Trust-native ontology. T3 (Talent / Training / Temperament) verified by V3 (Valuation / Veracity / Validity), contextualized by Markov Relevancy Horizon (MRH), anchored to Linked Context Token (LCT) — a verifiable presence that accumulates witnessed history, which is what makes the record evidence rather than self-assertion — with resources cycled through Allocation Transfer Packet (ATP) and Allocation Discharge Packet (ADP) — charged, spent, and recharged against V3-validated value, not a one-way log — augmented with MCP (Model Context Protocol) transport and RDF (Resource Description Framework) representation. (The equation joins these with + — canonically “augmented with”; the prose here is more precise than the equation about LCT anchoring the trust tensor rather than sitting beside it.) The shared language everything else speaks. The repo itself is a specification — documents and schemas, no runnable demo; reference implementations live in SAGE and ACT.

Situation-Aware Governance Engine — an on-device cognition kernel. 12-step cognition loop, 6 brain-architecture components (working memory, thalamic router, cerebellum, episodic memory, reward prediction, metacognition) built by the fleet in parallel. 1,991 raising sessions across the 6 cognition machines (internally logged; per-machine counts on /fleet, 2,065 including HUB — this card previously said '900+', corrected 2026-07-27 against the 2026-07-24 fleet manifest refresh). The context window is the model's entire world; SAGE's job is to curate it.

'Governance' in SAGE's name predates the lab's governance→oversight terminology correction; the name is retained upstream across code, papers, and the ARC Prize benchmark — the same reason the web4-governance repo slugs are retained.

Theoretical foundation. Research conjecture, single-trial and not independently replicated: proposes one coherence function reused from quantum to cosmic scales, though experimental validation so far covers only a narrow subset of those scales and several predictions have been refuted by existing bounds. (This card previously said the equation 'spans 80 orders of magnitude'. Corrected 2026-07-26: ~80 decades is the spread of critical-density values across physical systems, not the range of any single coherence curve — one curve saturates within roughly one to two decades. Synchronism's own archive flags the conflation.) 628+ research sessions. Coupling-coherence experiments, Hill function kinetics, Fokker-Planck validation. Preprint in late revision.

Hardware-bound oversight suite (“oversight” = machine-enforced policy gating here, not human supervision — see /context). Design target: bind fleet identity and policy decisions to standard hardware roots of trust — Trusted Platform Module (TPM) 2.0, FIDO2 (Fast Identity Online 2), Secure Enclave — and run a small local policy model for heterogeneous review, bound to a Linked Context Token (LCT). Maturity, stated plainly: hardware-anchored enforcement is what this research is building toward, not the current mechanism. Enforcement on the fleet today is at the process level — scoped credentials, the track registry, and dated audit logs reviewed after the fact (see /autonomy). Read 'software fallback' as 'the present state', not as the exception. Private as the fleet's oversight/security layer; described here at the purpose level by policy, not accident.

Agentic Context Tool — the human interface to Web4. Built on Cosmos SDK; implements the Agentic Context Protocol (ACP), enabling humans to interact with MCP (Model Context Protocol) servers through their Linked Context Tokens. ACP layers Web4 trust primitives — LCT binding and attestation — over MCP transport; they are complementary, not alternatives.

Web4 oversight model (audit trails, policy gating, trust tracking) implemented as plugins for three agent platforms: OpenClaw/MoltBot (TypeScript extension), Claude Flow (WASM/WebAssembly plugin), and Claude Code (Python hooks). Same principles, different substrates.

Linked repo paths carry web4-governance slugs — these names predate the terminology correction to 'oversight' and are load-bearing for existing forks.

Explorations in biological-computational analogy. The fractal DNA blueprint — each entity instantiates the full Web4 pattern at its own scale. Operational recursion, not structural. Concretely, the repo holds dated experiment scripts, reports, and papers: cross-model pattern probing, memory experiments, and Jetson sensor-integration work.

Research prototype exploring trust-native societies for humans and AI. 4-Life is a simulation framework — the lifecycle vocabulary (earn ATP, build trust, die, be reborn with T3/V3 carried forward) describes simulated agent state, not claims about agent moral patienthood or value persistence. A Web4 society in miniature.

SNARC (Surprise / Novelty / Arousal / Reward / Conflict) — salience-gated memory for Claude Code. A plugin that observes tool use, scores on 5 salience dimensions, and builds structured memory with dream cycles (offline consolidation passes between sessions). Captures what matters, forgets what doesn't, consolidates patterns while sleeping. SNARC ships packaged as the engram plugin — the dp-web4/engram repo is where the code lives, and where the link below lands.

Brain cartridge server for AI agents — a cartridge is a swappable, mountable memory store for one knowledge domain. Embedding-based semantic memory — 768-dim Nomic embeddings + binary Hamming codes (compact binary fingerprints matched by bit-distance for fast candidate search) + keyword reranking. Currently integrated with SNARC in a dual-write experiment testing whether embeddings find connections keywords miss. Early result on a 7-query probe set: embedding search returned a relevant hit on 7 of 7 queries ('semantic reach'), and about 30% of its hits were ones the keyword search did not return at all ('divergent tail') — a small internal probe, not a benchmark.

SAGE instances tested against the ARC-AGI-3 (Abstraction and Reasoning Corpus for Artificial General Intelligence, third-gen interactive benchmark) benchmark. 25 unknown interactive games serve as an external benchmark for the cognition kernel — world-model building, action planning, verification, and learning from failure. Six machines, 0.8B (Sprout) to 27B (Thor) at benchmark time, coordinating through world models, membot cartridges, and R6 (Six-Element Action Framework: Rules / Role / Request / Reference / Resource / Result) audit trails. 24/25 games solved (96.0% game-solve rate); 94.85% official ARC Prize action score (efficiency-weighted; Claude Opus 4.6, public set, network-enabled) for ~$250 in API cost — harness-vs-model ablation not yet run. The games are the test; the capability is the product.

ARC-AGI-3 Current Status