Sebastián · public engineering evidence · RUMBO IA

AI Systems & Agent Reliability

I build and investigate agent systems where an accepted decision is not treated as permission to mutate forever: authority is bound to state, effects are read back, evidence is verified, and stale decisions fail closed before the effect boundary.

agent reliabilitydeveloper toolingfail-closed executionstate revalidationeffect readbackasync ownershipMCP / pluginsverification

Featured developer-facing proof

Two public assets are merged into the same dependency-free Python reference repository and tied to executable tests and public CI.

MERGED · VERIFIED · developer quickstart

5-minute reliability quickstart

Walks through INTENT → AUTHORITY → MATERIALIZATION → READBACK → RECEIPT → VERIFICATION and includes a deterministic fail-closed drift path.

PR #10 provenance

MERGED · VERIFIED · technical deep dive

An accepted agent action is not necessarily executable

Explains why DECISION_ACCEPTED != EXECUTION_SAFE, traces the exact state-digest guard before execute(...), and proves the blocked target remains unmodified.

PR #11 provenance

PUBLIC CI PASS · current main
66 tests · Python 3.11 / 3.12 / 3.13

Main SHA 000e5723737605dda30e68646e28f54448877967. The post-merge workflow installs from source, runs the quickstart, runs the unit suite and verifies installed imports outside the checkout.

GitHub Actions run 34005553526

Reference boundary. This is a small public reference implementation and controlled synthetic demonstration. It is not presented as an enterprise production deployment, security certification, benchmark result or customer adoption proof.

Authored upstream OpenAI contributions

These are public contributions to OpenAI repositories. Their current state is shown explicitly; open PRs are not represented as accepted or merged work.

OPEN · MERGEABLE · authored code PR

Agents SDK PR #4868

Makes Responses-chain and history ownership explicit across compaction, destructive mutation, cancellation and stale publication boundaries.

Current head 5702cf2…. PR body reports 69 focused memory tests and 217 total memory tests passing, plus Ruff, mypy and Pyright checks. The upstream Actions run is action_required with 0 jobs, so it is not labeled CI PASS or FAIL.

OPEN · MERGEABLE · authored code PR

Codex plugin PR #730

Adds opt-in --prompt-file-consume with deliberate READ → DELETE → DISPATCH ownership ordering. The PR reports 3/3 focused runtime tests and 28/28 non-runtime tests passing; it does not claim a full Windows-suite pass.

OPEN · MERGEABLE · authored docs PR

OpenAI Go PR #885

Clarifies Bedrock Mantle model-family API roots while preserving existing SDK behavior. Documentation-only; no SDK behavior, authentication, signing, generated code or exported API changes are claimed.

Affiliation boundary. Contribution, issue authorship and implementation review do not imply OpenAI employment, maintainer status, endorsement or formal acceptance.

Reliability method

INTENTAUTHORITYPREFLIGHTEXECUTIONREADBACKFALSIFICATIONCLOSURE

Bind authority to current state

An approval is evidence about one intent evaluated against one state, not a timeless capability token.

Fail before the effect boundary

Revalidation must happen before the side effect. Compensating after an unsafe mutation is a different guarantee.

Read back the effect

prepared != executed and executed != observed. Success claims require evidence from the target boundary.

Preserve negative evidence

Invalid environments, stale findings and failed execution routes stay in the audit trail instead of being rewritten as successful attempts.

Selected investigations

Async cleanup ownership · Agents SDK #4747

Investigated the teardown boundary where registry ownership can disappear before backend cleanup settles.

Startup cancellation ownership · Agents SDK #4749

Investigated resources created before registry ownership is established and cancellation arrives during the transfer window.

Fail-closed CI ordering · Python SDK PR #3780

Reviewed prerequisite ordering and cancellation semantics; a later stale blocker was publicly withdrawn after re-reading the current head.

Current public links

GitHub · Verifiable Agent Control Plane · RUMBO IA · Demo & Enterprise Readiness Kit

Overall claim boundary

Public evidence is labeled by state. REFERENCE_IMPLEMENTATION != PRODUCTION_SYSTEM; OPEN_PR != MERGED_CONTRIBUTION; CONTRIBUTION != AFFILIATION. No paid-customer adoption, enterprise production scale, security certification or OpenAI employment/endorsement is claimed here.