deterministic enforcement layer · SLP-8
DEMO LIVE
The enforcement layer for AI agents.
AgenticRail holds an agent to its declared step order, refuses replays and skipped steps, and seals each completed sequence with a signed receipt.
CLIENT
your AI agent
→
WRAPPER
auth + normalise
→
GATE
harden + validate
→
CORE
enforce + seal
→
DECISION
ALLOW / DENY / HALT
→
RECEIPT
signed proof
—
TOTAL DECISIONS
every request, signed
—
ALLOWED
clean executions, in order
—
BLOCKED
refused at the gate
—
SEALED CYCLES
complete 8-step sequences
STOPPED
Recent refusals, signed
Loading…
CLEARED
Recent approvals, signed
Loading…
STEP ORDER IS LAW
Agents cannot skip steps or execute out of sequence. The rail enforces strict ordering at the infrastructure level — not by policy, not by prompt.
REPLAY IS STRUCTURALLY IMPOSSIBLE
Every request carries a unique nonce checked against the Durable Object. Duplicates are rejected before execution reaches any business logic.
EVERY DECISION IS SIGNED
ALLOW and DENY receipts are Ed25519-signed and written to R2. Anyone can verify one without asking us, and an altered receipt breaks its own signature and the chain.
SEQUENCE PROOF
A complete 8-step enforcement chain — real data from R2
Loading…
CRYPTOGRAPHIC PROOF
Every receipt is signed at write time — verify any event against the gate, live
Loading…