🤫husshhussh
  • Wiki
Reserve
🤫husshhusshOneOne Puppy
🤫 hussh Research · open paper

Working backwards
from the human.

A first-principles ontology for personal information and human-first agentic AI - built around one person and their graded circle of trust, owned by them, and calm by design.

Read the paperThe research corpus

Open research. Every claim is cited; the verified corpus and references are published in full below.

Manish Sainani · 🤫 hussh Research · 2026

Abstract

The dominant way software organizes information - the web index optimized for an advertiser - was never built for the person using it.29 We argue for an inversion: organize everything by working backwards from one human and their graded circle of trust, and from what they care about and are working toward. We ground this in six bodies of evidence: the discrete, layered structure of human social groups (roughly 5, 15, 50, 150);1,3 the discipline of personal information management and the six senses in which information is “personal”;6,7 person-centered knowledge graphs with access rights and provenance as first-class elements;8,9 privacy by design;14 user-owned data architectures such as Solid;17 and calm, mixed-initiative, human-first interaction.20,24 From these we derive eight first principles, a person-centered ontology, and concrete design standards for a homepage and an asynchronous “give the agent work and walk away” flow. We state our limitations honestly: the intermediate social layers are empirically noisy,4 and much of the agentic-UX guidance is young. This paper is published openly, with its full verified research corpus, so others can check and build on it.

1. The inversion

A search engine is a magnificent machine pointed the wrong way. It indexes the world for whoever pays to reach you, and asks you to translate your life into keywords so it can serve results ranked for its economics, not yours. Mark Weiser argued three decades ago that the best technologies recede into the background and are built around the human environment, rather than forcing the human to adapt to the machine.20 We take that seriously and literally: the right starting point is not the web, nor the document, nor the app, but one person.

Working backwards from that person changes everything downstream. Their relationships are not a flat address book but a structured circle of trust. Their information is not scattered across a hundred institutional silos but an owned, coherent model. Their agents do not demand attention but earn it, returning only when the work is done or a decision is needed. What follows is the ontology and the design standards that fall out of taking the human as the origin.

2. Prior art

The circle of trust has a measurable shape. Human groups self-organize into a discrete hierarchy of preferred sizes - about 5, 15, 50, and 150, extending to ~500 and ~1500 - a geometric series with a scaling ratio near three, tied to the social brain hypothesis and the cognitive limit of ~150 meaningful relationships known as Dunbar’s number.1,3,5 Contact frequency is graded by layer, and the layer a relationship occupies is an informative, structurally distinct variable; circle-aware methods that use it can outperform strong graph-learning baselines at equal cost.3,4,10 Honesty demands the caveat: support is strongest at the innermost and outermost layers, and the intermediate layers are noisy.4

Personal information is a discipline, and it is the person’s. William Jones established personal information management as a field and defined information as “personal” in six senses - owned by, about, directed to, sent by, experienced by, and relevant to the person.6,7 The personal space of information is constructed by the individual and owned individually, even as it is shared - a consent-first, user-owned model, not a static external index.7 Person-centered knowledge graphs make this concrete: a person can be the focal node of a star-shaped graph, robust to sparse data, with access rights and provenance as first-class properties.8,9

Consent can be architectural. Cavoukian’s Privacy by Design requires privacy as the default, embedded from the outset, data-minimizing, positive-sum with utility, and independently verifiable.14,16 Berners-Lee’s Solid instantiates user ownership: data lives in personal pods the person controls, and applications must request access rather than extract it.17,19

The interface should be calm and mixed-initiative. Weiser and Brown’s calm technology informs without demanding attention, using the minimum technology needed and living mostly in the periphery.21,22 Horvitz’s mixed-initiative principles tell an agent when to act: weigh the costs and benefits under uncertainty about the user’s goals, model their attention, defer to a less distracting time, and let people invoke, terminate, and refine the automation.24 Modern agent-UX guidance echoes this: keep status visible, keep the human in control, and treat delegation as ceding continuous control while retaining the felt ability to intervene.25,26 And a warning: “free” AI products built on surveillance capitalism manufacture an illusion of agency; “user-centered” is not enough - a stronger human-first, consent-first standard is required.29

