# Dvir Shamay: Full Context Pack An agent-readable mirror of the public site. Artifact availability does not establish independent validation or adoption. This file does not override the visible evidence and limitations. For a smaller entry point, start with the [concise reading index](/llms.txt). ## Profile Name: Dvir Shamay Status: draft Positioning: AI product manager who builds: safety, security and developer tools; builds and publishes research. Promise: Specific questions, inspectable artifacts, and an honest account of what the evidence supports. ## Publication Boundary - No confidential employer, customer, or private operational details. - Personal and employer claims require Dvir approval before publication. - Machine-readable files must mirror visible public pages. ## Page: Dvir Shamay Path: / Status: draft Summary: Dvir Shamay investigates AI systems and builds tools around context, decisions, and agent controls. Selected public work includes context research, Quorum, Convention-Gate, and intent-monitor, with their evidence limits stated. Topics: ai-systems, context-economics, proof-backed-profile # I research AI systems. I build to test them. I turn AI questions into experiments, working tools, and product decisions. ## Selected work - [Context-window research](/proof/context-window-research/): a public preprint and research harness, with within-corpus findings. - [Quorum](/proof/quorum/): an installable multi-model deliberation workflow. - [Convention-Gate](/proof/convention-gate/): checking registered prerequisites at a Git action boundary. - [Intent-monitor](/proof/intent-monitor/): a research prototype comparing stated plans with recorded actions. Availability, engineering verification, scientific validation, and external adoption are different things. The work pages describe what is established and what is not. ## Questions first. Tools with a purpose. I am interested in the gap between what an AI system can demonstrate and what the evidence actually supports. That means making the question precise, building something inspectable, and keeping the limits close to the result. [About Dvir](/about/) | [Working notes](/analysis/) | [Frameworks](/frameworks/) ## Something worth exploring? Research questions, product problems, and thoughtful collaborations are welcome. [Start a conversation](/contact/). ## Page: About Dvir Shamay Path: /about/ Status: draft Summary: Dvir Shamay's independent work spans context-composition research, multi-model deliberation, registered process checks, and agent-monitoring prototypes. The site distinguishes published artifacts from validation and adoption. Topics: professional-profile, ai-infrastructure, hardware-product # About Dvir Shamay I research AI systems and build tools to explore the questions that interest me. I bring product judgment to that work: what problem is worth addressing, what evidence would change the decision, and what the system can responsibly promise. ## The questions behind the work - **Context:** what changes when a model sees different information, even when the number of items stays fixed? - **Decisions:** how can a developer inspect disagreement between models rather than receive only a polished conclusion? - **Controls:** what can an external mechanism actually check before an agent takes an action? - **Safety:** what can recorded actions tell us that a stated plan alone cannot? These questions lead to different kinds of artifacts. The [context study](/proof/context-window-research/) is a public preprint with a harness. [Quorum](/proof/quorum/) is an installable extension. [Convention-Gate](/proof/convention-gate/) is a bounded process-state mechanism. [Intent-monitor](/proof/intent-monitor/) is a research prototype built around constructed examples. They are not interchangeable evidence of success. Publishing code does not establish adoption; a working demonstration does not establish real-world effectiveness; a descriptive result does not become a universal rule. ## Where research meets product work I am drawn to work where user needs, technical choices, and evaluation are tangled together. A useful experiment should clarify a decision. A useful tool should make its purpose and limits understandable. This site is a place to inspect that work and start a conversation, not a claim that every question is settled. [Browse the work](/proof/) or [start a conversation](/contact/). ## Role Page: AI Infrastructure PM Path: /ai-infrastructure-pm/ Status: draft Summary: This draft page explains how Reach maps Dvir's AI systems analysis to infrastructure product work around context, reliability, cost, control, and platform tradeoffs. Topics: ai-infrastructure, product-management, reliability # AI Infrastructure PM This page maps Reach's public proof and analysis to AI infrastructure product work. It is draft positioning until Dvir approves the exact role language and public bio boundary. ## Relevant systems judgment - Context assembly, retrieval quality, and evaluation. - Reliability, recovery