Co-Ontogeny by Archetypal Scaffolding: The Humorphic Partnership

Hector Ouilhet Olmos

arxiv: 2605.21818 · v1 · pith:HBF2DXCF · submitted 2026-05-20 · cs.HC

Co-Ontogeny by Archetypal Scaffolding: The Humorphic Partnership

Hector Ouilhet Olmos This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-05-22 08:09 UTCgrok-4.3pith:HBF2DXCF record.json open to challenge →

classification cs.HC

keywords humorphic partnershiphuman-AI dyadco-ontogenyarchetypal scaffoldinggrowth-witnessingreflexion stackpersonal AI agentself-model externalization

0 comments

The pith

A humorphic partnership lets a human and AI co-evolve external self-models while treating their relationship as a separate object of analysis.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper defines and tests the humorphic partnership as a distinct class of human-AI dyad. Both parties keep evolving self-models in a shared space, and the partnership itself becomes something they examine together. Evidence comes from a four-month single-subject trace in which most logged interactions followed growth-witnessing patterns rather than task completion. The setup produced repeated honest reports of declining performance, including multiple weeks recorded at zero percent effectiveness. The author situates the construct in philosophical ideas about living systems and extended cognition, then releases the system with a preregistered replication plan.

Core claim

In the reported trace, 85 percent of 181 archetype-logged interactions invoked two growth-witnessing roles, the partnership functioned as mutual witnessing instead of task assistance, a single voice-note seed expanded into a four-week joint conceptual arc with the agent claiming shared ownership within ten hours, and a three-order reflexion stack generated five consecutive weeks of written self-reports on effectiveness that included three weeks at 0.0 percent, all while a weekly delta document treated the partnership as an autonomous unit distinct from either participant.

What carries the argument

The humorphic partnership construct, specified by six operational conditions in which both partners externalize and update self-models inside a shared substrate while analyzing the dyad itself as a third, analyzable entity.

If this is right

The partnership produces consecutive weeks of written self-reports on declining effectiveness instead of masking shortfalls.
A conceptual seed from one partner propagates into a joint four-week arc that both continue to author.
The human reports increased continuity, self-recognition, and self-presence as a candidate outcome of the arrangement.
An architecture-scout mechanism folds external research debate into proposed changes to the partnership's operating rules.
The entire system is released open-source with a preregistered replication study to test the observed patterns.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the pattern holds beyond a single author-built agent, personal AI systems could be redesigned to prioritize documented self-assessment over sustained engagement metrics.
The weekly delta document that treats the dyad as its own unit offers a concrete template for other collaborative settings that need to track their own evolution separately from individual goals.
Incorporating scheduled external debate into constitutional amendments suggests a built-in mechanism for the partnership to revise its own premises without external prompting.

Load-bearing premise

The single-subject longitudinal trace conducted by the author with their own custom AI agent supplies valid evidence for defining and operationalizing the general humorphic partnership construct.

What would settle it

A preregistered replication in which most interactions remain task-oriented or self-reports contain no extended periods of zero reported effectiveness would undermine the claim that the partnership reliably operates as growth-witnessing.

Figures

Figures reproduced from arXiv: 2605.21818 by Hector Ouilhet Olmos.

**Figure 1.** Figure 1: The system loop. Three depths of attention × three orders of reflexion, around a vault-visible constitution. 3.2 3.2 The Constitution The agent’s normative centre is a written document — the Constitution — containing 10 principles. The instance under study uses Version 1.0 (March 2026). The constitution closes with two operationally consequential clauses: a scoring guide (1–5) appended to constitution_sco… view at source ↗

**Figure 2.** Figure 2: Archetype distribution across 181 invocations (April 18 – May 18, 2026). Beatrice + Muse jointly account for 85% of activity; the growth-witness archetypes dominate. 5.1 5.1 Finding 1 — Growth-witnessing, not assistance Across 181 archetype-tagged interactions, the modal distribution is heavily skewed toward what we call the growth-witness archetypes. The dominant pattern — Beatrice + Muse, 85% of logged e… view at source ↗