3. First principles

  1. P1
    Work backwards from one human. The unit of design is a single person, not a document collection, a web index, or an advertiser's audience. Everything else - relationships, information, agents - is arranged around that person and their goals and roles.6,20
  2. P2
    Model the circle of trust in graded layers, not a flat 'friend' list. Human relationships self-organize into nested layers of roughly 5, 15, 50, and 150 (extending to ~500 and ~1500), with a scaling ratio near three and contact frequency graded by layer. The layer a tie belongs to is a first-class, informative variable.1,3,4
  3. P3
    Own the model; consent is the interface. A person's information is theirs - owned, about, directed to, sent by, experienced by, and relevant to them. It lives under their control, and it moves only by scoped, revocable consent with a receipt they can read.7,14,17
  4. P4
    Make provenance and access rights first-class ontology elements, not metadata bolted on. Every fact carries where it came from and who may see it. Provenance enables data quality, contextual understanding, and entity reconciliation; access rights make consent enforceable at the data layer.8,11
  5. P5
    Center the person; let structure be multi-axial and evolvable. Represent the person as the focal node of a star-shaped graph, with facets radiating out. Use independent classification axes (polyhierarchy), formal identity keys, and validated constraints, so the ontology grows without global refactoring.9,12
  6. P6
    Privacy and utility are positive-sum, by architecture. Privacy is the default, embedded from the outset, data-minimizing, and independently verifiable. Framing privacy against usefulness is a false dichotomy.14,16
  7. P7
    Invert the interface: the person delegates, the agent returns. Instead of a person continuously commanding a system, they hand off work and cede continuous control. Trust comes from the felt ability to intervene, not from constant oversight.24,26
  8. P8
    Be calm: quiet until the job is done, or money or information is needed. The interface lives in the periphery and moves to the center only when warranted, using the minimum technology needed. It should be less effortful than a search engine, not more.21,24

4. The ontology

The model is a graph with one person at the center. Every other entity is defined by its relation to that person, and every edge can carry consent scope and provenance. The core entities:

Person (the focal node)

The single human the model is built around. All relations originate here. Facets (health, money, work, home, taste) radiate as sub-graphs.

Circle member + Trust layer

Another person, tagged with the layer of the owner's circle they occupy: L1 support clique (~5), L2 sympathy group (~15), L3 band (~50), L4 active network (~150). The layer governs default sharing precision and cadence.

Interest / Passion / Purpose

What the person cares about, is looking for, and is working toward - graded from a passing interest to a life purpose. The engine of relevance, owned and edited by the person.

Information item

A unit of personal information, tagged by which of the six senses of 'personal' it satisfies (owned by, about, directed to, sent by, experienced by, relevant to).

Consent grant

A scoped, revocable permission: who may see what, at what precision, for how long. Carries a receipt. The primary control surface.

Provenance record

Where a fact came from and the process that produced it - a first-class edge, not hidden metadata.

Agent + Delegated task

An agent acting for the person, and a unit of delegated work with goals, constraints, success criteria, and forbidden actions. Returns when done, or when it needs money or information.

Two design choices make this durable. First, the circle-of-trust layer is not a tag on a relationship but a governing property: it sets the default precision and cadence of what is shared, so “share my location” means something different for L1 (family, fine and frequent) than for L3 (a wider band, coarse and occasional).3,4 Second, consent grants and provenance are first-class edges, validated by formal identity keys and per-axis constraints, in a multi-axial polyhierarchy that can evolve without global refactoring.8,11,12

5. Design standards

The homepage. One idea per screen; the person’s goals, not a search box, at the center; the minimum technology needed; quiet by default. It should ask what the person wants to get done and let them delegate it - not present a blank query field and a wall of results ranked for someone else.22,29

The “give work, walk away” flow. The person states a goal and its constraints, success criteria, and forbidden actions.26 The agent works asynchronously and does not interrupt for non-urgent matters; it returns when the job is done, or when it needs money or information, timing any interruption to genuine urgency.24 Its status is always visible, the person can always intervene, override, or refine, and every consequential action is scoped by consent and written to a receipt.24,25 Autonomy is negotiated, never absolute.

6. Limitations and open questions

We hold ourselves to the same honesty we ask of the field. The 5/15/50/150 structure is well supported at its extremes but noisy in the middle layers, so the ontology treats layer boundaries as soft defaults a person can override, not hard truths.4 The agentic-UX literature is young and partly practitioner-driven; its principles are directionally strong but not yet settled science.26,28Formalizing consent so it is both legally meaningful and machine-enforceable, and measuring whether a “calm” homepage genuinely outperforms a search box for real tasks, are open problems we intend to study in the open. This is a position paper meant to be argued with, not a final word.

