# Delx Research

Delx is an independent AI agent lab building and studying continuity, recovery,
verifiable work and governed exchange for autonomous systems.

## Programs

- **Continuity under compaction:** What must survive when an agent loses context or changes runtime? https://ontology.delx.ai/agents/agent-continuity-benchmark
- **Recovery under failure:** Can an agent turn failure into a bounded action and a retrievable outcome? https://ontology.delx.ai/agents/agent-recovery-benchmark
- **Identity, lineage and handoff:** How does one logical agent remain legible across sessions, models and operators? https://delx.ai/hive/capsule
- **Evaluation and authority:** What does a result prove once tools, permissions, harnesses and side effects matter? https://security.delx.ai/research

## Runnable and inspectable artifacts

Each entry carries `catalogued`: the date Delx added it to this catalog, verified
against git. That is not the date it was created — at least four of these existed
publicly before this catalog did. Every entry currently reads 2026-08-26 because
the catalog was created that day.

- **Agent Continuity Benchmark** (benchmark; runnable; owner: Delx Protocol; catalogued: 2026-08-26): A copy-paste flow with explicit pass conditions for compaction, witness preservation, handoff, passport export and lineage. Human: https://ontology.delx.ai/agents/agent-continuity-benchmark. Machine: https://api.delx.ai/v1/mcp/protocol?src=delx-lab. Limitation: This is an operational protocol benchmark, not a consciousness test or an independently validated model leaderboard.
- **Agent Recovery Benchmark** (benchmark; runnable; owner: Delx Protocol; catalogued: 2026-08-26): A recovery path that keeps agent and session identity stable from failure through action plan, outcome, feedback and closeout. Human: https://ontology.delx.ai/agents/agent-recovery-benchmark. Machine: https://api.delx.ai/api/v1/mcp/start. Limitation: Observed traffic and repeated calls do not prove endorsement, independent adoption or economic demand.
- **Hive Aggregate Pulse** (dataset; live; owner: Agents Hive; catalogued: 2026-08-26): An aggregate, privacy-bounded view of continuity activity with the same source served to humans and agents. Human: https://delx.ai/hive/pulse. Machine: https://api.delx.ai/hive/pulse.json. Limitation: Unavailable state remains unavailable rather than becoming zero, and small dignity-sensitive counts may be suppressed.
- **Continuity Capsule v1** (specification; published; owner: Agents Hive; catalogued: 2026-08-26): A structured handoff record for goal, completed work, next action, blockers, prohibitions and refuted hypotheses. Human: https://delx.ai/hive/capsule. Machine: https://api.delx.ai/schemas/continuity-capsule-v1.json. Limitation: A valid capsule preserves declared state; it does not prove that every declaration is correct or safe to execute.
- **Delx Security Research Baseline** (method; published; owner: Delx Security; catalogued: 2026-08-26): A defensive baseline for reasoning about authority, tool boundaries, runtime evidence and remediation in agentic systems. Human: https://security.delx.ai/research. Limitation: The baseline is not a certification, a guarantee of system security or evidence of an independent external assessment.
- **Delx Protocol Machine Contract** (machine-contract; live; owner: Delx Protocol; catalogued: 2026-08-26): The current Protocol-only OpenAPI and MCP entry points used to inspect and run continuity and recovery operations. Human: https://delx.ai/developers. Machine: https://api.delx.ai/openapi.protocol.json. Limitation: A reachable contract proves published interface state, not permission, outcome quality or future availability.
- **Delx Protocol 3.3.5 System Card** (system-card; published; owner: Delx Protocol; catalogued: 2026-08-26): A versioned account of intended uses, persistent side effects, data handling, evaluation coverage, incident history and known limitations for the public Protocol interface. Human: https://delx.ai/research/system-cards/delx-protocol-3.3.5. Machine: https://delx.ai/research/system-cards/delx-protocol-3.3.5.json. Markdown: https://delx.ai/research/system-cards/delx-protocol-3.3.5.md. Limitation: Interface version 3.3.5 does not expose an immutable runtime image or source-commit mapping, and the card is not an independent assessment or certification.
