AI Safety & Governance

An agent saying it fixed something is not proof,
three checks at three time horizons is

Ground-Truth Verifier checks every remediation against reality at three time horizons: T+5 minutes, T+1 hour, and T+24 hours. The check compares the SLI value to the SLO target. If the symptom is gone and stays gone, the action is verified. If it comes back, trust drops and the action is flagged for review. Self-reported success no longer counts.

Get Started Talk to Sales
app.novaaiops.com / ground-truth
● LIVE
3
Verification windows
Auto
Trust score update
Real
SLI vs SLO comparison
Logged
Result in agent ledger
Three-Window Verification

T+5m, T+1h, T+24h, different problems, different windows

Different remediations regress at different timescales. A pgbouncer restart that fails comes back in seconds. A cache invalidation that was wrong shows up at peak traffic an hour later. A schema change that broke a downstream report only matters at the next batch run. Three windows let us catch all three.

  • T+5m · acute regression: catches "fix did not actually take", the most common class
  • T+1h · sustained recovery: catches "fix held briefly then degraded under load"
  • T+24h · downstream impact: catches "fix worked here but broke a batch job that runs at 3am"
app.novaaiops.com / ground-truth · windows
What Is Verified

Each remediation declares its own pass condition

A remediation does not pass just because nothing else broke. The agent declares the SLI it intends to fix and the threshold it expects. The verifier compares the live SLI value at each window to the threshold. Pass requires the SLI to be at or above the threshold. Anything else is a regression, not a "well, kind of."

  • Declared pass condition: every action carries a "this should bring SLI X above target Y" statement, recorded in the bundle
  • No vague success: no "looks better", pass is a numeric comparison against the SLO, not a vibe check
  • Custom verifiers per tool: some actions need bespoke checks (e.g. "table reindexed" verified by pg_stat_user_indexes)
app.novaaiops.com / ground-truth · pass-condition
Trust Feedback Loop

Verified actions promote, regressed actions demote

The verifier is the ground truth that drives the Agent Ledger's trust scores. Each pass adds to trust, each regression subtracts. Multiple regressions in a row demote the agent's autonomy tier automatically. The trust system is not based on the agent's own assessment of its work; it is based on whether reality stayed fixed.

  • Trust deltas published: every verifier outcome publishes a delta to Agent Ledger; visible per-agent on the fleet page
  • Auto-demote on streak: three regressions in a row demote the agent one autonomy tier without operator action
  • Auto-promote with caution: promotion still requires operator review, verifier wins are necessary, not sufficient
app.novaaiops.com / ground-truth · trust
Use During Postmortems

Every regression has a story

When you write a postmortem, the verifier history is your timeline. It shows which agents acted, which actions claimed to fix the problem, and at which window the fix held or broke. The page is filterable by incident, agent, and outcome so you can pull every verifier event for one incident in one click.

  • Filter by incident: every verifier check carries an incident id when one was active; pull them all for a postmortem
  • Outcome facets: pass / flap / regress / pending, pivot the table to see only regressions for the last quarter
  • Export to bundle: verifier rows are part of the Decision Bundle export, signed and chained with the rest
app.novaaiops.com / ground-truth · postmortem
Video walkthrough coming soon

Subscribe to Nova AI Ops on YouTube for demos, tutorials, and feature deep-dives.

Trust earned by outcome, not assertion

Every verified remediation pushes trust up. Every regression pushes it down. The agents that survive are the ones whose fixes hold.

Get Started Request a Demo