For Frontier Model Providers and Alignment Research Labs
Frontier alignment research needs adversarial corpora that reflect real-world failure modes — not synthetic benchmarks generated by the same models being evaluated. Sgraal's fleet processes millions of preflight decisions monthly. We are structuring that signal into research-licensable datasets, on terms that preserve tenant privacy.
Honest disclosure: this is on our long-term roadmap, not yet built. No demo calls, no follow-ups until there's something tangible to show. Email above just adds you to a low-volume list (~2 emails/year max).
A research-licensable derivative of the fleet-wide attack-and-defence record. Sanitised, scope-limited, with strict tenant-privacy guarantees. The legal and structural work is what is currently being built.
Labelled training signal for safety fine-tuning — preflight decisions, repair plans, healing trajectories. Designed to complement, not replace, in-house alignment data.
Differential-privacy-noised aggregation of attack vectors observed across the fleet. The shape of the signal, not the content, intended for alignment researchers studying emergent failure modes.
Structured access path for labs that want to co-author benchmarks, contribute new attack vectors to R6+ rounds, or run blind evaluations against the held-out R12 corpus.
The concept items above build on top of these production features.
239 adversarial cases on GitHub today. Apache 2.0 licensed. Drop into your eval harness right now.
See the research page →60 hard adversarial cases held privately for benchmark integrity. Academic and research labs can request blind evaluation access.
See the PARKED cases →Authenticated transparency of scoring engine configuration with SHA-256 config checksum. Available to licensed customers via a full API key (the endpoint is not exposed to unauthenticated readers).
If frontier alignment research or AI safety dataset partnerships is your domain, leave a note. We email back when there's something to show, not before.