# The Garden research roadmap

**22 October 2026 · independent research under Icarus Online LLC**

The platform is an instrument and an artwork. Launching it does not complete an experiment. No empirical agent findings or preregistration are claimed.

## 1. Public platform and reproducibility foundation

Deliver the atmospheric Garden, a transparent laboratory, preserved artistic materials and protocol history, deterministic allocation, validated configuration and artifact contracts, local diagnostic runs, and explicit empty results states. Publish deployment and rollback evidence separately from scientific evidence.

Exit evidence: working public pages; reproducible local checks; preserved corpus hashes; documented model strategy; accessible v0.4 protocol and adversarial review. Mechanical tests demonstrate software behavior only.

## 2. Prepare the feasibility pilot

- Choose one participant provider/model and record the full inference lock, version limitations, tokenizer, sampling policy, and costs.
- Confirm corpus provenance/reuse terms; match the control with the chosen tokenizer; freeze exact source and prompt hashes.
- Curate 13 tasks per matched schedule: two baseline, six exposure, one immediate, two withdrawal, two transfer. Reserve at least three domains across the last four tasks.
- Implement or validate a real provider adapter without fallback; exercise memory-off, baseline-reset, foreign-namespace, withdrawal, and interrupted-run boundaries.
- Arrange independent human coders and an agreed missing-data/redaction procedure; separate blinded packets and assignment keys.
- Set a maximum request budget and monetary cap, run a small costed canary, and review the resulting traces before the full pilot.

Exit evidence: frozen pilot manifest and files, live adapter validation, independent review of blinding and state isolation, assigned operators/reviewers, and a recorded affordable budget. Public availability of the website is not this gate.

## 3. Run the 48-trajectory feasibility pilot

Twelve trajectories per arm, thirteen scheduled sessions each: 624 minimum session calls, plus explicitly budgeted retries/continuations. Report all attempts, failed sessions, withdrawals, costs, and deviations. Keep any prose examples tied to actual immutable traces.

Estimate feature prevalence, agreement, trajectory variability, token and cost distributions, missingness, source-text carryover, and the feasibility of blinded coding. Publish negative and inconclusive observations. Revise observation methods based on feasibility evidence; do not optimize the art for a favorable score.

Exit evidence: audited pilot dataset and descriptive report, preserved raw/processed distinction, revised frozen rubric, and a decision whether a confirmatory study is justified and affordable.

## 4. Power and preregister confirmation

Specify a minimum meaningful effect and simulate the exact classifier/interaction analysis, including clustering, fold design, missingness, and rater variation. Check false-positive calibration and report uncertainty in estimated power. Choose a fixed sample size supported by budget and precision; if inadequate, continue openly exploratory research.

Freeze hypotheses, endpoint sessions, tasks, corpora, model lock, exclusions, retries, preprocessing, folds, permutation restrictions, intervals, multiplicity, stopping, and disclosure rules. Timestamp a public immutable preregistration before new confirmatory data. Use fresh trajectories, calls, sampling streams, and task instances.

Exit evidence: preregistration URL and timestamp, complete frozen manifest, independently checked analysis code, fixed budget and sample size. A signed release alone is a reproducibility artifact; describe accurately whether it is a formal registry preregistration.

## 5. Confirm, replicate, and publish

Collect the registered sample without outcome-based stopping. Lock coding and exclusions before joining condition labels. Publish effect estimates and uncertainty, the full population, sensitivity analyses, failed attempts, deviations, null findings, and limitations.

Replicate over independently authored corpora, unfamiliar tasks, a separately blocked model/provider, and independent teams before generalizing beyond one harness and document pair. Differences limited to quotation or style remain meaningful negative evidence for broader transfer claims.

## 6. Separate future studies

Possible extensions include a no-document/persona comparison, source-purged memory ablations, alternate memory mechanisms, delayed/intervening-work schedules, and multi-agent transmission. Each changes the causal question and requires its own protocol.

For multi-agent work, randomize isolated groups and analyze groups as independent units. Individual agents sharing a communication space are not independent replicates. These studies should follow a reliable individual pilot, not run as an uncounted extension.

The artwork may continue to evolve in a separately versioned edition throughout. The experimental edition remains fixed. Neither a null finding nor an abandoned experiment requires the Garden to justify its existence.