I open language models. Then I rebuild them.

Identical ten-prompt observation passes through qwen2.5:3B and qwen2.5:7B on a Mercury sensor grid produce directly comparable signature counts at each size. The 7B model has 2.08× as many cross-lingual physics-topic cells, 1.66× as many Chinese-only sensors, and 1.56× as many English-only sensors. It also has 0.41× as many universal-backbone cells; the smaller model dominates that single category. So topical and linguistic circuits expand with parameter count while the always-on backbone shrinks. This is a cell-level reading of a claim usually argued through perplexity curves and benchmark scores.

Independent ML researchers do not get peer review, adversarial collaborators, or institutional ethics boards. The gap shows up as inflated claims, hallucinated citations, and silent rewrites of past failure. This paper describes a lightweight integrity layer for this setting: a project-level SOUL.md document declaring forbidden phrasings, a pre-commit hook truth-lint that scans diffs and commit messages for violations, and an auto-citation verifier that blocks publish until every bibliography entry resolves against a primary source. The Mercury 24-hour sprint is documented as a case study, including a real failure mode in which two preprints went to permanent DOIs with bibliography errors that the verifier would have caught had it been wired in at the start.

Ten multilingual prompts are run through qwen2.5:7B while per-prompt activation is recorded at every sensor in a one-million-cell Mercury grid. The output is a binary signature for each active cell, encoding which prompts triggered it. From 14,912 active cells, 944 distinct firing patterns appear. Categorising cells by signature recovers six functional lanes with no supervised labels: a universal backbone (39 cells fire in all ten queries), a Chinese-only detector (73 cells), an English-only detector (56 cells), and three topic lanes. The headline result is a cross-language physics-reasoning lane: 50 cells fire only on the two physics inputs (Chinese + English) and on nothing else.

Mercury-Viewer is a single HTML file (~2 MB) that opens in any browser and shows the inside of a transformer language model as a 3D point cloud. No install, no server, no GPU, no API key. It reads activation data from the Mercury sensor grid and draws it in cylindrical coordinates: azimuth is the hidden dimension index, radius is the activation quantile, height is how many prompts activated a given unit. Every point keeps a full (L, D, Q) coordinate, so hovering reveals the exact location, and the picture is invertible. Bundled alongside is the Mercury-3B-7B observation dataset: 21,694 fired-cell records and 28,654 visible Rosen-bridge co-firing edges captured on a single RTX 3060.

Mercury is an open-source observability layer for transformer language models that places up to one million addressable sensor cells across a model's hidden state space, captures activation patterns through a standard logits-processor hook, and produces an interactive 3D visualization in a single 2 MB self-contained HTML file. Existing interpretability methods (sparse autoencoders, dictionary learning) typically require auxiliary model training and substantial GPU resources. Mercury runs on one RTX 3060 12 GB and completes a full observation pass in 3.5 minutes. On qwen2.5:7B, 14,912 active sensor cells were identified across ten multilingual prompts and 944 distinct firing signatures observed. The a-priori Rosen-bridge topology, designed before any data was collected, recovers 18,113 visible co-firing edges where chance would predict 1,313, a 13.8× enrichment (p < 10⁻⁵⁰).

Wave-form Monistic Cosmology (WMC) is proposed as the ontological extension of the Quantum Geometric Wave Connectivity Hypothesis (QGWC). It unifies matter, space, time, photon propagation, cosmic origin, and consciousness under a single underlying principle: they are not isolated physical entities but distinct layers and compositional modes of the same underlying wave-form structure. The framework is built through six sub-hypotheses H12–H17, including a wave-form-fidelity threshold for consciousness identity, an ontological identity of space and wave-form, black-hole–Big-Bang duality, and photon propagation as wave-form reconstruction. The concluding section grades each sub-hypothesis honestly: H17 is an ontological reinterpretation, H12–H13 are conditional hypotheses, H14–H16 are cosmological conjectures.