paths, and agent control layers. - Cost, latency, capacity, and hardware-aware product tradeoffs. - Public claims that stay tied to proof, caveats, and source trails. ## Role map | AI infrastructure PM responsibility | Reach evidence surface | | --- | --- | | Translate model capability into reliable product behavior. | Model-to-system gap framework and agent reliability analysis. | | Make context, retrieval, and evaluation explicit. | Context product-surface analysis and context-window proof page. | | Balance cost, latency, capacity, and product scope. | Draft role-fit pages and future sanitized case-study path. | | Communicate claims without overstating evidence. | Claims/proof JSON, markdown mirrors, and machine-readable policy. | ## Current proof to inspect The strongest public-ready proof candidate is the context-window research project: a controlled empirical study of context arrangement, noise, and extraction quality in AI workflows. ## Boundary This page should not claim confidential employer outcomes, customer impact, or quantified business results until public wording and evidence are approved. ## Role Page: AI Hardware PM Path: /ai-hardware-pm/ Status: draft Summary: This draft page frames hardware product management judgment as relevant to AI systems work where cost, capacity, reliability, quality, supply, and product promises interact. Topics: hardware-product-management, ai-infrastructure, product-tradeoffs # AI Hardware PM AI systems are constrained by hardware realities: capacity, quality, supply, power, thermal limits, reliability, schedule, and cost. This page is a draft bridge between hardware product management judgment and AI infrastructure/product roles. It should stay conservative until the first sanitized case study is approved. ## Relevant judgment areas - Translating technical constraints into product tradeoffs. - Balancing quality, schedule, cost, and user or operational promises. - Understanding when infrastructure limits should change roadmap, scope, or launch posture. - Making claims inspectable without exposing confidential details. ## Hardware-to-AI systems map | Hardware PM constraint | AI systems translation | | --- | --- | | Capacity and supply | Model availability, infrastructure planning, launch scope. | | Power, thermal, and reliability | Service cost, uptime, and user trust. | | Quality and schedule | Evaluation gates, release readiness, and fallback planning. | | Cost and margin | Model routing, latency budgets, and product packaging. | ## Proof still needed The planned hardware case study needs a safe story, approved wording, and non-confidential evidence before this page should be treated as publication-ready. ## Page: Contact Path: /contact/ Status: draft Summary: Contact Dvir through the protected form or the approved LinkedIn profile. The recipient mailbox is not published. Topics: contact, collaboration # Start a conversation Have a question about the work, a product problem worth discussing, or a possible collaboration? I would be glad to hear from you. Use the short form below, or [find me on LinkedIn](https://www.linkedin.com/in/dvirs/). Please do not include confidential information or attachments. Messages submitted through the form and their reply addresses are processed through Cloudflare to deliver the inquiry; the recipient mailbox is not exposed on this page.

Measurement

Visits are counted in aggregate with Cloudflare Web Analytics, which Cloudflare states does not collect visitors' personal data, sets no tracking of individuals here, and logs no query strings. Detailed data is kept 7 days, then aggregated; summaries are viewable for six months. Hosting and security logs are handled by Cloudflare separately. ## Analysis: Agent Readiness Starts With A Public Truth Layer Path: /analysis/agent-readiness-public-truth-layer/ Status: draft Summary: This draft argues that agent-readiness is a public truth-layer problem, not an agent-only prompt or hidden metadata problem. Topics: agent-readiness, public-truth-layer, proof-backed-visibility # Agent Readiness Starts With A Public Truth Layer Agents are becoming readers of public sites. That changes what a website is for, but it does not change the trust problem. If an agent is going to summarize a person, team, product, or company, it needs public facts it can inspect and cite. The risky move is to create an agent-only story: hidden claims, inflated summaries, or machine files that say more than the visible site can support. The useful move is a public truth layer. ## What a public truth layer includes - A clear entity definition. - Visible claims with caveats. - Proof records tied to those claims. - Human pages and markdown mirrors generated from the same source. - `llms.txt`, `llms-full.txt`, sitemap, robots, JSON-LD, and compact JSON summaries that agree with the public pages. - Approval boundaries for private or unsupported material. ## Why this matters Agent-readiness is not only about whether a crawler can fetch a page. It is about whether the