Change Control · AI Change-Control System
A new model, a tweaked prompt, a fresh integration every change is diffed, routed to the right approvers, and logged before it goes live.
billing-agent · prompt.diff - refund up to $500 without approval + refund up to $50; escalate above policy check: finance-strict → requires risk + finance reviewers: risk ✓ finance ✓ eng pending status: awaiting 1 approval not yet live
Audit-ready for teams at
Prompts, model routes, tools, and permissions are versioned. See exactly what changed, line by line.
Policy decides who signs off. Risk, finance, or eng are pulled in automatically by the rules that apply.
Every approval, block, and rollback is signed and streamed to your SIEM. Evidence without the screenshots.
Instant rollback
Every deployed version is retained. If a prompt regresses or a model misbehaves, roll back to the last approved state immediately and the rollback is logged like any other change.
Versioning & rollbackFrameworks built in
Change-control records map cleanly to SOC 2, ISO 27001, and the EU AI Act. Export a control narrative for any agent and any window your auditors get the evidence in their format.
See report templatesThe riskiest deploy in most AI programs is the tiny prompt tweak that ships at 6pm. Change Control kills that failure mode by treating every surface below as a reviewable unit no exceptions list.
Any character, system prompts included. A one-word edit can move a spend limit or drop a guardrail, so every character diff gets a reviewer.
Provider, model version, temperature and friends. Swapping to a cheaper route changes behavior; it goes through review like anything else.
Adding a tool is adding a capability, and it is reviewed like one with the scopes it requests spelled out in the diff.
Data boundaries, spend limits, access grants. Widening any of them routes to the owners of whatever sits behind them.
Policy edits are the highest-risk change in the system, so they route to the strictest reviewer set by default. Nobody quietly rewrites the rules.
Environments & progressive rollout
Changes move draft to staging to canary to live, and every promotion is itself a policy-gated change with its own approvers. Canary runs the new version for a bounded slice of traffic and halts automatically on regression signals live means it earned it.
Environments & rolloutEvidence on demand
Control narratives are generated per agent and per window; every record carries the diff, the policy version, the approvers, and timestamps. Retention is configurable to your regulatory window, and exports ship as PDF for auditors or JSONL for your GRC tooling. The framework mapping above is the summary this is the record-level detail underneath it.
Narrative templates How the log stays tamper-evident$ pactflow export --agent billing-agent \ --window 2026-Q2 --format control-narrative collecting 214 changes · 214 approvals · 3 rollbacks mapping SOC 2 CC8.1 · ISO 27001 A.8.32 · EU AI Act art. 12 ✓ billing-agent_2026-Q2.pdf written (jsonl alongside)