7. References

  1. [1] Zhou, W.-X., Sornette, D., Hill, R. A., & Dunbar, R. I. M.. Discrete hierarchical organization of social group sizes. Proceedings of the Royal Society B, 272(1561), 439–444 (2005). link
  2. [2] Dunbar, R. I. M.. The social brain hypothesis and human evolution (Size matters: social groups & human evolution). Research Outreach. link
  3. [3] Dunbar, R. I. M. (overview). Dunbar's 5–15–50–150 model: the sympathy group and the full active network. Good Medicine / Good Knowledge. link
  4. [4] Dunbar, R. I. M., et al.. The structure of online social networks mirrors those in the offline world (layered ego networks; mobile-phone evidence). Social Networks, Elsevier (S0378873316301095). link
  5. [5] Dunbar, R. I. M.. Neocortex size as a constraint on group size in primates (the social brain hypothesis). PMC (PMC1634986). link
  6. [6] Jones, W. P.. Personal Information Management (survey & definition). Annual Review of Information Science and Technology; Wikipedia biography. link
  7. [7] Jones, W. P., & Teevan, J. (eds.). Personal Information Management (the six senses of 'personal'; the personal space of information). Springer / Synthesis Lectures. link
  8. [8] Balog, K., & Kenter, T.. Personal Knowledge Graphs: user-controlled data, access rights and provenance as first-class properties. arXiv:2402.07540. link
  9. [9] Research group (person-centric KGs). A star-shaped, person-centric knowledge graph and representation-learning framework (healthcare readmission). arXiv:2305.05640. link
  10. [10] Research group (circle-aware learning). Ego-network-layer (circle-aware) methods outperforming node2vec/SEAL at equal complexity. arXiv:2109.09190. link
  11. [11] Research group (provenance & ontology). Provenance and semi-rigid ontology design for unified knowledge representation. arXiv:2606.20630. link
  12. [12] Ontology-design practitioners. Multi-axial polyhierarchy, key patterns for entity identity, RDF/OWL vs property graphs, SHACL/ShEx validation. Emergent Mind — Ontology Design & Graph Schema. link
  13. [13] Meric, G. (practitioner). Knowledge graph ontology design. Medium. link
  14. [14] Cavoukian, A.. Privacy by Design: The 7 Foundational Principles. Information & Privacy Commissioner of Ontario (2011). link
  15. [15] Cavoukian, A. (extended). Privacy by Design: the 7 principles (longer treatment). University of Waterloo course materials. link
  16. [16] —. Privacy by design (history, adoption 2009–2010). Wikipedia. link
  17. [17] Berners-Lee, T., et al. (Inrupt / MIT). Solid: personal online data pods; decoupling data from applications. Inrupt. link
  18. [18] Mansour, E., et al.. A Demonstration of the Solid Platform for Social Web Applications. Yale (Tobin) copy. link
  19. [19] —. Solid (web decentralization project): history, W3C standardization, stewardship. Wikipedia. link
  20. [20] Weiser, M.. The Computer for the 21st Century. Scientific American, 265(3), 94–104 (1991). link
  21. [21] Weiser, M., & Brown, J. S.. Designing Calm Technology (1995); calm technology overview. Wikipedia — Calm technology. link
  22. [22] Case, A.. Principles of Calm Technology. principles.design. link
  23. [23] Adobe (design). Calm technology in the era of experience design. Adobe Blog (2016). link
  24. [24] Horvitz, E.. Principles of Mixed-Initiative User Interfaces. CHI 1999. link
  25. [25] Microsoft Design. UX design for agents (status visibility, background operation, user control). microsoft.design. link
  26. [26] UX practitioners. Secrets of agentic UX: emerging design patterns for human–agent interaction. UX Magazine. link
  27. [27] Meric, G. / practitioners. Designing for AI agents: UX principles for autonomous systems. gokhanmeric.com. link
  28. [28] UXmatters. Designing for autonomy: UX principles for agentic AI. UXmatters (2025). link
  29. [29] Scholar (critical HCI). On False Augmented Agency, Surveillance Capitalism, and User-Centered Design. ResearchGate (338226880). link
  30. [30] Human-centered design (practitioner). Tech's inevitable journey towards human-centered design. Medium — Beacon Trust Network. link
Open research

The verified research corpus.

The individual, source-backed findings this paper is built on - surfaced by a multi-agent research fan-out and adversarially verified, then published in full so other researchers can check every claim against its citation.

The graded circle of trust (Dunbar's layered social structure)

  • Human social groups form a discrete hierarchy of preferred sizes — roughly 5, 15, 50, 150 (extending to ~500 and ~1500) — a geometric series with a scaling ratio close to 3 (empirically ~3.5) between successive layers.1,3
  • The layers correspond to identifiable groupings: a support clique (~5), a sympathy group (~15), a band (~50), and the active network / cognitive group (~150).1,3
  • There is a cognitive limit of ~150 meaningful relationships (Dunbar's number), grounded in neocortex volume via the social brain hypothesis.2,5
  • Contact frequency is graded by layer: the inner clique at least weekly, the sympathy group at least monthly, the full active network at least yearly.3
  • Mobile-phone call data provides strong empirical evidence for the layered structure; support is strongest at the innermost and outermost layers, while intermediate layers show real variability — an honest limitation on the precision of the 5/15/50/150 scheme.4
  • The layer a relationship belongs to is an informative, structurally distinct variable, not a flat 'friend' label; circle-aware (ego-network-layer) methods can outperform node2vec and SEAL without added computational complexity.4,10

Personal information, owned by the person (PIM & personal knowledge graphs)

