Source-linked AI summary
The Interface Is Downstream: Designing the Terms of Human-Agent Collaboration
Hector Ouilhet Olmos
TL;DR
Alicia's pilot exposed a provenance failure: a retrieved synthesis supplied an interpretation that the visible citation chain obscured. The paper audits attention, evidence, action, and learning in a persistent agent, but the pilot establishes no model advantage and cannot estimate performance.
Problem
Alicia's fluent response obscured the retrieved synthesis that supplied its interpretation, leaving the cited source note to appear as direct support.
Source: §1 · pp. 1–2Method
The paper audits Alicia's existing prompts and retrieval, evidence paths, permissions, and preference exports through four case-derived dimensions.
Source: §3 · p. 3Results
The pilot's apparent model comparison was confounded, and a matched control erased the apparent model effect while the provenance failure remained.
Source: §4 · p. 5Takeaways & Limitations
Humorphic design should make the conditions that produced an encounter and the person's recourse reviewable before and after it.
Source: §6 · p. 7
Abstract
from arXiv · showhide
Before an agent responds or acts, much of the experience has already been designed. Memory and retrieval shape what it notices. Evidence rules shape what it may claim. Permissions shape what it can do. Learning rules shape what it carries into the next encounter. The argument comes from Alicia, a personal agent I've built and used since January 2026. A fine-tuning pilot produced no defensible model-performance result. It exposed a provenance failure: Alicia repeated an interpretation from a retrieved synthesis, cited a source note credited by that synthesis, and left the synthesis out of the visible chain. In a model-blind review of thirty-three citations, one reviewer judged that the retrieved intermediary supplied the claim in sixteen relayed citations and part of it in four. Five of twelve citations to directly retrieved targets lacked support in the target excerpt supplied for review. These judgments remain unadjudicated, and the packet is not public. I call the shared setting a humorphic environment: a persistent computational setting that translates a human practice into software. The first Humorphism paper translated partnership. This paper translates the studio, the room where practice happens. The failure prompted an audit of attention, evidence, action, and learning, with a review artifact and available recourse for each. The test is whether the person can inspect and contest what shaped the teammate's behavior. The output is downstream. Correction, consent, and learning carry the collaborative interface back upstream.
1. The terms are set upstream
A provenance failure showed that a fluent response can conceal which retrieved document supplied its interpretation. The paper argues that attention, evidence, action, and learning are shaped upstream, while correction and other recourse carry the interface back upstream.
1. The terms are set upstream
- Alicia's Spanish response cited notes on Damasio and Rubin that existed in the vault but had not been retrieved for the prompt.The response translated claims from a synthesis, retaining two links while omitting the synthesis from the visible chain.Source: §1 · p. 1
- The retrieved synthesis supplied the interpretation, but Alicia cited a credited book note that contained no support for the specific claim.The note held book metadata and a broad bridge, not a passage about courage or somatic-marker patients.Source: §1 · p. 2
- The failure became visible only by tracing the claim backward through retrieval and authorship, despite a fluent response and a real citation target.By the time the response reached the interface, the consequential choices had already happened.Source: §1 · p. 2
- Attention shapes what the teammate notices; evidence, action boundaries, and learning shape what it may claim, do, and carry forward.These upstream decisions precede the output surface.Source: §1 · p. 2
- The studio models collaboration as shared work in which another person's attempts can reveal what one has not noticed.The art-class example links others' attempts to evidence for judging one's own drawing.Source: §1 · p. 2
2. What this adds
The paper extends Humorphism from translating partnership to translating the environment in which practice happens. It proposes reviewing whether a human practice's relational obligations survive in software, including the person's recourse when they do not.
2. What this adds
- The paper adds a test to Humorphism: whether obligations of the human practice survived their translation into software.It builds on a prior shift from isolated UX interactions toward the conditions where trust accumulates.Source: §2 · p. 2
- A humorphic environment names a human practice and its relational obligations, the software behavior carrying each obligation, and the person's recourse if translation fails.The paper applies these obligations together to a persistent personal agent, examining retrieval, memory, tool authority, and feedback.Source: §2 · pp. 2–3
- The review is one relationship environment derived from a named human practice, rather than a general account of all teamwork or agent systems.Source: §2 · p. 3
- Directability concerns persistent state transitions: whether a person can inspect, reverse, or contest an interaction's admission to memory or training.The argument draws on autobiographical design inquiry.Source: §2 · p. 3
- The account is autobiographical, with longitudinal access but interpretation concentrated in Alicia's builder, participant, evaluator, and category maker.No external researcher derived the four dimensions; one reviewer coded the citation chains without model identities or the author's prior classifications.Source: §2 · p. 3
3. Four case-derived dimensions
The audit applies four overlapping lenses—attention, evidence, action, and learning—to existing parts of Alicia, while distinguishing participant recourse from builder or model-owner powers. Its dimensions describe this case; other systems may require additional dimensions.
