Experimental Rust research software for explicit authority boundaries, negative knowledge, traceable state transitions, and reproducible falsification POCs.
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.
Shqip. ESS-MAI është kërkim eksperimental në sistemet e inteligjencës artificiale (IA) me qasje governance-first: teston nëse qeverisja mund të veprojë si ligj konstitutiv i arsyetimit, para dhe gjatë tij, e jo si kontroll pas tij — hipotezë e deklaruar, jo rezultat. Projekti zhvillohet në Tiranë, Shqipëri, nga Bledar Gjata (Gjata Legacy). Katër POC të ekzekutueshme në Rust (003–006) janë publike (burimi kanonik: GitHub), bashkë me sukseset, dështimet, kundërshembujt dhe kufijtë e pretendimeve të tyre. Një test që kalon tregon vetëm atë që u testua; nuk e provon teorinë.
A static reading page. GitHub is the canonical source; the Hugging Face dataset mirrors the collection folder verbatim. Each capsule's own files take precedence over this summary. External links open in a new tab.
Status at a glance
Collection
ESS-MAI Executable Prior Art — POC Collection, version 1.1.1: four bounded proof-of-concept capsules, POC 003–006. Licence Apache-2.0.
ess-mai-executable-prior-art-v1.1.1-poc-only.zip SHA-256 d7a322bf2b09f82803201b2194b75fb433c3f44eafd98441c7eec72d5b673bee Recorded hash, as stated in the GitHub release notes; this page does not recompute it.
gjata-legacy/ess-mai-executable-prior-art (dataset, Apache-2.0) at HF commit b5a4643f468538bc9f828649131234a687284530. Git-blob comparison (2026-09-28/29): all 199 files under executable-prior-art/ are identical to GitHub main at 70478d6599db3623a3d97aba1a64a6db6d014626 and to the v1.1.1 tag commit above. The dataset root adds README.md, MIRROR_PROVENANCE.md, MIRROR_MANIFEST_REVERIFIED.sha256 and .gitattributes.
Scope
This page and the Hugging Face mirror cover only the released POC capsules. The public GitHub repository also contains a broader experimental workspace. The full private ESS-MAI v1.8.9 core is not public and is not part of the collection.
Executed results as recorded in the collection's manifest.json
POC
Class · version
Recorded result
POC 003
SYSTEM_POC · 0.2.0
84 passed; 3/3 empty cold-start gap reproduced; 1/1 exact-pair positive control
POC 004
THEORY_POC · 0.2.0
19 passed; 5/5 success and counterexamples reproduced
POC 005
THEORY_POC · 1.1.1
native 8/8; detached 15/15; 3/3 identical process receipts; direct FFI knowledge write not materialised
POC 006
THEORY_POC · 1.1.1
detached 30/30; 3/3 identical process receipts; full negative receipt wire not materialised
Architecture hierarchy
ESS-MAI names one constitutional parent and three subordinate, bounded jurisdictions. The hierarchy is not flattened: LIGHT, QUANTUM and SHADOW are sibling jurisdictions delegated by GCL, not peers of GCL and not three sovereign systems.
GCL
Gjata Collapse Law
Constitutional parent authority: delegates and bounds subordinate roles without collapsing them into peers.
Delegated, bounded jurisdictions · in phase order
LIGHT
Coordination
Coordination and trace; does not decide.
Exercised in POC 003: the AlnurKarinaAthar Coordination state machine and the Besa selection client.
QUANTUM
Reasoning
Reasoning, pressure and elimination; produces evidence for Shadow and issues no verdict.
Exercised in POC 003: the UltimatumCollapseLaw Reasoning state machine and the Asht selection client. POC 006: sealed-history import and readiness gate.
SHADOW
Verification
Verdict and persistence, only within the jurisdiction delegated by GCL.
POC 003: a surgical, separate selection process, not the production Shadow core. POC 006 native tests: validation, transaction, deduplication and WAL.
Figure 1. Authority hierarchy. Role wording follows the historical architecture description in the repository README, which that README treats as an audit reference, not as runtime evidence; POC references come from the capsules. Role descriptions are design designations, established only to the extent each cited POC shows. Arrows show the phase order required by the POC 004 theory (Coordination ≤ Reasoning ≤ Verification); POC 004 records PHASE_ORDER = NOT_ENFORCED for the current ledger. Arrows do not transfer authority.