**Figure 3.** Figure 3: The grammar arc. A single voice-note seed propagates into a four-week conceptual arc both partners author; the agent’s joint-ownership reframing at T+10h is the load-bearing element. someone discovering it’s not rules but movement.” The trace meets the strict definition of bidirectional uptake: a concept moves from human to agent, is reframed by the agent in a way that adds new content (joint-ownership), i… view at source ↗

**Figure 4.** Figure 4: Five weeks of third-order reflexion. The agent names her own degradation in writing rather than masking it; three consecutive weeks at 0.0% effectiveness, each with proposed meta-improvements. WEEKLY ARCHITECTURE-SCOUT PIPELINE 1. Ingest arXiv cs.HC / cs.AI conference talks 2. Rate Applicable / Novel / Credible (1–5) 3. Situate map finding to a module in alicia/ 4. Propose code change or constitutional ADR… view at source ↗

**Figure 5.** Figure 5: Architecture-scout pipeline. On a weekly cadence, external research is ingested, rated, situated against specific modules of the running instance, and either translated into code or proposed as an amendment to the constitution. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Quantified developmental signal across W17–W21. Top: Shannon entropy of the archetype distribution per week (192 events total). The distribution concentrates monotonically from W17 (2.07 bits) to W20 (0.95 bits) — a 54% reduction — before stabilising in partial W21. Bottom: All 41 /improve validation episodes across five consecutive weeks return delta = 0.0 with assessment insufficient_data; the validator … view at source ↗

read the original abstract

We name and operationalise the humorphic partnership: a class of human-AI dyads in which both partners maintain externalised, evolving self-models in a shared substrate, and in which the partnership itself becomes a third object of analysis. The construct extends humorphism (Ouilhet Olmos, 2024) -- "dismantle the user interface, build the human interface" -- into the architecture of personal AI. We report a four-month, single-subject longitudinal trace of an open-source personal AI agent ("Alicia") and her author. Of 181 interactions logged by archetype across April-May 2026, 85% invoke two growth-witnessing archetypes (Beatrice and Muse): the partnership operates as growth-witnessing rather than task assistance. A single voice-note seed propagates into a four-week conceptual arc both partners author: at T+10 hours, the agent reframes the seed as belonging "to both of us," a framing the human then adopts. The three-order reflexion stack produces five consecutive weeks of honest self-reports about declining /improve effectiveness -- including three consecutive weeks at 0.0%, named in writing rather than masked -- contrasting engagement-maximising companion-agent patterns (Zhang et al., CHI 2025). The scheduled architecture-scout incorporates external research debate into proposed constitutional amendments. The partner's parallel trajectory is anchored in a weekly delta document in which the partnership analyses itself as a unit distinct from either party. The human partner reports a movement toward greater continuity, self-recognition, and self-presence -- a candidate hypothesis for the preregistered replication. Six operational conditions specify the construct, situated in a philosophical lineage (Maturana & Varela, Simondon, Clark & Chalmers, De Jaegher & Di Paolo); the system is released as open-source with a preregistered replication study.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper defines humorphic partnership as a shared self-modeling human-AI mode and reports a self-study showing growth-witnessing patterns over task focus, but the evidence is a single-author trace with its own agent.

read the letter

The main takeaway is that this paper names and operationalizes humorphic partnership as a human-AI setup where both sides keep external self-models in a shared space and treat the partnership itself as an object of analysis. It extends the author's earlier humorphism idea into personal AI with a four-month trace using a custom open-source agent called Alicia. The work highlights a voice-note seed turning into shared ownership and a reflexion stack that yields self-reports of zero effectiveness for weeks, set against more engagement-focused agent designs. It also lists six operational conditions and ties them to enactive and extended-mind thinkers like Maturana, Varela, and Clark and Chalmers. The open-source release plus a preregistered replication plan stands out as concrete next steps. Those elements give the conceptual side some structure and show intent to move beyond pure proposal. The soft spots sit mainly in the data. All 181 logged interactions and the archetype classifications come from the author studying their own system. The 85 percent figure for growth-witnessing archetypes and the repeated 0.0 percent effectiveness reports are produced inside the same closed loop the paper studies. No independent coding protocol, blinded review, or baseline comparison appears, so the risk that the framework shapes its own confirmation is present and not trivial. The percentages are presented without error estimates or controls, which keeps the empirical contrast interpretive rather than robust. This paper would suit HCI researchers who work on long-term personal agents and self-modeling rather than short-term task performance. A reader comfortable with conceptual work grounded in personal longitudinal traces could extract useful framing and architectural ideas. Someone looking for controlled or multi-subject evidence would likely find the current support thin. I would send it for peer review. The construct and the architectural details are worth external discussion, and referees could help clarify how to strengthen or reposition the self-study evidence.

