What does good human–computer interaction mean when the computer is no longer waiting for the human to operate it?
This is the intellectual layer of Gerardo’s work. Each body of work is labeled by its actual status so an argument, implementation, experiment, and evaluated finding are not treated as the same kind of evidence.
Human–AI Interaction
Agency, cognition, creativity, delegation, work, learning, memory, authorship, and accessibility.
Authority & Trust
Permission, provenance, identity, verification, execution, accountability, ownership, and recourse.
Human–Technology Futures
Embodied, ambient, spatial, decentralized, and other interaction models that change the human relationship with computation.
Bodies of work
Verifiably Human
What evidence justifies a claim of human creation, and what remains unknown?
Review the work →UI-GATES
Mechanisms and measurements for evidence in AI-assisted software work, including negative and inconclusive results. The dedicated repository and reference playbook remain available for inspection.
The Trust Stack
Identity, provenance, authenticity, authority, and capability are related signals with different limits.
Review the work →The Authority Layer
A design argument for connecting human-readable intent to bounded execution authority.
Review the work →What has been measured
These findings apply to the named UI-GATES workloads. They do not establish results for every agent or repository. Observed numbers do not override an experiment’s declared “no verdict.”
Gate and receipt overhead
The 16-task Workboard run observed 85.3% higher mean weighted tokens for ceremony versus control on 14 complete triples. Control accepted 16/16 tasks, ceremony 15/16, learned 14/16. Acceptance imbalance meant no formal verdict. Two pilots also observed higher cost.
Workboard experiment ↗Knowledge reuse
Learning’s incremental token effect was indistinguishable from zero in that run. No measured learning benefit was established.
Learning comparison ↗Hook-first savings
No saving demonstrated; no verdict. Control accepted 20/24 tasks and treatment 23/24. On 20 mutually accepted pairs, saving was −0.6% (95% interval −10.3% to +9.0%).
UI-GATES repository (hook-first experiment) ↗Audit checks
Tests cover rejection of vacuous commands such as bare git diff. An accepted command does not establish that the check was adequate for every claim.
Audit tests ↗Enforcement and record limits
Gate rules are instructions the model follows. Real enforcement remains the host’s sandbox and permission controls. Agent-writable records are tamper-evident to engine checks, not tamper-proof against a hostile agent.
Engine trust boundaries ↗Jev live approval
The project status report records live advisory approval, but no live APPROVE for a gate-class action. Live records are not published here, so this is a reported status, not a gate-class validation result.
WUN-19 status report (access required) ↗Agent handoff claim accuracy has not been evaluated. The token experiments checked code acceptance and costs, not whether the agents’ statements matched their actions. No real-agent learning certificate has passed (read the incomplete learning evaluation ↗).
Is checking verification claims a problem worth solving?
The question is whether reviewers repeatedly spend time resolving narrow claims such as “these tests passed on this revision,” when the evidence already existed at handoff. Existing CI output, diffs, and a concise handoff are alternatives a proposed checker must justify improving on.
The next step is a retrospective of 10–20 consecutive sessions, with human-reviewed labels and checking effort. Continue only if contradicted or effortful-to-confirm claims recur and existing evidence could resolve them. Stop if checking effort is rare, new verification would be required, or most claims remain ambiguous or unobservable. Missing records do not establish dishonesty.
No extractor, runner, evidence-packet feature, or replacement product is being announced. A positive retrospective would justify a prospective comparison that measures review effort and defect detection. Faster approval with worse defect detection is a failure.