I am Bledar Gjata, the architect and designer of ESS-MAI. I propose the AI Authority Entity as an independent executive competence with a defined duty, applicable law and jurisdiction, organised within a hierarchy. This is a proposed computational paradigm, offered for examination through code and experiments. It is not a claim of legal authority over people or a completed scientific proof of novelty.

One hierarchy, distinct competences

In the ESS-MAI proposal, GCL supplies the constitutional order. LIGHT coordinates the case; QUANTUM commissions and performs reasoning through subordinate competences; SHADOW verifies the current dossier and owns durable Vault writes. GCL is not another authority casting a vote.

The capsule contains 719 unchanged native source, configuration and fixture files from the identified AIGovern-Cyber snapshot, including its build dependency closure. This is broader than a minimal example: the source map identifies the exact functions exercised, and the provenance manifest records each file. The public repository has 949 files when documentation, evidence and the preprint are included.

A single authority can exercise different competences in one cycle. In the tested UCL instance, command_asht_routes advances XiYiLoaded to AshtScoped and issues distinct Historical and KiE commands. Repeating that commissioning operation on that instance is rejected. That guard does not establish universal exactly-once execution across arbitrary direct calls or new instances.

Five user turns, two final judgments

The experiment keeps one continuing case about Earth rotation in the ERA00 numerical model. The first two user turns lack required candidates or numerical evidence, so they produce operational HOLD/NoVerdict outcomes and questions.

The third turn supplies UT1 Julian-date parts 2400000.5 + 54388.0, a claim of 0.40228372400281387 radians, and a tolerance of 1e-12 radians. The model supports that bounded claim: KiS=1, KiC=0, followed by the native SHADOW pair (1,1) and durable T1/T2 records.

The fourth turn requests an opposing claim without supplying the numerical evidence. It remains a HOLD linked to the third turn’s durable record. The fifth supplies 0.0 radians under the same model inputs: KiS=0, KiC=1, and the native judgment is (0,0), with its own durable record.

These are five user turns, three operational HOLDs and two final native judgments. They are not five completed judgment cycles. The model comparison does not constitute a physical observation of Earth.

Why the other routes remain open

The historical ASHT route still reports MISSING_EXACT_POSITIVE. KiE completes only its own scoped mathematical obligation; it does not manufacture the historical route’s missing evidence. A completed model claim cannot silently discharge another competence’s requirements.

Missing evidence and scientific counterevidence also mean different things. The first, second and fourth turns lack enough material for a final judgment. The fifth supplies a numerical counterexample. Calling all four simply a failed experiment would erase that distinction.

Including the complete native source makes other routes inspectable; it does not prove that every route, input or hardware state is correct. External-agent interoperability is not demonstrated here.

What replay and Cargo actually showed

Exact T1/T2 replay returned the recorded response bytes without adding another durable Vault transaction. The WAL remained 123,245 bytes with the same SHA-256. Changing one bit of a request hash while retaining its identity produced PD_T1_REQUEST_CONFLICT and left that WAL unchanged. These are bounded replay results, not a guarantee for arbitrary external effects.

The full Cargo run returned exit 101: 1,821 tests passed, including four doctests; one failed; none were ignored. Release builds and continuity passed. Twenty focused tests also passed, but they overlap the full suite and are not twenty additional unique tests.

The failing test was blocked_stdin_obeys_deadline_and_retains_output. A separate probe reproduced a cold-start timing mechanism: fresh executable spawn calls exceeded the 750 ms budget before a main-entry marker appeared; warm starts retained the expected output. The underlying host cause is unknown. The original failure remains in the record; no native code, test or timeout was changed to hide it.

SHA-256 checks content integrity, not authority. The native source also contains an embedded HMAC key, which cannot establish authenticity against someone who has the source. General security, rollback protection without an external anchor, and complete materialisation of the paradigm remain unproven.

Reproduce, challenge and extend

The README provides the Windows/Rust 1.98.0/MinGW commands, exact inputs, source map and diagnostic logs. Independent reproduction, precise objections and attributed improvements are welcome. AI tools assisted implementation, examination and writing under my architectural direction and review.

• POC002 code and reproduction
• Hugging Face mirror
• Software and evidence DOI
• Research preprint
• PREreview request
• Originating ESS-MAI project
• ESS-MAI archive
• ORCID
• LinkedIn
• OpenAIRE discovery route

A review request is not a completed review; an OpenAIRE discovery link does not establish indexing. The capsule is published under Apache-2.0 with dependency notices preserved.

