AOS Kernel

Public demonstrator version: 0.1.1

AOS Kernel is a model-agnostic decision-evidence kernel. It converts bounded signals into deterministic operational decisions with replayable evidence. Decision-time evidence is the current claim; runtime assurance is the direction, not a shipped capability.

signal -> policy -> PASS / WARN / BLOCK -> JSONL record -> SHA-256 hash

This repository is the curated public kernel release surface. It contains the reference kernel, a minimal CLI, Trusted Output v0, Lean proof surface, contract tests, and compact evidence examples needed to inspect the public demonstrator.

Workflow products built on top of the kernel live separately, for example RafineriaAI/aos-workflow-gate.

After v0.1.1, this repository is in scope-locked maintenance mode. See docs/MAINTENANCE_POLICY.md.

What It Does

AOS does not replace scanners, guardrails, reviewers, operators, or final decision-makers. It sits after upstream tools and before downstream workflow use.

scanner / CI / evaluation signal -> AOS Kernel -> decision + evidence

Kernel capability	Public artifact
Deterministic PASS/WARN/BLOCK decision	core/aos_public_core.py
Replayable evidence packet	build_signal_evidence() and tests
Integrity anchor	SHA-256 digests in evidence and Trusted Output v0
Tamper detection	verify_signal_evidence() and aos trust verify
Abstract formal surface	lean/AOSPublicCore.lean
Publication boundary	SCOPE_OF_PROOF.md, COMMERCIAL_BOUNDARY.md

Practical Use Case

A scanner, eval, or CI job can produce a bounded signal that needs a repeatable handoff decision before downstream workflow use. AOS Kernel does not decide that an artifact is safe. It turns the bounded signal into a deterministic PASS / WARN / BLOCK decision and a replayable evidence record.

The practical aha moment: later reviewers can inspect the exact input digest, policy id, decision, and wrapper verification status instead of reconstructing a one-off CI judgment from logs.

Try It

python -m pip install -e .[dev]
aos demo --output-dir .tmp/aos-demo

Expected output:

AOS demo completed.
Verdict: WARN
Decision: REVIEW_REQUIRED
Signals: 1
Evidence: .tmp/aos-demo/check-record.jsonl
Summary: .tmp/aos-demo/check-summary.md

Verify the unsigned Trusted Output wrapper:

aos trust emit \
  --record .tmp/aos-demo/check-record.jsonl \
  --output .tmp/aos-demo/trusted-output.json

aos trust verify \
  --input .tmp/aos-demo/trusted-output.json \
  --record .tmp/aos-demo/check-record.jsonl

Current public verification status:

signature_status: UNSIGNED_NOT_OFFICIAL
official_aos_output: false

That is expected. Public Trusted Output v0 is structure-checkable and record-checkable, but not an official signed RafineriaAI/AOS verdict.

Validate

python tools/run_validation_gate.py --standard --skip-install

The standard gate runs formatting/lint checks, type checks, tests, integrity verification, and the public Lean surface build when Lean is available.

For release-surface audit precheck:

python tools/audit_precheck.py

What This Proves

The public kernel demonstrates that a bounded input can be converted into a stable decision and replayable evidence record, and that tampering with the record or wrapper is detectable by the included verifiers.

It does not prove production readiness, clinical or regulated-use readiness, security of a target repository, absence of vulnerabilities, compliance, SLA, commercial ROI, or Python-Lean refinement correctness.

Repository Boundary

RafineriaAI is the publishing brand and ownership boundary. AOS is the kernel. Products are concrete applications built on top of that kernel.

This repository is source-available and proprietary. It intentionally excludes private production policies, calibration logic, enterprise integrations, signing keys, deployment hardening, customer workflows, and commercial scoring.

Files

Path	Purpose
core/	Reference kernel implementation
aos_cli/	Minimal CLI and Trusted Output v0 utilities
lean/	Abstract public formal surface
tests/	Contract and tamper-detection tests
examples/reports/	Sanitized replay example
tools/run_validation_gate.py	Canonical local validation entrypoint
tools/audit_precheck.py	Release-surface audit precheck
evidence/integrity_manifest.json	Public integrity manifest

License

This repository is published under a proprietary demonstrator notice. Viewing the repository does not grant rights to copy, modify, distribute, host, commercialize, or create derivative works without written permission.

See LICENSE and NOTICE.