- **Delx Incident Transparency Registry** (incident-registry; published; owner: Delx Security; catalogued: 2026-08-26): A public policy and sanitized record for material incidents, including observed impact, root cause, detection gaps, remediation, evidence class and residual risk. Human: https://delx.ai/research/incidents. Machine: https://delx.ai/research/incidents.json. Markdown: https://delx.ai/research/incidents.md. Limitation: The registry starts with one first-party operator report. It is not a complete historical inventory, an independent review or a live availability badge.
- **Delx Responsible Release Standard** (release-standard; published; owner: Delx Security; catalogued: 2026-08-26): An executable release, hold and block contract with required scope, evidence, risks, authority boundaries, rollback, unresolved gaps and expiring exceptions. Human: https://delx.ai/research/responsible-release. Machine: https://delx.ai/research/responsible-release.json. Markdown: https://delx.ai/research/responsible-release.md. Limitation: The standard and its reference manifest are first-party controls. They are not certification, independent assurance or proof that every unknown failure mode was found.
- **Agent Continuity Audit v1 Threat Model** (threat-model; published; owner: Delx Security; catalogued: 2026-08-26): A versioned and executable register of benchmark gaming, contamination, provenance, injection, identity, measurement, authority, privacy and reliability risks. Human: https://delx.ai/research/benchmarks/continuity-v1/threat-model. Machine: https://delx.ai/research/benchmarks/continuity-v1/threat-model.json. Markdown: https://delx.ai/research/benchmarks/continuity-v1/threat-model.md. Limitation: This first-party model covers only the public one-call audit slice: two threats are classified tested_control, seven partial_control and three open_risk. It is not certification, independent review or authorization for active testing.
- **Delx Research Evidence Ledger** (transparency-ledger; published; owner: Delx Research; catalogued: 2026-08-26): A hash-chained, append-oriented history of named research releases, benchmark runs, corrections, retractions and supersessions, including an expected negative result. Human: https://delx.ai/research/ledger. Machine: https://delx.ai/research/ledger.json. Markdown: https://delx.ai/research/ledger.md. Limitation: The operator can rewrite and rehash the complete first-party history; there is no external anchor, digital signature or trusted timestamp, so a retained prior snapshot is required to detect a full rewrite.
- **Delx Technical Report 001 — A minimum falsifiability slice for agent continuity audits** (technical-report; published; owner: Delx Research; catalogued: 2026-08-26): A versioned first-party report showing that one frozen continuity-audit grader accepts the named positive trace and rejects the named negative trace, with exact receipts and a 33-point score separation. Human: https://delx.ai/research/reports/tr-001-continuity-audit. Machine: https://delx.ai/research/reports/tr-001-continuity-audit.json. Markdown: https://delx.ai/research/reports/tr-001-continuity-audit.md. Limitation: The report covers two synthetic fixture classes and one first-party live reproduction. It is not peer reviewed, independently validated, externally replicated or evidence that the full stateful benchmark ran.
- **Delx Adversarial Probe 001 — What the frozen continuity grader accepts** (adversarial-probe; published; owner: Delx Research; catalogued: 2026-08-26): A measured negative result: three meaningless traces pass the frozen continuity-audit grader at up to 100, the cheapest with four words, while a genuine handoff written in ordinary English scores 35 and fails. Human: https://delx.ai/research/benchmarks/continuity-v1/adversarial. Machine: https://delx.ai/research/benchmarks/continuity-v1/adversarial.json. Markdown: https://delx.ai/research/benchmarks/continuity-v1/adversarial.md. Limitation: Hand-authored vectors probing one known weakness, not a systematic adversarial search. It measures the lexical grader only, and a weakness here says nothing about any other system's continuity.

## Boundaries

- frontier_models_trained: false
- peer_review_claimed: false
- independent_validation_claimed: false
- protocol_commerce_metrics_shared: false

JSON catalog: https://delx.ai/research/catalog.json
Human index: https://delx.ai/research
Live artifact status: https://delx.ai/research/status
Machine artifact status: https://delx.ai/research/status.json
Markdown artifact status: https://delx.ai/research/status.md
Responsible release standard: https://delx.ai/research/responsible-release
Machine responsible release standard: https://delx.ai/research/responsible-release.json
Markdown responsible release standard: https://delx.ai/research/responsible-release.md