I am Bledar Gjata, the architect and designer of ESS-MAI. POC011 is a public Rust research capsule derived from that work. Its five-cycle fixture returns scalar verification bits 0,0,0,1,1. Each zero records why the current action cannot be admitted.

Different duties, one current case

ESS-MAI proposes an AI Authority organised around a duty, executive competence, jurisdiction and governing law. Permission or refusal follows from performing that duty. A permission token alone is not the authority.

The intended constitutional organisation comprises GCL, Supreme, laws and hierarchical authorities. POC011 materializes a bounded sequence from this proposal. LIGHT binds and coordinates the current request. QUANTUM evaluates candidates and proposes an eligible action. SHADOW independently verifies the current proposal. These duties do not vote or inherit one another’s competence.

Several calls can perform different duties. Recomputing a proposal during verification does not, by itself, mean the proposal authority has been duplicated. The question is which operation exercises which competence and whether its conditions hold.

A model small enough to inspect

The domain is five integers, x = 0,1,2,3,4. Each candidate supplies coefficients for y = a*x+b and evidence samples. Gain is the sum of its five values; risk is their maximum absolute value.

Eligibility requires complete, matching samples and compliance with the hard risk limit. Eligible candidates are ranked by gain descending, risk ascending and candidate SHA-256 ascending. A larger gain cannot compensate for a failed mandatory condition.

The five inputs change the evidence:

1. The risk-limit clause is missing: bit 0 and an exact question.
2. Limit 10 is supplied, but samples are missing: bit 0, preserving the gaps.
3. Complete samples for 4x+1 give risk 17: bit 0, retaining counterexamples at x=3 and x=4.
4. Complete evidence for 2x+1 gives gain 25 and risk 9: bit 1 and one local effect.
5. A new action, 2x+2, gives gain 30 and risk 10: bit 1 and a second distinct effect.

The inadmissible competitor’s objections remain in cycles four and five. Missing evidence is not a mathematical refutation, and success does not erase earlier evidence or approve the next cycle.

Memory supports continuity, not inherited approval

LIGHT retrieves bounded references to earlier records in the same context. SHADOW checks the evidence currently presented. Storage replay validates journal integrity; it is separate from deciding whether today’s proposal meets today’s requirements.

D-BAB records actual preconditions and outcomes without granting competence. The borrowed 0-NOI primitive triggers the operation while leaving its checks with the caller.

The report and optional local effect are committed together, synchronized and read back before memory state becomes visible. Output plus questions is then returned. An output-delivery failure cannot undo a known durable commit. Copied FFI handles reference one consumable slot; the local effect also has a context/action identity.

What was tested

The recorded Windows execution passed 38 Rust tests in debug and the same 38 in release, with zero failures or ignored tests. A separately implemented Python verifier passed 17 self-tests. This is software cross-checking, not independent human review.

Adversarial cases include altered or rehashed traces, corrupted receipts, wrong handle bindings, concurrent consumption, conflicting effects and truncated journals. Eight copied handles have one successful consumption; 32 concurrent attempts have one winner.

Reproduction compatibility: the first hosted CI runs stopped before tests because Rust 1.99.1 deprecated AtomicU64::fetch_update and the capsule denies warnings. The recorded toolchain is Rust 1.98.0. A minimal workflow correction pins that toolchain; runtime logic and test assertions remain unchanged. Consult the CI compatibility record and Actions history for the verified outcome and exact revision.

From the capsule directory, with Rust 1.98.0 installed:

cargo +1.98.0 fetch --locked
cargo +1.98.0 test --locked --offline
cargo +1.98.0 test --locked --offline --release
cargo +1.98.0 run --locked --offline --release -- demo demo-run
python tools/verify.py demo-run --self-test

The initial fetch needs network access if pinned dependencies are absent. The demo creates a new directory, which must not already exist. Builds and tests create build and temporary files.

Provenance and boundaries

Selected canonical ESS-MAI Light token, handle and atomic-slot blocks are exact extractions with recorded hashes. The neutral 0-NOI primitive is extracted from the later AIGovern derivative. The finite model, witness integration, trace verifier, journal and restart lifecycle are documented adaptations.

SHA-256, capabilities, atomic consumption and journals are established techniques. My proposed contribution is their ESS-MAI-specific constitutional organisation and the bounded materialization presented for criticism; novelty and general architectural correctness remain open questions.

The RAM write/readback is real process-owned memory work, not physical energy measurement. Hashes do not authenticate an issuer. Replacing the entire journal with an older valid prefix remains undetectable without a trusted external anchor. The scalar is not native ESS-MAI/AIGovern-Cyber’s paired verdict, external exactly-once execution or a security certification.