3. Four case-derived dimensions
- Alicia's pilot produced no consequential action, so the author treated it as the start of a broader audit rather than a complete taxonomy.Source: §3 · p. 3
- The four lenses audit existing prompts and retrieval, evidence paths, tool permissions, and preference exports; other systems may need additional dimensions.Goal alignment, common ground, coordination, identity, and recovery are named as possible additions.Source: §3 · p. 3
- Attention is shaped by prompts, retrieval, and weights, while an attention catalogue is proposed to record each pattern's location, owner, and refusal boundary.Prompt-level rules can be opened and changed directly; weight-level patterns require inference from behavior and changes through data or a model release.Source: §3 · pp. 4–5
- An evidence contract preserves both the retrieved synthesis and its credited source, giving the person an action when the chain is wrong.The claim's supplier can differ from the source note the synthesis credited.Source: §3 · pp. 4–5
- The learning gate admits explicit blind A/B wins while excluding ties or neither choices, but revision and removal controls remain unimplemented.Exports omit reviewer, admission time, exclusion reason, and corpus version; cumulative exports also require trial-ID deduplication.Source: §3 · pp. 4–5
- A proposed control does not meet the acceptance criterion until participants can use it, unlike controls held by builders or model owners.Source: §3 · p. 4
- A weekly note redirected a writing decision, but the author cannot tell whether Alicia sharpened judgment or replaced part of it.The note described strong depth but weak peripheral vision, including missed movement while serving stated concerns.Source: §3 · p. 4
- Action boundaries distinguish knowing how to perform an action from having authority to perform it; Alicia requires confirmation before sending a drafted email.A consequence map covering reversibility, reach, authorization, recovery, and who bears error costs across tools is not yet maintained.Source: §3 · p. 5
4. An evidence contract in practice
The pilot’s apparent model effect disappeared under a matched control, while the provenance failure remained. Review traced how claims moved through retrieved intermediaries and whether credited excerpts supported them, motivating a more contestable evidence contract.
4. An evidence contract in practice
- The matched control erased the apparent model effect, while the provenance failure survived.The first comparison was confounded by different runtimes and quantization paths.Source: §4 · p. 5
- Sixteen relayed citations were judged intermediary-supplied and four partly supplied; five of twelve direct citations lacked support in the supplied target excerpt.Across twenty-eight inspected target excerpts, eight had full support, eight partial support, and twelve none.Source: §4 · p. 5
- The reviewer labeled twenty citations Relayed, twelve Direct, and one Unclear; all twenty relayed labels matched automated path candidates, but this was not independent reproduction.The public-packet gate was unmet, and the shortened-title judgment remained open.Source: §4 · p. 6
- Among twenty relayed citations, credited targets had full support in two, partial support in seven, none in six, and were uninspected in five; the one unclear target had no support.Among twelve direct citations, six had full support, one partial support, and five none; judgments concern supplied excerpts, not exhaustive reviews.Source: §4 · p. 6
- For the courage claim, the credited note lacked support for the specific interpretation, and the available recourse was manual synthesis editing; a dedicated invalidation control was unimplemented.The retrieved synthesis supplied the interpretation while the response cited the note it credited.Source: §4 · p. 6
- The evidence contract records claim supplier, credited source, intermediary authorship, retrieval status, hop count, and available correction, making the interpretation chain reviewable.The proposed future review records the claim supplier for every citation, including directly retrieved targets.Source: §4 · pp. 6–7
5. Ownership, limits, and the loop
Humorphic design must connect familiar technical mechanisms to the person’s ability to inspect, challenge, and change the conditions behind an agent’s behavior. Alicia remains a limited single-person case, and the studio translation is incomplete.
