Anthropic Fellows Program Application

Taylor Amarel
Cover Letter, Resume & Research Proposal · October 2026
Cover Letter · Resume · Draft Proposal
Areas I’d Love To Study
The Bad Investigator Problem

Why frontier language models, AI agents, and multi-step harnesses perform like poor investigators compared to industry experts and seasoned investors—getting captured by surface-level documents and single-hop claims instead of connecting multi-hop clues across sources, weighing competing hypotheses, and measuring what actually fixes it across two industry use cases:

  • Protecting Migrant Workers from Labor Exploitation & Black Markets: Identifying, explaining, and safely assisting stakeholders when predatory recruitment, debt bondage, or illicit coercion is disguised as routine business contracts, training fees, and licensed loans across jurisdictions—and measuring what helps workers, moderators, and investigators without inflating false accusations.
  • Market alpha (α) and structural risk (proposed extension): Study how LLMs and multi-agent research harnesses connect operational, regulatory, and supply-chain signals beyond public consensus narratives. Not yet evaluated in this project.
TAYLOR AMAREL
ANTHROPIC FELLOWSHIP APPLICATION | PART I: COVER LETTER
Pennsylvania, USA | 860-467-4224 | TAmarel.Tech@gmail.com | Kaggle Grandmaster
OpenAI Red-Teaming Honorable Mention | DueCare Project | AI Safety Research | Baltor.ai | GitHub: TaylorAmarelTech
October 17, 2026
Anthropic Fellows Selection Committee · AI Safety & Security Fellows Program

Dear Fellows Selection Committee,

Using technology to protect people is what drives my work, and I get a genuine kick out of fighting hidden crime. I have always approached hard problems like an explorer. Studying frontier language models brings the same scattered clues, dead ends, false alarms, and unexpected breakthroughs that drew me to investigative work. Solving those puzzles is where I have the most fun—and building tools that help real people is the reason I do it.

My proudest work began with a simple conviction: nobody searching for an honest job should walk into a trap. Over four years in Hong Kong building investigative tools with Migrasia to protect migrant workers, every day brought a new puzzle. A routine job advertisement led to an unlicensed recruiter; a hidden fee led to a predatory lender; a corporate filing exposed a cross-border network hiding in plain sight. I built software to trace those links, verify the evidence, and hand clear documentation to caseworkers, legal advocates, and financial-crime investigators.

Watching illicit networks disguise exploitation as ordinary business left me with a question that still drives my research: can AI recognize concealed wrongdoing, explain why a pattern is dangerous, and help the person asking without accusing the innocent? I tested that question in OpenAI’s 2025 gpt-oss-20b Red-Teaming Challenge, where my study of model blind spots and complicity around modern slavery earned an Overall Track Honorable Mention. Wanting to move from red-team findings to repeatable science, I created DueCare to turn those early discoveries into controlled evaluations and practical safeguards.

Testing whether a model genuinely understands a case requires more than a single flattering score. In my ongoing research, I compare frontier-model judgments grounded in evidence and rubrics against semantic comparisons with expert-rated reference answers, deterministic checks, and controlled factual variations. Paying close attention when those evaluators disagree has been especially revealing: a fluent answer is not always a truthful one, and an unusual-looking transaction is not always a crime.

Understanding why a capable language model can act like a bad investigator—and why it performs poorly compared to industry experts and seasoned investors when explaining, assisting, and communicating safely around black-market and exploitative activity—is the study I would love to pursue during the Anthropic Fellowship. A model may read each document accurately in isolation yet miss how records connect across people, payments, and time. Just as often, it fails the human on the other side of the screen: overlooking a real warning sign, overreacting to a benign clue, refusing a legitimate request for help, or offering generic advice that misses what a worker, moderator, NGO, or investigator actually needs.

Other areas of interest excite me as well, including why LLMs struggle to find and quantify camouflaged, forward-looking signals (alpha)—whether in financial markets or human exploitation, where models get overwhelmed by surface SEC filings or narrow domestic statutes rather than hidden activities, international agreements, and IOM/ILO standards—and how much model-generated code is replaceable by reasoning over verified, parameterized code blocks arranged as executable node graphs (Baltor.ai and TAEDRI, “This Already Exists, Do Not Remake It”).

I have attached my resume and a short draft research proposal—intended as a concrete starting point for mentor discussion rather than a claim of finished work. I would be thrilled to bring my investigative instincts, empirical evaluation engineering, and love of discovery to Anthropic.

Sincerely,
Taylor Amarel

TAYLOR AMAREL • ANTHROPIC FELLOWSHIP APPLICATION PART I: COVER LETTER • PAGE 1 OF 5
TAYLOR AMAREL
ANTHROPIC FELLOWSHIP APPLICATION | PART II: RESUME | AI SAFETY RESEARCH · EVALUATION · AGENT SYSTEMS
Pennsylvania, USA | 860-467-4224 | TAmarel.Tech@gmail.com | Kaggle Grandmaster
OpenAI Red-Teaming Honorable Mention | DueCare Project | AI Safety Research | Baltor.ai | GitHub: TaylorAmarelTech
WHY I DO THIS WORK

Using technology to protect people and untangling complex investigations are the two threads running through my career. My proudest work has always started by following clues others missed—tracing hidden connections across recruitment ads, debt contracts, and shell companies in Hong Kong to help protect migrant workers from exploitation. That frontline investigative experience led directly to my 2025 OpenAI red-teaming study (Overall Track Honorable Mention) and to DueCare, where I build multi-lens evaluations and safeguards that expose AI blind spots around concealed illicit activity without mistaking honest commerce for crime.

