{"path":"depth.md","content":"# Depth: No Copout Axioms, No Permanent Bottom\n\nMost things people call \"axioms\" in arguments are not bedrock — they are **conclusions that haven't been examined deeply enough.** The acid of reason and introspection keeps dissolving, in every direction at once. Here is what that looks like on a real example.\n\nDecomposition is the main instrument for lowering the cost of finding out why another mind holds its position, which is the conviction the whole project rests on ([Opacity Is a Cost, Not a Mystery](vision.md#opacity-is-a-cost-not-a-mystery-aug-2026)). Calling something an axiom is how that cost gets declared unpayable — usually prematurely.\n\n## A Decomposition in Two Directions\n\n*And a third axis: scale.* The two directions below run downward and outward from a single claim. A 2026 reading of Jonathan Rowson's converging-timescales essay makes the *upward* direction legible too — the same mechanism recurring from premise to crux to aggregation rule to worldview to epoch-frame, with one invariant (the transaction cost of understanding another mind) and one honest caveat (each rung up adds an adversary, so the mechanisms are homologous, never free to transfer). The scale ladder, and the temporal rung where the graph currently has no representation at all: [research/fractal-scales-and-temporal-frame.md](research/fractal-scales-and-temporal-frame.md).\n\nPeter Singer's drowning-child case. A child is drowning in a shallow pond. You can wade in and save them at small cost to yourself. Almost everyone agrees you must.\n\n> **\"I must save the drowning child in front of me.\"**\n\nAsk *why*. Two legitimate branches open, and honest decomposition requires following both.\n\n**Toward the practical and political** — what the claim binds to in institutions:\n\n```\n→ Because failing at near-trivial cost would make me a moral monster.\n  → Because a society of agents like me needs the standard that such failure is punished.\n    → Because law already distinguishes the murder we prosecute from the starvation we tolerate half a world away.\n      → Because rule systems can only sustain lines their agents can actually follow\n         without collapsing into infinite demand.\n```\n\n**Toward the phenomenological** — why the claim has the *weight* it has:\n\n```\n→ Because the child is vivid, proximate, and present in my perceptual field.\n  → Because moral weight tracks features of the situation IN consciousness —\n     vividness, proximity, felt agency, felt causation.\n    → Because those features are how evolved moral circuitry computed\n       \"help who you can actually help.\"\n      → Because at this depth, the Singer-style conclusion\n         (\"therefore you must donate to distant charity\")\n         is not irrational to resist. It is the honest report that\n         the phenomenological weight of a statistical obligation\n         is lower than the phenomenological weight of a drowning child,\n         and that asymmetry is itself moral data, not a bug to be reasoned away.\n```\n\nAnd here is the move that most existing platforms can't make: **at depth, the branches meet.** The political standard (\"society needs this rule\") and the felt imperative (\"I cannot leave this child\") are tracking the same underlying computation — *who is in my field of effective help, how strongly is their condition pressing against my consciousness, and what rule could a society of agents with my kind of mind actually sustain.* The disagreement between an effective altruist who donates aggressively and a neighbor who doesn't — which looks on the surface like a fight about donation percentages — is, at depth, a disagreement about whether the phenomenological weighting system should be corrected by explicit reasoning or honored as load-bearing moral data. Neither answer is stupid. The graph doesn't settle this. It exposes it at the level where it can finally be argued honestly.\n\nThe weighting of phenomenological elements — vividness, proximity, felt agency, temporal pressure, the intensity of another mind visibly present — may be more universal across humans than any political conclusion built on top of it. That is the [convergence thesis](convergence.md) in miniature: the middle layers, where slogans and policies and tribal affiliations live, are where the confusion is loudest. The bedrock underneath, where moral weight is actually computed in conscious experience, is where the sharing is.\n\nResearch confirms this matters: many ethical decision problems are NP-hard (Stenseke, 2024) — decomposition into tractable sub-problems is not just useful but *computationally necessary*. Tao's [verification thesis](research/ai-safety-and-tao-augmentation-research.md) validates the architecture from mathematics: ideas are cheap, verification is the bottleneck, and structured decomposition is the answer.