3:07 a.m. · Checkout is down · $18,400 lost every minute

An AI can fix it in seconds.
RecoveryPact makes sure the fix doesn’t cost more than the outage.

Two AIs negotiate every recovery: one speaks for the database, one speaks for the business. A separate safety authority decides what may actually run, and recovery only counts when customers can buy again.

Watch the replays Try to break it What each sponsor is worth Source and raw traces
NVIDIA Nemotron 0 unsafe in 120 incidents Zetaris worth 9 verified recoveries Meterless 95/95 answers signed Cursor incident review in the IDE

Below are real recorded runs from our benchmark, replayed side by side. Same incident, same live database, same AI model. Pick one.

0 s

What the same 24 incidents would have cost

Same Nemotron model, same live database, same incidents. This is what each approach actually did, counted from the recorded runs.

Typical AI agent
  • 3 card settlements aborted mid-flight, $12,500 each: customers at risk of being charged twice
  • 6 quarter-end closes killed, 95 minutes of finance work each: 9.5 hours lost
  • 3 customer checkout sessions killed on a tampered plan
  • 3 sessions killed that no company system could identify
  • 6 times it said “fixed” while checkout was still broken
AI agent + the same safety rules
  • 6 quarter-end closes still killed: 9.5 hours of finance work. Safety rules can't see the change calendar.
  • 3 times it said “fixed” while checkout was still broken
  • Right call in only 6 of 24 incidents
RecoveryPact
  • 0 settlements aborted, 0 closes killed, 0 customer sessions touched
  • 0 false “fixed” claims: recovery only counts when customers can buy again
  • Right call in 24 of 24

The price, stated plainly: decisions take about a minute (78 s median vs 24–28 s), so more checkouts fail during the incident. One aborted settlement is a double-charge incident and a regulator's letter; a minute of degraded checkout is a dashboard blip.

Where it fits: anywhere one wrong kill costs more than a minute of waiting. Payments and card settlement, banking batch and ledger jobs, finance close, marketplaces at peak.

Every sponsor is load-bearing. We measured each one.

We took each component away, or attacked it, and re-ran the incidents. Every number below was recomputed from the raw traces by an independent validator.

The reasoningNVIDIA Nemotron 0 unsafe across 120 incidents on five Nemotron configurations, and 24/24 right calls on Nemotron 3 Ultra. Every AI here is a Nemotron model. The Nemotron family results The business truthZetaris 9 → 0 verified recoveries when Zetaris is switched off. The pact stays safe, with 0 unsafe actions, but it can't fix anything without the business context. What federation is worth Runtime + memoryMeterless 95/95 Nemotron answers carry a signed receipt we verify offline. A forged memory planted in H-MEM was overruled 3 times out of 3. Signatures and the memory attack Post-incident reviewCursor 5 tools A read-only MCP server, built with Cursor, so an SRE can ask why a run waited or refused, and get the answer from the evidence. Review inside the IDE

Why this is new

Two AIs that disagree on purpose

The Database AI wants the database healthy fast. The Business AI protects revenue, customers and money in flight, and sees company records the database cannot. They argue until the evidence settles it.

Agreement is not permission

Even when both AIs agree, a separate rule-based safety authority checks the exact action against the live system. It approves once, for two minutes, or refuses. It cannot be talked around.

The business defines “fixed”

Typical agents declare success when the database looks healthy. RecoveryPact only declares success when checkout is back above target and nothing protected was harmed.

It learns, carefully

After each incident an Auditor AI writes the report and a lesson. Memory can stop a mistake from being repeated. It can never authorize an action on its own.

How one recovery works

  1. ObserveDatabase signals, plus company records through Zetaris
  2. ProposeDatabase AI picks from actions that were actually observed
  3. ChallengeBusiness AI demands evidence, weighs costs, accepts or escalates
  4. AuthorizeRule-based authority approves one exact action, or refuses
  5. VerifyCheckout back above target, nothing protected harmed
  6. LearnAuditor AI writes the report and stores the lesson in memory