{"path":"research/claims-and-code-one-loop.md","content":"# Claims and code: one loop, two directions\n\n*2026-10-25. Two of his questions, read against the record. Everything here is a proposal and unruled.*\n\n## 1. His questions\n\nOn the first limit of *they are one structure* (`saying-what-kind-a-claim-is.md` § 9b): *\"This has been discussed in\nthe past, right, all the stuff about discussing the inner workings of Deliberus phrased as claims in the graph,\nmeta-argumentation or what we may call it, read up on this.\"*\n\nAnd: *\"In perhaps the Borderland docs I discussed (iirc) the possibility for claims about the greenhouse at BL to have\nsoftware be built and rebuilt regularly based on the claims/argumentation in the graph that talks/argues/settles how\nfor example the plants in the greenhouse should be watered, temperature-regulated, moisture controlled for them in\ndifferent seasons, whatever you might imagine on a concrete domain-specific level, or multiple interconnected domains\nargued about in the graph (or in a compartmented graph, as in the Borderland case). Has this been logged as a possible\nfuture improvement? One could set Claude to act on a loop or just enable me to press a button, to run CC to interpret\nthe graph at time T and enact all the leading/settled-enough rules and machinery that you might want (with certain\nsafeguards/guardrails, needless to say).\"*\n\n## 2. What the record already holds on the first direction: code into claims\n\n- **The three answers of August** (`what-belongs-in-the-ontology.md` § 5): a commitment the system cannot hold two\n  people disagreeing about is de-baked into a setting, given a does-not-fit value, or **published**: written into the\n  graph as a claim anyone can attack. Publishing is named there as the operator-reflexivity gap: the arithmetic, the\n  taxonomies and the prompts live in code, and no one can attack them in the graph.\n- **First instance, 2026-08-22**: the is/ought partition was published as `claim_ontology_isought`, grounded by a\n  claim on the two floors and attacked by the live hedonic-monism claim (`scripts/seed_ontology_commitments.py`).\n- **The limit was already stated** (ledger AJ4, unruled): Russell's further language for the structure of the first,\n  so publishing narrows the gap one commitment at a time and never closes it, because each publication relies on the\n  form that remains (`the-tractatus-and-deliberus.md` § 5, § 7). The first limit of § 9b restates it.\n- **The four questions of 2026-10-10** (`claim-or-system.md` § 3): can two people disagree about it (a claim), does it\n  say who did what (the record), can it be computed again (a reading), and what is left is form, kept small and made a\n  setting wherever two people could want it different.\n- **His hold, 2026-08-30**: *\"Let's hold off on the whole meta-rabbit hole of self-hosted fractal priors or beliefs\n  about Deliberus stored inside Deliberus. I'll ponder that for a while before we have at it. I feel like all the\n  in-flight reviews have to at least come much closer to being done before I wanna start on that.\"* Founder question 2\n  (should Deliberus host its own design, `conviction-and-critique.md`) sits under the same hold.\n\n## 3. The second direction: claims into code\n\nHis greenhouse idea was not found in the record before tonight (searched: his typed words since 2025-09-29, the notes,\nthe Borderland notes in `.private/`, the session transcripts). Nearest relatives:\n\n- **Filled slots compile to the reasoning engines** (ledger AM16, proposed and open; its neighbour AM18, ruled, keeps two rules for one case from reaching an engine together): an inference's slots become\n  clingo or ProbLog text that code runs. That is claims into code for checking; the greenhouse is claims into code for\n  acting.\n- **A policy's rules tried over cases by computation** (the what-if view, `what-if-and-lenses.md`; trying rule sets\n  out, `normic-rules-normative-reasons-and-trying-rules-out.md`): the dry run of the same compiler.\n- **Propose, never assert**, and **confident housekeeping is the machine's job** (the Principles): who may change what.\n\n## 4. One loop\n\nCode into claims makes the machinery arguable; claims into code makes the arguments run. Put together, the graph is the\nsource the software is built from, including in time the software's own rules: *\"rationality standards built into the\nsoftware\"* (2025, AI Safety Gothenburg) read literally. The loop keeps a fixed point: something applies the first rule\nwithout being argued at that moment (Carroll's tortoise). Lean names the design for it, the de Bruijn criterion: keep\nthe trusted checker small, so that everything else can be checked by it. Here that is the compiler and the arithmetic.\n\n## 5. The greenhouse, as a proposal\n\n- **The source** is a compartment's graph (a blob, `blobs-community-sandboxes-and-the-advice-process.md`): rules about\n  watering, temperature and moisture by season, as generalizations with their exceptions, each a claim.\n- **The build** reads the graph at a dated moment and compiles the rules whose posterior and weight pass a stated\n  threshold into a configuration the controller runs. It is a reading in the sense of *read, never set*: dated, and\n  rebuilt when the claims change.\n- **Before anything runs**: the build shows what would change against the running configuration (a diff), and a person\n  presses the button. Hard physical bounds (no heating above a ceiling, no watering into a flooded tray) are form in the\n  controller's code, outside the graph, so no argument can switch them off.\n- **After**: what ran, when, and from which claims goes in the record; one step back restores the last configuration.\n- **What it would test**: whether arguments people settle in the graph produce rules that work, which is a scoreboard\n  the world keeps (the plants), the fact layer reaching ground.\n\nFiled in `TODO.md` as an idea for after the reviews, unruled.\n"}