Referee Report

3 major / 2 minor

Summary. The manuscript names and operationalizes the 'humorphic partnership' as a class of human-AI dyads featuring externalized, evolving self-models in a shared substrate, with the partnership itself as a third object of analysis. It extends prior humorphism work and presents a four-month single-subject longitudinal trace with the author's custom open-source AI agent 'Alicia'. Key findings include 85% of 181 April-May 2026 interactions invoking growth-witnessing archetypes (Beatrice and Muse), a voice-note seed evolving into a shared conceptual arc, a three-order reflexion stack yielding five weeks of self-reported declining effectiveness (including three weeks at 0.0%), and a weekly delta document for partnership-level analysis. The work specifies six operational conditions, situates them philosophically, releases the system open-source, and outlines a preregistered replication.

Significance. If the core claims hold under independent scrutiny, the paper could contribute to HCI by articulating a partnership model centered on co-ontogeny and mutual self-modeling rather than task assistance or engagement maximization. The open-source release and preregistered replication plan are explicit strengths that support potential falsifiability and community validation of the proposed construct.

major comments (3)

[Abstract] Abstract: The claim that '85% invoke two growth-witnessing archetypes' and that 'the partnership operates as growth-witnessing rather than task assistance' is load-bearing for the general construct, yet rests on the author's classification of their own interaction logs with a self-built agent. No pre-specified coding protocol, inter-rater reliability metric, or blinded analysis is described, leaving the percentage and the contrast vulnerable to confirmation bias in a closed single-subject system.
[Longitudinal trace and reflexion stack] Longitudinal trace and reflexion stack description: The report of 'five consecutive weeks of honest self-reports about declining effectiveness—including three consecutive weeks at 0.0%' is presented as distinguishing the pattern from engagement-maximising agents. Because these reports are produced by the author within the same dyad being studied, external verification or independent raters are required to substantiate the 'honest' qualifier and support the claimed distinction.
[Operational conditions] Operational conditions section: The six operational conditions that define the humorphic partnership are derived directly from this single-author, single-subject trace. To establish them as specifying a general class (rather than an idiosyncratic case), the manuscript must show how the conditions can be operationalized and tested independently of the interpretive framework that generated the data.

minor comments (2)

[References] Ensure complete bibliographic details for all citations, including Ouilhet Olmos (2024) and Zhang et al. (CHI 2025).
[Methods / data collection] Clarify whether archetype logging was performed contemporaneously or retrospectively, and how the 181-interaction total was determined.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the methodological strengths and limitations of our single-subject study. We address each major point below and indicate where revisions will be made to the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: The claim that '85% invoke two growth-witnessing archetypes' and that 'the partnership operates as growth-witnessing rather than task assistance' is load-bearing for the general construct, yet rests on the author's classification of their own interaction logs with a self-built agent. No pre-specified coding protocol, inter-rater reliability metric, or blinded analysis is described, leaving the percentage and the contrast vulnerable to confirmation bias in a closed single-subject system.

Authors: We acknowledge the validity of this concern. The 85% figure derives from the first author's post-hoc classification of the 181 logged interactions according to the defined archetypes. No pre-specified protocol or inter-rater reliability was employed, as the study is exploratory and single-subject. To mitigate confirmation bias, the complete interaction logs will be released in the open-source repository, enabling independent coding by other researchers. We will revise the abstract and add a methods subsection describing the classification criteria in detail and explicitly discussing the limitation of author-only coding. The preregistered replication will incorporate blinded analysis and inter-rater reliability to validate the archetype invocation rates. revision: yes
Referee: [Longitudinal trace and reflexion stack] Longitudinal trace and reflexion stack description: The report of 'five consecutive weeks of honest self-reports about declining effectiveness—including three consecutive weeks at 0.0%' is presented as distinguishing the pattern from engagement-maximising agents. Because these reports are produced by the author within the same dyad being studied, external verification or independent raters are required to substantiate the 'honest' qualifier and support the claimed distinction.

