Learn
Read plain-language concepts, papers, formulas, results, and limitations.
Start with the research mapOpen Research Community
AI can answer. Can it learn from what happens next?
This community studies AI systems that act, change their environment, receive feedback, remember failure, and carry evidence into the next decision.
learnusecontributepartner
ActWhat evidence must exist before an AI action proceeds?
ObserveWhat changes when an agent becomes part of the world it predicts?
LearnDoes a correction become durable, scoped behavior?
AuditCan another person inspect the rule, result, and limit?
Four Routes
Enter at the level that matches your question.
No affiliation or private access is required. Read first, run an artifact, improve something small, or bring a concrete reliability problem.
Use
Inspect benchmarks, datasets, schemas, validators, demos, and minimal references.
Open Hugging Face artifactsContribute
Ask a question, add a safe example, improve a document, or publish an experiment.
Choose a discussion routePartner
Bring a bounded evaluation problem where failure recurrence or action evidence has real cost.
Start a private inquiryResearch Projects
Four projects, each built around a concrete question.
Open the paper, protocol, schema, example, or reference implementation that matches the question you want to investigate.
ProjectQuestionAvailable materialsWhat they do not establish
ReflexBenchDoes competence survive when an agent changes the world it observes?Protocol and paper artifactResearch evaluation, not deployment certification
Cognitive ImmunityDoes a verified failure become a scoped, auditable defense against recurrence?Minimal public interfacesProduction memory and policy remain private
WisdomBenchCan an evaluator distinguish answer quality from judgment under uncertainty?Schema, scorer, examplesBenchmark conditions bound the result
Proof-Carrying ActionWhat proof should accompany a consequential action request?Protocol and toy referenceNo private execution integrations
Ouroboros PapersHow do the concepts and evidence records connect across the research program?Paper and DOI indexArchive and navigation, not product support
Versioned Releases
Versioned artifacts with commands, manifests, and explicit limits.
Each release fixes the exact files, version, and verification commands needed to cite or run it. None is a production safety certificate or the complete product.
ReleaseContentsChecksLimits
Proof-Carrying Action v0.1.0Proof-envelope schema and deterministic repair demoDemo check and pytestToy protocol, not production authorization
ReflexBench v2.0.0-public.120-scenario observer-depth artifactPublic artifact auditDevelopment benchmark, not deployment validation
WisdomBench v0.1.0-public.1Tasks, metrics, score records, and run-card contractAnalysis, schema validation, and unit testsDoes not establish general wisdom or safety
SOVEREIGN Public Interfaces v0.1.0Cognitive Immunity and structured-memory interfacesLifecycle demo and unit testsMinimal interfaces, not the complete product
Current Research Conversation
Two KDD 2026 workshop papers. Two ways an agent can fail after it becomes consequential.
ReflexBench evaluates whether agent reasoning survives deeper observer-participant feedback. Across 20 scenarios, six domains, four observer-depth levels, and nine public models, all tested models degraded at deeper levels.
Cognitive Immunity evaluates repeated severe-threshold failure after correction. Under the documented score-level protocol, the pooled repeat failure rate moved from 0.764 for No Memory to 0.650 for Cognitive Immunity; the paired seed-task delta was -0.106 with a 95% bootstrap interval of [-0.211, -0.003].
Boundary: both are accepted and presented non-archival KDD 2026 workshop papers. They are not KDD main-conference proceedings papers, production safety certification, or proof of general reliability.
First Contribution
Participation can start small and become rigorous.
1Read or run one public artifact.
2Ask one concrete question or describe one use case.
3Repair one document, link, fixture, or example.
4Publish a baseline, experiment, or extension.
5Maintain a bounded component or research thread.
Open-Core Boundary
Open enough to inspect and extend. Bounded enough to protect people and product.
Public
Published concepts and formulas, protocols, schemas, validators, small fixtures, minimal references, safe demos, documentation, and negative results.
Release review
Large datasets, detailed traces, near-production references, security-relevant examples, and components whose combination may reveal implementation advantage.
Private
Production orchestration, exact weights and thresholds, private prompts and data, customer systems, tuning history, deployment automation, and unreleased research.
Questions
What this community is, and what it is not.
Who can contribute?
Researchers, engineers, students, operators, and readers. Useful work includes questions, use cases, documentation, safe fixtures, baselines, experiments, and code.
Must every contribution reproduce an experiment or dispute a conclusion?
No. Contributions can clarify a concept, connect related materials, improve usability, add an example, or describe a real problem. Counterexamples and conclusion checks are two options among many.
Does open research mean the complete product is open source?
No. Public protocols, interfaces, validators, and minimal references are separated from private production implementation, operational parameters, customer material, and proprietary learning loops.
Does a paper or benchmark certify production safety?
No. Each result is bounded by its stated protocol, data, version, and limitations.
Community Standard
Bring a real question. Leave behind a public artifact another person can use.
The artifact may be a clearer explanation, a safe fixture, a comparison, a negative result, a corrected boundary, or a working extension.