00 / overview - research program
The cognition research behind the instrument.
AI throughput compounds. Human deliberate attention is a fixed-capacity channel. Everything zeno measures sits in that gap - and every number on this site traces back to a peer-reviewed or pre-registered source.
Pre-registered & frozen
The RTLX-S validation and SCED field study were pre-registered before data collection. The design is frozen; the analysis plan is fixed.
Claims are mapped
Every research claim on this site is enumerated in claims-map.json, and updates ship via feed.xml.
Local-first & auditable
Capture runs on your machine by default - local-first and auditable, your cognition never leaves your control. Report issues via security.txt.
01 / papers - the program
Three papers. One thesis.
Everything below feeds one argument: the bottleneck in AI-assisted work is the human supervising it. The thesis paper states the case; the two protocols put it under pre-registered test.
The Bandwidth Gap
AI throughput compounds while the human supervising it does not - the widening gap is supervision cost. The thesis paper states the case for measuring it and traces every claim to the evidence base.
Read the preview →
02 / concurrency - the curve
Attention peaks near four.
Effective output rises with parallelism, then the cost of holding it all in your head takes over. Across domains that look unrelated - supervised agents, monitored screens, working-memory chunks, live conversations - the peak clusters at a small, strikingly similar number. Pick a domain and watch the shape hold.
- Domain
- AI agents under supervision
- Peak at
- N = 4measuring
- Signal
- RTLX-S supervision load climbs steeply once you pass ~4 parallel agents.
Curve data lives in one shared file, grounded in the R1-R6 evidence review; agents and conversations are evidence-shaped and per-user calibrated.
Our working hypothesis: the peak is a property of human attention, not of the thing being supervised. zeno's job is to measure where your peak sits, on your real workload, instead of guessing.
03 / prosoche - protocol v1.0 preview
A four-week attention trial, pre-registered.
The protocol tests whether prosoche, a Stoic attention discipline centered on impression-checking and deliberate assent during action, performs differently from matched mindfulness and productivity controls in AI-supervisor work. The pre-specified primary contrast: prosoche vs mindfulness on NASA-TLX change during a standardized AI-mediated task.
04 / sclt - theory v1.0 preview
Cognitive load theory, extended to agent crews.
SCLT argues that classical cognitive load theory under-specifies the reality of supervising AI agent crews. Supervisors are not primarily learning domain schemas: they are calibrating trust, sampling outputs, integrating heterogenous deliverables, and deciding intervention timing under uncertainty. Four new load constructs carry that work - trust calibration, oversight, integration, and intervention decision load.
05 / method - how we do research
Method before data.
Four rules govern every study in the program. They are boring on purpose: rigor is the product.
06 / foundations - evidence base
Fourteen foundations. Every claim sourced.
zeno's model is not a metaphor. It stands on six decades of workload, attention, and automation research - from the workload scale NASA validated in 1988 to a 2025 randomized controlled trial of AI-assisted experts.
- 01
NASA-TLX workload scale
The validated workload standard zeno adapts. Hart & Staveland 1988
cited in: Whitepaper V · SCLT 4 · Prosoche 5
- 02
Raw-TLX validity
License to drop pairwise weighting. Byers, Bittner & Hill 1989
- 03
Cognitive load theory
Working memory is capacity-limited. Sweller 1988
cited in: Whitepaper VI · SCLT 1
- 04
Yerkes-Dodson law
The inverted-U; the concurrency curve's ancestor. Yerkes & Dodson 1908
- 05
Multiple Resource Theory
Parallel tasks interfere super-additively. Wickens 2008
- 06
Supervisory control
Managing agents is supervisory control. Sheridan & Verplank 1978
cited in: SCLT 2
- 07
Levels of automation
Automation transforms work and adds coordination demands. Parasuraman, Sheridan & Wickens 2000
- 08
Ironies of Automation
Automating the easy parts leaves the hardest residual. Bainbridge 1983
- 09
Vigilance is hard work
Monitoring is not free attention. Warm, Parasuraman & Matthews 2008
- 10
The lumberjack effect
More automation, worse failure recovery. Onnasch et al. 2014
- 11
Automation complacency
Emerges under multi-task load; afflicts experts. Parasuraman & Manzey 2010
- 12
Interruption recovery cost
~10-15 min to resume programming after a break. Parnin & Rugaber 2009
- 13
AI made experts 19% slower
A randomized controlled trial - while developers felt faster. Becker et al. 2025, arXiv:2507.09089
cited in: Whitepaper IV · Prosoche refs
- 14
Human-AI combos often underperform
Meta-analysis: combinations are frequently worse than the best alone. Vaccaro, Almaatouq & Malone 2024, Nature Human Behaviour
07 / pre-release - waitlist
Every claim sourced. Waves are opening.
This instrument is in pre-release. Waves exist because the instrument is calibrated per cohort, and every early user gets direct access to the researcher.