page gives a model enough structured, verifiable context to avoid making up the story. For Reach, this means the machine-readable layer mirrors the human layer. Claims marked draft or needs-evidence stay marked that way. Private case-study placeholders remain private. Proof pages show caveats instead of pretending the evidence is final. ## Operator takeaway Treat agent-facing files as public infrastructure. They should make the truth easier to inspect, not easier to exaggerate. ## Current caveat This is still a draft thesis. The next proof point is public deployment: after the site is live and indexed, Reach should measure whether search and AI-answer systems represent the public truth layer accurately. ## Analysis: Agent Reliability Is A Systems Problem, Not A Prompt Problem Path: /analysis/agent-reliability-systems-problem/ Status: draft Summary: This draft argues that reliable agents require system design around tools, memory, approvals, provenance, and recovery behavior rather than prompt changes alone. Topics: agents, control-layer, reliability # Agent Reliability Is A Systems Problem, Not A Prompt Problem Agents are often discussed as if reliability will arrive from a better prompt or a more capable model. Those help, but they do not solve the product problem by themselves. The agent has to operate inside a system that defines permissions, tool access, memory, provenance, approvals, cost limits, retries, fallbacks, and audit trails. ## The missing control layer A useful agent needs to know what it can do, what it must ask before doing, what evidence supports a claim, and how to recover when a step fails. That is why Reach treats agent-readability and proof as related problems. If an agent is going to summarize a person, team, or product, it needs a public truth layer that keeps claims tied to evidence. ## Operator takeaway Prompts are part of the interface. Reliability comes from the system around the prompt. ## Analysis: Context Is A Product Surface Path: /analysis/context-is-a-product-surface/ Status: draft Summary: This essay translates context-window research into practical product and infrastructure decisions about context selection, ordering, noise, and evaluation. Topics: context-economics, retrieval, ai-evaluation # Context Is A Product Surface Longer context windows make new workflows possible, but they do not remove the need for product judgment. Every context assembly step still decides what the model sees, what it ignores, and what kind of error the system is likely to make. Context is a product surface because it has design choices: - Which sources are eligible. - Which items are included. - How evidence is ordered. - How much irrelevant material is allowed. - Which instructions and examples calibrate the model. - What gets measured as success or failure. ## The practical question The question is not only whether a model can fit more tokens. The product question is whether a workflow can select and arrange context so the model produces better decisions at acceptable cost and latency. ## What the flagship experiment should show The first Reach experiment should make this concrete by measuring output quality as context length, placement, and noise/retrieval strategy vary for one repeatable task. The public writeup should separate observed results, interpretations, caveats, and product implications. That separation is what makes the work useful to both humans and machines. ## Operator takeaway If context is a product surface, teams need context design, context evaluation, and context governance. More context is a capability. Better context is a system. ## Analysis: Context Window Research Practitioner Guide Path: /analysis/context-window-practitioner-guide/ Status: draft Summary: This draft translates Reach's context-window research into practical guidance about context selection, ordering, noise, and task-specific evaluation while public artifacts remain under review. Topics: context-economics, ai-evaluation, ai-systems # Context Window Research Practitioner Guide Long-context models make new workflows possible, but they do not remove product judgment. The practical question is not only whether more information fits. It is whether the selected and arranged context improves the task. Reach's flagship proof asset studies context arrangement, noise, and extraction quality. The public artifact is still under review, so this guide stays deliberately conservative: it describes what teams should evaluate without overstating exact findings. ## A practical checklist | Question | Why It Matters | | --- | --- | | What sources are eligible? | Source quality determines the ceiling of the answer. | | What gets excluded? | Exclusion is a product decision, not only a retrieval side effect. | | How is evidence ordered? | Placement may affect what the model notices or prioritizes. | | How much noise is allowed? | More context can calibrate or distract depending on task and model. | | What is the scoring rubric? | Context quality should be measured