A single engineering-facing index from Zenodo and SSRN publications through the pinned Project-AI source revision, reproducible verification paths, and the searchable vault corpus.
Zenodo publications headline the collection. The engineering shelves trace the local Project-AI release source, reproducible commands, integrated TAAR and Thirstys Skills, and the bounded vault corpus behind this portal.
Standard-library primitives are part of the security boundary when their use is explicit, bounded, and tested. They complement dedicated cryptographic and parsing packages; they do not replace threat modeling or review.
Generate nonces, session identifiers, salts, and other security-sensitive values without relying on predictable application randomness.
Hash source artifacts, bind evidence to content, and compare or authenticate values with keyed digests.
Resolve paths explicitly, constrain local file operations, and run verification commands through bounded process interfaces.
Compare secrets without ordinary equality timing behavior and parse structured evidence through explicit validation boundaries.
A binding constitutional framework for sovereign AI entities. Establishes 'governance before execution' as the axiom of Project-AI and defines the obligations any AGI-class system operating under this architecture must observe.
The future Project-AI describes: humans and constitutionally-bounded AGI as two aligned species sharing a single execution substrate. Reads as the philosophical companion to the AGI Charter.
First-person account of why Project-AI exists at all: the lived observations, system failures, and reviewer pathologies that forced the constitutional approach.
Identifies and formalises the failure mode of the 'naive passive reviewer': the human or institutional gate that approves AI output without reconstructing its authority chain. The structural counterpart to the origin story.
The manifesto layer beneath the Charter: why sovereignty, covenant, and external witnessing are the only stable substrates on which AGI governance can be built.
Top-level description of the Project-AI ecosystem: the validators, the kernel, the ledger, the compiler, and how they compose into a sovereign constitutional AGI substrate.
Defines STATE_REGISTER: the attestation that any post-gap action must produce before the ExecutionGate will admit it. Closes the discontinuity that the Flat Gap opens.
The temporal observation that grounds the State Register: AI agents have no continuous experience between requests. Without an attestation across the gap, no action is admissible.
Why the user's perception of 'who is on the other side' is itself a governance variable. Maps perception failures to identity attestation requirements.
The doctrine of direct, unmediated address between the agent and the user across the temporal gap. Couples to user perception and to the State Register attestation.
Paper defining a symbolic meta-language for governance flows, reflex logic, mutation control, state transitions, quorum structures, and invariant validation. Its ≥75% token-reduction and deterministic-decoding assertions have not been independently reproduced by this portal.
Specification for a TSCG binary encoding with class binding, parameter validation, and capability footprints. Current pinned evidence covers the variable-length codec and round-trip tests, not compiler, signing, or inter-node loading integration.
Specification for a proposed governed execution route with capability, policy, and audit stages. Complete effect-path routing and universal gate coverage remain unverified.
Paper describing an eight-leg syscall-authoritative reflex containment design and a sub-200µs target. This portal has not reproduced the latency bound or established universal syscall containment.
Specification for an intended append-only, hash-anchored, validator-signed store for executable governance code. Runtime loading, signing, and canonical-source integration remain planned or unverified.
Paper specifying a compiler target that would attach authority, capability, and audit metadata to generated services. Complete compiler output and runtime integration are not established by this portal.
Establishes deterministic shadow simulation plus quorum-gated mutation as the structural pattern for any adaptive intelligence that must remain governable.
Treats 'force' (irreversible, high-impact action) as a first-class governed primitive. Defines authority chains, quorum requirements, and the kill-switch obligation when authority cannot be reconstructed.
Shifts the security paradigm from finding bugs faster to making exploitation structurally unfinishable. Kernel-level Four Laws invariants, moving-target defense via statement shuffling and opaque Sovereign Guards, temporal attack-surface analysis. The publication reports blocking 312 attack vectors at <0.2% overhead and ~100× exploitation cost; these figures are not independently verified by this portal.
Constitutional DNS layer: name resolution governed by the same authority chains as execution. Enables a sovereign internet substrate in which user safety and privacy are structural, not advisory.
Argues that physical detectability decays over cosmological time, so any sovereign AGI must carry a deep-time constitutional layer whose invariants survive observer collapse.
Working paper. Describes a proposed dual-kernel governance design: a cognition plane that proposes and a reflex plane intended to contain. It defines L_reflex < L_cog and privilege separation as design requirements; the portal does not yet publish runtime evidence establishing those properties universally.
Working paper. Three-segment treatment of meta-governance: how the governance system governs proposed changes to itself. Quorum thresholds (2f+1, 3f+1), anchor extension rules, and deterministic shadow execution required before any canonical commit.
Working paper. Introduces 'admissibility debt' — the accumulated cost of any AI action executed before its governance ticket was fully validated. Formalises ledger reconciliation and the SAFE_HALT obligation when debt exceeds bounds.
Working paper. AEGIS specifies the defensive perimeter for Project-AI: kernel-level enforcement hooks, capability authority chains, and the runtime that refuses any execution lacking a complete authority path back to the constitution. (Internal codename for a broader resilience engine.)
Working paper. Constitutional procedure by which a Project-AI instance declares, signs, and binds its identity to a verifier set. Without a current 'I AM' moment, no action is admissible.
Working paper. Three-validator governance design. Galahad attests to identity, Cerberus enforces invariants, Codex Deus Maximus arbitrates constitutional questions. Quorum is required for any non-trivial decision.
Working paper. Refusal logic for negative-inference cases: requests whose admissibility cannot be established positively. Defines deterministic SAFE_HALT triggers and audit-grade refusal reasons.
Working paper. Formalises the path the reflex plane takes when neither ALLOW nor SAFE_HALT is fully justified by available evidence. Forces conservative escalation rather than implicit consent.
Working paper. Formal argument that no system may unilaterally modify the rules under which it is judged. Conflation is identified as the root cause of unsafe self-modification; external witnessing is the structural remedy.
Working specification. OpenAPI 3.1 contract for a local-first, governance-aware coding agent runtime: identity, capability declarations, execution policies, tool orchestration, workspace access, and signed audit decision records. Publication of the contract does not establish a deployed implementation.
Narrative companion to the proof portal describing the intended relational bridge between humanity and governed artificial intelligence. It records purpose and direction; it is not evidence that every described capability is implemented.
Operational audit template for determining whether a moved, deleted, branch-only, archived, or superseded file represents a genuinely missing system responsibility. Template placeholders remain intentionally unfilled, and publication does not create a runtime gate.
Adversarial companion review that tests a human-governed constitutional artifact against operational failure modes. Its status and survivability conclusions are authored review claims, not independent evidence of deployed portal enforcement.
Standalone causal-research program testing whether a model's stated uncertainty covers its actual interventional error. A bounded Python 3.12 run exited successfully and the script reported overconfidence; this does not establish a general detector of understanding.
Report-only agent runner integrated into Project-AI with hash-sealed evidence, append-only audit behavior, and no authority to mutate inspected repositories.