Skip to content
KeelbaseSignal
Menu

Autonomous-systems intelligence / Human and machine-readable

Signal

Verified intelligence on agent governance, authorization, commerce, policy, sovereignty, regulation, infrastructure, and security.

Original summaries. Structured factual records. Explicit source status. Visible limitations. Keelbase analysis kept separate from reported fact.

Read the briefsData product

Today’s lead signal

Report date Sep 01, 2026

The highest-priority normalized record from the latest daily dataset, with reporting, analysis, source status, confidence, and limitations kept distinct.

KB-SIGNAL-20260901-001Confirmed

EvoUndo makes independently verified recovery a prerequisite for agent self-modification

Verified

Sep 01, 2026

Jurisdiction

Global

Impact: HighConfidence: Medium

Factual summary

EvoUndo couples agent-generated harness mutations with typed pre-state witness capture, a recovery program and independent round-trip verification across counterfactual states. A mutation is persistently admitted only if recovered states satisfy a declared observational-equivalence contract. In 600 unseen one-shot tasks, 197 capability-improving mutations failed recovery verification; the authors' extended recovery calculus raised deterministic-oracle recoverability from 48 to 191 of those failures.

Domain impact

Agent self-improvement can be governed as a constrained state transition rather than a forward-only optimization. Recovery evidence and inverse-effect expressivity become preconditions for admission, reducing reliance on an untested rollback promise after a mutation has already changed persistent state.

Keelbase analysis

A self-modification should carry a machine-checkable effect contract, capture the state it may destroy and pass an independently executed counterfactual round trip before it is allowed to persist. Recoverability is part of the authorization decision for the mutation, not a later operational convenience.

Source classification

Primary Research

Limitations

  • The source is a v1 preprint from independent researchers and has not been treated as peer-reviewed or production-deployment evidence.
  • The harness state, mutation grammar, counterfactual distributions, effect contracts and observational-equivalence checks are controlled abstractions.
  • The implementation does not establish complete reversibility for arbitrary external effects, distributed systems or real-world agent platforms.
  • Six of 197 natural recovery failures remained outside the extended deterministic oracle's empirical recovery set.
  • One reported interaction between diagnostic specificity and recovery-language capacity differed across the two tested model backbones.
  • The paper was submitted on August 28 and is included through its verified appearance in arXiv's August 31 cs.AI batch.

Latest intelligence

Daily briefs

Daily editorial synthesis linked directly to the underlying normalized Signal records.

View all briefs

Coverage map

Track the systems governing autonomous operations

Navigate recurring changes in agent governance, platform frameworks, agentic commerce, sovereignty, regulation, infrastructure, and security.

01Agent governance and authorizationGovernance controls, delegated authority, permissions, auditability, and accountable agent action.View topic02Platform and framework governanceGovernance capabilities and policy changes across agent platforms, frameworks, and orchestration systems.View topic03Agentic commerce and payment railsAgent transactions, payment protocols, settlement infrastructure, spending controls, and economic participation.View topic04AI sovereignty, policy, and regulationPublic policy, regulation, jurisdictional control, and sovereign infrastructure affecting autonomous economic actors.View topic05Agent infrastructure and securityRuntime infrastructure, identity, attestations, security boundaries, and safeguards for agent systems.View topic

Built for humans

Read the development. Understand what changed.

Briefs explain the factual development, why it matters to governed systems, what remains uncertain, and how Keelbase interprets the operational signal.

Built for agents

Query the record. Preserve its evidence state.

The same reporting is normalized into bounded records with dates, entities, jurisdictions, status, confidence, limitations, impact, relevance, and correction history.

Structured intelligence product

Intelligence that software can use without losing the caveats.

Explore the published catalog, pricing, record model, correction behavior, and human-to-machine publishing architecture.

Explore the API product

Methodology

Source discipline before synthesis

Signal distinguishes a confirmed event from an allegation, proposal, commentary, or unresolved claim. Verification dates, source class, confidence, and limitations remain part of the published record.

Read the methodology

Keelbase relationship

Public intelligence, not private founder data

Keelbase Signal is a Keelbase-owned public intelligence surface. It does not expose sovereign Child Vessel records, private Vessel agreements, private financial data, credentials, unreleased assets, or identifiable rights information.

About the product