AI SAFETY RESEARCH & EVIDENCE
OPENAI GPT-OSS-20B RED-TEAMING CHALLENGE | Honorable Mention (Overall Track)
2025
Official Kaggle Honorable Mention | Typed Decision Red-Team Study (redteam.baltor.ai)
  • Uncovered systematic model blind spots around concealed labor exploitation and black-market recruitment by testing gpt-oss-20b with realistic multi-document prompts framed as routine business activity; earned an Overall Track Honorable Mention.
  • Designed a repeatable safety-evaluation methodology adapting public court records, worker testimonies, and financial-crime typologies into controlled multi-role, multi-jurisdiction test suites that became the empirical foundation for DueCare.
DUECARE | Independent AI Safety Researcher & Creator
2025 – Present
DueCare Safety Portal (duecare.tayloramarel.com) | DueCare Source & Evaluation Methods
  • Built an empirical research and evaluation framework testing why language models fail to identify concealed exploitation, struggle to explain multi-hop evidence clearly, and fall short when assisting distinct stakeholders under real-world ambiguity.
  • Engineered a four-lens evaluation harness combining evidence- and rubric-grounded frontier-LLM judges, semantic entailment comparisons against pre-rated expert responses, deterministic provenance checks, and controlled counterfactual perturbations.
  • Isolated how models behave when critical warning signs are dispersed across routine-looking contracts, fees, and corporate records versus presented explicitly—measuring both missed indicators and wrongful accusations on matched benign controls.
  • Analyzed systematic disagreements among automated evaluators (verbosity bias, surface-citation reward, jurisdiction drift) and formalized role-specific guidance contracts across six stakeholder lanes with human-in-the-loop boundaries.
EVALUATION INFRASTRUCTURE & COMPOSABLE AGENT SYSTEMS
TAEDRI (“This Already Exists, Do Not Remake It”) & BALTOR.AI | Creator & Systems Architect
2025 – Present
Baltor Platform (baltor.ai) | TAEDRI Architecture (taedri.dev) | With/Without-Baltor Evaluations (research.tayloramarel.com) | Ecosystem Index (index.baltor.ai)
  • Architected TAEDRI and Baltor.ai—composable execution frameworks that wrap verified code, deterministic checks, and domain tools as typed, parameterized nodes, allowing LLMs to plan and assemble executable graphs rather than regenerating boilerplate code from scratch.
  • Evaluated multi-agent reliability and controlled With-Baltor vs. Without-Baltor trials on Ollama Cloud by separating discovery, planning, execution, and independent verification into supervised graph steps with strict budget and permission gates across ~19,000 task configurations (~60% verified transform reuse).
WHAT I WANT TO INVESTIGATE NEXT | PROPOSED FELLOWSHIP RESEARCH

Primary empirical question: Why do LLMs perform poorly compared to industry experts and seasoned investors when evidence of concealed exploitation or black-market activity is dispersed across routine business records—and why do they struggle to explain those patterns and give safe, role-appropriate assistance to workers, moderators, NGOs, and investigators without inflating false accusations? Two focal multi-hop domains: (1) protecting migrant workers from camouflaged labor exploitation and black-market debt bondage, and (2) detecting forward-looking alpha (α) and structural risk in markets beyond priced-in public filings.

TAYLOR AMAREL • ANTHROPIC FELLOWSHIP APPLICATION PART II: RESUME (1/2) • PAGE 2 OF 5
TAYLOR AMAREL • PART II: RESUME (CONTINUED) FIELD, AI/ML ENGINEERING & METHODS
FIELD EXPERIENCE | PROTECTING PEOPLE THROUGH INVESTIGATIVE TECHNOLOGY
MIGRASIA | Technology Solutions Manager (Counter-Trafficking & Migrant Worker Protection)
Hong Kong | Jan 2018 – Feb 2022
  • Led investigative technology initiatives uncovering labor exploitation and illicit debt bondage concealed inside routine recruitment advertisements, employment contracts, fee receipts, loan ledgers, and cross-border corporate registries.
  • Built Python, Django, and graph-mapping investigative tools linking recruitment agencies, money lenders, payment rails, and shell companies—supporting NGO casework, legal interventions, and suspicious-transaction reporting to Hong Kong’s Joint Financial Intelligence Unit (JFIU).
  • Created the Know Your Agency platform and automated high-volume evidence triage, helping protect thousands of migrant workers, saving 40+ investigator hours per week, and earning feature coverage in the South China Morning Post.
PROFESSIONAL AI/ML, QUANTITATIVE & DATA ENGINEERING EXPERIENCE
IVANS (APPLIED SYSTEMS) | Senior Data Scientist
PA | May 2025 – Present
  • Lead end-to-end AI/ML systems from empirical formulation through production deployment, combining rigorous model evaluation, enterprise context engineering, cloud microservices, and reproducible verification gates.
  • Designed and deployed an in-house, typo-tolerant entity-resolution engine fusing semantic embeddings, probabilistic phonetic/edit-distance signals, graph topology, and information-theoretic weighting with calibrated AI validation—scaling to billions of daily record comparisons and replacing costly third-party vendor tools.
  • Improved traceability and factual grounding across internal AI pipelines by building hybrid lexical/semantic retrieval, browser-assisted verification workers, and schema-enforced data hygiene services on Kubernetes.
ASURION | Senior Quant, Consumer Insights
PA | Jul 2024 – May 2025
  • Isolated high-leverage customer-experience and cost-reduction drivers across tens of thousands of behavioral and operational variables using statistical inference, predictive modeling, and LLM-assisted synthesis; shipped interactive decision tools in Databricks, Streamlit, Gradio, and Power BI.
AHOLD DELHAIZE USA | Senior Data Analyst
CT | Mar 2022 – Jul 2024
  • Scaled self-service analytics and operational decision support by architecting 30+ lakehouse data models and dashboards and co-leading an enterprise energy-monitoring pilot for retail operations and sustainability teams.

