Distinct questions, shared publication
Research
Two workstreams. Their hypotheses, evidence, and experimental status remain separate.
01 / Persistent agents
Experience outside a single context
Investigate external memory, pending investigations, context continuity, self-model claims, and source provenance in a bounded controller. The apparatus supports scripted engineering checks and live local-model experiments. Scripted outcomes test the apparatus; they do not demonstrate model learning.
- What exists
- Local evidence includes bounded delivery tests, forecast pilots, source interpretation diagnostics, and memory comparisons. Some attempts failed or were invalidated. Backfill review is in progress.
- Public evidence
- A completed delivery calibration with no valid actions, with selected derived evidence.
- Open questions
- When does access to experience improve task performance beyond additional context, computation, tools, or supplied task structure? Can apparent benefits survive matching and interventions?
- Limits
- Small samples, structured tasks, delivery failures, changing configurations between separate protocols, and unequal realized compute restrict generalization. There is no consciousness score.
02 / Developmental computation
Construction is not yet utility
An independent review distinguishes executable construction, inherited behavior during growth, and a measured learning advantage. Compact recipes and heterogeneous modules do not by themselves establish that advantage.
- Status
- Research/specification milestone completed 14 October 2026. Two draft protocols; a bounded two-cell engineering witness with 17 reported checks. Learning experiments unexecuted. The historical check results have not been independently rerun for this website.
- Primary comparator
- A conventional parameterized modular builder, with matched supplied information and resource accounting.
- Next question
- Can a specified construction or growth mechanism outperform ordinary building and function-preserving growth under matched resources?
- Publication
- The full assessment and witness await a separate public-safety and evidence-bundle review. This workstream is not integrated with the persistent-agent trial.