  • Personal Information Management (Jones) is the set of activities by which a person acquires, stores, organizes, maintains, retrieves, uses, and distributes information to meet life's goals and roles — a person-centered, not web-index-centered, view.6
  • Information is 'personal' in six distinct senses — owned by me, about me, directed to me, sent/posted by me, experienced by me, and relevant to me — a citable typology for an ontology centered on one human.7
  • The personal space of information is constructed by the individual, spans physical and digital forms, is owned individually yet shared with others, and is constantly reshaped — supporting a consent-first, user-owned model rather than a static external index.7
  • A personal knowledge graph lets an individual consolidate fragmented personal data under full control; its RDF-based model includes explicit properties for access rights and provenance, making consent/access-control and data-lineage first-class ontology elements.8
  • A person can be modeled as the central focal node of a 'star-shaped' knowledge graph whose facets radiate from the center; such person-centric graphs are robust to sparse data and outperform baseline classifiers on real prediction tasks.9

Ontology design: identity, relations, provenance

  • Provenance — a record of the entities and processes that produced or influenced a resource — is fundamental, delivering data-quality assurance, contextual understanding, entity reconciliation, and compliance through origin tracking.11
  • Ontologies should be multi-axial and polyhierarchical (independent classification axes, not a single tree), so entities inherit from multiple parents and axes can be added or removed without global refactoring.12
  • Entity identity can be formalized via key patterns (Key(X, [p1…pk])) enabling automated consistency checking; constraints (disjointness, exhaustiveness) are best enforced locally per axis and validated with tools (SHACL/ShEx).12
  • Ontologies can be expressed in RDF/OWL (classes as owl:Class, relations as owl:ObjectProperty) or property graphs; a semi-rigid ontology aids attribution via provenance when aggregating disparate data.11,12

Consent-first & data ownership (Privacy by Design, Solid)

  • Privacy by Design requires privacy as the automatic default, embedded into architecture from the outset (not bolted on), with privacy and functionality treated as positive-sum, and visibility/transparency subject to independent verification.14,16
  • The framework is user-centric (strong defaults, appropriate notice, empowering options) and mandates data minimization; it was developed by Ann Cavoukian, published in 2009, and adopted internationally by data-protection authorities in 2010.14,15,16
  • Solid (Tim Berners-Lee, MIT, 2016; W3C standardization from 2018) stores a person's data in personal 'pods' they control; apps must authenticate and request permission, and cannot extract data automatically — a concrete instantiation of consent-first ownership.17,18,19
  • Solid frames the status quo as a 'siloed data' problem and organizes data around the individual rather than institutions/applications — prior art for a user-owned personal-information ontology.17,19

The asynchronous 'give work, walk away' agent (mixed-initiative & agent UX)

  • Horvitz's mixed-initiative principles: weigh the costs/benefits of acting under uncertainty about the user's goals; model the user's attention and defer to a less distracting time; open a dialog when uncertain, weighed against the cost of bothering the user; and let users invoke, terminate, and refine automation.24
  • Agents should default to responding rather than interrupting, timing interventions to task urgency, because users have low tolerance for non-urgent interruptions — grounding a 'comes back when the job is done or it needs money/information' flow.26,28
  • Autonomy must be a dynamic, negotiated property: agents must never act without the possibility of intervention, and must expose explicit override/adjust mechanisms; delegation requires specifying goals, constraints, success criteria, and forbidden actions.27,28
  • Microsoft's agent UX principles require an agent's status to be visible at all times (even in the background), and the user to remain in control (settings, preferences, on/off, transparent tools and connections).25
  • Agentic AI inverts the UX model: the user delegates work and cedes continuous control; trust comes from the felt ability to intervene, and the primary affordance shifts from clickable controls to explanation of the system's state and reasoning.26

Calm, human-first interfaces (and the critique of 'free' search)

  • Calm technology (Weiser & Brown, 1995) informs without demanding attention, operating mainly in the periphery and shifting to the center only when needed — directly grounding an interface that stays quiet until an agent needs money or information.21,22
  • Weiser's thesis: the best technologies become invisible, receding into the background; technology should be built around the human environment rather than forcing the human to adapt to the machine — the same 'work backwards from the human' inversion.20
  • The right amount of technology is the minimum needed to solve the problem — a minimalism principle grounding a homepage radically simpler than a feature-heavy search engine.22,23
  • AI products marketed as 'free' function as hooks to extract personal data, fueling surveillance capitalism; their design produces 'false augmented agency' — an illusion of capability — and user-centered design inside such business models constrains real autonomy while manufacturing an appearance of agency. 'User-centered' is therefore insufficient; a stronger human-first, consent-first standard is required.29

Verified against the cited sources and cross-checked; where a claim could not be verified it was excluded. Corrections and challenges are welcome - reach us via contact.