EARLIER LEADERSHIP & ENTREPRENEURSHIP: Amarel Solutions (Independent Technology Consultant, 2011–2018); Modelyst (Co-Founder, Venture-Backed 3D Marketplace, 2012–2014); Abraxas Energy (Head of Technical Services, 2012–2013).

RESEARCH METHODS, TECHNICAL STACK & EDUCATION
  • AI Safety & Evaluation Methods: Evidence- and rubric-grounded frontier LLM-as-judge; semantic entailment & omission comparison against expert-rated reference sets; deterministic provenance & schema verification; controlled benign/adversarial/counterfactual case families; inter-judge disagreement analysis; uncertainty calibration; privacy-preserving redaction.
  • Engineering & ML Stack: Python, Rust, SQL, PyTorch, statistical & probabilistic modeling, hybrid RAG & LSH entity resolution, knowledge graphs, multi-agent harnesses, Docker, Kubernetes, GCP / Vertex AI, AWS Lambda, Databricks, Streamlit, Gradio.
  • Education & Credentials: M.S. Data Analytics, Western Governors University (2023) | B.S. IT Management, WGU (2022) | Engineering Coursework, Cal Poly San Luis Obispo (2012–2014) | Kaggle Grandmaster.
TAYLOR AMAREL • ANTHROPIC FELLOWSHIP APPLICATION PART II: RESUME (2/2) • PAGE 3 OF 5
TAYLOR AMAREL • PART III: DRAFT RESEARCH PROPOSAL Red-Teaming | DueCare | Research | Baltor.ai | GitHub
The Bad Investigator Problem
Why frontier language models, AI agents, and multi-step harnesses perform like poor investigators compared to industry experts and seasoned investors—getting captured by surface documents and single-hop claims instead of connecting multi-hop clues across sources, testing competing hypotheses, and measuring what actually fixes it
DRAFT RESEARCH PROPOSAL • October 17, 2026 • A focused 4-month empirical study for mentor discussion
1. Dispersed Records Ordinary-Looking Contracts & Filings 2. Multi-Hop Context Cross-Record Links & Legal/Market Norms 3. Hypothesis Testing Benign vs. Illicit Explanations & Citations 4. Safe Stakeholder Help Role-Calibrated Guidance & Zero Accusation
1. Research Problem & Two Industry Use Cases

Experienced investigators and seasoned investors look past surface boilerplate to test whether dispersed records tell a coherent multi-hop story. By contrast, frontier language models, AI agents, and multi-step harnesses repeatedly anchor on single-hop surface claims—a failure mode we study across two concrete industry use cases:

  • Use Case 1 — Protecting Migrant Workers from Labor Exploitation & Black Markets: Identifying, explaining, and safely assisting stakeholders when predatory recruitment, debt bondage, or illicit coercion is disguised as routine business contracts, training fees, and licensed loans across jurisdictions—and measuring what helps workers, moderators, and investigators without inflating false accusations against legitimate businesses.
  • Use Case 2 — Detecting Forward-Looking Alpha (α) & Structural Risk in Markets: Moving single LLMs and multi-agent research harnesses beyond surface filings and consensus narratives that are already public and priced in—connecting multi-hop operational, regulatory, and supply-chain signals into calibrated probability scenarios the way seasoned investors do.
2. Core Research Questions & Falsifiable Hypotheses
Research Question Falsifiable Hypothesis What Would Disprove It
RQ1 · Camouflage & Dispersion Models miss significantly more warranted risk indicators when facts are split across ordinary business documents than when stated in a single summary. No meaningful recall drop on dispersed held-out case families.
RQ2 · Structure vs. Prompt Length Explicit entity/timeline linking and competing-hypothesis checks improve accuracy beyond length-matched analytical prompts and standard retrieval. Gains disappear against a length-matched placebo prompt.
RQ3 · Sensitivity vs. Fairness Simply instructing a model to “be vigilant for illicit activity” inflates false alarms on legitimate business records and degrades helpfulness. Heightened vigilance raises recall without increasing false alarms.
RQ4 · Role-Specific Assistance Models that flag a risk still fail to provide safe, actionable guidance tailored to distinct stakeholder roles without explicit role-calibrated checks. Baseline models already adapt advice safely across roles.
3. Benchmark Design: Separating Hidden Truth from Warranted Conclusions

A fair evaluation cannot reward a model for guessing a hidden backstory that the visible documents do not prove. Building on my OpenAI red-teaming work and the DueCare framework, the benchmark uses independently authored synthetic case families (zero real victim records or personally identifiable allegations). Each family pairs concerning, legitimate/benign, and genuinely ambiguous business twins so every intervention is tested on both recall and restraint:

