The deterministic control plane for regulated AI decisions
Point your existing OpenTelemetry exporter at Mesh — the same line you send to Datadog — and every AI decision becomes a tamper-evident, offline-verifiable record, governed by a deterministic gate that can block a bad decision before it executes. Lending-grade auditability and structural explainability, not “ask the model to explain itself.”
OTEL_EXPORTER_OTLP_ENDPOINT=https://vasperamesh.com/api/otelCapture every decision. Govern the ones that matter. Prove all of it.
Connect in one line, see what your agents actually do, gate the decisions deterministically, and hand a regulator proof they can verify offline.
Connect in one line
Point your existing OpenTelemetry exporter at Mesh — the same line you send to Datadog. Every agent decision is captured: tool calls and structured outputs, not just generations. No SDK rewrite, no per-decision code, no rip-and-replace.
See what your agents actually do
A system map, auto-derived from the ledger, shows every agent, action, and model — and flags the decisions the LLM is making with no deterministic gate checking it. Understand your system before you try to govern it.
Govern — and block — the decision, not the model
A deterministic gate runs your policy over the actual decision values — a refund amount, a loan approval, a position size — and returns pass, hold, or escalate. Start in shadow, move to advisory, then enforcing: in enforcing mode a hold or escalate is blocked before it executes and routed to a human, inline in the request path. Structural, reproducible reasons; never the model narrating itself.
A hash-chained, externally-anchored ledger
Every decision is one immutable, hash-chained row, stamped with the policy version in force and witnessed on Bitcoin (or your WORM store). Only a 32-byte hash leaves your boundary — no content, no PII. No row, no action.
Verify without trusting us
A standalone offline verifier — importing nothing from our code — re-derives every hash from the published spec and cross-checks the external witness. Your examiner runs it, no login and no vendor dashboard required.
Lending-grade, out of the box
A credit-decision gate derives the ECOA / Reg B principal reason codes deterministically and generates the §1002.9 adverse-action notice from the recorded row — never “the model said so.” Framework-mapped evidence packs for Fannie Mae LL-2026-04 (effective Aug 6, 2026), Freddie Mac Bulletin 2025-16 (§1302.8), ECOA / Reg B, EU AI Act, ISO 42001, and NIST — every control backed by the cryptographic ledger, not self-attestation.
Don't take our word for anything
This is our live production ledger, verified by a standalone script that imports nothing from our code. Run it yourself, right now.
# Fetch the verifier (200 lines, zero imports from VasperaMesh)
$ curl -sO https://vasperamesh.com/verify-mesh-ledger.mjs
$ npm install canonicalize # RFC 8785 JSON canonicalization — the only dependency
# Re-derive every hash, check every link, cross-check the Bitcoin witness
$ node verify-mesh-ledger.mjs --base https://vasperamesh.com
PASS — chain intact.
head: seq N, rowHash 3b9c949e5b25d09f…
anchors: all consistent with the external witness.
latest anchor: verifiable on Bitcoin with any OpenTimestamps client.Every row commits to the previous one; every head is periodically stamped into Bitcoin. Tamper with one field of one historical decision and the verifier names the exact row — no matter who did it, including us.
Get a tamper-evident audit trail of the AI you already run
Shadow mode starts with a base-URL change — no gating, no integration project, no risk to production. When you and your examiners are ready, graduate to advisory, then enforcing — where the gate blocks a bad decision before it executes — one workflow at a time.