Evidence behind the design.

In May 2026, 70.2% of sampled individual Codex users requested at least one task estimated above one hour of human work. Random 0.1% user sample; effort is model-estimated, not measured runtime or successful completion.

Estimated autonomy is higher in Claude Code than chat/Cowork for 26 of 31 output categories. Sessions: 10 April–10 June 2026; classifier scale 1–5. This is a product-surface association, not proof that a different interface causes better delegation.

Repeat cross-occupation use rose from 13.1% to 25.9% across about 6,200 consistently observed US users, April–July 2026. The eligible repeated-task pool grows over time. This is descriptive use, not a workforce-wide or agent-delegation rate.

Enterprise usage and a survey of 9,000 workers across almost 100 enterprises describe reusable, integrated workflows. About 20% of Enterprise messages went through Custom GPTs or Projects in recent months. Vendor-customer evidence; these features are not equivalent to autonomous agents.

Respondents report agent scaling in 40% of larger organisations and 22% of smaller ones. Organisational size and survey participation matter; these are reported practices, not a census or evidence of Aksara demand.

53% report agents in selected production workflows or broad organisational use. More than 800 developers and engineering leaders across six Southeast Asian countries and India. Not SEA-only, all workplaces, or nationally representative country estimates.

37% reported delegating tasks independently to agents; 46% cited privacy concerns. September 2025 fieldwork: 2,132 Australian/New Zealand knowledge workers, weighted online panel. Different measure and population from Agoda.

Official PDF, Appendix 10, printed page 76: Indonesia 72.78%, Papua Tengah 28.71%, Papua Pegunungan 6.76%. Susenas 2024; ICT questions cover ages 5+, internet use in the preceding three months. Publication: 29 August 2025. Rates measure use, not uptime, speed, or demand. Provincial boundaries must remain explicit.

The four Cs are connectivity, compute, context, and competency. High-income economies account for 87% of notable models and 91% of AI venture funding while representing 17% of global population. More than 40% of ChatGPT traffic came from middle-income countries in mid-2025. These describe infrastructure and use, not product effectiveness.

Asia-Pacific urban/rural internet use: 88%/66%; worldwide 85%/58%. Population estimates combine surveys, modelling, and nowcasts. Regions overlap with income groups. These averages cannot stand in for a particular province or measure connection reliability.

2025 Digital Government Index data: 35 of 36 surveyed OECD governments use AI somewhere; 6 of 36 have an open algorithm register. Formal transparency standards: 11 of 36. Germany and the US did not participate. Published 15 June 2026; this is not Indonesian government practice or a finding about any particular system.

44% expressed confidence that AI would assess public-benefit applications only when safe and trustworthy. This is public perception from the 2025 Trust Survey, published June 2026. It does not measure system safety, local trust, or actual service outcomes.

The Indonesian remote-village accountability programme summary reports maths outcomes rising from 37% to 49%, and language from 37% to 50%, after a year. This was not an AI intervention. It informs questions about accessible information and accountability, not an AI benefit claim.

Collective Benefit, Authority to Control, Responsibility, Ethics. A normative lens for custodianship, consent, access, and permitted reuse. Community data is not automatically public; implementation and outcomes require their own evidence.

Frames AI through human choices and agency. Useful as a design principle; not empirical proof of our product, local devices, or any proposed deployment.

An existing participatory-software precedent for traceable proposals, revisions, meetings, and accountability. Architectural reference, not a proposed replacement for Strata or an AI evaluation.

Six-week, guided Copilot/GPT-4 intervention for first-year senior secondary students. Composite assessment gain: 0.31 SD; English-only gain: 0.23 SD. The composite also included AI/digital knowledge. Not offline, Papua, or evidence for the proposed devices.

Summarises a large gap between real-household robot tasks and controlled simulation. Benchmark success is setting-specific; it does not establish safe autonomous physical actions. Observation and reviewed display are separate from robotic control.

Reviews sensor, network, data, AI, and privacy architecture, with a corpus of 298 articles and searches through September 2025. A broad literature synthesis, not field evidence for Abstraksi’s surfaces or environmental sensing.

Benchmarks small models on single edge devices and distributed clusters. Finds some favourable edge cases and others requiring cloud fallback. Resource, latency, and cost trade-offs are platform-specific; no universal offline guarantee follows.