5. Ownership, limits, and the loop
- Engineering can verify enforced permissions, provenance records, and dataset versions, while UX must examine their effects on initiative, consent, recovery, and whether people can challenge interpretations.The design also asks why an interaction was allowed to change the teammate.Source: §5 · p. 7
- Alicia is a single-person case built from a dense private archive, and the pilot establishes no model advantage or performance estimate.Its prompts exposed a broken comparison and a provenance failure.Source: §5 · p. 7
- Alicia has one clearly participant-facing recourse—email confirmation—while attention and evidence mainly depend on builder or file-owner access.Learning offers a participant-facing choice without later revision or removal; role overlap made the system appear more contestable than it is for an ordinary participant.Source: §5 · p. 7
- The studio translation lacks a durable field of neighboring work through which another person’s attempts can train perception.The case also does not establish improved human judgment or creativity; the weekly scene permits either sharpened perception or an adopted agent interpretation.Source: §5 · p. 7
- Distinguishing strengthened from displaced judgment requires longitudinal evidence, free-text reasons, independent coding, and tasks that reveal either outcome.Inspection and contest need a human-facing surface; correction, consent, and refusal make the relationship a loop.Source: §5 · p. 7
6. Conclusion
Humorphism extends from translating partnership into software to translating the environment where practice happens. The design aim is to make the relationship’s terms reviewable before and after an exchange.
6. Conclusion
- Extending Humorphism from partnership to the environment would make the terms of that partnership reviewable before and after the visible exchange.The environment translates the studio, the room where practice happens.Source: §6 · p. 7
- The conclusion shifts the design focus from the visible exchange to the conditions shaping it and the person’s ability to review them.This is the paper’s proposed extension of partnership-centered Humorphism to the environment.Source: §6 · p. 7
- Humorphic design owns both the conditions that produced an encounter and the recourse available to the person inside it.The interface is downstream; the relationship is a loop.Source: §6 · p. 7
Appendix A. Pilot method and controls
The pilot’s small comparison was not able to establish a model-performance result. Matched controls removed the apparent citation difference, while audits exposed data-quality and review-protocol limits.
Appendix A. Pilot method and controls
- The pilot fine-tuned official Qwen 3.5 9B with rank-32 LoRA for eighteen steps on 164 exact-text vault examples, using 148 for training and 16 for validation.“Exact vault text” describes selection, not human authorship.Source: Appendix A · p. 8
- The three-prompt comparison could not isolate the adapter because runtime, quantization, artifact size, and sampling changed together.Stock Qwen completed seven citations, all to retrieved documents; the tuned arm completed seven and directly grounded two.Source: Appendix A · p. 8
- Under matched MLX settings, the apparent citation difference disappeared, and counts shifted sharply under the second seed; the prompts support no performance claim.The control held architecture, tokenizer, chat template, sampler, output limit, and retrieved context identical, with matched hashes and stable case-derived seeds in the second setting.Source: Appendix A · p. 8
- A malformed agent-generated vault note entered retrieval, training, the sole contradiction validation example, and the first audit; seven of eleven contradiction analyses were malformed.Six malformed analyses entered training.Source: Appendix A · p. 8
- The model-blind reviewer labeled twenty citations Relayed, twelve Direct, and one Unclear, but the comparison was with an automated path audit, not inter-rater agreement.Because the author had not locked a complete item-level sheet before receiving the reviewer’s, no agreement percentage or kappa was reported; the unclear title remained unadjudicated.Source: Appendix A · p. 8
Appendix B. Data availability and publication boundary
The publication boundary excludes private source material and leaves the audit aggregate unreproducible from public materials. The paper specifies what it can publish and what would require privacy clearance.
Appendix B. Data availability and publication boundary
- The paper names only the author's notes, Alicia-generated syntheses, and published works; it excludes identifiable private third-party material.The private vault and training corpus contain personal and third-party material and cannot be released.Source: Appendix B · p. 8
- Audit code and raw run records remain local; the reviewer returned an unadjudicated coding sheet, and the missing locked author sheet prevents full agreement calculation.The reviewer received a thirty-three-item blinded packet; one unclear item remains unadjudicated.Source: Appendix B · p. 8
- No public companion package accompanies the preprint, and the aggregate cannot be reproduced publicly; releasing a redacted packet requires privacy clearance.Clearance would cover prompts, run records, code, and citation excerpts.Source: Appendix B · p. 9
- Technical model and runtime names remain in the academic preprint to interpret the control; an editorial version should generalize them and link to the cleared preprint.Source: Appendix B · p. 9