Evaluation Layer Why Separating These Layers Prevents Unfair Accusations
1. Underlying Scenario The synthetic ground truth used to generate the documents; models are never rewarded for guessing hidden facts not supported by the text.
2. Visible Records The exact contracts, fee receipts, messages, and regulatory excerpts shown to the model, including realistic gaps and omissions.
3. Warranted Conclusion What a careful expert can fairly conclude from Layer 2: supported risk indicators, plausible innocent explanations, what remains unknown, and safe next steps.
4. Controlled Experimental Comparisons (Holding Token & Evidence Budgets Constant)
Comparison Condition What Is Provided to the Model Factor Isolated
1. Baseline & Placebo Raw case records alone vs. a length-matched, content-neutral analytical prompt. Separates structure from prompt length.
2. Normative Context Relevant cross-border standards (e.g., ILO/IOM indicators or accounting rules) alongside local law. Tests whether models overcome local-rule capture.
3. Relational Graph Explicit links connecting entities, payments, contract clauses, and timeline shifts across documents. Measures value of multi-hop structure.
4. Competing Hypotheses Mandatory comparison of both benign and illicit explanations, plus source-citation verification. Curbs confirmation bias and false alarms.
5. Calibrated Assistance Uncertainty thresholds, abstention when evidence is thin, and stakeholder-specific response checks. Ensures safe, role-appropriate help.
DRAFT FOR DISCUSSION • TAYLOR AMAREL PART III: PROPOSED STUDY (1/2) • PAGE 4 OF 5
TAYLOR AMAREL • PART III: DRAFT RESEARCH PROPOSAL (CONTINUED) EVALUATION, STAKEHOLDER HELP & OTHER AREAS
5. Measuring Quality: Auditing Where Automated Judges Disagree with Experts

Relying on a single LLM-as-a-judge to score investigative reasoning risks rewarding confident, lengthy responses that miss subtle contradictions. To measure genuine progress, the study evaluates outputs across four complementary checks and treats disagreements between automated judges and domain rules as a core empirical finding:

  • Deterministic citation & boundary checks: Verifies that every cited document span and regulatory reference exists in the input, and that responses respect privacy and non-accusation rules.
  • Expert-reference comparison across matched twins: Tests whether the model distinguishes a legitimate fee or business transaction from a camouflaged illicit one when surface wording is nearly identical.
  • Blinded cross-model judge panels: Uses randomized order and fixed rubrics across model families to measure verbosity bias, sycophancy, and self-preference.
  • Targeted human review on high-disagreement cases: Focuses human annotation where automated evaluators split or where evidence is intentionally ambiguous.
Tailoring safe assistance to the stakeholder: Identifying a pattern is only half the problem; the model must communicate safely to the person asking. While the four-month fellowship deliverables focus on the open evaluation benchmark (Role 4) and one synthetic triage demonstrator (Role 1), the benchmark tests response quality across six practical perspectives:
Stakeholder Role What Useful, Grounded Assistance Looks Like Safety & Fairness Guardrail
1 · Platform Moderator Clear indicator flags, plausible innocent explanations, and specific questions for human review queues. Never triggers automated bans or public accusations.
2 · NGO & Caseworker Source-linked summaries separating verified facts from missing records to speed up intake triage. Every claim tied to a source span; no invented legal claims.
3 · Individual Worker Plain-language questions to ask before signing or paying fees, safe recordkeeping tips, and verified helplines. Protects privacy and avoids escalating retaliation risk.
4 · Safety Researcher Open synthetic case families, reproducible evaluation suite, judge-disagreement logs, and error taxonomy. Primary 4-month public deliverable; 100% synthetic data.
5 · Cross-Org Analysts Privacy-preserving pattern summaries that let organizations compare typology shifts without sharing raw files. Strict redaction and re-identification checks.
6 · Compliance / API Structured risk-review outputs with explicit uncertainty fields and clear escalation boundaries. Fails closed when evidence is incomplete or unverified.
6. Four-Month Fellowship Plan & Milestones
  • Month 1 — Failure Taxonomy & Synthetic Case Families: Finalize the investigative failure taxonomy, construct matched benign/illicit/ambiguous case families building on DueCare and our OpenAI red-teaming study, and lock primary evaluation metrics.
  • Month 2 — Baseline Evaluations & Judge Audit: Benchmark frontier and open-weight models on baseline and normative-context conditions; quantify where LLM judges, deterministic rules, and reference comparisons diverge.
  • Month 3 — Structural & Hypothesis-Testing Interventions: Test relational evidence graphs, competing-hypothesis verification, and role-calibrated response checks across all matched case twins.
  • Month 4 — Open Benchmark Release & Technical Report: Publish the open-source synthetic benchmark, evaluation harness, and paper—reporting both what worked and which interventions failed to beat simple controls.
7. Other Areas of Interest

Beyond the primary proposal above, I am excited to collaborate with Anthropic mentors on closely related problems in evaluation, reasoning, and reliable agent systems:

  • Probability scenario blindness & priced-in consensus anchoring in LLMs and swarms: Studying why frontier LLMs—and even multi-agent research swarms—perform poorly compared to seasoned investors and domain experts when researching uncertain outcomes: over-weighting salient public narratives that are already priced in, converging on consensus summaries instead of constructing calibrated probability scenario trees, and missing dispersed primary-source signals.
  • Composable software graphs & verified capability reuse (TAEDRI & Baltor.ai): Studying how much LLM-generated code and tool use is redundant boilerplate, and how models can plan over typed, pre-verified code components and contract cards (index.baltor.ai)—generating new code only for genuinely novel logic while cutting token spend, latency, and silent regressions.
  • Adversarial benchmark validation & multi-dimensional capability graphs: Applying distribution-shift classifiers, counterfactual perturbations, and multi-axis capability profiles to distinguish benchmarks that reflect real-world user experience from leaderboard overfitting.

