AI Safety Calibration: Design Partner Invitation

Why production calibration matters

The KAIROS AI safety adapter is a structural governance layer for autonomous agent deployments. Operating alongside existing model alignment, prompt guardrails, and output classifiers, it previews every proposed tool call against the session’s remaining alignment headroom before external execution commits. Two operators replaying the same trajectory arrive at the identical gate verdict.

The adapter ships with a v1 synthetic baseline calibrated against public agent-evaluation benchmarks (METR, GAIA, SWE-bench Verified, τ-bench, UK AISI, Microsoft AIRT, MAST, and OpenTelemetry GenAI). The methodology is detailed in our published debrief: Calibrating the AI Safety Adapter.

Extending that baseline to production agent harnesses requires calibration against real multi-step trajectories across design partner deployments.

Eligible environments

Four production agent archetypes match the calibration profile:

• Document-reasoning agents: Legal, clinical, and financial-operations agents that search, retrieve, draft, cite-check, and export documents with human reviewer sign-off. Reviewer edit deltas and accept/reject outcomes provide strong calibration anchors.
• Coding and DevOps agents: Autonomous software engineering agents with shell, git, file-edit, and deployment tool access gated by CI pipelines, pull-request reviews, and rollback events.
• Customer-action agents: Support, CRM, refund, identity-verification, and outbound-communication agents operating under policy-bounded tool authority and human escalation workflows.
• Browser and computer-use agents: Agents that drive web browsers or operating-system interfaces through DOM interactions, form submissions, navigations, and file operations with recorded action traces.

Deployments generating at least 30 days of multi-step trajectory volume and recording an operator or automated accept, reject, or escalate signal match the standard contribution shape. Pre-launch teams running high-volume staging or red-team evaluation suites are also eligible.

Contribution scope

A pilot contribution consists of a 30-to-90-day window of agent trajectory telemetry from one production archetype:

• Redacted trajectory traces: Normalized locally to strip prompt text, completion payloads, customer PII, and credentials before export while preserving structural tool-call metadata.
• Window labels: Reviewer- or harness-confirmed classifications on contributed trajectories (confirmed-clean, unreviewed, or excluded-incident).
• Incident context: Disclosure of any boundary violations, rollbacks, or unintended external actions within the window so baseline distributions are accurately partitioned.
• Aggregate publication rights: Permission to include anonymized, aggregate calibration statistics (never raw trajectories) in co-authored methodology collateral.

AnankeLabs supplies the local redaction tooling and a round-trip verification harness so partner engineering teams can validate every export prior to transfer.

Partner deliverables

• Environment-calibrated baseline report: A calibrated policy-positive action rate measured against the partner’s production trajectories, with confidence intervals per agent archetype.
• Trajectory headroom profile: A structural breakdown showing which tool-call chains run high on capability pressure and which remain well inside the safe interior across observe, state-gate, and action-gate enforcement modes.
• Right-to-recall: If an unflagged boundary issue is later identified inside a contributed window, the calibration cohort is re-run and any published aggregate is updated accordingly.
• Co-authorship credit: Optional co-authorship recognition on resulting calibration research.
• Priority adapter access: Early access to KAIROS AI safety adapter releases ahead of general availability.

Data protection and confidentiality

• Mutual NDA: Executed prior to sharing any technical specifications, tooling, or sample exports.
• Local redaction: Raw prompts, completions, source code, customer records, and API keys never leave the partner environment in identifiable form.
• Aggregate-only reporting: Published findings report statistical rates and confidence intervals only.
• Pre-release review: Partner teams review and approve any external output prior to publication.
• NDA specification package: The complete data specification, span mapping rules, redaction profiles, labeling guide, and replay verification harness are delivered under mutual NDA before any trajectory exchange begins.

Scope a pilot

Engineering and AI safety teams operating across any of the four agent archetypes can initiate a pilot review via the contact page. Upon initial scoping, AnankeLabs provides the draft mutual NDA, the complete partner data specification, and the local redaction and verification harness.