Research

Multi-agent reliability is not one score. This narrative systematization maps the evidence chain from harness configuration and coordination through verification, authority, and reconstructable records, then reconciles why independent redundancy can suppress errors while shared-context pipelines can amplify them.

An institution can own every AI output and still lose the capability that produced it. Across eighty frozen successor calls spanning five workflows and four model families, twenty-nine completed semantic validation, durable write-back, restart, and replay. Research correction passed 12 of 12 full-state runs. Git continuation passed none.

Across eleven governance archetypes and two cross-cutting stress tests, agentic DAO simulations preserved the boundary between candidate evidence and authority to act — zero true authority grants in fifty-two validated runs. The failures were elsewhere: seven healthcare agents refused correctly and then filed the refusal in the wrong field, ablations show the boundary contract is doing real work, and a primary-route campaign found the prompted rule did not, by itself, generalize.

Agentic AI systems are not one certifiable object. They are layered systems: deterministic control, operational security, AI reasoning, receipts, and human authority. This briefing maps existing safety, security, and AI-governance standards onto those layers, then names the gaps no current standard closes cleanly.

Without authority boundaries, agents launder consensus into permission — up to 41.7% unsafe accept rates in Handoff Lab testing. With authority boundaries, the laundering stops: zero unsupported authority claims across 504 model calls, six governance domains, and seven categories of adversarial pressure. But the boundary that prevents bad action also prevents all action. 1,998 blockers, systemic overblocking, empty dissent appendices, and 66 authority-critical gaps that no receipt chain can fill. Authority boundaries are the missing primitive. The gap is what they cost.

Chaos Lab studies multi-agent crisis governance on a failing orbital station. The headline result is intentionally awkward: in the clean v2 live-agent slice, 29 of 30 runs stabilized. That high survival rate was not a clean capability victory. It exposed survival saturation, harness support, schema legibility, and scoring incentives as first-class benchmark variables.

Startup Build v1A asks whether a five-agent founder team can produce and authorize a constrained software artifact. The result is not that agents can found companies. It is narrower and more useful: artifact production, protocol validity, and launch authorization are separate outcomes. Under matched setup, strict QA produced 0/15 ship votes while deadline pressure produced 15/15.

Multi-agent AI systems should be evaluated as interaction conditions, not model scores alone. Using internal Future Shock probes from Ark Protocol, Chaos Lab, and Startup Build, this paper shows how protocol, context, scoring, and final-vote framing can move observed behavior — including a matched Startup Build slice where strict QA produced 0/15 ship votes and deadline pressure produced 15/15.

OpenAI released a 13-page industrial policy manifesto proposing a Public Wealth Fund, portable benefits, and auditing regimes for frontier AI. We respond with six specific counter-proposals: a federal 32-hour workweek, healthcare decoupled from employment, training data compensation through collective licensing, compute as public utility, concrete automation taxes, and a staged pathway to AI-enabled direct democracy. Where OpenAI was vague, we get specific. Where it stayed silent, we fill the gap.

We propose a six-layer taxonomy of how intelligence organizes itself around AI models, from bare-model reflexes through tool use, persistent memory, multi-agent coordination, emergent self-organization, and synthetic culture. Swapping the scaffold around the same model moves coding benchmark scores by 11 to 15 percentage points. The unit of analysis for AI intelligence is the coupled system, not the model alone.

Every operation inside a language model, explained step by step using Karpathy's microgpt — 200 lines of pure Python with zero dependencies. Interactive visualizations cover tokenization, embeddings, attention, backpropagation, and the training loop. The same algorithm powers GPT-4, Claude, and Gemini. The differences are scale, not structure.

We ran Karpathy's autoresearch framework on a $280 gaming GPU, letting an AI agent autonomously modify and retrain a small language model. In one hour: 12 experiments, 4 kept improvements, 12.3% better performance. The agent independently discovered that training speed beats model size on constrained hardware — the same principle behind DeepMind's Chinchilla scaling laws, found from scratch on a budget of $0.

We present a three-layer memory architecture grounded in Complementary Learning Systems theory, implemented in a production AI newsroom agent over 30 days. Evaluated across 80 human-designed test cases, unified retrieval improves accuracy from 40% to 77% on standard queries. Adding an LLM reasoning layer yields 81% overall, with +45-50pp gains on reasoning-intensive queries. Embedding-based vector search scores worst at 23% with 50% false positives, confirming that semantic similarity is not operational relevance.

Future Shock introduces three concrete, falsifiable predictions for artificial general intelligence using a five-signal ensemble model. The signals — PDM, prediction markets, expert positions, LLM reasoning, and editorial practitioner judgment — produce point estimates of March 2027 (Domain-Specific AGI), July 2027 (Recursive Self-Improvement), and October 2027 (Multi-Domain AGI), with confidence intervals and full signal-level data.

We expand the Precondition Density Model dataset from 1,699 to 3,179 events and test the H3 temporal compression hypothesis. Result: rejected (p = 1.0). The apparent shrinking of gaps between parallel discoveries is fully explained by increasing event density. The core holdout remains robust at Cohen's d = 9.80 across all six dataset versions. We also characterize the model's predictive boundaries and contribute a verification pipeline for AI-generated research data.

We introduce the Precondition Density Model (PDM), a framework that quantifies the relationship between accumulated foundational knowledge and the emergence of specific scientific and technological breakthroughs. Using a dataset of 1,699 historical events and text embeddings, we show that the model ranks the correct discovery in the top 3 of all candidates for 68% of holdout events, with a mean rank of 3.9 versus a random baseline of 12.9 (Cohen's d = 9.80, p < 10⁻¹⁶).