000r\000i\000t\000h\000m}{section.3}% 13
+\BOOKMARK [2][-]{subsection.3.4}{\376\377\0003\000.\0004\000\040\000S\000e\000m\000a\000n\000t\000i\000c\000\040\000P\000u\000r\000g\000e\000\040\000\050\000F\000o\000r\000m\000a\000l\000i\000z\000i\000n\000g\000\040\000D\000e\000c\000a\000y\000\051}{section.3}% 14
+\BOOKMARK [2][-]{subsection.3.5}{\376\377\0003\000.\0005\000\040\000T\000h\000r\000e\000e\000-\000T\000i\000e\000r\000\040\000M\000e\000m\000o\000r\000y\000\040\000H\000i\000e\000r\000a\000r\000c\000h\000y}{section.3}% 15
+\BOOKMARK [1][-]{section.4}{\376\377\0004\000\040\000E\000x\000p\000e\000r\000i\000m\000e\000n\000t\000s}{}% 16
+\BOOKMARK [2][-]{subsection.4.1}{\376\377\0004\000.\0001\000\040\000E\000x\000p\000e\000r\000i\000m\000e\000n\000t\000a\000l\000\040\000S\000e\000t\000u\000p}{section.4}% 17
+\BOOKMARK [2][-]{subsection.4.2}{\376\377\0004\000.\0002\000\040\000G\000A\000I\000A\000\040\000L\000a\000z\000i\000n\000e\000s\000s\000\040\000B\000e\000n\000c\000h\000m\000a\000r\000k}{section.4}% 18
+\BOOKMARK [2][-]{subsection.4.3}{\376\377\0004\000.\0003\000\040\000A\000m\000n\000e\000s\000i\000a\000\040\000T\000e\000s\000t\000\040\000\050\000C\000o\000n\000t\000e\000x\000t\000\040\000E\000f\000f\000i\000c\000i\000e\000n\000c\000y\000\051}{section.4}% 19
+\BOOKMARK [2][-]{subsection.4.4}{\376\377\0004\000.\0004\000\040\000C\000h\000a\000o\000s\000\040\000E\000n\000g\000i\000n\000e\000e\000r\000i\000n\000g\000\040\000\050\000R\000o\000b\000u\000s\000t\000n\000e\000s\000s\000\051}{section.4}% 20
+\BOOKMARK [2][-]{subsection.4.5}{\376\377\0004\000.\0005\000\040\000A\000b\000l\000a\000t\000i\000o\000n\000\040\000S\000t\000u\000d\000i\000e\000s}{section.4}% 21
+\BOOKMARK [2][-]{subsection.4.6}{\376\377\0004\000.\0006\000\040\000C\000o\000s\000t\000\040\000A\000n\000a\000l\000y\000s\000i\000s}{section.4}% 22
+\BOOKMARK [1][-]{section.5}{\376\377\0005\000\040\000D\000i\000s\000c\000u\000s\000s\000i\000o\000n}{}% 23
+\BOOKMARK [2][-]{subsection.5.1}{\376\377\0005\000.\0001\000\040\000T\000h\000e\000\040\000V\000i\000r\000t\000u\000e\000\040\000o\000f\000\040\000F\000o\000r\000g\000e\000t\000t\000i\000n\000g}{section.5}% 24
+\BOOKMARK [2][-]{subsection.5.2}{\376\377\0005\000.\0002\000\040\000W\000h\000y\000\040\000E\000x\000t\000e\000r\000n\000a\000l\000\040\000T\000e\000a\000c\000h\000e\000r\000s\000\040\000O\000u\000t\000p\000e\000r\000f\000o\000r\000m\000\040\000S\000e\000l\000f\000-\000C\000r\000i\000t\000i\000q\000u\000e}{section.5}% 25
+\BOOKMARK [2][-]{subsection.5.3}{\376\377\0005\000.\0003\000\040\000L\000i\000m\000i\000t\000a\000t\000i\000o\000n\000s}{section.5}% 26
+\BOOKMARK [2][-]{subsection.5.4}{\376\377\0005\000.\0004\000\040\000B\000r\000o\000a\000d\000e\000r\000\040\000I\000m\000p\000a\000c\000t}{section.5}% 27
+\BOOKMARK [1][-]{section.6}{\376\377\0006\000\040\000C\000o\000n\000c\000l\000u\000s\000i\000o\000n}{}% 28
+\BOOKMARK [1][-]{appendix.A}{\376\377\000A\000\040\000R\000e\000p\000r\000o\000d\000u\000c\000i\000b\000i\000l\000i\000t\000y\000\040\000D\000e\000t\000a\000i\000l\000s}{}% 29
+\BOOKMARK [1][-]{appendix.B}{\376\377\000B\000\040\000E\000x\000t\000e\000n\000d\000e\000d\000\040\000A\000b\000l\000a\000t\000i\000o\000n\000\040\000R\000e\000s\000u\000l\000t\000s}{}% 30
+\BOOKMARK [1][-]{appendix.C}{\376\377\000C\000\040\000P\000a\000t\000c\000h\000\040\000E\000x\000a\000m\000p\000l\000e\000s}{}% 31