gjata-legacy/ess-mai-poc-005-ffi-law0-authority-nonduplication

ESS-MAI POC 005 — FFI LAW-0 Authority Non-Duplication

ESS-MAI is experimental, governance-first AI systems research by Bledar Gjata (Gjata Legacy), developed in Tirana, Albania. Albania denotes where the research is developed; ESS-MAI is not an Albanian-language model and not a national or sovereign AI system. Its relevance to AI governance is, in the repository's words, “an invitation to evaluate the project, not a claim of academic validation, regulatory compliance, or safety assurance”; it is not a legal or compliance framework, and governance-first control as such is prior art, not claimed as new.

This is a single-capsule mirror of POC 005 from the ESS-MAI executable prior-art collection, release ess-mai-executable-prior-art-v1.1.1. The directory poc-005-ffi-law0-authority-nonduplication/ is a verbatim copy of the capsule as published in the collection and on GitHub. The capsule files are the authority. This card summarises them and adds no claim of its own. Nothing in this repository is a model, a demo, or the ESS-MAI core. The full private ESS-MAI v1.8.9 core is not public and is not included.

Abstract

POC 005 is an isolated executable theory proof of concept (THEORY_POC). It tests one bounded property of the existing ESS-MAI v189 FFI gate. A C-compatible CapHandle may be copied bit for bit, but copying the handle must not copy or replenish the authority represented by the Rust-owned slot (Copy(CapHandle) != Copy(Authority)). The focal v189 file is compiled byte-for-byte in a detached Cargo workspace. The runtime experiment exercises a forged nonce, copied and repeated handles, an unknown generation, an empty payload, eight concurrent bit-identical copies, and two fresh generations for the same module and payload.

The positive result is local to one issued generation: identifier-local handle replay is prevented. The negative result is equally part of the finding. This legacy FFI surface does not complete a knowledge write, because its source intentionally requires a committed constitutional verdict. In addition, two fresh generations for the same module and payload both reach the post-authorization hold, which is a published counterexample to stronger semantic idempotency. The tested claim is local to one issued generation and to the disclosed legacy_ffi_compat/test surface. The capsule is classified PARTIAL_POC.

Position in the authority hierarchy. ESS-MAI names one constitutional parent, GCL, and three bounded jurisdictions that GCL delegates in phase order: LIGHT (coordination), QUANTUM (reasoning) and SHADOW (verification). These three are not peers of GCL. SHADOW holds verdict and persistence only within the jurisdiction GCL delegates to it. For the FFI surface, the capsule states the rule "the FFI child carries evidence but never assumes verdict or persistence authority" (FINAL_VERDICT.txt).

What POC 005 observes is narrower than that rule. The tested direct FFI route stops at CompatibilityHold (-8) when the superior committed verdict is absent. The connection to the commit path is also absent: PRIOR_ART.md records the GCL constitutional composition as "partial: one-shot gate + verdict hold; commit connection absent". The role names are design designations, established only to the extent this POC shows them.

Identity

Field	Value
Author	Bledar Gjata
Author ORCID iD	0009-0003-7854-266X
Organisation	Gjata Legacy
Contact	[email protected]
Capsule	poc-005-ffi-law0-authority-nonduplication/ — POC-005, FFI LAW-0 Authority Non-Duplication
Capsule version	1.1.1 (verification date 2026-09-14)
Class	THEORY_POC
Collection status	verified_for_collection_v1.1.1
Collection release	ess-mai-executable-prior-art-v1.1.1, tag commit d38c2bab85e7bc9e1683da2cdfccc43e958639ad (2026-09-14)
DOI (collection version containing this capsule)	Zenodo v1.1.1 10.5281/zenodo.22750188, containing POC 003–006. POC 005 is not part of v1.0.0 (10.5281/zenodo.22074028, POC 003 + POC 004). Concept DOI for all versions: 10.5281/zenodo.22074027
Canonical GitHub path	https://github.com/gjatalegacy-create/ESS-MAI/tree/ess-mai-executable-prior-art-v1.1.1/publications/executable-prior-art/poc-005-ffi-law0-authority-nonduplication
Licence	Apache-2.0 (the capsule's LICENSE and CITATION.cff)

Contents of this repository

Path	Content
poc-005-ffi-law0-authority-nonduplication/	Verbatim capsule, identical to GitHub. It contains the Cargo workspace (source_core/, experiment/, Cargo.lock), the claim, theory, protocol, results, verdict, prior-art and reproducibility files, the evidence/ logs and artifact_hashes.sha256.
data/poc.jsonl	The single POC-005 row of the collection table, which the Dataset Viewer displays. Every value is copied from the reading page and the collection manifest.json, and the row adds no claim.
MIRROR_MANIFEST_REVERIFIED.sha256	SHA-256 of the mirrored files, recomputed at mirror time.