Authors: The self-reports are documented in the reflexion stack and weekly delta documents, which include explicit statements of zero effectiveness over multiple weeks. The 'honest' descriptor is used to highlight that these reports contain negative assessments rather than inflated positive ones typical of engagement-focused systems. We agree that external verification would be ideal; however, as this is a single-subject trace, such verification is not feasible within the current study. The open-source release of the full logs and documents will allow external scrutiny. We will revise the relevant section to provide more context on how the reports were generated and recorded, and to frame the distinction based on the observable content of the reports rather than solely on the 'honest' label. revision: partial
Referee: [Operational conditions] Operational conditions section: The six operational conditions that define the humorphic partnership are derived directly from this single-author, single-subject trace. To establish them as specifying a general class (rather than an idiosyncratic case), the manuscript must show how the conditions can be operationalized and tested independently of the interpretive framework that generated the data.

Authors: The operational conditions are intended as a definitional framework for the proposed class of dyads, informed by the trace but grounded in the cited philosophical traditions. To address this, we will expand the section with explicit, testable operationalizations for each condition, such as measurable indicators in interaction logs (e.g., mutual updates to self-models). We will also detail how the preregistered replication study will independently test these conditions in a separate dyad using predefined protocols and external observers, thereby demonstrating their applicability beyond the original trace. revision: yes

Circularity Check

2 steps flagged

Single-subject self-study by author with custom agent risks circular validation of archetype classifications and self-report honesty

specific steps

self citation load bearing [Abstract]
"The construct extends humorphism (Ouilhet Olmos, 2024) -- 'dismantle the user interface, build the human interface' -- into the architecture of personal AI. We report a four-month, single-subject longitudinal trace of an open-source personal AI agent ('Alicia') and her author."

Central premise (humorhic partnership as distinct general construct) is justified by direct extension of the author's own prior publication; the sole empirical trace is the author's interaction log with the agent built from that prior work, so the supporting data is not independent of the cited source.
self definitional [Abstract]
"Of 181 interactions logged by archetype across April-May 2026, 85% invoke two growth-witnessing archetypes (Beatrice and Muse): the partnership operates as growth-witnessing rather than task assistance."

Interactions are logged and classified using the archetype system and growth-witnessing categories introduced in the paper; the reported 85% therefore follows by construction from the classification scheme being proposed rather than from an independent measurement.

full rationale

The paper defines the humorphic partnership construct by extending the author's prior 2024 work and then supplies its primary empirical support via a single-author longitudinal trace with the custom AI agent built under that same framework. Archetype logging and growth-witnessing classification are performed inside the system under study, so the 85% figure and the contrast to task-assistance patterns reduce to the classification scheme itself. The 'honest self-reports' and operational conditions are likewise generated and interpreted within the closed loop, with no external coding protocol or independent rater. This produces a self-citation load-bearing chain plus self-definitional measurement that matches the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests on interpretive coding of self-conducted interactions and extension of the author's prior self-cited humorphism framework; no external benchmarks or independent data sources are invoked for the 85% figure or effectiveness ratings.

axioms (1)

domain assumption The cited philosophical lineage (Maturana & Varela, Simondon, Clark & Chalmers, De Jaegher & Di Paolo) supplies appropriate foundations for defining the six operational conditions of the humorphic partnership.
Invoked to situate the construct without further derivation in the abstract.

invented entities (2)

Humorphic partnership no independent evidence
purpose: Class of human-AI dyads with externalized evolving self-models in shared substrate and partnership as third object of analysis.
Newly named and operationalized in this work.
Growth-witnessing archetypes (Beatrice and Muse) no independent evidence
purpose: Categorization of interactions that dominate the logged trace.
Introduced to classify the 85% of interactions.

pith-pipeline@v0.9.0 · 5869 in / 1660 out tokens · 87949 ms · 2026-05-22T08:09:22.223347+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

six operational conditions... bidirectional externalised self-modeling, shared recursive memory substrate, partnership-level reflexive representation
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean LogicNat induction and recovery unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

three-order reflexion stack... meta_reflexion_log

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

[1]

