Clippable Research

Bits and atoms research for marketing that learns

We publish design notes on domain-embedded agents, and we run atoms research studies aimed at measuring real-world attention. Software alone is not the moat. Proprietary outcome data is.

Two tracks, one lab

Bits

Software & agent research

Domain-embedded agents with memory outside the prompt, tool orchestration, governance, and shared state across surfaces, so marketing work has owners, spend limits, and an audit trail.

Atoms

Physical-world research studies

Not just software. We pursue atoms research: instruments and studies that measure real-world attention and conversion, the offline black box the ad industry still cannot close.

If the goal is a moat, measure what others cannot

Honest framing: atoms research only compounds when it produces scarce labels, attention and conversion events joined to creatives, placements, and brands. That is the weak spot in marketing science we are building toward.

Hardware copies. Datasets compound.
Atoms research only builds a moat if it produces data nobody else has. Devices get replicated. Proprietary attention and outcome events tied to creatives and placements do not.
Online is solved-ish. Offline is not.
Click and view-through attribution are imperfect but workable. Foot-traffic panels and DOOH cameras sell rough presence. Almost nobody joins physical dwell to campaign outcomes at agent resolution.
The question worth pursuing
Can we measure real-world attention and conversion in a way the ad industry currently can’t, and feed that signal into the same learning loops we use for online creatives?

Near-term research directions

Computer vision attention measurement

Cheap camera or sensor units that estimate passersby, stops, looks, and dwell time at posters and displays, then join those events to creative and location metadata.

Privacy-safe by design

On-device processing. No faces stored. Aggregates only: counts, dwell histograms, optional coarse proxies that never leave the device as identifiable imagery.

Closed-loop learning signals

If agents train on “this creative here produced this dwell and these scans,” that dataset plugs into the same RL / outcome thesis we already argue for online, now covering the physical world.

Full agenda: Atoms Research and the Marketing Measurement Moat.

Publications

Research articles

Research Note #002 · July 24, 2026

Atoms Research and the Marketing Measurement Moat

Why proprietary physical-attention data beats hardware, and what to measure next

Atoms ResearchPhysical AttentionEvaluation & MeasurementSocial & Marketing AI

If the goal is a durable moat, hardware is the wrong object. Devices get copied. Proprietary datasets compound. This note argues that Clippable Research should pursue atoms research, measurement of real-world attention and conversion that the ad industry still treats as a black box, and connect those signals to the same learning loops we already run for online creatives. Online attribution is a solved-ish problem. Offline is not. Rough foot-traffic panels and billboard cameras exist; almost nobody has closed the loop from physical dwell and engagement back to campaign outcomes at the resolution marketing agents need. We outline a concrete direction: privacy-safe, on-device computer vision for passersby, stops, looks, and dwell time, linked to creative and location metadata so agents can learn which work actually holds attention in the physical world.

Read article

Research Note #001 · June 4, 2026

Architecting Domain-Embedded AI Agents

Memory, Tools, Governance, and Multi-Surface Orchestration

Agent ArchitectureLLM SystemsHuman-in-the-Loop

Most production “agents” today are still chat sessions with extra API calls. That is enough for drafting copy; it is not enough when work has owners, deadlines, spend limits, and an audit log. This research note lays out a practical architecture for domain-embedded agents: the model reads and writes through product schemas, keeps memory outside the prompt, routes side effects through tools, and leaves humans on the hook for irreversible steps. The argument is grounded in published work on tool-augmented language models and in operational constraints from social-media programs, where coordination cost often exceeds model capability. The note is descriptive engineering guidance, not a benchmark study.