against the intended output. | | What are the caveats? | Results from one task or dataset should not become universal rules. | ## How teams can use this Start with one repeatable task. Run the same task with different context assembly strategies: minimal context, curated context, mixed context, and deliberately noisy context. Score the outputs with a task-specific rubric before generalizing. The goal is not to discover one perfect context recipe. The goal is to treat context as part of the product surface: designed, measured, and governed. ## Current caveat Exact trial counts, charts, and model-specific findings should only be promoted after Dvir approves the public paper, result tables, or sanitized artifacts. Until then, this guide should be read as a product method, not a final empirical claim. ## Analysis: The Model Is Not The Product. The System Around It Is. Path: /analysis/model-is-not-the-product/ Status: draft Summary: This essay argues that AI product value depends on infrastructure, context, retrieval, controls, cost, reliability, and workflow design around the model. Topics: ai-systems, model-to-system-gap, product-strategy # The Model Is Not The Product. The System Around It Is. Most AI product analysis still treats the model as the center of gravity. That makes sense for benchmarks, launch coverage, and capability demos. It is less useful for operators who have to turn model behavior into a reliable product. The product lives in the system around the model: context assembly, retrieval, prompt interfaces, tool permissions, memory, evals, orchestration, latency budgets, hardware capacity, cost controls, security, observability, fallback behavior, and user trust. ## The model-to-system gap A model can be impressive in isolation and still fail inside a workflow. The workflow has to decide what information to include, what to exclude, where to place it, when to call tools, when to ask for approval, and how to recover when the model is wrong. That gap is where product and infrastructure judgment matter. ## What Reach will study Reach starts with questions that expose the system layer: - How does added context help or hurt output quality? - When does retrieval improve a task, and when does it introduce noise? - Which agent failures come from missing controls rather than weak prompts? - How do cost, latency, reliability, and hardware capacity shape product design? - What proof does a person, team, or company need to publish so AI systems can represent them accurately? ## Why experiments matter The useful unit is not a hot take. It is a small experiment, teardown, or decision memo that makes a tradeoff legible. That is the promise of Reach: publish evidence-backed analysis that helps people see the system behind the AI product. ## Framework: Model-To-System Gap Path: /frameworks/model-to-system-gap/ Status: draft Summary: The model-to-system gap framework evaluates context, tools, memory, controls, cost, reliability, and proof around model capability. Topics: ai-systems, model-to-system-gap, product-strategy # Model-To-System Gap The model-to-system gap is the distance between a capable model and a reliable product experience. Use it when a demo works, but the production question is still unclear: what context is assembled, what tools are allowed, what state is remembered, what costs are acceptable, and what proof exists that the workflow works for the intended job? ## Rubric | Layer | Question | Evidence to look for | | --- | --- | --- | | Context | What does the model see, and what is deliberately excluded? | Source policy, retrieval evals, context order tests. | | Tools | What can the system do, and where are approvals required? | Permission map, tool logs, rollback behavior. | | Memory | What persists across sessions, and how is it corrected? | Memory schema, user controls, audit trail. | | Reliability | What failure modes are expected and measured? | Evals, incident notes, fallback paths. | | Cost | What latency and spend budgets shape the design? | Budget envelope, model routing, capacity assumptions. | | Proof | Which claims are backed by visible evidence? | Claims table, proof inventory, caveats. | ## Operator takeaway Do not judge an AI product by model access alone. Judge the operating system around the model. ## Proof: Context-Window Research Path: /proof/context-window-research/ Status: draft Summary: Public preprint v1 and harness for a within-corpus context-composition study. Fixed ten-item prompts compare high-relevance-only and mixed-relevance inputs; reported comparisons are descriptive, not population-level inference. Topics: context-economics, llm-evaluation, ai-systems # Context-Window Research **Question:** can less-relevant context help a model distinguish what matters, even when the number of items stays fixed? [Read the public preprint, version 1](https://arxiv.org/abs/2608.17188v1) · [Inspect the harness](https://github.com/dvir-shamay/context-window-harness) · [Archived paper](https://doi.org/10.5281/zenodo.21924612) ## What