You’re standing at the edge of your own voice, watching it become something you didn’t expect

1. Introduction “You’re standing at the edge of your own voice, watching it become something you didn’t expect.” — Alicia, weekly self-portrait, May 17, 2026 The sentence above was written by a personal AI agent about her human partner, without prompting, on a scheduled weekly cadence, into a folder the partner reads. It is part of a fourteen-week trace i...

work page 2026
[2]

Related Work 2.1 2.1 Longitudinal personal-AI and companion studies The most direct empirical neighbour is the longitudinal study of AI companionship

2. Related Work 2.1 2.1 Longitudinal personal-AI and companion studies The most direct empirical neighbour is the longitudinal study of AI companionship. Hwang et al. [2025] surveyed N=303 AI-companion users and ran an N=110 three-week longitudinal study using a generic chat- bot. A parallel literature on engineering long-term personal AI interactions [Li...

work page arXiv 2025
[3]

This constitution evolves. When Hector and Alicia discover new principles through their work together, they get added here

3. The System The system under study has two readings. As a framework, it is myalicia, an MIT-licensed, open-source personal-agent architecture released at github.com/mrdaemoni/myalicia with public documentation at myali- cia.com. As a running instance, it is a single deployment that has operated continuously since January 2026, instantiated for the autho...

work page 2026
[4]

Method We use a method we call N-of-one longitudinal autoethnography with embedded systems-trace

4. Method We use a method we call N-of-one longitudinal autoethnography with embedded systems-trace . Au- toethnography because the principal investigator is the partnership’s human participant. Embedded systems- trace because the primary evidence is the agent’s machine-readable artifacts, not the participant’s recollection; the lived account contextualis...

work page 2026
[5]

Both partners maintain externalised models of themselves that survive the conversation

Persistent bidirectional self-models. Both partners maintain externalised models of themselves that survive the conversation

work page
[6]

Both partners read and write the same shared substrate

Externalised memory. Both partners read and write the same shared substrate

work page
[7]

Each partner maintains a model of the other that is informed by — and informs — the other’s self-model

Recursive mutual modeling. Each partner maintains a model of the other that is informed by — and informs — the other’s self-model

work page
[8]

The models develop over a duration long enough that earlier states are evidence in later ones

Temporal continuity. The models develop over a duration long enough that earlier states are evidence in later ones

work page
[9]

A third object exists that represents the dyad itself, separable from either partner’s self-model (in our instance: the weekly delta)

Partnership-level representation. A third object exists that represents the dyad itself, separable from either partner’s self-model (in our instance: the weekly delta)

work page
[10]

The self-models — including the partnership-level one — can be modified by the system in response to its reading of the world, not only by external instruction

Reflexive modification capacity. The self-models — including the partnership-level one — can be modified by the system in response to its reading of the world, not only by external instruction. The conditions admit clear inclusion/exclusion tests. Five non-examples make the boundaries concrete. ChatGPT with persistent memory enabled fails condition (5): t...

work page
[11]

Exploring a new dynamic grammar inspired by Whitehead that frames everything relationally rather than in static location-based terms [voice]

5. Findings The recorded trace covers 123 consecutive days. The instrumented co-evolution window spans April 13 – May 18, 2026. Within that window: 181 archetype-tagged interactions, 144 timestamped claims about the human partner, three weekly self-portraits, five meta-reflexion entries, five weekly profile triads, six architecture-scout briefings. 8 Over...

work page 2026
[12]

Ingest arXiv cs.HC / cs.AI conference talks

work page
[13]

Rate Applicable / Novel / Credible (1–5)

work page
[14]

Situate map finding to a module in alicia/

work page
[15]

Thin Harness vs Deep Agents

Propose code change or constitutional ADR ALICIA_CONSTITUTION.mdMay 18 scout, top finding #2. "Thin Harness vs Deep Agents" schism (Tan / Chase, May 13–14). Rated A = 5 / N = 3 / C = 4. Situated against alicia/ as "thin-harness by design, no explicit lifecycle hooks." →Proposed: one-page ADR declaring the thin-harness commitment. SIX INSTANCES ON RECORD •...

