Alex Tavern
- Problem
- Distinctive decision
- Evidence
INITIAL COORDINATE / COORDENADA INICIAL
The same work, read from your side of the screen.
Language can be changed at any time / O idioma pode ser alterado a qualquer momento
AGENTIC SYSTEMS / VISUAL AI / CONTROL PLANES
Complex ideas become usable when every decision has an owner, every boundary can be inspected, and the interface makes the system legible.
Explore the systems ↓SELECTED SYSTEMS / ONE CONNECTED PRACTICE
Select a node to inspect the problem it owns, the decision that shapes it, and the evidence behind the claim.
REPEATED DECISIONS / ENGINEERING POSITION
Let models interpret and propose. Let typed contracts, state machines, and deterministic guards decide what becomes real.
Repeated runs, negative controls, replayable logs, and explicit limitations turn an interesting result into an engineering claim.
A difficult workflow becomes useful when people can inspect, edit, recover, and understand the state they are changing.
PAPERS / CASE STUDIES / VALIDATION REPORTS
Architecture claims are published with their measurements, negative controls, failure cases, and limits. These are engineering field notes, not peer-reviewed papers unless explicitly labeled as such.
A cache-aware cost study with explicit attribution limits and normalized provider counterfactuals.
↗ VALIDATION REPORT / 2026Witnessed actions, whispers, two-level recall assertions, and blind continuity review.
↗ CASE STUDY / 2026A deterministic retry-then-redact guard tested against informational-payload leaks.
↗ CONTROLLED EXPERIMENT / 2026A reproduction attempt that rejected the reported failure and exposed a different architectural gap.
↗FINAL DOCUMENT / CURRICULUM VITAE
The résumé comes last because the work is its context. Every role, project, and figure below traces back to verified work — where a client is confidential, the details are withheld, never the engineering.