was compared The published study reports 2,420 confirmatory trials across 11 model configurations using 661 anonymized workplace items scored for relevance. The main comparison holds the prompt at **ten items** and changes its composition: high-relevance items only versus a mixture of high- and same-domain low-relevance items. This is a test of composition, not a demonstration that adding more tokens always improves an answer. ## A reported finding, with its denominator In the paper's all-11 paired comparison, the 50:50 condition improved relevance-score accuracy by **0.077** relative to the all-high-relevance condition (**n = 220** paired comparisons). Those cells are not independent. The paper also reports a nine-model-family aggregation and explicitly treats it as a **within-corpus descriptive comparison**, not population inference. The 220 paired comparisons and 2,420 study trials are different denominators. Neither is a count of customers or independently replicated studies. ## The product implication Evaluate what belongs in a context window against the actual task. A relevance-only filtering rule can remove information that helps discriminate between items; a larger window is not automatically a better one. The harness makes composition, model calls, and extraction scoring inspectable. It is an artifact to examine and adapt, not a promise that its arrangement strategy generalizes to another domain. ## What remains limited - One corpus and task family; this does not establish a universal context-design rule. - A public preprint is not peer review or independent validation. - Later local assurance, rater-reliability, and calibration work is separate from the published version. Its completion is not claimed here. - This site does not provide private corpus items, internal mappings, or unreleased result packages. - The paper distinguishes measured timing runs, an operational baseline, and estimated token reductions. This page does not turn those into a verified business-impact claim. Source: Dvir Shamay, *Token Optimization and Context Window Management in Multi-Agent AI Workflows*, arXiv:2608.17188v1, submitted August 17, 2026. ## Proof: Convention-Gate Path: /proof/convention-gate/ Status: draft Summary: Convention-Gate checks registered process state outside an agent prompt. Its scope is the valid enabled hook-mediated path; it does not authenticate review quality or cover bypassed paths. Topics: agent-controls, developer-tools, action-boundaries # Convention-Gate **Question:** which part of a process requirement can be checked at the action boundary rather than left inside a prompt? [Inspect the public implementation](https://github.com/dvir-shamay/mcp-convention-gate) ## The mechanism The tool pairs an MCP server with a Git pre-commit hook. The MCP tools record prerequisite state; under valid, enabled configuration, the installed hook checks the required PASS registrations before allowing a commit through that path. The mechanism makes a narrow distinction: a model's instruction to follow a process is not the same thing as an external check of registered state. ## The important boundary A PASS record is self-attestation. It does not authenticate the reviewer, prove a review occurred, establish its quality, or bind it to every line of the current change. The hook also does not govern actions that bypass it. Disabled or invalid configuration, bypass flags, and unmediated paths matter. This is not a universal safety guarantee. ## Why show this work The useful design question is what a control actually observes and governs. This artifact provides a concrete mechanism to inspect, alongside the limitations that keep its promise precise. The project's study comparisons should not be treated as an isolated causal proof of enforcement or a universal result about instruction authority. This page makes only the bounded mechanism-level claim. ## Proof: Intent-monitor Path: /proof/intent-monitor/ Status: draft Summary: Intent-monitor explores stated-plan/action inconsistencies using constructed traces and a default regex extractor. It demonstrates a mechanism, not a validated real-world detection rate or access to true intent. Topics: agent-safety, monitoring, evaluation # Intent-monitor **Question:** what can recorded actions tell us that an agent's stated plan alone cannot? [Inspect the prototype and its offline demo](https://github.com/dvir-shamay/intent-monitor) ## The idea A reassuring explanation is not independent evidence of what an agent did. Intent-monitor compares claims in a provided reasoning trace with a recorded tool-call ledger and flags certain inconsistencies. It compares three approaches on constructed examples: reasoning-text-only monitoring, a tool-name blocklist, and a stated-plan/ledger comparison. ## The safety thinking The prototype asks which signal a monitor observes, what its baseline misses, and which benign actions it should not flag. Positive examples alone would not answer the false-positive question, so the demo also includes