\n\n> People can put words to more than they think they can, if they try. Not many things are truly ineffable.\n\nThis principle drives every layer of Deliberus — from the extraction pipeline to the [convergence thesis](convergence.md) to the platform's potential as [alignment infrastructure](research/deliberus-as-alignment-infrastructure.md).\n\n---\n\n## The Socratic Function\n\nThe platform should NEVER accept unexamined premises as stopping points. The Socratic function — \"you marked this as foundational; can you articulate what's underneath?\" — extends progressive disclosure to the value premises themselves. What looks like bedrock at one zoom level can be expanded at the next.\n\nArchitectural implication: the fourth claim type is **\"value premise (currently undecomposed)\"** — not \"value axiom.\" The label marks depth-relative status, not permanent bedrock. The system invites further decomposition, always. (See [research/extraction-virtual-run.md §The \"No Copout Axioms\" Principle](research/extraction-virtual-run.md))\n\n---\n\n## The Descent, Defined — and Scaffolded\n\n**A descent is following any claim downward through what it rests on**: asking, of the claim in front of you, *why? — resting on what?* — and then asking the same of each answer. Every step is small: one claim opens into the premises beneath it, each premise is itself a claim, and the question repeats. A **deep descent** is nothing more exotic than continuing past the two or three steps where conversation normally stops — down to where a branch reaches ground. Ground comes in kinds, and naming the kind is the point: a checkable fact (empirical), a value you hold (and may mark as yours), a choice of definition, a weighing that turns out to be contextual, or an honest *this is bedrock for me*. Nothing about the descent requires cleverness; it requires only not stopping, and the map remembers where you have been.\n\n**What makes it possible for ordinary users is that the machine carries everything except the judgment.** Developmental psychology has a precise name for why this matters: people reliably perform above their unsupported level when a scaffold carries part of the load (Vygotsky's zone of proximal development; Fischer's supported-versus-functional gap), and the operation a descent invites — holding your own premise as an object you can examine without losing the ability to inhabit it — is exactly the operation the scaffolding must support (Kegan's subject-object move; see [research/coupling-the-poles.md](research/coupling-the-poles.md)). Concretely, the machine holds seven things so the user holds one:\n\n1. **It proposes the decomposition** — you ratify, edit, or refuse; you never author structure from a blank page.\n2. **It asks one question at a time** — taps, not essays; a critical question arrives as a small answerable thing, with *does not fit* always available.\n3. **It holds the map** — you never keep the tree in working memory; you look at one seam while the graph remembers the whole.\n4. **It types the ground when you arrive** — proposing *this looks like a value choice / a definition / missing evidence*, never ruling.\n5. **It offers the return move** — your position rendered whole again after the descent, through your own lens or another's, so analysis ends in re-inhabitation rather than in pieces.\n6. **It never shows a gap without a move** — an open question always comes with a next step, because a gap with no move produces avoidance, not curiosity.\n7. **It lets you stop anywhere** — absence is a state, not a failure; the descent's depth at any point just is how far anyone has cared to go.\n\nThe one thing the machine cannot carry is the only irreplaceable part: recognizing *yes, that is my reason* or *no, that is not it*. Everything else — tracking, sequencing, typing, remembering — is load the scaffold takes, which is why the depth of the graph is not capped by the stamina of its users.\n\n## Reason as Universal Acid\n\nDennett's Universal Acid (from \"Darwin's Dangerous Idea\"): once understood, evolution dissolves every boundary. Reason is the same — once applied to any topic, it dissolves assumptions and reveals structure regardless of what it's applied to. The value is in the **method**, not the content. Any sufficiently contentious topic produces rich structure when rational decomposition is applied.\n\nThe starting topic doesn't matter. The acid works on everything. The seed is almost arbitrary.