“No child phase may mint the parent authority or assume a sibling's jurisdiction.” — POC 004, CLAIM_BOUNDARY.md
Knowledge outcomes stay distinct under GCL: constructive K+ reinforces or expands supported routes; rigorous-negative K− excludes rigorously verified failure routes; HOLD is neither. POC 006 executes only the K− re-entry half; K+ is source-evidenced and natively tested.
“Quantum” names a role. It does not claim quantum hardware or quantum computation.
The four POCs
Each card summarises the capsule's own README.md, claim, results and verdict files. The status string is copied from manifest.json and is part of the claim.
POC 003 — System cold-start reachability
SYSTEM_POC · capsule 0.2.0 · released in collection v1.0.0
From generation zero — an empty Shadow knowledge state — can the bounded real authority path (GCL → Light/Besa → separate Shadow selection → Quantum/Asht → separate Shadow selection) reach the post-Asht final-Shadow point?
What was shown
Locked offline Cargo build; 84/84 disclosed tests pass (11 GCL, 13 Shadow-contract, 44 Quantum, 10 system unit, 6 integration).
Wrong-phase GCL directives are rejected by the extracted Light and Quantum state machines.
Besa accepts a complete empty selection through a separate Shadow process; Asht then fails closed because no exact positive candidate exists (3/3 fresh runs).
An exact-pair positive control (one positive, one negative candidate) passes the same Asht gate (1/1), isolating missing generation-zero evidence as the bounded causal variable.
What was not shown
From the empty state the post-Asht probe is not reached: the cold-start reachability gap is reproduced, not closed.
No production Shadow judgment, transaction, WAL commit or terminal receipt is executed; the positive control is not a production fix.
The proposed GCL-authorized one-time genesis route is future work; it is not implemented or claimed sufficient.
THEORY_POC · capsule 0.2.0 · released in collection v1.0.0
PARTIAL_MATERIALIZATION
Question tested
Does the locally enforced LAW-0 rule (each caller-reported step must satisfy after <= before) establish global continuity of the candidate space across steps, phases and process boundaries?
A tuple is accepted only when after <= before; the direct expansion 2 -> 5 is rejected without appending a step.
The reference chain 16 -> 3 -> 2 -> 1 -> 1 -> 0 is accepted and classified as collapsed.
A harness-only, domain-separated binding of the existing receipt ID and the LAW-0 digest changes when the LAW-0 digest is altered and detects the substitution — an advancement probe, not production integration.
What was not shown (failures and counterexamples)
16 -> 1 followed by 100 -> 50 is accepted; the ledger's current state grows from 1 to 50.
A terminal Verification 1 -> 0 can be followed by Coordination 100 -> 50, reopening a collapsed ledger.
1 -> -1 and +infinity -> +infinity are accepted; the report omits the injected before=100.
Changing only the LAW-0 close report leaves the current Shadow receipt identifier unchanged, because LAW-0 is not an input to it.
Current Light-to-Quantum and Quantum-to-Shadow wires carry no continuous LAW-0 transcript (static source mapping, not executed).
“Uncertainty” here is a caller-reported candidate-space count, not Shannon entropy, probability or truth.
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?
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.
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.
Durable consumption across restart, an idempotent knowledge sink and end-to-end FFI-to-Vault persistence are not materialised.
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.
Does the existing v189 negative-knowledge import and filtering code bring valid, required, sealed history into READY state and hard-block a semantically matching later candidate, while preventing reasoning when the required history is missing, corrupt or downgraded?
What was shown
Strict release build; detached tests 30/30; 3/3 independent process receipts identical. Native v189 runs 12/12, 19/19 and 1/1 are separate runs, not summed.
Valid sealed history imports one entry as READY; the matching later candidate is hard-blocked with score 0.000.
Missing (DEGRADED), corrupt (NOT_READY) and downgraded (NOT_READY) required history prevents PRO activation.
v189 source and focused native tests keep constructive K+ and rigorous-negative K− as distinct write branches under one GCL/Shadow verdict boundary.
What was not shown (failures and counterexamples)
The cycle response carries negative_persisted as a Boolean, not a typed, independently verifiable negative-commit receipt (source inspection, not runtime-measured).
Detached terminate/reopen recovery through the authoritative Shadow Vault and the next Quantum process is not executed; WAL recovery passes only at native component scope (1/1).
Symmetric detached positive-history (K+) re-entry is not executed.
Negative knowledge is not claimed as a novel concept, and the FNV seal is not claimed to be collision-resistant.
