What GeraWitness Will Be: The Safety Layer for Agent Actions
Published 21 April 2026 · 6 min read
The product in a paragraph
Agents are going to execute real-world actions on behalf of users. Some of those actions will be high-stakes: large payments, medical bookings, legal commitments, estate decisions. Trusting the model alone for all of those is the wrong risk profile. GeraNexus transactions flagged as high-risk route through GeraWitness, where a trained reviewer sees the proposed action, the context, and the consent chain, and approves or rejects before commit.
Risk tiers
- T0 — auto: standard bookings, low-value transactions. No human review.
- T1 — sampled: random 1-5% sample reviewed for calibration. Does not block commit.
- T2 — mandatory pre-commit: high-value, medical, legal, irreversible. Human must approve before commit.
- T3 — hard stop: categorically refused classes. Model cannot override.
Who the reviewers are
Full-time Gera trained reviewers in multiple time zones, plus a vetted contractor pool for surge capacity. Reviewers see the proposed action, the user’s consent token, the agent’s reasoning trace (where available), and the relevant category- specific policy. They approve, reject, or ask for more context (which delays commit but doesn’t kill it).
What GeraWitness is not
- Not a dispute resolution service. Disputes happen after the fact; GeraWitness is pre-commit.
- Not content moderation. Moderation is model-level; GeraWitness is action-level.
- Not training. The reviewer signal does not train the underlying LLMs.
How it fits with GeraNexus
GeraNexus emits a risk tag with every transaction. Tier-T2 and above gate through GeraWitness. The review result is added to the signed receipt trail, so downstream audit shows a reviewer approved the action.
Timeline
- Q3 2026 — policy drafting, tier taxonomy, reviewer workflow design.
- 2027 — pilot within the Gera network only.
- 2028 — open to third-party agent platforms via SDK.
- 2030 — general availability with compliance-grade SLAs.
How to follow along
Design drafts on this blog and /research. We particularly want to hear from content-moderation and trust-and-safety practitioners — a lot of the lessons transfer but some don’t. Waitlist at /#waitlist.
Help design agent safety that scales.
Join the waitlist