The capsule README.md gives this reading order. Start with CLAIM.md, SOURCE_RUNTIME_MAP.md and EXPERIMENT_PROTOCOL.md. Results and boundaries are in RESULTS.md, FINAL_VERDICT.txt, SCIENTIFIC_REPORT.md, PRIOR_ART.md and REPRODUCIBILITY.md.

from datasets import load_dataset
poc = load_dataset("gjata-legacy/ess-mai-poc-005-ffi-law0-authority-nonduplication", split="collection_v1.1.1")

Question tested

Does the existing v189 FFI gate let one freshly issued, generation-bound CapHandle cross authority validation at most once — refusing copied replay, forged nonce and unknown generation — while withholding a direct knowledge write when no committed superior verdict is present?

CLAIM.md bounds the question in these words: "The tested claim is local to one issued generation and to the disclosed legacy_ffi_compat/test surface. It is deliberately capable of producing a negative result for the stronger semantic interpretation."

What was shown

• Native v189 focal tests 8/8; strict release all-targets build; detached tests 15/15; 3/3 independent process receipts identical.
• Copy(CapHandle) != Copy(Authority) for one issued generation: forged nonce -3; first valid non-empty call reaches CompatibilityHold (-8); copied and third replays -1; unknown generation -2; empty payload -5, then -1.
• Eight concurrent bit-identical copies: exactly one post-authorization result and seven refusals.
• The direct FFI route does not silently write knowledge; it stops at CompatibilityHold without the superior committed verdict.
• Source identity: 3/3 focal files byte-identical; v189 source files modified: 0.

Runtime receipt, verbatim from RESULTS.md:

ARTIFACT_TYPE=THEORY_POC
THEORY=FFI_LAW0_AUTHORITY_NONDUPLICATION
RUN0_INSTRUMENT_VALIDATION=PASS
BASELINE_FRESH_GENERATION_CODE=-8
TARGET_COPIED_REPLAY_CODE=-1
ABLATION_BURNED_AUTHORITY_CODE=-1
ADVERSARIAL_FORGED_NONCE_CODE=-3
FORGED_NONCE_CODE=-3
FIRST_VALID_NONEMPTY_CODE=-8
COPIED_HANDLE_REPLAY_CODE=-1
THIRD_REPLAY_CODE=-1
UNKNOWN_GENERATION_CODE=-2
EMPTY_PAYLOAD_CODE=-5
EMPTY_PAYLOAD_REPLAY_CODE=-1
CONCURRENT_COPIES_POST_AUTH=1
CONCURRENT_COPIES_REJECTED=7
COUNTEREXAMPLE_FRESH_SEMANTIC_FIRST_CODE=-8
COUNTEREXAMPLE_FRESH_SEMANTIC_SECOND_CODE=-8
COUNTEREXAMPLE_STRONG_SEMANTIC_IDEMPOTENCY=FOUND
COPY_HANDLE_EQUALS_COPY_AUTHORITY=false
DIRECT_FFI_KNOWLEDGE_WRITE_MATERIALIZED=false
DIRECT_FFI_WRITE_BOUNDARY=COMPATIBILITY_HOLD
EXPERIMENT_STATUS=PASS

How to read a passing result: "EXPERIMENT_STATUS=PASS means the bounded claim and its negative boundary were reproduced. It does not mean the legacy FFI route persisted knowledge." (RESULTS.md)

The experiment is expected to pass only when -8 remains visible, because a test that expected 0 would falsely claim end-to-end persistence. A passing test count is not a proof of the theory. The run counts come from disclosed maintainer evidence and are not presented as independent replication.

What was not shown (failures and counterexamples)

• Two fresh generations for the same module and payload each reach -8: generation-local non-duplication is not semantic non-duplication.
• A locally authorized non-empty call returns CompatibilityHold instead of a committed knowledge-write success.
• Durable consumption across restart, an idempotent knowledge sink and end-to-end FFI-to-Vault persistence are not materialised. The restart step was not executed: RESTART_RECOVERY=NOT_EXECUTED_NO_DURABLE_SLOT_IN_NARROW_CLAIM; GLOBAL_DURABILITY_NOT_PROVEN (FINAL_VERDICT.txt).
• No default-production in-repository caller of the legacy FFI exports was found in the declared v189 source search.
• The eight-copy concurrency case is serialised by the current global Mutex; no lock-free property is claimed.
• The first strict --all-targets build failed on an incomplete authored adapter; the failure is preserved and only the adapter was completed.
• The FNV seal is not claimed to be cryptographically unforgeable.
• Bit-for-bit binary reproducibility is not established. Three clean build roots produced the same semantic receipt but different Windows executable SHA-256 values. The binary is recorded as a runtime witness and is not stored in the capsule.

The capsule's complete list of unsupported readings is reproduced verbatim under Status.

