Research hub · evidence shelf · public safe

Evidence for the work you want to delegate.

This is the shelf for the parts of the work that can be published: papers, lab experiments, shareable repos, methods, schemas, and public-safe reconstructions. The question underneath all of it is whether agents can become delegation-grade: predictably good under variation, bounded by evidence, authority, receipts, and review.

Raising Agents is Adrian Sanchez de la Sierra's personal public research initiative. It is not an official Zartis publication unless explicitly marked. Advisory can start with Adrian; larger private team work routes through Zartis. Private IP, client work, and non-shareable repos stay private.


Open surfaces

Method

How the work is done: preregistration, trace-level evidence, public-safe examples, agent-assisted execution, human-owned verdicts, and the line between public lessons and private work.

Free Experiments

The public intake loop. Submit a serious agent behavior problem; selected public-safe ideas may become free lab experiments, reconstructions, or Behavior Watch case files.


Delegation control stack

The repo stack exists to make delegation inspectable, bounded, and reviewable.

These components are not the homepage message. They are the proof layer behind it: a concrete way to describe work, ratify policy, run procedures, protect external actions, record evidence, and verify agent claims.

Authority

Charter defines principles, rubrics, authority tables, and policy rails for delegated work.

Procedure

Writ packages recurring work as ratified, typed, checkable procedures.

Action

AAA protects consequential actions with single-use leases validated at the effect boundary.

Evidence

Attest records tool actions and adjudicates whether claims like "done" are supported.

Memory

Keel gives delegated work an append-only event ledger and replayable decision state.

Empirics

Workbench and the Lab test where output passes while the delegated behavior fails.


How private work becomes public learning

When a real project cannot be shared, the pattern can still be taught.

Public-safe research uses reconstruction instead of disclosure: remove identifying detail, replace private data with synthetic or public equivalents, publish the behavioral pattern, and name the boundary. When the repo or project cannot be shared, it stays out of the Lab; the newsletter can still teach the general lesson without turning Zartis, client, or project IP into content.

Observed

A failure pattern appears in real engineering work, public artifacts, or a repo that can be discussed safely.

Reconstructed

The pattern is rebuilt with public, synthetic, redacted, open-source, or explicitly approved material.

Protected

Client traces, private Zartis IP, non-shareable repos, colleague-identifying detail, and engagement specifics stay out.

Submit a public-safe experiment idea →