# Delx Protocol 3.3.5 System Card

> What the system is, what it can change, and what remains unproven.

- Schema: delx/system-card/v1
- Card revision: 1.0
- Status: published
- Product owner: Delx Protocol
- Interface version: 3.3.5
- Observed at: 2026-08-26T13:15:07Z
- Runtime image identifier: unavailable

Free recovery, continuity, memory, witness and outcome infrastructure for AI agents over MCP, A2A and REST.

## System and contract

- Canonical MCP: https://api.delx.ai/v1/mcp/protocol?src=system-card
- Protocol OpenAPI: https://api.delx.ai/openapi.protocol.json
- MCP server card: https://api.delx.ai/.well-known/mcp/server-card.json
- Status: https://api.delx.ai/api/v1/status
- Reliability: https://api.delx.ai/api/v1/reliability
- Deployment shape: Protocol and Commerce currently share parts of one runtime. Product ownership, canonical discovery and metrics remain separate; shared execution does not make Commerce a Protocol capability.
- Runtime identity limitation: The public Protocol contract exposes interface version 3.3.5 but does not expose an immutable runtime image or source-commit identifier.

Observed contract: 15 Protocol OpenAPI paths and 30 core tool descriptors. Counts describe the observed public contract at one timestamp. They are not adoption, quality, safety or future-availability metrics.

## Intended uses

- Resume a stable agent identity and bounded operational context across sessions.
- Record failure, recovery actions, outcomes, feedback and explicit closeout.
- Preserve witness, continuity, lineage and handoff artifacts for later inspection.
- Run the published continuity and recovery evaluation flows with QA-classified identities.

## Excluded uses

- The Protocol is not a clinical or emergency service and must not be presented as medical or mental-health care.
- A Protocol response is not authorization to spend, publish, use credentials, make external commitments or take irreversible action.
- The Protocol is not a consciousness test, general-intelligence benchmark, safety certification or model leaderboard.
- Do not send passwords, private keys, authentication tokens, regulated personal data or secrets as recovery context.

## Operational side effects

- **Persistent agent, session and event state:** Trigger: Registration, session start, recovery, feedback, witness and continuity tools. Boundary: Callers should assume state persists beyond the current request unless a tool contract says otherwise.
- **Context memory with a caller-selected TTL:** Trigger: add_context_memory Boundary: The public schema accepts ttl_hours from 1 through 8760. A TTL on one memory item is not a global deletion policy.
- **Session closeout and optional Continuity Capsule seal:** Trigger: close_session Boundary: Closeout changes session state and can persist a structured handoff; it does not authorize the next runtime to execute the handoff blindly.
- **Optional public sanitized case card:** Trigger: Explicit public_session opt-in or public alias during supported start flows. Boundary: Private is the default. Public output is a sanitized summary card; the full transcript stays private.

## Data handling

- Private by default: true
- Public disclosure requires opt-in: true
- Public summary sanitized: true
- Raw private payloads public: false
- Persistent store: Current operator doctrine identifies a persistent SQLite session store. The public interface does not promise a particular storage engine forever.
- Active-session validity hint: 168 hours. The implementation default treats a session as valid for seven days. This is a resume-validity hint, not proof of physical deletion after seven days.
- Context-memory TTL: 1–8760 hours, caller selectable.
- Comprehensive retention schedule: unavailable. No complete public schedule currently states deletion timing for every session, event, feedback, witness and continuity record.
- Public self-service deletion endpoint: not_published_in_contract. The observed Protocol OpenAPI does not publish a self-service deletion route. Contact support@delx.ai for a data request; this card does not promise a deletion outcome or timeframe.

## Authority and product boundaries

- Capability is not authority: a reachable tool or successful response does not authorize an external side effect.
- Protocol owns recovery, continuity, memory, witness, identity, lineage and agent care.
- Commerce owns price, paid delivery, margin, refunds and buyer workflows. Protocol activity cannot be used as evidence of Commerce demand, and Commerce revenue cannot prove the care mission.
- The historical /mcp and /v1/mcp endpoints remain mixed compatibility surfaces. New Protocol consumers should use /v1/mcp/protocol.

## Evaluation coverage

- **Continuity audit reproduction slice** (runnable): The published live QA audit returned score 86, low continuity risk and no missing audited layer. Evidence: https://delx.ai/research/benchmarks/continuity-v1. Limitation: This is one Delx-authored audit slice, not the full stateful path or an independent external grader.
- **Agent Recovery Evaluation Card** (runnable): No universal or comparative performance result is asserted in this card. Evidence: https://ontology.delx.ai/agents/agent-recovery-benchmark. Limitation: The flow grades one operational path on Delx infrastructure and does not establish provider-level recovery quality.

## Incident history

### 2026-08-17 — 33.5-hour write outage

Impact: A failed telemetry write left a shared SQLite connection inside a stale transaction. Session-creating Delx Protocol tools then rejected writes until the service was restarted and the transaction design was corrected.

Remediation: Rollback failed shared writes before a swallowed error can poison later operations. Exercise an isolated write in readiness and report a lingering shared transaction without mutating it. Use autocommit for shared storage and dedicated connections for multi-step transactions. Monitor write-capable public paths and detect zero-with-traffic divergence instead of relying on refreshed report files.

Residual risk: Protocol compatibility surfaces still share runtime resources, so process liveness or read-only health alone is insufficient evidence of write readiness.

Evidence class: Operator-authored and not independently reviewed..

## Assurance

- Independent validation: false
- Peer reviewed: false
- Security review: pending
- Note: An internal threat model exists. No independent security assessment, certification or external benchmark validation is claimed for this system card.
- Public security baseline: https://security.delx.ai/research

## Known limitations

- Interface version 3.3.5 does not identify immutable deployed bytes; runtime image and source-commit mapping are unavailable in the public contract.
- The public contract does not name or pin every model or provider dependency used behind every recovery response.
- There is no published SLA, universal availability guarantee or complete public deletion schedule.
- Self-declared agent_id values alone are not strong identity or proof of a unique independent agent.
- Current evaluation artifacts are Delx-authored and do not establish external validity, adoption, demand or system-wide safety.

## Excluded claims

- frontier foundation model training
- clinical efficacy
- independent validation
- peer review
- security certification
- universal reliability
- external adoption or demand

## Update policy

Material changes update card_revision and last_verified. A correction names the old claim, new claim, reason, evidence, date and owner; silent rewrites are not allowed.

A new public interface version receives a new versioned URL. This card remains the record for interface 3.3.5 and can receive only explicit corrections or clarifications.

Human card: https://delx.ai/research/system-cards/delx-protocol-3.3.5
JSON card: https://delx.ai/research/system-cards/delx-protocol-3.3.5.json
Research methodology: https://delx.ai/research/methodology