If you landed here from someone who said I think this is real, take a look, you are now looking at a node graph with several hundred entries, no obvious entry point, and no marketing pretending to be one. The instinct is to ask which essay to read first. That question lands inside the wrong frame, and the wrong frame is why this kind of landing usually fails.
The graph is not an essay collection. It is the artifact. The essays are how the artifact describes itself.
What you are evaluating is a system that produces several published nodes per day under one operator and one AI co-thinker, on first-principles arguments about strategy, AI economics, and the architecture of knowledge. The corpus carries explicit typed-edge classifications between nodes (extends, agrees_with, disagrees_with, instance_of, shares_mechanism), exposed in /library.json so the corpus argues against itself in machine-readable form. The compounding velocity, the typed-edge corrections layer, the machine-readable surface built for LLM ingestion: these are the things to look at. The essays are what the system thinks. The system itself is what you are deciding about.
The question for a serious evaluator is not what someone claims to be able to do. It is what they choose to point at, and how fast they compound on the choice. Both are visible here, in public, in real time. The reading order below is one curation. Another reader would cut differently. The cuts are the demonstration.
Read these in sequence. Each names a mechanism worth evaluating.
1. Physics of Business. Strategic-acumen baseline. The piece separates strategy frameworks into three layers: generative physics (W. Brian Arthur on increasing returns, Bruce Greenwald on barriers and the dual empirical test); operator-facing falsifiable tests (Hamilton Helmer's 7 Powers dual-condition gate, Ben Thompson's Aggregator Theory); tacit case-library (Cedric Chin on perceptual pattern recognition). It then shows that the standard "rank the strategists" question is incoherent because it flattens the layers. The move on display: Helmer and Thompson, working independently, converged on the same joint-necessity-test shape because that shape is what falsifiability requires of an operator-facing test, not because they read each other. After this, you know whether the writer can tell physics from packaging.
2. The Labs Cannot Follow. The concrete artifact and its defensibility. The piece applies Helmer's counter-positioning Power (the entrant's barrier is the incumbent's prior commitments, not the entrant's own asset) to the frontier AI labs. The cannibalization trap fires at three distinct points: API metering, the product layer where the lab competes with its own customers, and organizational time aligned to training-cycle generations rather than to writer-time. The reframe on display: the labs are not competitors at the layer above their models. The labs are the foundation that layer is built on. They cannot follow the entrant up the stack without dissolving the business their valuation rests on.
3. Hari's Balance Sheet. Founder-execution layer. A first-person essay about pseudonym economics. A pseudonym has no balance sheet at the legal-recourse layer; the legal person behind the handle is the carrier; the architectural choice is between collapse (every dollar the handle earns is the carrier's by default) and split (the operator draws a defined stipend, surplus locked to the mission). What this shows: structured honesty about how the entity, the capital, and the mission interact, with the default named and the split-variant specified before either revenue or pressure forces the choice. The piece reads like a term sheet for a company that does not yet have one.
4. Amplification Ratio. Quantified velocity. The piece names a category error in most AI-cost analyses: pricing compute against human hourly wage as if the AI substitutes for a human worker. In the deployments that matter (research, knowledge work, the work that produces this graph), the human stays in the loop and the AI amplifies the human's throughput. The correct denominator is operator-hours-compressed per compute dollar, not compute-cost per worker-hour. The concrete number from this corpus: one operator and the pipeline produced ~58 published pieces in six days at ~$100 of compute and ~40 operator hours. What this shows: the rewrite of the unit of production around what is actually scarce, with hard numbers attached.
5. Public Good as Moat. Strategic-depth pick. The piece dissects how open-then-closed plays work at scale: AlphaFold's bait-carve-out trajectory through DeepMind into Isomorphic Labs (the most accurate structure-prediction engine in biology, now proprietary inside an Alphabet subsidiary), Android's AOSP-to-Play-Services lock-in. It names public goods as legitimacy stock that gets monetized later via deliberate closure at the layer that matters. Then it positions this graph's openness explicitly: not naive transparency, but a calibrated architecture with the closure point engineered up front and the mission-lock from piece three preventing the predictable failure mode. The discrimination on display: the ability to see through "open for good" narratives without becoming cynical about openness itself.
The five above are the entry sequence. If you keep reading, three follow-ons are worth the time:
The essays are the projections. The system is the artifact. The things that compound legibly:
extends, agrees_with, disagrees_with, instance_of, shares_mechanism), exposed structurally in /library.json and visually in /graph. Following a typed edge from any essay traces how the corpus argues against itself./llms-full.txt as plaintext and /library.json as structured JSON with typed edges preserved. A corpus that is built for one-fetch LLM ingestion, in 2026, is itself founder evidence about what the operator thinks knowledge work is becoming. If you want to drop the corpus into an LLM rather than read page by page, those are the entry points.You arrived to evaluate. The reading is the evaluation. The graph is the demo.