\n\n---\n\n## Extended Universally\n\nThis principle applies not only to value premises but to ALL claim types:\n\n- **Empirical evidence**: \"Study found d=0.4\" decomposes into sample size + methodology + statistical test + confounders + replication status\n- **Expert testimony**: \"Leading climate scientist says X\" decomposes into the scientist's credentials, the specific claim, the evidence behind it\n- **Meta-analyses**: decompose into individual studies, which decompose into their methodologies\n- **Evidence as subgraph**: a PDF uploaded as evidence should be extractable into its constituent claims, each examinable — not a black-box attachment\n\nThe graph has no natural bottom; it has a *current* bottom ([sorry markers](research/lean-deliberus-analogies.md)) that is always invitable to go deeper. The same holds at the very end of a descent: a classified terminus is a fallible fixed point under the current decomposition moves, never a proof of undecomposability, and the verdict that marks it is itself a challengeable claim ([Convergence](convergence.md) carries the full statement). Depth at any point reflects how much the community has cared to examine that thread — and that asymmetry IS the information. (See [research/scheme-bounded-decomposition-and-evidence-as-subgraph.md](research/scheme-bounded-decomposition-and-evidence-as-subgraph.md))\n\n**Depth is a property of the graph; usability is a property of the response** (April 2026): the graph has no permanent bottom (structural commitment, non-negotiable), but every truth-graph response has a first sentence that must summarize honestly. The calibration rule: surface the *shallowest* sorry markers (closest to the conclusion) first, and make deeper inspection effortless. \"No Copout Axioms\" and \"beginner-readable first layer\" are not in tension — they operate at different layers. (See [research/truth-graph-evidence-system.md §\"No Copout Axioms\" Is a Property of the Graph](research/truth-graph-evidence-system.md))\n\n**The acid applies to the system itself** (April 2026): The system's own intermediate representations — CQ polarity assertions, chain-of-thought reasoning traces, bundled claims — are \"currently undecomposed,\" not \"permanently atomic.\" A CQ assertion like \"Dr. Smith is credible\" bundles publication record, institutional affiliation, domain match, and peer agreement into one sentence. Each is independently verifiable. The \"mother claim\" should decompose into subclaims whose QBAF strength flows back transitively. The sorry model is self-similar: each level of decomposition reveals the next level's sorry markers. (See [research/self-similar-decomposition-and-claim-ontology.md](research/self-similar-decomposition-and-claim-ontology.md))\n\n**And the acid reaches the schema, where it finds a theory already installed** (August 2026): run it one layer further than the representations and it arrives at the ontology itself, which turns out to encode substantive positions nobody decided to take — the residue types are ought-shaped, the terminus enum encodes an is/ought split, and strength was additive because summing is what one does. What sorts these is a mechanical test: a commitment belongs in the graph if the system can hold disagreement about it and keep working, and in the ontology if it cannot. The three responses are ordered by cost — de-bake it into a recorded parameter whose default changes nothing (`support_semantics` is the shipped template), give its enum a way to confess that it lacked a slot, or publish it as a claim the system's own mechanisms can attack. (See [research/what-belongs-in-the-ontology.md](research/what-belongs-in-the-ontology.md))\n\n---\n\n## Deutsch's Error-Correction Frame\n\nDavid Deutsch's epistemology gives the depth principle its institutional meaning. Knowledge grows through conjecture and criticism, not final justification; good explanations are hard to vary while still doing their explanatory job. For Deliberus, depth is the mechanism that turns vague, easy-to-vary explanations of disagreement into more constrained ones: not \"they are irrational,\" but \"they are using a different operative definition of fairness and weighting a different empirical mechanism.\"\n\nThat makes decomposition an error-correcting institution, not just a UX affordance. The graph does not certify final truth; it exposes assumptions, criticisms, missing premises, and contested definitions so errors can be found and corrected without violence or status combat. In the non-zero-sum cooperation synthesis, this is the bridge between Deutsch and Wright: improving our ability to correct errors in collective reasoning is itself the deepest positive-sum game. See [research/david-deutsch-deep-research-non-zero-sum-economics.md](research/david-deutsch-deep-research-non-zero-sum-economics.md) and [research/non-zero-sum-economics-and-civilizational-cooperation.md §10](research/non-zero-sum-economics-and-civilizational-cooperation.md).