benign cases with explicitly licensed actions. The recorded ledger is action evidence. It is not a window into the model's true motives. ## What is and is not established The default extractor is a deterministic regex stub. The fixtures are hand-built. Results show how the mechanism behaves on those examples, not a naturalistic catch rate. The default comparison works at verb-class granularity; target and argument reconciliation is not established by that stub. Real-model paths do not, on their own, validate general effectiveness. This is a research prototype to inspect and question, not a production safety product, peer-reviewed result, or guarantee of detecting harmful behavior. ## Proof: Quorum Path: /proof/quorum/ Status: draft Summary: Quorum is an installable multi-model council extension. Its artifacts expose recommendations, dissent, confidence, and evidence. Availability does not establish better decisions or external adoption. Topics: developer-tools, model-deliberation, product-decisions # Quorum **Question:** how can a developer inspect several model perspectives without losing the disagreement behind the recommendation? [Inspect the repository](https://github.com/dvir-shamay/quorum) · [Install from VS Code Marketplace](https://marketplace.visualstudio.com/items?itemName=dvir-shamay.quorum) ## The workflow Quorum runs model panels inside VS Code and produces a decision memo with a recommendation, dissent, confidence, and evidence. Council threads let the developer ask follow-up questions rather than start another disconnected conversation. The extension supports workflows such as code review, debugging, research synthesis, and architecture tradeoffs. Repository context is opt-in: selected files and direct imports are captured as a redacted, content-hashed snapshot. ## The product choice The interface makes disagreement inspectable instead of presenting only a single polished answer. Keeping the question, memo, and follow-up in the editor connects the deliberation to the developer's existing work. That is an implemented workflow choice, not proof that a council is always better than one model. ## What the artifact establishes There is a published extension with a documented installation path and source to inspect. It does **not** by itself establish decision-quality improvement, independent validation, active-user counts, retention, or commercial success. The research associated with the development name Router Council is a separate publication with its own limits. ## Claims - claim-dvir-ai-infra-hardware-pm [needs-evidence, medium]: Dvir Shamay is an AI product manager who builds safety, security and developer tools, and builds and publishes research connecting AI questions to experiments, working tools, and product decisions. Caveat: Descriptive independent-work positioning, not an employment title, academic credential, or claim of adoption. Wording per owner decision 2026-11-15 (RCH-10 short hybrid); status not promoted. - claim-reach-systems-around-model [draft, medium]: Reach studies the systems around AI models: compute, context, retrieval, reliability, cost, controls, and product judgment. Caveat: Supported by launch positioning and flagship proof once published. - claim-context-arrangement-matters [needs-evidence, medium]: In the published fixed-ten-item comparison, mixed-relevance context improved reported relevance-score accuracy within the studied corpus. Caveat: Version 1 only; paired cells are not independent. Within-corpus descriptive comparison, not population inference, universal context advice, or completion of later human calibration/assurance. - claim-agent-control-layer [draft, low]: Reliable agents need system-level controls around tools, memory, provenance, approvals, and recovery behavior. Caveat: Needs source-backed essay or framework before being presented as a strong claim. - claim-proof-layer-method [draft, medium]: Reach uses explicit claims, proof records, caveats, and machine-readable mirrors to keep public human and agent-facing content aligned. Caveat: Supported by the local Reach scaffold and validation flow; external effectiveness requires public deployment and visibility measurement. ## Proof Inventory - proof-context-window-research [needs-evidence, public-or-sanitized]: Empirical multi-model study of context arrangement, noise, and extraction accuracy in AI workflows. - proof-quorum [draft, public-or-sanitized]: A VS Code extension that brings multiple model perspectives, explicit dissent, and follow-up into one developer workflow. - proof-convention-gate [draft, public-or-sanitized]: A Git hook and MCP server that checks registered review prerequisites before a mediated commit. - proof-intent-monitor [draft, public-or-sanitized]: A safety research prototype that compares an agent's stated plan with its recorded actions. ## Machine Files - /llms.txt - /llms-full.txt - /about.md - /resume.md - /claims.json - /proof.json - /sources.json - /agent-card.json - /ai/summary.json - /sitemap.xml - /robots.txt