iXentBench™

Automated reasoning analysis via LLM-as-a-Judge

Screens Prompt Injection & Strategy for language, safety and relevance

Greys out Benchmark, Laboratory and Arena in the platform so nobody starts a new game while you deploy. Games already running are not interrupted. Cogito Ludens rooms are closed separately, from the Cogito Ludens tab.

Room total = board + one AI price per model in play. Cost is what the labs charge us; margin is what is left after that. Game structure (seats, players, humans, AIs) is fixed in code and cannot be changed here.

Named shortcuts to replay.ixentlabs.com. Look up a certificate above to create one.

⚠ Categories: frontier (commercial APIs), open (open source), custom (unverified). Provider is used by the SDK to select the correct library.

Cost fields. Prices are USD per million tokens, as billed by the lab. Out $/M covers output and thinking tokens — most labs bill them at the same rate. Prompt caching is never assumed: any discount it produces is margin in our favour, never a figure we price against. TPS is the model's observed generation speed and Sec/move its average time per move: together they cap what a model can spend within a room's clock. A model whose average exceeds a room's clock would lose turns there — the room editor flags it. Rooms makes the model selectable in Cogito Ludens — it requires all four cost fields and a lab name. Reviewed is stamped by the server whenever a price changes; you never type it.

Resend free plan: 100 emails per day and 3,000 per month. Over the limit nothing is lost — emails stay pending until they are retried.

Cloud Scheduler — terminal, exceptional cases only

Cloud Scheduler calls the backend every 15 min (free). Day to day, use only the switch above. Pause Cloud Scheduler only in exceptional cases (changing region, rebuilding the service) and resume it right after.

Stops the automatic retry

gcloud scheduler jobs pause ixentbench-email-retry --location=us-central1

Resumes the automatic retry

gcloud scheduler jobs resume ixentbench-email-retry --location=us-central1

⚠ While Cloud Scheduler is paused, this card still shows the interval above, but nothing is retried: “Last automatic run” stops advancing.

Sent automatically once an hour to every active subscription that renews in 17 days or less, with the amount of the next invoice. Use this to send one now for a specific subscription — to test it, or if an automatic one failed.

⟳ Retry continues from the step that failed: it never sends a second acknowledgement or a second credit note. NEEDS REVIEW means the refund could not be done safely by code: do it from the Stripe dashboard (credit note with refund, memo from template 6.5) and then use ✔ Resolve to record what you did.

Refunds required by law or decided by us (for example, compensation for an incident). Withdrawals and failed payments have their own flow — do not refund them here. Every refund is a credit note that Stripe emails to the customer, with the reason in the memo.