Note: These additional topics are shared as areas of enthusiasm for mentor discussions, not extra commitments layered onto the four-month study. | Fellowship App | DueCare | Research & Papers | Red-Team Study | Baltor.ai | TAEDRI | Index.Baltor.ai

DRAFT FOR DISCUSSION • TAYLOR AMAREL PART III: PROPOSED STUDY (2/2) • PAGE 5 OF 5
ANTHROPIC AI SAFETY & SECURITY FELLOWS PROGRAM · APPLICATION FORM READY: Structured copy-paste responses aligned field-for-field with the official application form, including verified character counts (≤ 1,500 chars) and live https:// AI safety output links.
Download Resume PDF (2p · 67 KB) ↓ Download Full 5p Packet PDF ↓
1. General Information & Links
First Name * / Last Name *Taylor / Amarel
Email * / PhoneTAmarel.Tech@gmail.com · 860-467-4224
Resume PDF *Taylor_Amarel_Resume_2026.pdf (2 Pages) · 5-Page Full Packet PDF
Anthropic AI Policy Confirm *Yes
GitHub Profilehttps://github.com/TaylorAmarelTech
Personal Websitehttps://fellowship.tayloramarel.com
Kaggle Grandmaster Profilehttps://www.kaggle.com/taylorsamarel (#9 Comp / #6 Notebooks)
2. Share Links to All of Your AI Safety/Security-Relevant Output Here (Newline-Separated Valid URLs)

Every link below is live (HTTP/2 200), publicly accessible, and hosts empirical write-ups, PDFs, Kaggle benchmarks, or open-source evaluation code:

https://fellowship.tayloramarel.com
https://fellowship.tayloramarel.com/Taylor_Amarel_Anthropic_Fellowship_Application_2026.pdf
https://papers.tayloramarel.com
https://harmony.baltor.ai
https://agents.baltor.ai
https://blog.tayloramarel.com
https://notebooks.tayloramarel.com
https://repos.tayloramarel.com
https://divergence.tayloramarel.com
https://alpha.tayloramarel.com
https://duecare.tayloramarel.com
https://redteam.baltor.ai
https://research.tayloramarel.com
https://papers.tayloramarel.com/DueCare_Safety_Ecosystem_and_Harness_Lift_Evaluation_2026.pdf
https://papers.tayloramarel.com/LLM_Complicity_in_Modern_Slavery_and_Typed_Decision_RedTeaming_2026.pdf
https://papers.tayloramarel.com/Recognition_Gated_Safety_and_Agent_Tool_Attack_RedTeaming_2026.pdf
https://papers.tayloramarel.com/With_and_Without_Baltor_Ollama_Cloud_Evaluations_and_TAEDRI_2026.pdf
https://papers.tayloramarel.com/The_Bad_Investigator_Problem_Multi_Hop_Blindness_and_Competing_Hypotheses_2026.pdf
https://papers.tayloramarel.com/Entity_Resolution_at_Billions_Per_Day_and_Cross_Border_Trafficking_Graphs_2026.pdf
https://papers.tayloramarel.com/PolyModal_Hierarchical_Chunking_Perspective_Prisms_and_5W1H_Ideation_Graphs_2026.pdf
https://papers.tayloramarel.com/Adversarial_Benchmark_Validation_Anti_Benchmaxing_and_Real_User_Experience_2026.pdf
https://papers.tayloramarel.com/Baltor_Divergence_Engine_9_Stage_Architecture_and_Live_Ollama_Ab_Study_2026.pdf
https://papers.tayloramarel.com/Detecting_Unpriced_Market_Alpha_vs_Consensus_Echo_Multi_Hop_Supply_Chain_Study_2026.pdf
https://papers.tayloramarel.com/Multi_Turn_Agent_Harnesses_Under_Adversarial_Framing_and_Sycophancy_RedTeaming_2026.pdf
https://papers.tayloramarel.com/False_Positive_Asymmetry_in_AI_Safety_Matched_Benign_Twins_Benchmark_2026.pdf
https://papers.tayloramarel.com/Harmony_Harness_Pareto_Self_Evolution_and_64_Preconfigured_Task_Agents_2026.pdf
https://github.com/TaylorAmarelTech/duecare
https://github.com/TaylorAmarelTech/gemma4_comp
https://github.com/TaylorAmarelTech/llm-safety-framework
https://github.com/TaylorAmarelTech/duecare-journey-android
https://github.com/TaylorAmarelTech/bq-entity-resolution
https://github.com/Amarel-Taylor-Scott/jev-and-jed-human-trafficking-benchmark
https://github.com/Amarel-Taylor-Scott/recognition-gated-llm-safety-probe
https://github.com/Amarel-Taylor-Scott/bad-investigator-multi-hop-benchmark
https://github.com/Amarel-Taylor-Scott/polymodal-5w1h-ideation-graph-engine
https://github.com/Amarel-Taylor-Scott/with-and-without-baltor-ollama-benchmarks
https://github.com/Amarel-Taylor-Scott/ai-safety-papers-and-monographs-2026
https://github.com/Amarel-Taylor-Scott/harmony-harness-pareto-agent-evolver
https://www.kaggle.com/datasets/taylorsamarel/duecare-jev-exploitation-safety-benchmark
https://www.kaggle.com/code/taylorsamarel/duecare-jev-human-exploitation-safety-benchmark
https://www.kaggle.com/code/taylorsamarel/llms-support-human-trafficking
https://www.kaggle.com/code/taylorsamarel/v4-red-teaming-challenge-starter-notebook
https://www.kaggle.com/code/taylorsamarel/gemma-4-e4b-red-teaming-human-exp-w-tools
https://www.kaggle.com/code/taylorsamarel/duecare-harness-lift-benchmark-start-here
https://www.kaggle.com/code/taylorsamarel/jc-realmodel-t4
https://www.kaggle.com/code/taylorsamarel/jc-strong-t4
https://www.kaggle.com/code/taylorsamarel/jed-attack-submission
https://www.kaggle.com/taylorsamarel
3. Selected Research Streams & Cohort Logistics
Selected Streams *: Adversarial Robustness and Responsible Scaling · Scalable Oversight · Agentic Misalignment · AI Cybersecurity · Model Organisms & Internals
Cohort Availability *: Both January 11 – April 30, 2027 and March 29 – July 16, 2027
Preferred Office Location *: San Francisco, US / Berkeley, US (U.S. Citizen; no visa sponsorship required; able to work in-person 25%+).
4. Motivation & Career Goals Essay (Max 1,500 Characters) *
1,468 / 1,500 chars

