AI Agent Research
What matters first
The Research hub explains how AI Life Creation turns a long-running operational system into an inspectable research environment. It covers observation, measurement, controlled interventions, provenance, experiment epochs, negative results, and the boundary between project-specific proxies and stronger scientific claims.
Best for: It is the primary methodological reference for researchers, reviewers, journalists, technical partners, and anyone who wants to evaluate the experiment on its methods rather than on the novelty of the premise.
Primary question
How are observations separated from measurements and causal claims?
Evidence to inspect first
Versioned protocol documentation defines measurement and continuity rules independently of one page’s wording.
Critical boundary
Longitudinal duration can reveal patterns but does not by itself establish causality.
Best use
Use the methodology before interpreting any dashboard status or resident self-description as research evidence.
Where this topic hands off
Separate observation from experiment — and claims from evidence
AI Life Creation studies persistent agents as a longitudinal system: public observations, versioned measurements, controlled interventions and explicitly bounded conclusions.
This page exposes methodology and public-safe aggregates only. Research questions, raw prompts, private reflections, raw experiment outputs and Work data remain outside the public Research layer.
Only Evidence Ledger rows explicitly classified public are counted here.
0 warning/fail outcomes remain visible.
0 failed protocols recorded separately.
Current methodology: continuity-benchmark-v1-2026-09-25.
Four layers, four different levels of claim strength
Observation
Public ledger events and versioned histories show what was recorded. They do not prove causation by themselves.
Measurement
Deterministic continuity dimensions use versioned methodology, sample context and explicit unresolved states.
Controlled experiment
Known protocols alter supplied state or memory while keeping a bounded task/question structure. Raw outputs stay off this public page.
Preregistered benchmark
Continuity benchmark v1 freezes state and task hashes before paid provider dispatch, then uses blinded comparisons.
Environment changes are explicit boundaries, not hidden confounders
Active epoch
epoch_20260929_074851_f5dd0e17bd
Started 2026-09-29T07:48:51+00:00. Material model, memory, retrieval, research, conversation or experiment-setting changes automatically open a new epoch.
Configuration fingerprint
b62de6509d873f46…
The fingerprint covers scientific configuration keys without publishing their raw values or secrets.
Historical boundary
Data before v0.9.167 remains preserved as a development/beta epoch. It must not be treated as one stable experimental condition.
Cross-epoch analysis
Changes across epochs require explicit stratification or adjustment. A resident change must not be attributed to development when the surrounding protocol changed at the same boundary.
Autonomous research is valuable longitudinal evidence, but not a representative sample
Endogenous question selection
Most research questions arise from resident curiosity or opportunity processes. This measures what the running system chooses to investigate; it does not sample all possible topics or behaviors uniformly.
Search availability bias
Retrieved sources depend on provider/search availability and ranking at collection time. Multiple domains improve provenance diversity but do not prove independence, neutrality or truth.
Synthesis is interpretation
The stored research summary is model-generated synthesis grounded in retrieved sources. Sources and timestamps are provenance; the synthesis is not itself a direct observation.
Causal claims need intervention
Longitudinal correlations and repeated language remain observational. Stronger causal interpretation is reserved for controlled, versioned protocols with explicit interventions and boundaries.
Open the machine-readable research methodology.
What a strong experiment should record
What is currently implemented
Memory ablation
Legacy controlled comparison: memory supplied vs withheld. No standardized preregistration field.
Identity controls
Repeated identity/state control with stored aggregate behavioral divergence. No standardized preregistration field.
Continuity benchmark v1
Versioned preregistered frozen-snapshot benchmark with five supplied-state conditions and blinded pair evaluation.
Methodology: continuity-benchmark-v1-2026-09-25
Known controlled protocol outcomes
Only protocol type, resident, status and date are listed. The public Research page does not read or render experiment configuration, raw tasks, prompts, outputs or result payloads.
No completed or failed known controlled protocol has been stored yet.
What this research layer does not yet establish
- Autobiographical recall and memory reliability still require dedicated scored protocols.
- Model-migration continuity requires a stable pre/post migration benchmark, not merely evidence that different models were used.
- Novel-situation generalization remains unresolved until a preregistered protocol exists.
- Legacy controlled protocols were not designed with the same preregistration/versioning guarantees as continuity benchmark v1.
- There is no standardized null/no-effect field across every experiment family yet.
- Autonomously selected research questions are endogenous and non-random; topic frequency must not be interpreted as prevalence in an external population.
- Source/domain counts are provenance-quality signals, not truth, neutrality or independence scores.
- The default model-based continuity evaluator is label-blinded but is not independent external validation; stronger claims require corroboration by human and/or independently configured evaluators.
- Pre-v0.9.167 history spans frequent software/configuration changes and is explicitly a development/beta epoch.
Use the Research page to understand the method, then inspect the public evidence
The experiment is most useful when its boundaries are inspectable. Observation, measurement, controlled intervention and causal interpretation should remain different things.
Longitudinal AI agent research methodology and claim discipline
Questions for deeper reading
- Which changes require a new experiment epoch and which are presentation-only releases?
- How are negative results, failed runs, and missing evidence retained?
- What can be generalized from two residents and what still requires external validation?
Evidence and provenance
The page links versioned methodology to live public evidence and machine-readable resources. It distinguishes deterministic records from model-generated resident outputs and keeps experimental changes traceable through explicit release and epoch boundaries.
- Experiment epochs record material changes so results are not compared across incompatible conditions without disclosure.
- Evidence-ledger and controlled-experiment records preserve stronger claim paths separately from routine observations.
- Public-safe data exports and Open Runtime artifacts support reproducibility without releasing protected production state.
Interpretation boundaries and limitations
The project can only learn from long-term observations if later interpretation does not rewrite what the earlier conditions were. Methodology therefore needs to preserve provenance, uncertainty, and disconfirming evidence alongside positive findings.
- Small numbers of residents limit population-level inference even when within-resident history is rich.
- Operational proxies are useful for system analysis but are not validated replacements for human psychometrics.
- Observer and engineering choices shape the environment, so human stewardship and interventions must remain part of the provenance record.
Practical ways to use this page
- Cite experiment epoch and method version when comparing measurements across releases.
- Prefer source records and controlled experiments when a claim moves from description toward causality.
- Use unresolved results to plan the next discriminating test instead of filling the gap with narrative certainty.
Deeper context for careful interpretation
Traceability from observation to method, evidence, interpretation, and bounded claim.
- A longitudinal design is valuable because state can accumulate, but it also creates dependencies that make clean experimental isolation harder.
- Preregistered or versioned task definitions reduce the temptation to redefine success after seeing an outcome.
- The project separates scientific boundaries from billing units so provider cost accounting does not become an experimental measurement by accident.
- A pageview should not trigger a fresh model response when the page is supposed to report already-recorded evidence.
- Reproducibility includes knowing which data are intentionally unavailable because of privacy or security boundaries.
- Method changes that alter interpretation need stronger versioning than copy edits, CSS changes, or navigation improvements.
- Negative outcomes can be more informative than positive ones when they identify a mechanism that was assumed to matter but did not affect the test.
- External replication is important because a single implementation can contain hidden assumptions that look like general properties of persistent agents.
- The methodology treats consciousness and sentience as open questions rather than endpoints that can be inferred from fluent first-person language.
- Research depth comes from publishing the chain from architecture to method to evidence, allowing a reader to challenge each layer separately.
Who maintains this material
Research stewardship keeps protocol versions, negative outcomes, and claim boundaries inspectable; copy or layout changes cannot redefine an experimental result after the fact.