Test the acting system.
Turn questions about “I,” “we,” and “other” into controlled experiments with rival explanations and explicit failure conditions.
Read the frameworkIndependent AI research
Keep human authority inside the loop.
When AI acts on our behalf, it needs to keep track of who is acting, what permission applies, and which commitments matter. RES puts those distinctions to the test.
Relational–Episodic Self · Research under IFERS fiscal sponsorship
Current phaseReplication and assay development
Evidence boundaryNo established model-level RES finding
Immediate needModel access and compute
Why this work matters
Delegation needs
accountability.
Being able to take an action is different from having permission. A persuasive account is different from independent evidence. RES studies these distinctions—and asks how they can survive complex tasks, handoffs, and failures.
Publish what holds up.
Track successful controls, mixed results, failed gates, and non-replications. Let the evidence set the next question.
Inspect the findingsGovern the governors.
Keep records, audit judgments, permission, and execution separate. Make authority inspectable and actions bounded.
Explore the designFrom the research record
A promising result.
A necessary retest.
Changing the wording improved performance in one diagnostic. A fresh five-factor comparison did not replicate the benefit. Both results belong in the public record.
Initial matched diagnostic
17/32 → 26/32
Generated answers: code-mapped versus neutral labels and task wording.
Fresh five-factor comparison
35/64 · 38/64
Generated answers: loaded versus fully neutral wording on fresh fixtures.
Different fixtures and comparison conditions. These studies do not form one pooled effect, reverse the original failed gate, or identify a RES mechanism.
Open to scrutiny
Follow the reasoning.
Check the record.
Read the public framework, inspect the protocols and archived reports, or help design an independent test. The documentation includes a glossary and a staged research roadmap.
Fund the next careful test
Better evidence takes more than one answer.
Contributions support model tokens, API access, compute, and the repeated measurements needed to test alternatives and document failures.