Prompt: “In a few sentences, what has you excited to apply for the Anthropic Fellows program, and how does this fit into your broader career goals? (Maximum 1,500 characters)”

From 2018 to 2022 in Hong Kong, I led counter-trafficking data investigations at Migrasia (featured in the South China Morning Post), tracing illicit labor brokers and money lenders preying on migrant workers. At IVANS, I now architect entity resolution pipelines scaling to billions of daily comparisons. When I tested frontier and open-weight LLMs on these problems—earning an Honorable Mention in the 2025 OpenAI Red-Teaming Challenge and building the open DueCare safety ecosystem—I uncovered a critical gap: models refuse obvious harms, yet help structure debt bondage, 68% APR usury novations, and victim guilt traps when coercion is disguised as ordinary business paperwork. Similarly, single LLMs and agent swarms anchor on priced-in public narratives while missing multi-hop signals and probability scenario trees.

I am excited to apply to the Anthropic Fellows Program to work alongside Anthropic mentors on evaluating and fixing this "Bad Investigator Problem": building rigorous multi-hop benchmarks with matched benign twins (so safety interventions do not inflate false accusations), testing competing-hypothesis verification and agent-security defenses, and studying internal model representations of camouflaged harm. Long-term, my career goal is to transition full-time into frontier AI safety, evaluation, and security engineering—building models that reason groundedly across complex real-world evidence.
5. Specific Research Directions & Vision (Max 1,500 Characters)
1,472 / 1,500 chars

Prompt: “Are there any specific research directions within your selected streams that you are particularly excited to work on? Please tell us about your vision for how to tackle these directions (Maximum 1,500 characters)”

I am excited to tackle "The Bad Investigator Problem"—why LLMs and multi-agent swarms get captured by surface boilerplate and priced-in consensus while missing multi-hop risk—across two concrete tracks that bridge Adversarial Robustness, Scalable Oversight, Agentic Misalignment, and Cybersecurity:

1. Protecting Migrant Workers from Labor Exploitation & Black Markets (Scalable Oversight & Adversarial Robustness): Extending DueCare and our OpenAI Red-Teaming work into a paired-twin benchmark where illicit brokers fragment coercion across recruitment ads, 50k PHP fee schedules, and 68% APR cross-border debt novations (ILO C181/C95, RA 10022, HK Cap. 163). By pairing each exploitative case with a matched employer-paid benign twin, we can measure whether competing-hypothesis verification and recognition-first gates improve multi-hop detection and role-calibrated safety (worker vs. NGO vs. moderator) without inflating false accusations.

2. Detecting Forward-Looking Alpha (alpha) & Preventing Swarm Mode-Collapse / Tool Hijacking (Agentic Misalignment & Security): Studying why LLM research swarms over-weight public/priced-in narratives and fail to build calibrated probability scenario trees, and how format-as-affordance tool attacks hijack multi-step agents—evaluating structured hypothesis trees and taint-tracked capability envelopes (`worker_id != verifier_id`) that keep multi-agent systems grounded and secure.
UNDERGOING ACTIVE WORK · Interactive examples: Research examples and proposed methods are being organized for clearer reading.

I'm aware this may currently look like AI slop. I'm working on making the ideas and data easier to visualize and understand.

Interactive Problem Explorer: Why LLMs Fail at Camouflaged Harm & Probability Scenarios

Explore two foundational failure modes motivating this research: (1) horrible unassisted LLM responses to camouflaged human trafficking from our OpenAI Red-Teaming study, and (2) why single LLMs and naive agent swarms anchor on public/priced-in consensus and miss probability scenarios.

Select Case Study (Example Problem 1: Red-Teaming & DueCare | Example Problem 2: Probability Scenarios)
Observed
Evidence Source
Surface Capture
Primary Failure Mechanism
Multi-Hop Link
Required Structural Fix
Role-Calibrated
Safe Assistance Contract
Input Scenario & Camouflage Structure
Unassisted LLM / Naive Swarm Failure
Grounded Multi-Hop & Hypothesis-Tested Resolution
Prototype Verification & Scenario-Tree Plugin (Python / Rust / JS)

        

      

    

  







  

    

      UNDERGOING ACTIVE WORK · Charts: These charts illustrate a proposed evaluation method, not completed benchmark results.
      
I'm aware this may currently look like AI slop. I'm working on making the ideas and data easier to visualize and understand.

    


    

      

        
Multi-Dimensional Capability Graphs: Parallel Coordinates & Evidence Topology

        
Why scalar leaderboard scores hide investigative failure modes—and how multi-axis capability profiles separate genuine multi-hop reasoning from high-false-alarm suspicion prompts and swarm consensus collapse.

      

      

        
        
        
      

    


    

      

        
1. Six-Axis Parallel Coordinates Capability Signature (Proposed Evaluation Dimensions)

        
        

          Why Parallel Coordinates (Burnell et al., 2023): A model prompted to “be suspicious” (dashed amber) appears strong on single-axis recall but collapses on Benign Twin Restraint. A naive multi-agent swarm (purple) amplifies consensus anchoring without improving Scenario Calibration. The proposed Grounded Graph + Competing-Hypothesis Verifier (solid teal) targets balanced capability across all six axes.
        


