HOW IT WORKS

How obsvr governs every AI decision

From the moment an LLM call is made to the second it reaches your ledger - every step is intercepted, enforced, signed, and committed.

THE PIPELINE

Four stages. Zero gaps.

Intercept

The obsvr SDK intercepts every LLM call at the framework level - OpenAI, Anthropic, LangChain, Vercel AI SDK. No code changes beyond a single import.

Enforce Policy

The policy engine evaluates the call against your rules: PII detection, tool allowlists, budget limits, content filters. Deterministic. No AI judging AI.

Sign

Each event is HMAC-SHA256 signed with your key. The signature covers the full payload - prompt, response, policy result, timestamp. Any tampering breaks the chain.

Commit to Ledger

Signed events are committed to an append-only ledger. Merkle-tree anchored. Optionally mirrored to S3 Object Lock and GitHub for independent verification.

THE DIFFERENCE

Without obsvr vs. with obsvr

Without obsvr

LLM calls happen with no visibility
PII flows freely into third-party APIs
No record of what the AI actually did
Policy enforcement is probabilistic at best
Audit prep takes weeks of manual log assembly
No way to prove compliance after the fact

With obsvr

Every call intercepted before it reaches the LLM
PII detected and redacted in real-time
Cryptographically signed audit trail for every event
Deterministic policy enforcement - no guessing
Compliance evidence exports in one click
Non-repudiable proof of every decision

DATA HANDLING

What obsvr sees vs. what it doesn't

What reaches obsvr

Signed event metadata (timestamps, model, token count)
Policy evaluation results (PERMIT/DENY + reason)
Content hashes (SHA-256 of prompt and response)
Merkle tree position and chain anchors

What stays local

Raw prompt content
Raw response content
API keys and credentials
User session data and PII

Decision timeline

Every LLM call and agent run, fully audited and secured

audit_pipeline.log01/07

Intercept

Proxied at init - your code stays untouched

proxied at init · zero call-site changes

What you get

Governance tooling for AI systems

Policy Engine & Management

Every LLM call and agent run is checked against your rules before and after the model call - without changing your integration. The rules are versioned and content-addressed, and every change to them is a signed, diffable event in the same Merkle chain as your logs.

Shadow mode
Org policy packs & inheritance
Signed policy-change history

Policy EvaluatorPROMPT RECEIVED

Keyword Block"SSN" in promptBLOCKED

PII Pre-callEmail detectedREDACTED

Topic Allow-listFinance ✓ALLOWED

Post-call Scan-SKIPPED

Evaluated in 0.3msBLOCKED

Agentic Governance

obsvr extends call-level policy enforcement to full agent runs. Every tool invocation, every step, and the final run output are evaluated against your policy - so multi-step agents are as governable as a single LLM call.

Tool restrictions
Step limits & runaway-loop detection
Kill switch

Agent Run Tracerun_cr_a1b2c3d4

CrewAI

3 / 10 stepsENFORCED

Investigate & Respond

When a policy fires or an agent misbehaves, obsvr turns the signed events into a structured incident - root cause, blast radius, and the exact next step - so you are not grepping logs at 2am. Pivot on any user, tool, or model, then export a signed incident report.