ESS-MAI investigates AI systems in which reasoning, epistemic state, evidence, verification, authority boundaries, negative knowledge and execution constraints are explicitly represented by the architecture rather than treated only as emergent properties of a foundation model.
The project uses executable proof-of-concept capsules, adversarial tests, negative results, source mapping and reproducible evidence to test bounded architectural claims.
Experimental failure is retained as evidence; architecture is not rewritten merely to make an implementation pass.
A success establishes only what the disclosed experiment actually demonstrates. A failure is preserved as a typed boundary or unresolved gap. Advancement means identifying the smallest architecture-preserving connection, constraint or experiment that can close that gap.
Experimental sequence
Applied to POC 005 and POC 006 in collection v1.1.1 (standard-method closure). Each capsule's FINAL_VERDICT.txt records the outcome of every step, including steps that were not executed (for example, POC 005 RESTART_RECOVERY=NOT_EXECUTED_NO_DURABLE_SLOT_IN_NARROW_CLAIM).
theory -> current source -> production mechanism -> isolated experiment -> baseline -> target -> ablation -> adversarial failure -> counterexample -> recovery boundary -> repeatability -> verdict
What a passing test means
A green Cargo run means the experiment behaved as asserted, including when the asserted observation is a fail-closed outcome or a counterexample. It never converts an expected failure into architectural success. A passing test count is not a proof of the theory.
Status vocabulary
MATERIALIZED
Source, runtime behavior, test and artifact are all present for the stated bounded property.
PARTIAL
At least one required layer or boundary is absent.
DOCUMENTED_ONLY
A claim appears in documentation without a matching disclosed executable path.
NOT_MATERIALIZED
The expected connection is absent, or a counterexample demonstrates that the stronger claim does not hold.
Claim boundary
Classification of POC 005 and POC 006, verbatim from the v1.1.1 release notes (“Both capsules are classified:”). POC 003 and POC 004 keep their own status strings (PARTIAL_SYSTEM_MATERIALIZATION_WITH_CAUSAL_CONTROL; PARTIAL_MATERIALIZATION) and are bounded by their own CLAIM_BOUNDARY.md.
This seals only the disclosed POC questions. It does not declare the complete ESS-MAI v189 architecture closed, production-certified or fully materialised.
For every artifact summarised on this page:
This artifact demonstrates the stated behavior only within the tested executable scope.
It does not establish full ESS-MAI production closure or universal validity of the underlying theory.
Reproducibility
For POC 005 and POC 006, behavioural reproducibility is claimed: identical code and Cargo inputs produced identical semantic receipts. Bit-for-bit binary reproducibility is not claimed: Windows executable SHA-256 values differed across clean build roots. POC 003 and POC 004 record repeated runs (3/3 empty cold-start gap reproduced; 5/5 success and counterexamples reproduced) and make no binary-identity claim. Run counts come from disclosed maintainer evidence and are not presented as independent replication.
Non-claims
No formal verification or machine-checked proof; the collection provides executable tests and counterexamples.
No security certification, production readiness or external validation.
No exhaustive patent search, and no novelty, patentability or freedom-to-operate opinion.
The full private ESS-MAI v1.8.9 core is not included and is not represented as present.
Context, claim and assumptions — hypotheses, not results
Status: HYPOTHESES, NOT RESULTS. The architect's theory (NuraLogic; Governance-First Law Cognitive Architecture, G-FLCA; AI DOMINLAW) states hypotheses with explicit reject conditions. Nothing in POC 003–006 establishes it.
NuraLogic is the doctrine, G-FLCA the candidate family, and ESS-MAI under Gjata Collapse Law (GCL) its first reference materialisation. The central question, in the author's words:
“Can evidence admission, reasoning transition, authority, knowledge formation and closure be bound to one executable law, such that no cognitive state becomes authoritative outside that law, and such that missing evidence or missing executable capability withdraws authority instead of being completed by fluency?”
The rule: “Indicia open investigation. Evidence opens progression. Zero evidence stops authority.” The order GCL > LIGHT > QUANTUM > SHADOW is never flattened. It is constitutional precedence (GCL is the parent law) followed by the phase order of three bounded jurisdictions (LIGHT → QUANTUM → SHADOW); it does not mean LIGHT holds more verdict authority than SHADOW.
PARTIAL_PUBLIC_EVIDENCE: a public POC exercises a bounded part, for the tested scope only; it never means the hypothesis is supported. NOT_YET_TESTED_PUBLICLY: no public artifact exercises it. No POC proves a hypothesis.
