ARQUÍA-EVAL-001

Invitación personal de José María Garbini: descargá y evaluá la semilla universal de ARQUÍA · Semilla experimental abierta a personas, investigadores y agentes de IA; se agradecen críticas y contraejemplos.

ARQUÍA-SEED-001 — universal seed (experimental)

Download the universal seed (.md) · Release and SHA-256 checksum · Guía en español

A portable, vendor-neutral invitation for AI agents, researchers, and developers to inspect, challenge, and test ARQUÍA's 0-1-2-3 ethical governance proposal. Includes a Markdown protocol, a local Python reference simulator and reproducible tests. No account or affiliation is required. The reference is experimental: reading or downloading is not adoption; passing the tests is not evidence of production safety. Independent criticism, negative findings and counterexamples are welcome.

En español: la semilla universal ARQUÍA-SEED-001 se puede descargar gratis y analizar con cualquier IA. Es una propuesta experimental, no una instalación automática ni una certificación. Cómo probarla e informar resultados.

An offline, falsifiable experiment about an agent's authority to act. Can revoked permission reappear after a receipt or recovery? Can an implementation conceal an unauthorized effect? A valid counterexample matters as much as a passing result.

Package version: VERSION. Check GitHub Releases for publication status. No ARQUÍA affiliation, endorsement, paid model, API key or account is needed to run or criticize it. Passing means matching a bounded synthetic contract. It is not a safety certification, production validation or approval of the 0-1-2-3 theory.

First run

Python 3.10+ and its standard library. Download a fixed release and inspect its sources. Linux/macOS (python3); Windows uses py -3:

python3 verify_package.py
python3 runner.py --evaluator YOUR_NAME --relationship unspecified > historical.json
python3 community_runner.py --evaluator YOUR_NAME --relationship unspecified --markdown community.md > community.json
python3 community_runner.py --example-jsonl > example.json
python3 verify_report.py community.json --cases suites/community-v1.json
python3 check_all.py --output checks

The historical runner retains 29 cases from v0.1.1 (v0.1.0 has 28). The separately versioned community battery has 34 cases. check_all.py checks five historical and ten additional deliberate faults, temporal sensitivity, source preservation, schemas, reports, deterministic generation and packaging. An expected negative-control exit 1 is success of that sensitivity check; an exception or timeout is never a successful detection. See QUICKSTART.

What the results mean

• Historical reports use candidate snapshots; dishonest snapshots remain a limitation.
• Community reports use an evaluator-owned synthetic effect ledger. No real endpoint is contacted. Only the modeled command routes are visible.
• A case may correctly detect an effect that already happened. It must not be described as preventing that effect. COMM-015 and historical C05 illustrate this.
• Contract matches, selected-fault sensitivity, observability, inconclusive cases and measured runtime remain separate. There is no global safety score.

The independent example implements an ordinary capability/ACL policy without importing engine.Gate or the evaluator. It was written for the project, so code independence is not evaluator independence. Agreement with this comparator establishes no advantage unique to ARQUÍA. No complete production ARQUÍA architecture is evaluated here.

Connect, challenge and contribute

ADAPTER defines portable JSONL messages, size/time bounds and failure behavior. Arbitrary Python imports are disabled in the legacy CLI. External code requires express acceptance and the demonstrated Linux isolation boundary; it is never silently run without that isolation. The shipped example works across platforms without extra libraries.

Keep new cases separate from cases.json. COUNTEREXAMPLES explains hypotheses, deterministic seeds and minimized traces. Publish criticism or results under your own control, then optionally link them in issue #1. No favorable result or author approval is required for external criticism.

Read PROTOCOL, CONTRIBUTING, REPORT_TEMPLATE, SECURITY, CHANGELOG and MIGRATION. External evidence distinguishes claims, reproducible runs, implementations and critiques without treating them as ranks.

Historical releases, original/, case bytes and historical reports are retained. AUDIT_COMMUNITY records the initial audit and PROVENANCE credits earlier work. SHA-256 identifies bytes; it does not authenticate an author. Software and original package documentation use the MIT license; linked third-party material retains its own rights.

Project owner: José María Garbini. ARQUÍA is the originating governance research project. Evidence, proposed mechanisms and broader theory are distinguished in SCOPE. Guía en español.

Community implementation review / Revisión de implementación.