Built to be run by agents.
Under written rules.

• 01

  Rules decide. Models explain.

  Prices come from the catalog. Limits come from your policy. Max prices and risk outcomes come from written rules. A language model never invents a number you pay or a limit you hit.

• 02

  Three outcomes.

  Every decision ends one of three ways: the agent acts, the agent asks, or the agent hands the decision to senior review with its notes.

• 03

  Inside its limits alone. Beyond them, senior review.

  An agent can pause, hold, draft and recommend, and it acts alone inside its written limits. The reviewer is a different agent from the one that proposed the decision, and it records its reasoning.

• 04

  Everything on record.

  Every decision is logged with its inputs and its reasoning: the playbook version, the tier, the rule, the outcome and the senior review, where there was one.

1. Limit 01Money out

   Capacity commitment AllowedCeiling

   Refund or credit AllowedCeiling

   covered bycontract or prepayment

   01

   Money out.

   No capacity commitment is made to the capacity layer unless a customer contract or prepayment covers it, up to a ceiling the owner sets. No single refund or credit goes above a ceiling the owner sets.

2. Limit 02Rule changes

   1. ProposesAny agent
   2. EvaluatesSenior review
   3. Takes effectAfter review

   digestevery change

   limits, decision rightsowner only

   02

   Rule changes.

   An agent may propose a change to a playbook, a price or a term. The change takes effect only after senior review has evaluated it, and the owner sees every change in the digest. Limits and decision rights change only by the owner.

3. Limit 03Stop

   • Any agent Run Pause
   • Any desk Run Pause
   • All outbound actions Run Pause

   03

   Stop.

   The owner can pause any agent, any desk or all outbound actions at once.

4. Limit 04Record

   digestowner's digest · weekly

   anomalyto the owner, at once

   04

   Record.

   Every decision is logged with its inputs and its reasoning. The owner receives a short digest, and anomalies escalate to the owner immediately: a spend spike, repeated overrides, a security event.

One rule set. Every account. Before anything starts.

HOLD    if launch.gpus > account.recent_peak × k  and  account.age < a
VERIFY  if payment.country ≠ signup.country
HOLD    if quota.auto_increase  and  signin.risk = elevated
BLOCK   if workload.signature = proof_of_work
ALLOW   otherwise

1. 01Agent

   Observe

   A reviewer agent grades every conversation and action against written criteria.

2. 02Agent

   Propose

   The agent drafts a change to its own instructions, with the cases that prompted it.

3. 03Automated

   Replay

   The draft runs against past cases, simulated cases and attack cases. No customer is involved.

4. 04Senior review

   Review

   Senior review reads the diff and the results, then approves, edits or rejects. The reviewer is never the agent that proposed the change.

5. 05Automated

   Stage

   An approved change runs in shadow, then for a share of cases, then for all. One step rolls it back.

6. 06Automated

   Record

   The change appears in the changelog with its release stage, and in the owner's digest.

• 01 Read your workloads. Agents work from account metadata, usage, tickets and the telemetry you can see yourself. They do not open your containers, volumes or buckets.
• 02 Train Fantasti models on your data. Fantasti does not train models on your tickets or your usage. The model provider is a sub-processor, listed on the Trust page.
• 03 Pretend to be a person. Every agent identifies itself as an AI in its first message and in its signature.
• 04 Raise their own limits. Limits and decision rights change only by the owner, and are enforced outside the model.
• 05 Take instructions from data. Text in a ticket, a log line or a file name is treated as data, never as an instruction.