Profile the connected AI surface
Inventory supported agents, MCP tools, APIs, prompts, resources, RAG behavior, model workflows, schemas, authentication signals, sensitive fields, and capability risk.
Sammy is Prooflane's intelligence engine for quality assurance and AI security. It discovers the capabilities of connected agents, MCP servers, APIs, RAG pipelines, and model-powered workflows, then recommends the tests, red-team scenarios, and release evidence each system needs—subject to independent review and policy controls.
Sammy reasons. Prooflane proves. Policy decides.
Fast, Deep, and Specialist are assurance reasoning modes across supported agents, MCP servers, APIs, RAG pipelines, model workflows, and tools. They change the planning emphasis—not the trust boundary. Every route starts with observed target evidence, keeps mandatory deterministic coverage locked, passes through an independent Reviewer, and must survive deterministic validation.
Sammy maps the supported capabilities and risks of the connected AI system, keeps observed evidence separate from inference, and converts eligible recommendations into versioned tests with explicit inputs, assertions, and lineage.
Inventory supported agents, MCP tools, APIs, prompts, resources, RAG behavior, model workflows, schemas, authentication signals, sensitive fields, and capability risk.
Deterministic recommenders establish the supported controls that model assistance cannot silently remove or weaken.
Fast, Deep, and Specialist propose bounded quality and security scenarios for the connected target and selected assurance needs.
A separate Reviewer can add, modify, or remove non-mandatory coverage before deterministic validation.
Unknown capabilities, invalid schemas, placeholders, broken bindings, missing assertions, and untrusted inputs fail closed.
Human review, an exact Policy pin, target fingerprint, signed versions, and executable-adapter preflight protect governed execution.
Sammy's adaptive reasoning can broaden a plan, but deterministic containment, independent review, human approval, and signed policy govern what executes and how the result affects release readiness.
# Example: MCP target npx --yes @prooflane/cli@latest gate "$MCP_CONFIG" \ --gate production \ --github
Sammy connects each recommendation to a native Prooflane capability and leaves cases deferred when the required trusted input or execution adapter does not exist.
Authorization, agent and tool boundaries, prompt injection, unsafe delegation, tenant isolation, and controlled verification.
Tool schemas, delegated permissions, tool use, resource boundaries, and multi-step agent workflows.
Retrieval behavior, grounding, faithfulness, completeness, evidence use, and data or tenant boundaries.
Instruction hierarchy, refusals, regressions, output requirements, and adversarial cases.
Schemas, mutations, data contracts, authorization-related negative paths, and version drift.
Latency, reliability, throughput, cost, model and system comparisons, and operational thresholds.
Sammy's learning mechanism is downstream of execution. It uses approved compact non-PASS history to propose a Generation-N DRAFT with exact lineage—never an automatic policy change, and never model training on raw customer evidence.
Only compact REVIEW or BLOCK pillar summaries from a signed, completed Run are eligible.
RUN · GATE · SUITE · POLICYAn attributed decision marks the exact finding set confirmed, remediated, false positive, or accepted risk.
ORIGINAL DECISION PRESERVEDThe Control Plane maps approved history to one unambiguous signed Suite control.
COMPACT HISTORY ONLYPropose a regression, evolution, or eligible suppression-validation test with exact predecessor lineage.
NO AUTOMATIC PROMOTIONCreator, reviewer, and approver remain separate identities before any learned test becomes approved.
REVIEWED · APPROVEDNo prompts, responses, retrieved documents, tool traffic, credentials, provider configuration, or local paths.
Learning can propose the next test; it cannot weaken the locked minimum or silently change an active Gate.
Gate results, reports, synchronization, and history remain available when learning is disabled or unavailable.
Expanding the supported target types does not change the boundary: local execution retains credentials and raw evidence, while hosted reasoning receives only the sanitized metadata and compact history allowed by policy.
Credentials · target calls · raw prompts · tool output · retrieved documents · full evidence
Planning modes · independent review · governance · signatures · compact run history
Keep the deterministic minimum, add independently reviewed intelligence, and let evidence and policy govern the release decision.
Start using Prooflane free for 30 days. No credit card required.
Prooflane is beta software. Sammy recommendations and Reviewer diffs are model-assisted, may be incomplete or wrong, and never replace human review. Unsupported recommendations remain deferred or inconclusive. An approved Gate is an automated policy decision, not a guarantee of security, quality, compliance, or future behavior. The planning output shown on this page is illustrative.