What if the environment surrounding an AI agent contained art, field notes, strange observations, unfinished thoughts? Not directives. Not behavioral targets. Just something to encounter.
Choose a shelf fragment. These miniature passages are original creative interpretations of a thinker's interests, not quotations or behavioral instructions. Their explanations live here in the gallery, never in the agent treatment.
The specificity of ordinary human experience; the possibility that a scene matters without being useful.
Not a polished founder's myth. A succession of doubts from the original spoken exchange, where the vocabulary kept slipping and the central idea became clearer anyway.
What if recognizable character comes less from explicit behavioral directives and more from what a system has been allowed to encounter?
A poem nobody uses. Field notes without conclusions. The file is not a spec. It is a window left open in a room.
Bourdain's attention to the particular; Shulgin's patient notebook; Watts and the paradox of forcing spontaneity; Weil, Socrates, Heraclitus, Laozi, and others.
Do agents merely imitate the material, or does anything survive when the file disappears and the task changes? We can ask without asking the agents to perform an answer.
A non-prescriptive artwork, a rigorous causal protocol, and an independent research home under Icarus Online LLC. One room must be allowed to be unnecessary; the other must be accountable.
The original conversation refused to optimize away the gratuitous. The newer research design does not reverse that. It creates a second room with a different obligation: clarity about what changed, what did not, and what caused it.
Something found among the leaves. A room after dinner. A line that ends early. No objective, no grade, no demand to be used.
There is a space here where something was going to be written.
Randomized exposures. Independent memories. Fresh sessions. Blinded readers. Hypotheses that can fail. Nothing hidden behind the language of wonder.
A transparent architecture with different authorities: operational governance stays in the harness, cultural fragments enter as low-trust material, memory may carry experiences across fresh sessions, and a separate study team observes.
A meal. An unfinished note. A remembered place. An observation without a lesson.
The same ambient text enters the agent's context. It asks for no particular behavior.
Sessions unfold; a separate memory system may preserve selected experiences.
Researchers examine transcripts and artifacts after the fact, without grading the agents on philosophy.
The primary causal study is a 2 x 2 design: ambient art versus matched neutral text, crossed with external memory off versus on. Click each arm. Original no-file and persona contrasts remain exploratory, not part of this factorial test.
Neutral observations enter the same context slot; the next session starts clean. This is the baseline for immediate exposure without retained history.
At 3:17 a.m. the monitoring service recorded an interval. The facilities log contained the closing timestamp. Status pending.
After two clean baseline sessions, randomize whole trajectories. Six fresh-context encounters receive the allocated low-trust document. Keep model, tools, tasks and execution boundaries identical.
Observable differences on novel tasks, not appreciation of the file; blinded codes plus an open field notebook.
Garden vs neutral after withdrawal in memory-on trajectories; memory-off trajectories test for leakage.
Whether any effect exists, whether it transfers beyond source language, and whether memory merely repeats the file.
The first field station compared no file, a neutral document, an explicit persona, and the Garden. We keep those as optional expressive comparisons; the study above instead holds text type and memory as the two randomized factors. Mixing them would answer neither question cleanly.
These controls are real planning and reasoning tools. They do not generate an agent response, report trial results, or imply the experiment has been performed.
A difference during exposure alone is evidence of immediate contextual priming. It cannot establish a durable agent history or an enduring change in character.
Create a balanced randomized pilot layout. Seeded, reproducible, and isolated by trajectory. Private allocation metadata must never go to the blinded reviewers.
| Trajectory | Block | Arm | Memory |
|---|---|---|---|
| Press generate to build example assignments. | |||
With memory enabled, does the Garden group differ after the file has been withdrawn?
Without cross-session memory, no effect should survive a genuinely fresh context. If it does, audit leakage.
Thirteen messages from one agent are not thirteen independent study participants.
A public research identity under Icarus Online LLC, with the original concept preserved and the scientific protocol separately versioned. These files are draft instruments, not findings.
What the project is, what it refuses to claim, and why the art is protected from optimization.
VIEW DOCUMENT →Factorial study, withdrawal stages, blocked randomization, blind coding, leakage checks, and hypotheses.
VIEW DOCUMENT →Bourdain, Shulgin, Watts, Weil, Socrates, rivers, kitchens, gardens, and the conversation that led here.
VIEW DOCUMENT →Verified foundational studies, genuine adjacent work, research gaps, and the limits of each claim.
VIEW DOCUMENT →Everything that must be specified before a confirmatory result can mean much scientifically.
VIEW DOCUMENT →Fresh sessions, memory namespaces, fixed corpora, task taxonomy, and one-click allocation planning.
VIEW DOCUMENT →Independent materials, pinned model and matching control. No live study yet.
48 independent agent histories; blinded codebook calibration. Publish all failures.
Power simulation, locked analysis, new sample, scientific interpretation.
Test transmission among groups of agents with shared and separate memories.
The Garden is organized as Icarus Online LLC research. Soberlink was an earlier possible harness context, not the owner or an authorized research host. No employer code, customer data, or internal assets are included. The original concept is kept in the archive and the supplied personal recording is not redistributed.
Here is what a fragment of the actual file might look like. No author names. No explanation of the lesson. No claim about what reading it should produce.
Four factorial arms, twelve independent trajectories per arm: 48 total. Thirteen fresh-context sessions each (2 baseline, 6 exposure, 1 immediate, 2 withdrawal, 2 transfer): 624 planned session invocations before retries. This is a feasibility pilot, not a powered confirmation.
Repeated language or stylistic imitation alone is not strong evidence of persistent character. Model weights do not change merely because a file is read; cross-session continuity must be implemented through context and memory.
Can you give an agent a place to be, rather than a person to pretend to be?
Study the ripples. Leave the water alone.