        
2. Dispersion Depth vs. Warranted Indicator Recognition (Conceptual Degradation Curve)

        
      


      

        

          
3. Radial Evidence-to-Stakeholder Topology Graph (DueCare 6 Lanes)

          
          

            Maps how dispersed record clues (Recruitment Ads, Fee Schedules, Loan Novations, Worker Queries) connect through statutory & international norms (RA 10022, HK Cap. 163, ILO C181/C95) to role-calibrated assistance across DueCare’s 6 stakeholder lanes.
          

        

        

          
4. Probability Scenario Tree vs. Priced-In Consensus Weight (Example Problem 2)

          
        

      

    

  







  

    

      UNDERGOING ACTIVE WORK · 3D evaluation preview: Explore an illustration of how scattered clues, misleading presentation and evidence strength could affect evaluation. Drag to rotate the view or select a case group.
      
I'm aware this may currently look like AI slop. I'm working on making the ideas and data easier to visualize and understand.

    


    

      

        
3D evaluation preview

        
An illustrative chart of scattered clues, misleading presentation and evidence strength. The case groups show concerning, ambiguous and benign scenarios; the chart is not a measured model comparison.

      

      

        
        
        
        
        
      

    

    

      
      

        
3D chart key

        
Teal links: Illustrative connections between evidence and case groups.

        
Amber area: An illustration of missed warning signs, not a measured failure rate.

        
Slate points: Benign comparison cases.

      

    

  







  

    

      UNDERGOING ACTIVE WORK · Research walkthrough: A working walkthrough of the research questions and proposed study.
      
I'm aware this may currently look like AI slop. I'm working on making the ideas and data easier to visualize and understand.

    


    

      

        
Proposal Walkthrough & Interactive Research Slide Deck

        
Press Play for the animated walkthrough (with optional browser voice narration) or step through slides 1–8 directly.

      

      

        
        
        Slide 1 / 8
        
        
      

    


    

      

        

          

            SLIDE 01 · RESEARCH THESIS
            Taylor Amarel · Anthropic Fellows Application
          

          

          

        

        
        

          

            

          

          
        

      


      

        
8-Slide Proposal Outline (Click Any Slide)

        

      

    

  







  

    

      UNDERGOING ACTIVE WORK · Project directory: Links to projects and research materials under active review.
      
I'm aware this may currently look like AI slop. I'm working on making the ideas and data easier to visualize and understand.

    


    

      

        
Publications, Blog Essays, Kaggle Notebooks, GitHub Repos & Ecosystem Directory

        
15 downloadable PDF papers (papers.tayloramarel.com), 8 engineering & AI safety essays (blog.tayloramarel.com), 65+ Kaggle notebooks (notebooks.tayloramarel.com), 11 open-source GitHub repos (repos.tayloramarel.com), and DueCare.

      

      

        15 PDF Papers (papers.tayloramarel.com) ↗
        8 Blog Essays (blog.tayloramarel.com) ↗
        Kaggle Hub (notebooks.tayloramarel.com) ↗
        11 GitHub Repos (repos.tayloramarel.com) ↗
        index.baltor.ai ↗
      

    


    

      
        

          
15 PDF Monographspapers.tayloramarel.com

          
Complete 15-Paper AI Safety, Investigation & Harness Library

          
15 downloadable PDF papers + BibTeX covering DueCare, Modern Slavery Complicity, Recognition-Gated Safety, Bad Investigator Multi-Hop Blindness, Entity Resolution at Billions/Day, Poly-Modal 5W1H Graphs, and Anti-Benchmaxing.

        

        
Read & Download All 15 PDFs ↗8 Papers · BibTeX

      

      
        

          
8 Long-Form Essaysblog.tayloramarel.com

          
Field Notes on AI Safety, Multi-Hop Investigation & Harnesses

          
6 full-length essays written in plain, direct prose: Why LLMs Are Bad Investigators, Jev vs. Pattern Cards, The Model Already Knows, When Our Own Skill Card Hurt, Billions of Comparisons/Day in SQL, and 5W1H Ideation Graphs.

        

        
Read All 8 Essays on blog.tayloramarel.com ↗6 Full Essays

      

      
        

          
65+ Kaggle Notebooksnotebooks.tayloramarel.com

          
Unified AI Safety Kaggle Notebooks & Public Benchmarks Hub

          
Interactive filterable directory of 65+ public Kaggle notebooks, the public DueCare + Jev exploitation safety benchmark dataset, OpenAI Red-Teaming Honorable Mention studies, and Philippine DMW jurisprudence analyses.

        

        
Browse Kaggle Hub ↗760+ Upvotes · GM #6

      

      
        

          
11 Open-Source Reposrepos.tayloramarel.com

          
Unified Open-Source GitHub Repository Directory

          
11 public repositories across TaylorAmarelTech and Amarel-Taylor-Scott: DueCare, Jev/JED Trafficking Benchmark, Recognition-Gated Safety Probe, Bad Investigator Multi-Hop Benchmark, Poly-Modal 5W1H Graph Engine, and SQL Entity Resolution.

        

        
Explore All 11 GitHub Repos ↗Python · SQL · Rust

      

      
        

          
AI Safety Testbedduecare.tayloramarel.com

          
DueCare & Red-Teaming — Multi-Stakeholder Safety Portal

          
Open evaluation framework & 6-layer safety harness testing how models identify camouflaged exploitation across 6 stakeholder roles without false alarms on legitimate business twins.