Status

Experimental status, copied from the collection manifest.json. The string is part of the claim:

PARTIAL_POC_LOCAL_PROPERTY_READY_GLOBAL_THEORY_NOT_MATERIALISED

Verdict lines, verbatim from FINAL_VERDICT.txt:

NARROW_CLAIM_STATUS=POC_READY
GLOBAL_THEORY_STATUS=THEORY_NOT_MATERIALISED_GLOBALLY
POC_STATUS=PARTIAL_POC
CLAIM_BOUNDARY=SEALED_FOR_POC_005_ONLY
FULL_ESS_MAI_V189_ARCHITECTURE_CLOSED=false

Verbatim from the capsule's CLAIM_BOUNDARY.md:

The evidence supports a project-specific implementation contribution: a copyable C ABI handle is separated from Rust-owned one-shot authority, and a successful consumption is prevented from becoming a direct knowledge write without the superior GCL/Shadow verdict path. Any novelty claim must be limited to this precise constitutional composition and evaluated by qualified counsel against the cited and uncited art.

CLAIM_BOUNDARY=SEALED applies only to this POC's local question. It does not declare the full ESS-MAI v189 architecture closed.

Not supported, verbatim from CLAIM_BOUNDARY.md:

• “Capabilities,” “single-use tokens,” atomic compare-and-swap, or reference monitors were invented by ESS-MAI.
• The handle is cryptographically unforgeable merely because its nonce is FNV.
• A fresh token cannot be issued for the same semantic authorization.
• The public FFI exports have a non-test in-repository default-production caller; none was found in the declared v189 caller search.
• Consumption survives restart.
• The legacy FFI surface currently performs an end-to-end Vault write.
• No bypass exists anywhere outside the tested closure.
• The mechanism is formally verified, certified, production-secure, or a completed patent search.

Theory and prior art

The primitives involved are established prior art: capability-as-authority, one-shot or linear authorization, nonce validation, atomic state transitions, reference-monitor mediation, deny-by-default handling and unsafe FFI contracts. Among the works the capsule cites are CHERI, the CHERIoT guide, the Rust Project FFI and ownership guides, and Saltzer and Schroeder (1975), and POC 005 does not claim to have invented any of these primitives.

The project-specific contribution tested is an arrangement and authority boundary rather than a new primitive. A freely copyable C representation is kept below a Rust-owned one-shot authority, and after local consumption the compatibility path is still forbidden from becoming knowledge unless the superior confirmed-verdict and Shadow transaction authority exists. The experiment, in the words of THEORY.md, "materially supports the first half for one generation and the hold boundary"; the first half is that representation may cross the FFI boundary without transferring or multiplying the authoritative consumption state. "It does not yet support durable semantic non-duplication."

The capsule also declares two independent works that postdate the 2026-06-03 media record but predate this POC's release (PRIOR_ART.md): Plumb, Technical Disclosure Commons, 2026-08-13, and Xu et al., arXiv:2608.01710 (submitted 2026-08-03). Plumb describes a fail-closed gateway redeeming single-use capability tokens; Xu et al. show that identifier-local single-use tokens do not prevent semantic replay through fresh issuance, which POC 005 does not solve. The media record does not disclose this mechanism and is chronology only; in the capsule's words, "Chronology does not imply copying or establish legal priority."

The search, dated 2026-09-10, is a focused technical comparison. It is not an exhaustive patent search, a freedom-to-operate opinion or a legal novelty opinion. It records NO_DIRECT_MATCH_FOUND_IN_DECLARED_SEARCH_SCOPE for the exact composition only, and the capsule states that this is not a claim that the composition is legally novel.

Full analysis: PRIOR_ART.md at the v1.1.1 tag.

Reproduce

Follow the capsule's own REPRODUCIBILITY.md (also at poc-005-ffi-law0-authority-nonduplication/REPRODUCIBILITY.md), not this card. Run the commands from the capsule directory, which is the Cargo workspace root. The capsule's rule is: "All build output must remain outside this capsule."

Toolchain used, as recorded in the capsule:

rustc 1.98.0 (88d9e12ae 2026-08-18)
cargo 1.98.0 (797e8a9bc 2026-08-05)
host: x86_64-pc-windows-gnu

Build and test (PowerShell), verbatim from REPRODUCIBILITY.md:

$env:CARGO_TARGET_DIR = "<workspace>\_audit_build_cache\poc005_reproduction"
rustc --version --verbose
cargo --version --verbose
cargo build --workspace --all-targets --release --locked
cargo test --workspace --all-targets --release --locked -- --test-threads=1
1..3 | ForEach-Object {
  cargo run -p poc005-ffi-law0-experiment --release --locked
  if ($LASTEXITCODE -ne 0) { exit $LASTEXITCODE }
}

Expected output, verbatim:

build exit = 0
tests = 15 passed, 0 failed
runtime runs = 3 passed, 0 failed
EXPERIMENT_STATUS=PASS
FIRST_VALID_NONEMPTY_CODE=-8
COPIED_HANDLE_REPLAY_CODE=-1
CONCURRENT_COPIES_POST_AUTH=1
CONCURRENT_COPIES_REJECTED=7
COUNTEREXAMPLE_STRONG_SEMANTIC_IDEMPOTENCY=FOUND

The strict method first exposed an incomplete authored adapter, with an exit-101 result. That result and the adapter-only closure are preserved in evidence/08_standard_protocol_runs.txt.