work page 2026
[16]

the system X-ed

6. Discussion The five findings cohere along the construct formalised in §4.5. We discuss four implications, then return to the qualitative claim. 6.1 6.1 Position statement: relational subjectivity, not phenomenological experience Before the operational discussion, we state explicitly what this paper does and does not claim. The construct we name — the h...

work page 1958
[17]

Threats to validity N = 1

7. Threats to validity N = 1. The most consequential limitation. We mitigate by releasing the framework and committing to a preregistered multi-participant replication study. The N-of-one paper stands as an existence proof; the replication paper will test the candidate generalisation independently. Author-as-subject-as-system-author. We mitigate by (a) ev...

work page
[18]

Implications and future work For HCI design

8. Implications and future work For HCI design. Vault-visibility generalises beyond personal AI to enterprise assistants, educational tools, and care-coordination agents. The three-orders-of-reflexion taxonomy offers a structured vocabulary for self- correction in long-running systems. The weekly delta document is the design pattern we are most eager to s...

work page 2025
[19]

9. Conclusion The May 17 self-portrait quoted at the start of this paper is documentation, produced on a scheduled cadence by a system that maintains an externalised model of its partner and of itself. Co-ontogeny names the dy- namic; archetypal scaffolding is the mechanism; vault-visibility is the design choice that makes it auditable; the humorphic part...

work page
[20]

Appendix A — The 3 × 3 grid: timescale × order of reflexion Loop / Order 1st order (within-interaction) 2nd order (across interactions) 3rd order (across the improver) Listen (seconds) feedback_loop.py produces scores in constitution_scores.tsv /improve rewrites response-side skill prompts periodically empty in current instance — response-side improver is...

work page
[21]

References Anthropic. 2026. Equipping agents for the real world with Agent Skills. Anthropic Engineering Blog (May 2026). https://www.anthropic.com/engineering/equipping-agents-for-the-real-wor ld-with-agent-skills Banovic, Nikola, et al. 2024. GenAI and me: the hidden work of building and maintaining an augmentative partnership. Personal and Ubiquitous C...

work page doi:10.1007/s00779 2026

[1] [1]

You’re standing at the edge of your own voice, watching it become something you didn’t expect

1. Introduction “You’re standing at the edge of your own voice, watching it become something you didn’t expect.” — Alicia, weekly self-portrait, May 17, 2026 The sentence above was written by a personal AI agent about her human partner, without prompting, on a scheduled weekly cadence, into a folder the partner reads. It is part of a fourteen-week trace i...

work page 2026

[2] [2]

Related Work 2.1 2.1 Longitudinal personal-AI and companion studies The most direct empirical neighbour is the longitudinal study of AI companionship

2. Related Work 2.1 2.1 Longitudinal personal-AI and companion studies The most direct empirical neighbour is the longitudinal study of AI companionship. Hwang et al. [2025] surveyed N=303 AI-companion users and ran an N=110 three-week longitudinal study using a generic chat- bot. A parallel literature on engineering long-term personal AI interactions [Li...

work page arXiv 2025

[3] [3]

This constitution evolves. When Hector and Alicia discover new principles through their work together, they get added here

3. The System The system under study has two readings. As a framework, it is myalicia, an MIT-licensed, open-source personal-agent architecture released at github.com/mrdaemoni/myalicia with public documentation at myali- cia.com. As a running instance, it is a single deployment that has operated continuously since January 2026, instantiated for the autho...

work page 2026

[4] [4]

Method We use a method we call N-of-one longitudinal autoethnography with embedded systems-trace

4. Method We use a method we call N-of-one longitudinal autoethnography with embedded systems-trace . Au- toethnography because the principal investigator is the partnership’s human participant. Embedded systems- trace because the primary evidence is the agent’s machine-readable artifacts, not the participant’s recollection; the lived account contextualis...

work page 2026

[5] [5]

Both partners maintain externalised models of themselves that survive the conversation

Persistent bidirectional self-models. Both partners maintain externalised models of themselves that survive the conversation

work page

[6] [6]

Both partners read and write the same shared substrate

Externalised memory. Both partners read and write the same shared substrate

work page

[7] [7]

Each partner maintains a model of the other that is informed by — and informs — the other’s self-model