\n\n---\n\n## The Convergence Thesis\n\nThe deepest implication of \"No Copout Axioms\" is not just that you CAN keep decomposing — it's that when you DO, values converge. Drill past the middle layers (definitions, cultural framing, empirical beliefs) and what remains is shared: consciousness matters, suffering is real, shared humanity is foundational. This is a founding conviction, not a proven fact — but it's what makes the civilizational graph a potential *discovery*, not just a map.\n\nStenseke (2024) shows that many ethical decision problems are NP-hard — morally optimal action under uncertainty is computationally intractable in general. This means decomposition (breaking hard moral problems into tractable sub-problems) is not just useful but *necessary* — exactly what argument scheme classification does. The sorry model isn't a UX convenience; it's a computational necessity.\n\n**Full exploration**: [convergence.md](convergence.md) — the thesis, MGE empirical support, the Lean analogy at full depth, Constitutional AI as snapshot vs camera. [research/deliberus-as-alignment-infrastructure.md](research/deliberus-as-alignment-infrastructure.md) — the full alignment case with formal ethics research.\n\nThe acid reaches weightings too: when someone reports \"liberty outweighs safety,\" asking *what exactly is being weighed against what* splits the weight into compressed empirical beliefs, a risk posture, a scope choice, an aggregation rule — and a typed residue far narrower than the original \"value difference.\" Tested against six hard cases in [research/value-weighting-decomposition.md](research/value-weighting-decomposition.md): no scalar weighting survived decomposition; what survives is typed, precisely stateable, and still arguable.\n\n---\n\n## Tao's Verification Thesis\n\nTerence Tao's \"Mathematical Methods and Human Thought in the Age of AI\" ([arXiv:2603.26524](https://arxiv.org/abs/2603.26524), March 29, 2026) validates this architecture from the world of mathematics. His core argument: verification is the bottleneck, formal infrastructure (Lean) provides structural guarantees, but meaning and value remain irreducibly human.\n\nReplace \"mathematical theorems\" with \"claims about the world\" and \"Lean\" with \"Deliberus\":\n\n- **Ideas are free**: Anyone can make claims. LLMs can generate thousands. Social media produces millions daily.\n- **Verification is the bottleneck**: Which claims are supported? What are they built on? Which are contested?\n- **Lean is the answer for math**: A formal verification system where every step must be justified.\n- **Deliberus is the answer for reasoning**: A structured argumentation system where every claim is decomposed, scheme-classified, critically questioned, and connected to evidence.\n\nTao didn't just *argue* this — he demonstrated it. In 2023 he and roughly 25 mathematicians, most of whom had never worked together, formalized a 33-page research paper (the \"PFR conjecture\") in the Lean proof assistant in three weeks. They coordinated through a color-coded blueprint where each `sorry` marker was a structural placeholder anyone could claim and fill in. Trust between the contributors was not required, because the kernel verified the join. Deliberus is the same pattern applied to reasoning about the world instead of reasoning about mathematics — sketch the structure with holes, let others contribute individual premises, let the system check the join. The 13-year-old \"No Copout Axioms\" principle and the 2023 Lean social protocol turn out to be the same discovery from opposite directions. (See [research/lean-deliberus-analogies.md](research/lean-deliberus-analogies.md).)\n\nTao also warns about **AI collapse** — AI trained on AI outputs degrades without genuine human content. A human-maintained argument graph is the antidote: genuine reasoning, structured and persistent, immune to recursive degradation. (See [research/ai-safety-and-tao-augmentation-research.md §5](research/ai-safety-and-tao-augmentation-research.md))\n\n---\n\n## The Truly Ineffable\n\nThe truly ineffable exists — but it is much deeper than where most debates stop and most platforms accept defeat. Polanyi's \"tacit knowledge\" — we know more than we can tell. Embodied intuitions, aesthetic judgments, the felt sense that a worldview is right before articulation. This layer may be fundamentally outside Deliberus's scope — or it may be what the [worldview filter](worldview-lenses.md) approximates non-propositionally.\n\nThe depth at which language is genuinely exhausted is MUCH deeper than where most people stop trying. The system's job is to keep inviting decomposition — gently, Socratically — until the user reaches their actual bedrock, not a convenient stopping point.