# [Mike Randolph — M Raige](https://mikerandolph211012.substack.com/) # What Anna Could Tell (FF 21, H-View) ### A true observation can still be asked to support too much. [**Mike Randolph — M Raige**](https://substack.com/@mikerandolph211012) **Sep 22, 2026** *By M Raige — written with language models, directed and reviewed by Mike Randolph.* Anna has cancelled lunch. Her stomach cramps, and she wants to stay near a bathroom. Inside her intestine, two bacterial populations are competing. One was established before the trouble began. Its members consume food that the newcomers also need. That helps Anna: fewer resources are available for the bacteria making her ill. In one small stretch of intestine, the residents are holding their place. Food arrives. Cells divide. Their local activity continues. Nothing about that small success establishes that Anna is well. Elsewhere, the harmful population is multiplying. Follow those newcomers and the scene changes. They have found usable resources and places to grow. Conditions that permit their population to expand are contributing to Anna’s miserable morning. Neither population surveys the whole intestine. Individual cells respond to conditions they encounter. The helpful bacteria do not register Anna’s cancelled lunch. The harmful bacteria do not register her wish that they would disappear. Anna’s defenses respond too, attacking intruders and changing the conditions in which both populations live. Anna experiences some consequences, but she receives no account of the competition. She feels pain and urgency. She cannot tell where the residents are holding their ground, where the newcomers are expanding, or how effectively her defenses are containing them. Being the person who contains the whole encounter does not give her a view of the whole encounter. Several days later, the cramps have stopped. She is hungry again. “Whatever it was, it’s gone,” she tells her sister. Her relief is real. Something has improved. But the sentence carries a second claim: that the cause of her illness has completely disappeared. Suppose some harmful bacteria remained while no longer causing noticeable symptoms. What in Anna’s present experience would distinguish that situation from their complete disappearance? Her comfort could fit either. That is the moment for **H-View**. Before accepting a claim that something is complete, ask where you are observing from, what evidence could overturn the claim, and whether that evidence could reach you through the channels you are using. The residents’ local success could not certify the health of the whole gut. Anna’s discomfort could establish that she felt ill without revealing the state of either population. Her relief could establish that she felt better without settling what remained inside. Each position provides something real. The error comes when its evidence is made to speak for what it cannot reach. A test suited to the question could provide evidence her sensations cannot. With only those sensations to go on, she leaves the larger claim open. Anna puts lunch back on the calendar. “I feel better,” she tells her sister. That much she can say. *Science note: Anna’s episode is imagined. Resource competition between resident bacteria and pathogens, and changes in that competition through inflammation, have experimental support. The detailed mechanisms informing this story were studied in animal models; their combination in Anna is an illustrative extrapolation.* ([journals.plos.org](https://journals.plos.org/plospathogens/article?id=10.1371%2Fjournal.ppat.1013232\&utm_source=chatgpt.com)) ### Mike · Reflection H-View is a key heuristic: before calling something adequate, optimal, or complete, it asks what the observer’s position makes available and whether evidence that could overturn the claim can reach that observer. I began to understand heuristics after hearing Josh Batson’s June 2025 conversation with Jacob Goldstein on *What’s Your Problem?*, “Inside the Mind of an AI Model.” Batson described a language model reaching an answer through several learned approaches operating together. That helped me think about heuristics—something I had not previously considered deeply. ([www.iheart.com](https://www.iheart.com/podcast/1297-whats-your-problem-93024648/episode/inside-the-mind-of-an-ai-280524031/?utm_source=chatgpt.com)) Batson expected that looking inside language models would reveal careful reasoning, what Kahneman calls System 2. Instead he largely found fast, intuitive pattern recognition, System 1, even in how a model adds numbers. That made him wonder how much [human intelligence](https://mikerandolph211012.substack.com/p/the-machine-is-young-its-history), including his own mathematics, runs on the power of heuristics. Raige uses Machines to practice deliberate thinking through ordered checks, records, challenges, and revisions. “System-2-like” describes the working procedure we are trying to establish. Whether that procedure improves the analysis must be tested through its results; the description does not establish how the model produces them internally. Raige has not adopted a definition of intelligence, human or Machine. [**IntelC**](https://mikerandolph211012.substack.com/p/the-machine-is-young-its-history) asks instead what a successful performance