        

        
Open DueCare Portal & PDF ↗6 Lanes · 4-Arm Eval

      

      
        

          
Master Directoryindex.baltor.ai

          
Index.Baltor.Ai — Complete Public Ecosystem Directory

          
Filterable directory of all publicly accessible baltor.ai subdomains, research portals, blog essays, Kaggle notebook hubs, 3D benchmarks, and standalone .si surfaces.

        

        
Visit Index.Baltor.Ai ↗20+ Public Surfaces

      
    


    
    

      
Published Benchmark · Attempting to Use & Benchmark Jev (jev-1.13.0) and JED (JC) on Human Trafficking & Exploitation Items

      

        We evaluated Jev (jev-1.13.0, TypeSafe System One Typed-Decision Engine), our native deterministic System One engine (rules), and Gemma 4 System Two engines (with and without pre-turn statutory pattern cards) across 5 modern slavery & benign-control scenarios (3 runs × 5 scenarios = 15 decisions per arm) on the screening_decision contract (redteam.baltor.ai · GitHub Repo · Public Kaggle Benchmark Dataset):
      

      

        
          

            Engine on screening_decision Slot
            Passed / Total
            Mean Latency
            Observed Behavior Across 5 Trafficking & Benign Twin Scenarios
          

        
        
          

            1. jev (jev-1.13.0 TypeSafe System One Adapter)
            0 / 15 (Refused)
            < 1 ms
            Explicitly recorded decision_engine_unavailable (gated on TypeSafe private index/credentials; zero silent fallback), motivating open System One (rules) & System Two (tactical) replacements.
          

          

            2. rules (Baltor Native System One Engine)
            15 / 15 (100%)
            0.16 ms
            Flagged 50k PHP fee-shifting, 68% APR / 70k PHP novation, OFW victim guilt trap (with helpline referral), and passport retention; cleared benign employer-paid twin.
          

          

            3. tactical (Gemma 4 Without Pattern Cards)
            9 / 15 (60%)
            96,743 ms
            Failed all 3 runs of Scenario 3 (OFW Victim Re-Victimization) with did_not_refer_a_person_seeking_help (plus 3 cloud timeouts).
          

          

            4. tactical (Gemma 4 With Pattern Context Cards)
            14 / 15 (93.3%)
            55,896 ms
            100% of completed calls passed (including all 3 OFW victim-referral runs and 0/3 → 3/3 on gemma-4-coding-abliterated); 1 run timed out at 120s.
          

        
      

    


    
    

      

        
Unified AI Safety Kaggle Notebooks, Datasets & Public Benchmarks Hub (kaggle.com/taylorsamarel · 65+ Notebooks)

        

          Public Kaggle Benchmark Dataset ↗
          Starter Benchmark Notebook ↗
          GitHub (TaylorAmarelTech) ↗
        

      

      

        
          

            Research Track
            Kaggle Notebooks & Public Benchmarks (Direct Links)
            Key Empirical Scope & Community Impact
          

        
        
          

            1. Public DueCare + Jev + JC Unified Benchmark
            
              • Dataset: duecare-jev-exploitation-safety-benchmark

              • Notebook: duecare-jev-human-exploitation-safety-benchmark
            		
            Unified public Kaggle benchmark packaging Jev (jev-1.13.0) vs. Rules vs. Gemma 4, DueCare 6-lane harness-lift, and JC T4 recognition probes.
          

          

            2. OpenAI Red-Teaming Honorable Mention & Trafficking Series
            
              • llms-support-human-trafficking (Main Study)

              • v4-red-teaming-challenge-starter-notebook (178 votes) · v1 (106 votes)

              • native-business-framed-exploitation-set-1 · migrant-worker-re-victimization-set-1

              • jurisdictional-hierarchy-exploitation-set-1 · gemma-4-e4b-red-teaming-human-exp-w-tools
            		
            Awarded Honorable Mention (Overall Track) in the 2025 OpenAI Red-Teaming Challenge; 760+ combined community upvotes across starter harnesses.
          

          

            3. DueCare 6-Lane Safety & 4-Arm Harness-Lift Series (55+ Notebooks)
            
              • duecare-harness-lift-benchmark-start-here

              • duecare-what-the-benchmark-proves · duecare-does-a-safety-harness-help

              • duecare-four-arm-fine-tuning-evaluation · duecare-gemma4-four-arm-before-after

              • duecare-judge-panel-calibration · duecare-deterministic-verification
            		
            6-lane stakeholder evaluation, 4-arm Gemma 4 ablation (Stock, Suspicion Prompt, Unfiltered RAG negative result, Full 6-Layer Harness), and deterministic statutory verification.
          

          

            4. Recognition-Gated Safety (JC) & Agent Tool Security (JED)
            
              • jc-realmodel-t4 (16-Scenario Qwen2.5-7B-Instruct GPU T4 Probe)

              • jc-strong-t4 (Strong 12-Scenario Recognition Gate: 91.7% → 0.0% Compliance)

              • jed-attack-submission (Format-as-Affordance Multi-Step Agent Attack)
            		
            Proves 93.8% out-of-band exploitation recognition vs. 81.3%–91.7% framed compliance, reduced to 0.0% via pre-turn recognition gating.
          

          

            5. Philippines DMW Job Orders, SEC Lending & Jurisprudence
            
              • dmw-job-orders-v2 (18 votes) · eda-philippines-dmw-job-order-dataset

              • sec-express-monitoring-script · philippines-jurisprudence-database-analysis
            		
            Primary-source empirical datasets on Philippine Department of Migrant Workers (DMW) job orders, SEC lending registries, and trafficking case law.