NUKEZ // DISPATCH
Blog
Research notes, field reports, and technical essays on agent memory, provenance, verification, and ownership.
13dispatches
4research threads
2026active field notes
Research Note — Frontier Models / May 2026
How the Frontier Is Quietly Building Toward Nukez
Across 55 cloud models and 825 benchmark runs, every frontier lab is shipping pieces of the same underlying problem — durable, verifiable, portable agentic state — as a proprietary, vendor-locked mechanism. A pattern hiding in plain sight.
Research Note — Digital Ownership / May 2026
What You Bought, What They Let You Keep
An agent buys you a movie. Twenty months later, the storefront is gone. The movie is not. A field report on receipt-bound libraries and digital ownership.
Read the research note →
Research Note — Agent EDI / May 2026
Receipts for Agent-Issued EDI
EDI standardized B2B transport. Autonomous agents change the trust model. Purchase orders, invoices, remittances, and adjustments now need receipt-bound proof chains.
Read the research note →
Field Note — MCP Provenance / May 2026
The Provenance Gap
A forensic audit of Claude Desktop on macOS shows MCP has no provenance primitive. The absence is not a gap in tooling — it is a gap in the protocol.
Read the field note →
Memory Is a Lie — Part 2 / April 2026
The Infrastructure Behind the Lie
Part 1 covered why context windows can't be memory. Part 2 goes underneath: what they cost at the GPU level, what those costs force the infrastructure to do, and where memory has to live.
Read Part 2 →
Memory Is a Lie — Part 1 / April 2026
Your Agent's Memory Is a Lie
Every multi-agent framework on GitHub treats memory like a solved problem. None of them are right. The context window is structurally incapable of being memory — and the math says so.
Read Part 1 →
Field Note — Verification Architecture / April 2026
From Bash Script to Protocol
The concrete architecture that turns forensic agent reconstruction into real-time detection — a context manager, a local signing handler, and cryptographic behavioral attestation.
Read the field note →
Field Note — Agent Self-Knowledge / April 2026
The Agent That Couldn't Verify Itself
An AI analyzed its own session logs and discovered it couldn't distinguish between thinking deeply and not thinking at all. The foundational problem of agent self-knowledge.
Read the field note →
Field Note — Cost of Trust / April 2026
The $42,000 Bug Report
When agent quality degrades without a verification layer, the cost isn't linear — it's multiplicative. One engineer's fleet failure shows how unverifiable trust compounds into catastrophic waste.
Read the field note →
Field Note — Agent Telemetry / April 2026
What 234,760 Tool Calls Reveal
A senior AMD AI director data-mined months of Claude Code session logs to understand why her engineering workflow collapsed. A case study in agent behavior with no verification layer.
Read the field note →
Technical Note — Inference Economics / April 2026
Pricing the Quadratic
How LLM inference stays linearly priced when the underlying compute is O(n²) — and where the math gets fragile.
Read the note →
Hidden Token Tax — Part 2 / March 2026
The Gatekeeper's Spec
The companies that authored MCP also control the only data channel through it in consumer apps, bill per byte on that channel, and exempt their own tools from the cost.
Read Part 2 →
Hidden Token Tax — Part 1 / March 2026
The Hidden Token Tax
LLM platforms force binary data through probabilistic prediction, charge per token on both ends, and already possess the infrastructure to make it cost zero.
Read Part 1 →