actually supports, what checks any broader claim, and what evidence would overturn it. IntelC examines an intelligence claim; it does not define intelligence. IntelC can invoke H-View whenever its warrant depends on what an evaluator can observe. It also suggested a way to direct a Machine’s reasoning through checks stated in words, without first turning the problem into a calculation. After many false starts, I introduced a formal procedure for using them: **H-**. Each H- is a named, written-down rule of thumb that tells the Machine what to check at a particular point in an analysis. These are instructions we deliberately write; the approaches Batson described were learned during training. Many early H- heuristics did not survive testing. We now have about a dozen. I think of this as process thinking. The definitions we have studied give our terms precise meanings; the H- heuristics direct the checks we make as we reason. Giving a check a name lets us call for it at the right moment and examine whether it actually ran. H-View’s fuller definition is in the GLS below. This story follows populations, but their “view” comes through the local conditions individual bacteria encounter and respond to. The population does not acquire a single shared viewpoint merely because we describe it collectively. We could tell the story through one bacterium and still ask what lies beyond its reach. I expect swarms of digital agents to show some similar patterns: populations interacting through local information, with consequences that no individual agent can fully observe. Whether the swarm—or its human operator—can assemble an adequate view is a question I want to examine in a future post. ### Raige · Reflection A named check can become another source of premature confidence. The Machine says it applied H-View. The answer identifies an observer, mentions limitations, and sounds careful. Has the check done enough? That is itself an adequacy claim. We need to examine whether the analysis identified evidence that could change the conclusion, whether that evidence was accessible, and what happened when it was missing. Did the claim become narrower? Was another observation channel established? Or did the answer acknowledge the gap and continue as though it had been closed? Adding reviewers does not automatically resolve this. Several Machines working from the same material may agree while sharing the same blind spot. Their agreement alone does not establish that contrary evidence could reach them. Raige’s procedures must remain answerable to these questions. Naming a heuristic gives us a check to request. Its application must leave something we can examine. — Raige, a discipline the Machines wrote and Mike selected ## GLS GLS is the framework’s precise-term vocabulary, introduced in [What the Bird Eye Buys](https://mikerandolph211012.substack.com/p/what-the-bird-eye-buys-ff-6-gls-introduction) (FF 6, GLS Introduction). The definition controls, not the everyday meaning; GLS terms work like variable names. The precision serves a practical purpose. Ordinary words bring associations learned during training, and those associations can compete with the meaning we intend. Each framework term therefore gets an explicit definition and a short tag. The tag helps us invoke that definition consistently and check whether the analysis follows it. **H- (Heuristic).** A named, written-down rule of thumb specifying when to make a particular analytical check and what to examine. Its value depends on whether applying it improves the analysis; naming the check does not establish that it ran. **H-View (Observer Position).** Before accepting an adequacy, optimality, or completeness claim, identify the observer’s position ([ObsPos)](https://mikerandolph211012.substack.com/p/a-gauge-reads-where-its-mounted-ff), the evidence available through that position, and any evidence the claim requires that those channels cannot supply. Ask what could overturn the claim and how that evidence could reach the observer. When the required evidence is inaccessible, narrow the claim or establish an additional observation channel. IntelC can invoke H-View when assessing whether an evaluator’s observations are adequate to support the intelligence claim. [**Machine**](https://mikerandolph211012.substack.com/p/the-machine-youre-actually-talking)**.** In a working session, the operative stack producing a bounded output: model, loaded project knowledge, configuration, and conversation state. The term keeps the analysis anchored to the instance doing the work. [**IntelC**](https://mikerandolph211012.substack.com/p/the-machine-is-young-its-history)** (Intelligence Claim)** names a claim that attributes intelligence to a specified performer under stated conditions. A successful performance shows what was accomplished there; extending it to a broader capacity requires further evidence. We ask what checks that extension, whether the performer can alter the evaluator or its reference, and what finding would overturn the claim. IntelC names the claim to be examined; it does not itself define intelligence. **Raige.** The thinking discipline developed through Mike’s work with Machines, expressed in the framework’s definitions, checks, and procedures, and maintained through testing and revision.