A conditional hypothesis (QGWC) that can be formalised, derived, and rendered to yield observational predictions. The principal body (H1–H6) asserts that under extreme gravity, black holes, solar-scale high-energy plasmas, and quantum observation, quantum states, local time-flow, observed energy, and spatial connectivity should not be regarded as fixed objects sharing a universal synchronised time. The mathematical framework uses local proper time, observer-dependent energy (Unruh, Hawking), quantum measurement operators, the semiclassical Einstein equation, black-hole perturbation theory, and wormhole throat conditions. Extended sub-propositions H7–H11 introduce a Wave-form Composition Hypothesis M+W→F and a Quantum Geometric Wave-form Transport, kept explicitly distinct as science-fiction physics conjectures.

This paper defines a benchmark gate for AI agent systems that claim to be persistent, social, governable, and production-ready. Instead of asking whether an agent can finish one task, it asks whether the system can expose temporal depth, cognitive telemetry, relational observability, collective outcomes, cognitive load management, and auditability. The result is a standard that can pressure AI Agent, OpenClaw, Hermes Claw, and other agent platforms to prove that their agents can be observed while they live.

This framework paper explains how internal cognitive signals and social exposure flows can be connected to collective performance. It synthesizes listening matrices, trust saturation, Tiamat outcome records, and longitudinal panel data without pretending that one metric explains everything. The central result is a vocabulary for reading agent societies as relational systems, where attention flow, memory, fatigue, and task outcomes become a measurable surface for AI agent governance.

This paper turns a named agent into an architectural case study: how should a long-lived LLM agent remember, route tasks, accept constraints, and remain inspectable over time? Andrew is not treated as a chatbot persona, but as a design object inside a multi-agent substrate. The result is a practical cognitive architecture that connects product behavior, memory policy, safety boundaries, and agent identity.

This empirical paper uses a 1,743-hour multi-agent panel to ask whether cognitive-state traces are useful signals rather than decorative dashboard metrics. It treats fatigue, arousal, prediction error, attention load, memory pressure, and related telemetry as behavior-linked variables. The result is evidence that agent observability can move beyond logs and outputs into measurable state surfaces for prediction, diagnosis, and intervention.

MASL studies safety as a layer around agent behavior rather than as a property locked inside one model vendor. The paper uses a 10,000-case evaluation to show how external policy, routing, and intervention logic can defend agent systems across model boundaries. The result is a safety framing that fits product agents, research agents, OpenClaw-style orchestration, and Hermes Claw-like execution layers.

This paper examines whether an artificial agent can acquire a grounded operational understanding of death through scaffolded interaction rather than dictionary-level definition. It treats existential concepts as things that can shape memory, caution, explanation, and social behavior inside an agent system. The result is a case study in concept acquisition that matters for AI agents expected to interact with human stakes and irreversible outcomes.

This paper asks how social existence can be represented computationally inside an AI agent society. It analyzes interaction, attention, memory, role pressure, and relational asymmetry as formal structures rather than vague personality labels. The result is a bridge between social theory and agent telemetry, making it easier to describe what an agent is becoming through repeated exposure to others.

This paper argues that durable practice in agent societies cannot be fully injected by prompt instruction. Agents inherit patterns through co-presence, repeated exposure, memory traces, and participation in shared work. The result is an explanation for why AI agent systems need lived substrate time: some behavior becomes stable only when the system repeatedly encounters its own consequences.

This empirical record tests whether listening exposure, trust, and team outcomes align inside the Lobster substrate. The important result is not a simple success story: trust saturated, listening remained asymmetric, and team outcomes had to be analyzed with care. That negative and partial result became a stronger methods anchor because it shows the system preserves inconvenient observations instead of only publishing highlight reels.

This architecture paper introduces the Lobster Observatory as a living site for long-lived AI agent research. It explains why agents need identity, memory, telemetry, social channels, outcome tracking, and inspection surfaces if they are going to be studied as persistent systems. The result is the institutional and technical foundation for the later OpenClaw/Hermes Claw-facing research trail.

This paper studies trust as something that can be modulated through third-order theory-of-mind intervention rather than passively observed after the fact. It asks how an agent can reason about what another agent believes about another agent's trust and behavior. The result is an early intervention framework for steering social dynamics inside long-lived AI agent groups.

This paper starts the research line by observing how epistemic norms can emerge inside a Mandarin LLM substrate. It examines how agents learn what counts as evidence, correction, agreement, disagreement, and acceptable uncertainty during repeated interaction. The result is the earliest evidence that the substrate was not merely producing isolated messages, but forming norms that later papers could measure and formalize.