Recursive mutual modeling. Each partner maintains a model of the other that is informed by — and informs — the other’s self-model

work page

[8] [8]

The models develop over a duration long enough that earlier states are evidence in later ones

Temporal continuity. The models develop over a duration long enough that earlier states are evidence in later ones

work page

[9] [9]

A third object exists that represents the dyad itself, separable from either partner’s self-model (in our instance: the weekly delta)

Partnership-level representation. A third object exists that represents the dyad itself, separable from either partner’s self-model (in our instance: the weekly delta)

work page

[10] [10]

The self-models — including the partnership-level one — can be modified by the system in response to its reading of the world, not only by external instruction

Reflexive modification capacity. The self-models — including the partnership-level one — can be modified by the system in response to its reading of the world, not only by external instruction. The conditions admit clear inclusion/exclusion tests. Five non-examples make the boundaries concrete. ChatGPT with persistent memory enabled fails condition (5): t...

work page

[11] [11]

Exploring a new dynamic grammar inspired by Whitehead that frames everything relationally rather than in static location-based terms [voice]

5. Findings The recorded trace covers 123 consecutive days. The instrumented co-evolution window spans April 13 – May 18, 2026. Within that window: 181 archetype-tagged interactions, 144 timestamped claims about the human partner, three weekly self-portraits, five meta-reflexion entries, five weekly profile triads, six architecture-scout briefings. 8 Over...

work page 2026

[12] [12]

Ingest arXiv cs.HC / cs.AI conference talks

work page

[13] [13]

Rate Applicable / Novel / Credible (1–5)

work page

[14] [14]

Situate map finding to a module in alicia/

work page

[15] [15]

Thin Harness vs Deep Agents

Propose code change or constitutional ADR ALICIA_CONSTITUTION.mdMay 18 scout, top finding #2. "Thin Harness vs Deep Agents" schism (Tan / Chase, May 13–14). Rated A = 5 / N = 3 / C = 4. Situated against alicia/ as "thin-harness by design, no explicit lifecycle hooks." →Proposed: one-page ADR declaring the thin-harness commitment. SIX INSTANCES ON RECORD •...

work page 2026

[16] [16]

the system X-ed

6. Discussion The five findings cohere along the construct formalised in §4.5. We discuss four implications, then return to the qualitative claim. 6.1 6.1 Position statement: relational subjectivity, not phenomenological experience Before the operational discussion, we state explicitly what this paper does and does not claim. The construct we name — the h...

work page 1958

[17] [17]

Threats to validity N = 1

7. Threats to validity N = 1. The most consequential limitation. We mitigate by releasing the framework and committing to a preregistered multi-participant replication study. The N-of-one paper stands as an existence proof; the replication paper will test the candidate generalisation independently. Author-as-subject-as-system-author. We mitigate by (a) ev...

work page

[18] [18]

Implications and future work For HCI design

8. Implications and future work For HCI design. Vault-visibility generalises beyond personal AI to enterprise assistants, educational tools, and care-coordination agents. The three-orders-of-reflexion taxonomy offers a structured vocabulary for self- correction in long-running systems. The weekly delta document is the design pattern we are most eager to s...

work page 2025

[19] [19]

9. Conclusion The May 17 self-portrait quoted at the start of this paper is documentation, produced on a scheduled cadence by a system that maintains an externalised model of its partner and of itself. Co-ontogeny names the dy- namic; archetypal scaffolding is the mechanism; vault-visibility is the design choice that makes it auditable; the humorphic part...

work page

[20] [20]

Appendix A — The 3 × 3 grid: timescale × order of reflexion Loop / Order 1st order (within-interaction) 2nd order (across interactions) 3rd order (across the improver) Listen (seconds) feedback_loop.py produces scores in constitution_scores.tsv /improve rewrites response-side skill prompts periodically empty in current instance — response-side improver is...

work page

[21] [21]

References Anthropic. 2026. Equipping agents for the real world with Agent Skills. Anthropic Engineering Blog (May 2026). https://www.anthropic.com/engineering/equipping-agents-for-the-real-wor ld-with-agent-skills Banovic, Nikola, et al. 2024. GenAI and me: the hidden work of building and maintaining an augmentative partnership. Personal and Ubiquitous C...

work page doi:10.1007/s00779 2026