The capsule is Apache-2.0; the full private runtime is excluded. AI assisted implementation and writing under my architecture and direction. Reproducible counterexamples, corrections and attributed contributions are welcome.

Code, evidence and review protocol · Hugging Face mirror · Software DOI · Research preprint · PREreview request · OpenAIRE discovery route · ORCID · LinkedIn · Originating ESS-MAI project

A discovery link or review request does not establish indexing, completed review or endorsement.

A green test result means that the assertions executed in that run did not fail. The engineering question is what those assertions actually cover. A test can correctly demonstrate a refusal, reproduce a counterexample, or confirm that a stronger property remains unsupported.

That distinction organizes seven public ESS-MAI and AIGovern-Cyber experiments, POC003–009. They isolate questions about execution gates, reusable failure history, foreign-function interfaces and persistent state. They are runnable research capsules derived from a larger architecture; their results do not certify the private product.

The results below are previously recorded project observations. Preparing this article involved reading their public procedures, not running a new test campaign or independently replicating them. The family reproduction index provides pinned artifacts and reporting instructions.

These totals describe different suites and observation types. Combining them into one “proof score” would obscure their meaning. The public procedures also retain negative controls; a successful harness execution does not erase an adverse result.

Consider POC005. Refusing seven copies of a handle supports a local one-use claim. It does not prove that two newly issued handles cannot represent the same semantic request. Generation identity and operation identity are different invariants. The disclosed reissuance counterexample makes that distinction testable.

POC009 introduces another boundary. Its supplied Earth Rotation Angle model can support one proposition and contradict another under specified inputs. That comparison uses an existing numerical method. Repeating the calculation does not supply another independent physical observation, and storing its result does not transfer native knowledge authority.

Start with one exact continuity experiment

POC008 offers a concrete comparison: after each supported input, are the restored process's checkpoint bytes identical to uninterrupted execution? Comparing every prefix helps locate a divergence that an end-only comparison could hide.

Use Windows PowerShell with Git, Rust/Cargo and Python 3. Rust 1.98.0 is the recorded toolchain, not a claimed minimum version. Use a disposable directory; both poc008-review and poc008-review-demo below must be absent. Run commands individually, check $LASTEXITCODE after each external program, and stop on any nonzero exit.

git clone https://github.com/gjatalegacy-create/AIGovern-Cyber-POC008-Continuity.git poc008-review
Set-Location poc008-review
git checkout --detach 84c40ea2679fcbe29ad10faa451a41c576e202f8
git rev-parse HEAD
rustc --version --verbose
cargo --version --verbose
$env:CARGO_TARGET_DIR = Join-Path $env:TEMP ('poc008-review-' + [guid]::NewGuid().ToString('N'))
cargo test --locked
cargo build --release --locked
python tools/demo.py --binary "$env:CARGO_TARGET_DIR\release\poc008-continuity.exe" --output-dir ../poc008-review-demo

Confirm the printed commit matches the checkout above, then follow its README and claim boundary. Initial dependency downloads may be necessary. The recorded expectation is 35 passing tests; the demonstration separately launches five processes and saves their outputs. The 65-turn comparison belongs to the test suite, not that five-process demonstration.

Preserve the first differing checkpoint, operation identity or retry result. Record a dependency-download failure as an environment problem before interpreting it as a program defect. Do not alter inputs or assertions simply to obtain green output.

A passing continuity run still leaves an important adverse control: a coherent older WAL can be accepted. Replay consistency does not establish authenticated freshness. Likewise, process restart tests do not demonstrate arbitrary power-loss survival. POC008's ordinary inspect path can initialize storage or repair an eligible incomplete tail, so inspect a copy when preserving evidence.

Make the review claim-sized

Choose one invariant and report the exact commit, environment, commands, exit codes and earliest mismatch. Negative results are useful when attached to a reproducible claim. Use the structured reproduction form; declare your relationship to the project. No favorable conclusion or acceptance of the broader theory is required.

AI tools assisted implementation, tests and writing, including this article. I direct the architecture and final design decisions and review code and execution through shell checks. Author review, AI agreement and automated CI are distinct from independent human review; no independent replication is claimed here.

Relevant resources

• GitHub: seven-experiment reproduction index
• Hugging Face: POC008 artifact mirror
• LinkedIn: Bledar Gjata
• Zenodo: versioned POC008 software
• OpenAIRE: POC008 software metadata
• ORCID: author identity

These channels identify source, distribution, archives and the author. Hosting, indexing and identifiers do not establish peer review or security certification.