Eleven of the 22 assumptions (nine rows); the status refers to the ESS-MAI materialisation, not to the established principle.
Assumption
Grounded in
Status
A9 · Zero evidence stops authority; missing evidence is never positive evidence
Under the discipline, each experiment is to end in one evidence-backed state, or stay NO EVIDENCE / UNPROVEN. No public capsule runs this five-state classification yet (A20: NOT_YET_TESTED_PUBLICLY).
GO
“scoped evidence satisfied; only that scope advances.”
RESTART
“failure preserved; root cause; bounded correction; re-entry at the same obligation. No erasure. No pretend-retry.”
UNSAFE
“contain; requalify only against a named recoverable gap; independent validated flows may continue.”
UNVALID
“premise not falsified; means insufficient”: a recorded frontier with its re-entry condition.
INVALIDATED
The premise is defeated; the route ends and is kept as Negative Knowledge.
NO EVIDENCE / UNPROVEN
Uncertainty preserved; nothing promoted.
How risk is accepted
The risk is scientific: a hypothesis may be false, partly true or not yet realisable. Each claim is published with the observation that would defeat it, each outcome is to be routed to one of the states above, and failures stay in the record for others to inherit. In the author's words: “The hypothesis may fail.”
AI DOMINLAW, as a hypothesis
AI DOMINLAW (“User Domination over the Dominant Law in the System”) names a possible class of locally governed systems in which the governing law and the data stay on the user's or institution's device. It is a hypothesis, not a fact (“Naming it does not create it.”), and it depends on A17, which has no public test yet. It would lose its technical ground if lawful authority proved impossible to keep inside any declared local envelope, or if A17 were rejected (silent drop or silent externalisation of critical steps). G-FLCA itself is proposed as an architecture that can help any intelligent system; an ESS-MAI-style core could also run inside cloud-hosted systems as an engine under contract. Neither use has been tested publicly.
The cultural concept of Besa motivates an executable research question about whether a system can be required to prove that it fulfilled an accepted obligation. In POC 003, Besa is only the name of a Light-side selection state machine; this page reports no result on the obligation question.
Prior art and public record
Each capsule's PRIOR_ART.md separates established foundations, the ESS-MAI contribution and the limits of its search; POC 005 and POC 006 also declare adjacent and later independent work. These are focused technical comparisons, not exhaustive patent searches, novelty opinions or freedom-to-operate analyses.
Declared prior art per capsule. Each row header links to that capsule's PRIOR_ART.md at the v1.1.1 tag.
9 — Mitchell 1977; Mackworth 1977; Cousot and Cousot 1977; Kildall 1973; Strom and Yemini 1986; Lee, Jones and Ben-Amram 2001; Necula 1997; RFC 9162; Shannon 1948
Candidate-set reduction, monotone ordered reasoning and proof-object transport are established. The limited review found no exact description of the named composition; this is stated not to be a novelty conclusion.
6 external — CHERI (Watson et al. 2015); CHERIoT guide; Rust Project FFI and ownership guides; Saltzer and Schroeder 1975; Plumb, Technical Disclosure Commons 2026-08-13; Xu et al., arXiv:2608.01710 — plus the 2026-06-03 media record
NO_DIRECT_MATCH_FOUND_IN_DECLARED_SEARCH_SCOPE for the exact composition; the individual mechanisms are established. Semantic replay through fresh issuance, identified by later independent work, is not solved by POC 005.
7 external — Wang, arXiv:2606.21024, and its reference implementation; Constitutive Governance, arXiv:2605.24538; LATTICE (2026); Riesbeck 1981; Hammond 1986; Ba et al. 2024 — plus the 2026-06-03 media record
The broad negative-knowledge propositions are declared existing theory, with a direct match in Wang's preprint. NO_DIRECT_MATCH_FOUND_IN_DECLARED_SEARCH_SCOPE for the full tested composition.
Used by the collection only as public conceptual attribution of “Dija Negative” to Bledar Gjata and ESS-MAI; not implementation evidence. See the public disclosure trace.
Media records establish chronology and attribution; they are not technical evidence. Later independent work is declared without alleging copying or making a universal priority claim, and no priority claim is made here.
Where ESS-MAI appears
The author's own profiles and the public pages where ESS-MAI appears. Media and social posts establish chronology and attribution; they are not technical evidence. The technical evidence is the POC collection.
Agreement with the theory is not required. The highest-value 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.