03 / AGENT WORKFLOWS · EXECUTABLE EVIDENCE
Continuum.
Save a system investigation so another engineer can reproduce the simulation.
6 services · 180 requests · 16 requests/sec · seed 42 · no scheduled fault
Candidate change
Edit a proposal yourself or ask a local model. Only service worker counts can change; arbitrary code is never executed.
Memory that earns its place
The capsule contains the exact scenario, proposal, limits, seed, engine version, before/after traces, and computed check results. Changes to the editable inputs invalidate the review. Successor capsules retain the previous digest; they do not silently rewrite history.
Hashes establish content integrity, not authorship or scientific truth. Browser checkpoints are local storage, not a durable agent execution service. JSON and Markdown exports are portable files, not a live connection to another agent.
Where AI belongs
AI helps propose a change. A bounded simulator produces the evidence. A human decides what should proceed. The local adapter uses structured model outputs; the provenance follows the separation of artifacts, activities, and actors described in W3C PROV. This prototype does not claim AGI, automated deployment, or full PROV compliance.
The person behind the project
A note from Luis.
Reproducible failure modes are part of my AI evaluation work. Here, the useful handoff is a set of inputs, a proposed change, and a simulation someone else can rerun. Reproducing the model does not prove a deployment is correct.
Who it helps
Engineers handing a system investigation to another person or agent.
Try this
Review a proposed change, run the simulation, and export its capsule so the inputs can be reopened.
Build the underlying system scenario