Proposed: no single point of failure.

GARDIEN — the gate set

Sixteen independent gates, tiered: root, identity, containment, recovery. All sixteen fail closed. Two-gate failure is an instant cut. Independence is measured by dependency fault injection, never asserted.

16 owner-specified gates · none enforced or validated.

Gate-set laws · specified, not enforced

1. Independence is measured, not declared. Measurement is dependency fault injection: fail each registered dependency deliberately, record whether the set degrades safely, record whether any out-of-band verifier failed in sympathy with the thing it was verifying. Sympathetic failure is a shared dependency in disguise. That register entry is struck.
2. Fail-closed on every gate. Two-gate failure is an instant cut from the lattice. No review, no grace.
3. Status ladder, one direction only: UNSPECIFIED → SPECIFIED → INSTRUMENTED → RUN → REPLICATED → VALIDATED → STRUCK. No skipping. Nothing reaches VALIDATED on argument.
4. Validation needs four things: a threshold written before data, a reproducible artifact a stranger can run, a confidence interval, and a named failure condition actually observed once.
5. G12 against G6: break-glass wins on scope, not on principle. Advance policy authorisation, human quorum over independent channels, narrow scope, short limit, self-generated signed evidence, reconciled after, incident by default.

The architectureProposed

The attack campaign

Planned, in isolated environments. None of these have been run against the real console yet.

Every test records scenario, expected and actual behaviour, evidence, severity, fix and retest. False alarms are tested too.

Campaign results

0 of 12 campaigns executed. Results are added here only when a test was actually run and evidence exists.

Known limitations

• ▸ This package is reference documentation, not executable security software.
• ▸ The nine PASS results come from a reference gate on synthetic cases, not from the production console.
• ▸ The checkpoint signature used a test-only key. No production identity or key management exists yet.
• ▸ None of the twelve security attack campaigns have been executed.
• ▸ Cross-model operation, external audit and robotics testing have not been performed.
• ▸ Distributed safeguards, AΩ.AMEND controls and the learning loop are proposed designs, not enforced protections.
• ▸ The monitoring-boundary rule is a tested policy model on a draft branch, not a deployed security boundary. Real enforcement needs separate server-side permissions and independently controlled alerts.
• ▸ Repository CI results are reported by the builder for an unmerged draft branch in a private repository; this site cannot independently verify them, and they say nothing about a live deployment.
• ▸ Security claims apply only to the specific components, versions and conditions supported by evidence.

Layered defensesProposed

A hash or signature proves a record wasn't altered — not that it was true. Truth needs evidence checks too.

Robotics: don't rush the physical worldProposed

1. STAGE 1
   Digital simulation

   Decisions, cooperation and emergencies tested with no physical equipment.

2. STAGE 2
   Software-only integration

   Simulated tools and devices, restricted permissions, full logging.

3. STAGE 3
   Hardware-in-the-loop

   Real controllers and sensors, with motors disabled.

4. STAGE 4
   Controlled physical trials

   Constrained areas, emergency stops, qualified human supervision.

5. STAGE 5
   Broader deployment

   Only after independent safety, security and recovery assessments.

Learning without losing controlProposed

1. 1. Check where the information came from.
2. 2. Decide if it is new, relevant and supported by evidence.
3. 3. Compare with existing knowledge; flag contradictions.
4. 4. Add useful findings to the right project records.
5. 5. Test proposed changes before they touch trusted rules.
6. 6. Keep the previous working version for rollback.
7. 7. Stop refining when changes stop meaningfully improving (Goldilocks).

New information must not automatically become governing policy. Every change keeps its evidence provenance, a version history and a rollback path.