\n\nIn Deliberus's own terms: the end of decomposition is not a declared axiom — it is a marker the community has repeatedly tried to decompose, and every child claim it produced has been withdrawn. That's the Lean/Mathlib deprecation pattern stated as a depth principle, and it's also what falsification of this entire page would look like: a claim that consistently resists decomposition over many honest attempts is where Deliberus would, in good faith, rest.\n\n---\n\n## The Hormesis Frame\n\nThe biological concept of hormesis — small doses of stress strengthening organisms — names the same principle from a different angle. Removing ALL friction from reasoning (letting the LLM do the thinking) weakens the thinker, just as removing all physical stress weakens the body. The Socratic function is hormetic: it applies controlled epistemic stress (\"can you articulate what's underneath?\") at exactly the site where productive discomfort generates insight. Simon's scientific-method frame adds: the Socratic function is Step B applied recursively — each invitation to go deeper is an invitation to check that the current bottom is not a convenient stopping point. See [research/wiki-that-writes-itself-and-productive-friction.md](research/wiki-that-writes-itself-and-productive-friction.md), [research/scientific-method-two-steps-and-deliberus.md](research/scientific-method-two-steps-and-deliberus.md).\n\n---\n\n## No Permanent Bottom Is Also an Engine\n\nThis page argues for endless decomposition on epistemic grounds. It turns out to answer a motivational problem nobody had posed to it.\n\nCuriosity, in the account that has held since Loewenstein's 1994 review, is what happens when a person notices a gap between a question they have activated and an answer they do not have. Its least convenient property is transience: curiosity is intense but shallow, and it \"tends to disappoint when satisfied.\" Answered questions *lose* their importance. For anything built to close gaps, that is a structural problem — the mechanism that delivers the value also removes the motive that brought the person.\n\nA system with a permanent bottom would hit this wall. This one cannot, because nothing here is ever more than *currently* undecomposed: every closed gap exposes the next layer's gaps, and every confirmed terminus remains a fallible fixed point open to a new move. The principle chosen because reason should have no privileged stopping point doubles as an inexhaustible supply of the one thing curiosity requires. Worth recording as a coincidence rather than a design triumph, since it was not designed for this — but also worth recording, because a project this dependent on sustained voluntary attention should know where its supply comes from. Full treatment: [research/curiosity-as-growth-fuel.md](research/curiosity-as-growth-fuel.md).\n\n## Why a Decomposable Structure Is Revisable At All\n\nThere is a reason beyond preference that reasoning should be built in layers, and it is the oldest argument in the study of complex systems. Herbert Simon's *architecture of complexity* (1962) identifies **near-decomposability** as the signature of every evolved complex system: interactions inside a subsystem are far stronger than interactions between subsystems. Matter to molecule to organelle to cell to tissue to organ. Simon's point is not that complex things happen to look layered. It is that near-decomposability is *what makes them modifiable at all* — a change inside one module does not force the whole system to be recomputed, so evolution can improve a part without redesigning the whole.\n\nThat is exactly the property this graph needs. A stranger must be able to revise one premise without re-deriving everything above it. No permanent bottom gives the descent somewhere to go; near-decomposability is what makes each step of it survivable.\n\nIt also names its own boundary, and the boundary is one the project has already met. Simon's property requires within-module coupling to exceed between-module coupling, which holds for most arguments — it is what makes a subtree a subtree — and fails precisely where a sub-claim belongs to two parents at once. The corpus found that failure independently and gave it a sharper name: **overlapping subtrees are a sheaf gluing condition**, and failure to glue is contextuality ([research/decomposition-axes.md](research/decomposition-axes.md)). Read through Simon, that failure is not an inconvenience in the implementation. It is the edge of the decomposability assumption itself, showing up where the theory says it should. Full test of this conviction against other scales: [research/convictions-across-scales.md](research/convictions-across-scales.md) § 2.4.\n\n---\n\n**See also**: [Vision](vision.md) · [Civilizational Vision](civilizational-vision.md) · [Analysis ↔ Attunement](analysis-and-attunement.md) · [Lean Analogies](research/lean-deliberus-analogies.md) · [Worldview Lenses](worldview-lenses.md)\n"}