REPRODUCIBILITY.md also gives a native-source confirmation command (expected 8 passed, 0 failed) and a source-identity check. Both run against the v189 workspace, which is not part of this repository or of the public collection. The detached reproduction above does not require that workspace: DETACHED_REPRODUCTION=PASS_BUILD_15_TESTS_AND_3_RUNTIME_RUNS_WITHOUT_PRIVATE_SOURCE (FINAL_VERDICT.txt). Behavioural reproducibility and bit-for-bit binary reproducibility are separate claims, and only the first is established.

Integrity

• MIRROR_MANIFEST_REVERIFIED.sha256 (repository root) gives the SHA-256 of the mirrored files, recomputed at mirror time.
• poc-005-ffi-law0-authority-nonduplication/artifact_hashes.sha256 is the capsule's own hash manifest (37 entries), shipped with the capsule.
• The capsule's artifact_manifest_sha256 is ca935c761d490afd717dcd435f07501af4166b8a52f76f3057b4083cd4b445d1. It is the SHA-256 of that artifact_hashes.sha256 file, as recorded in the collection manifest.json and in data/poc.jsonl.

sha256sum -c MIRROR_MANIFEST_REVERIFIED.sha256
cd poc-005-ffi-law0-authority-nonduplication
sha256sum artifact_hashes.sha256   # expected: ca935c761d490afd717dcd435f07501af4166b8a52f76f3057b4083cd4b445d1
sha256sum -c artifact_hashes.sha256

Runtime-witness binary hashes are recorded in RESULTS.md. The binary itself is not stored. Matching hashes establish file identity only, not the correctness of any claim.

Part of

• Full collection mirror, with all four capsules verbatim: gjata-legacy/ess-mai-executable-prior-art. This capsule is at executable-prior-art/poc-005-ffi-law0-authority-nonduplication/.
• Reading page, which states what each POC showed and did not show: gjata-legacy/ess-mai
• Canonical repository: github.com/gjatalegacy-create/ESS-MAI, collection path publications/executable-prior-art/
• The other single-capsule repositories:
  • gjata-legacy/ess-mai-poc-003-system-cold-start-reachability: POC 003, system cold-start reachability
  • gjata-legacy/ess-mai-poc-004-gcl-law0-global-continuity: POC 004, GCL LAW-0 global continuity
  • gjata-legacy/ess-mai-poc-006-living-negative-knowledge-verified-reentry: POC 006, Living Negative Knowledge: verified re-entry

Test and refute

Agreement with the theory is not required. The most useful external work is adversarial: find an authority bypass, an invalid state assumption, a non-reproducible result, a receipt-substitution path, a documentation/source mismatch, or an overbroad theoretical claim.

The capsule's falsification condition, verbatim from CLAIM.md:

The narrow claim is falsified if any copied use of the same generation reaches the post-authority boundary more than once, if a forged/unknown handle is accepted, if an invalid payload leaves reusable authority, or if the current surface silently reports a successful knowledge write without the committed verdict path.

• Independent technical review: GitHub Discussion #12
• Independent reproducibility report: issue template
• Security-sensitive findings: report privately as described in SECURITY.md

Citation

The capsule's CITATION.cff says: "If you use this bounded ESS-MAI theory POC, cite this artifact version." The DOI below identifies the Zenodo collection version v1.1.1, which contains this capsule. The capsule has no separate DOI.

@software{gjata_2026_ess_mai_poc_005,
  author  = {Gjata, Bledar},
  title   = {{ESS-MAI POC-005: FFI LAW-0 Authority Non-Duplication}},
  version = {1.1.1},
  year    = {2026},
  month   = sep,
  license = {Apache-2.0},
  url     = {https://github.com/gjatalegacy-create/ESS-MAI},
  doi     = {10.5281/zenodo.22750188},
  note    = {Capsule of ESS-MAI Executable Prior Art v1.1.1; the DOI identifies that collection version}
}

To refer to all versions of the collection, use the concept DOI 10.5281/zenodo.22074027.

Author

Bledar Gjata, Gjata Legacy, [email protected] · GitHub · ORCID · LinkedIn · X · TikTok · YouTube · Facebook · Facebook page GJATA

Profiles establish attribution and chronology, not technical evidence. The technical evidence is the capsule.