sovereign-sdk-airlock 1.4.0

Model-neutral outbound governance boundary for sovereign perimeter inspection, policy enforcement, and evidence generation

sovereign-sdk-airlock

Model-neutral outbound governance boundary for the Sovereign Systems SDK.

Inspects, evaluates, minimises, and records structured payloads before they cross a sovereign perimeter and enter an external computational system.

---

Architecture

sovereign-sdk-airlock implements the four-component lifecycle defined in the Airlock Specification (SAR-0004):

Component	Responsibility
Boundary Interception	Capture normalised outbound payloads; transport-agnostic
Policy Engine	Evaluate local YAML governance rules (raw, fields, telemetry scopes)
Sieve Convergence	Deterministic context minimisation via sovereign-sdk-sieve
Evidence Generation	Immutable ForensicReceipt committed to sovereign-sdk-ledger

---

Quick Start

import asyncio
from sovereign_ledger import SovereignLedger
from sovereign_airlock import AirlockBoundary, normalize_openai

ledger = SovereignLedger(".keys/sovereign_audit.db")
boundary = AirlockBoundary(
    policy_path="policy.yaml",
    signing_key=".keys/",
    ledger=ledger,
)

request = {
    "model": "gpt-4o",
    "messages": [{"role": "user", "content": "Analyse the quarterly report."}],
}

async def main():
    result = await boundary.process(normalize_openai(request))
    print(result.sieved_content)
    print(result.telemetry.tax_savings_percentage, "% Prose Tax savings")
    print(result.receipt["metadata"]["payload_hash"])  # raw pre-sieve content hash

asyncio.run(main())

---

Policy Configuration

version: "1.0"

global:
  max_token_ceiling: 40000
  prose_tax_warning_threshold: 0.35

rules:

  - name: block_private_keys
    scope: raw
    pattern: "-----BEGIN PRIVATE KEY-----"
    action: deny

  - name: guard_internal_namespaces
    scope: fields
    fields:
      - messages.content
      - tools.description
    pattern: "internal\\.sovereign\\.local"
    action: warn

  - name: excessive_context
    scope: telemetry
    metric: raw_tokens
    threshold: 30000
    action: warn

Policy Actions

Action	Behaviour
allow	Explicit pass-through — no effect on verdict
warn	Appends to AirlockResult.policy_warnings; transmission continues
deny	Raises AirlockPolicyViolation; payload blocked

Exception Inspection

When a post-sieve deny rule fires, any warn-action messages accumulated during the pre-sieve pass are preserved on the exception for diagnostic inspection:

try:
    result = await boundary.process(payload)
except AirlockPolicyViolation as exc:
    print(exc)           # deny violation message
    print(exc.warnings)  # pre-sieve warn messages (list[str], may be empty)
    raise

---

Transport Normalisation

Airlock governs boundaries, not transports. Use the factory functions to convert provider-specific request formats into the provider-neutral NormalizedPayload:

from sovereign_airlock import normalize_openai, normalize_anthropic, normalize_raw

# OpenAI-compatible chat completion request
payload = normalize_openai({"messages": [...], "model": "gpt-4o"})

# Anthropic-compatible messages API request
payload = normalize_anthropic({"messages": [...], "system": "...", "model": "claude-sonnet-4-6"})

# Raw string
payload = normalize_raw("plain text to govern")

---

Invariants

• Zero external network dependencies — all operations execute on local silicon.
• Ledger write failure is non-fatal; outbound transmission is never blocked by a storage anomaly.
• Context minimisation uses sovereign-sdk-sieve exclusively — no model-based compression.
• All policy evaluation is deterministic and auditable.