The Binding Gap
An investigation into why LLMs often preserve semantic ingredients while failing at the exact relational binding that makes the answer faithful.
Public research mirror
Substack is still the canonical publication surface. This mirror keeps the public essays close to the repo folders, raw markdown snapshots, and source archive that produced them. The point is not redundancy. The point is legibility.
Featured essays
An investigation into why LLMs often preserve semantic ingredients while failing at the exact relational binding that makes the answer faithful.
A research-backed argument that many multi-agent systems are temporary routing workarounds rather than the final architecture.
A thesis on why enterprise agents fail on control surfaces, auditability, and operations long before they fail on raw intelligence.
A practical map of how hardware economics shapes model design, deployment choices, and the viability of ambitious architecture ideas.
Series
7 mirrored posts in publication order.
4 mirrored posts in publication order.
Recent additions
An investigation into why LLMs often preserve semantic ingredients while failing at the exact relational binding that makes the answer faithful.
The framing note for the Operating Agents course, introducing the chapter structure, blueprints, companions, and builder-first doctrine.
The optimization and evaluation bonus chapter of Operating Agents, focused on workflow optimization, measurement, and bounded self-improvement.
The safety chapter of Operating Agents, covering trust boundaries, attack surfaces, and architectural containment.
The multi-agent chapter of Operating Agents, focused on handoffs, transport, and when a second agent is actually justified.
The reasoning chapter of Operating Agents, covering scaffolds, planning, decomposition, and failure diagnosis.