{"path":"research/deep-descent-experiment.md","content":"# The Deep-Descent Experiment\n\n*July 10 2026. Two everyday claims, each pushed by hand all the way to bedrock, to see which rabbit hole decomposition actually leads down and what kind of thing is at the bottom. Topics chosen deliberately non-meta and concrete (a white lie; the badness of a stubbed toe), per the substance-over-meta testing principle: complexity is the medium of discovery, but the surface should be something anyone recognizes, not a philosophy debate about the platform's own domain. Both entry claims were extracted on the live platform; the descents were hand-walked from the real graph roots after the Gemini balance went too thin for the large decompose calls. Companion artifacts: [white-lies-descent.md](deep-descent-experiment/white-lies-descent.md), [pain-badness-descent.md](deep-descent-experiment/pain-badness-descent.md). Continues the H-series of [frontier-extraction-experiment.md](frontier-extraction-experiment.md).*\n\n## The two bottoms are different in kind, and that is the finding\n\n| | Descent 1: the white lie | Descent 2: the badness of pain |\n|---|---|---|\n| Surface type | normative (ethics) | normative (but really about consciousness) |\n| What happened on the way down | the clash **shrank** | the claim **changed category** |\n| Terminus | fittingness shading into permissive-zone (parity) | **open is-bedrock / hard-problem** |\n| Residue fraction | small (most dissolved) | irreducible, but not a value residue at all |\n| Convergence wager | one point **for** it, on everyday material | orthogonal: the descent left the ought-domain the wager is about |\n\n**Descent 1 shrank.** An everyday moral intuition that presents as kindness-versus-honesty turned out to be mostly a stipulation switching off honesty's strongest reason, plus an empirical trust-risk, plus a sideways integrity branch, plus a **conceptual confusion about what kindness includes** (a lie that protects also isolates, so it was never purely kind). What genuinely survived was a small, precisely stateable fittingness-parity leaf: protect a beloved's peace, or honor their standing as one who can bear their own life. This is a data point *for* the convergence wager on object-level material, and a live demonstration that a resolution operator **other than decomposition** (reframing the covering value, exactly what the weighing scheme's CQ7 asks for) is what did most of the dissolving.\n\n**Descent 2 changed category.** \"Physical pain is bad\" wears ought-grammar but is largely an *is*-claim about consciousness. Its functional reading dissolves into biology but is blocked from full dissolution by **pain asymbolia** (felt pain without minding it, proving the awfulness and the avoidance-disposition are two things). Its phenomenological reading descends into valence, through three serious decompositions (representational, functional, phenomenal-realist / the founder's own QRI and Symmetry-Theory-of-Valence lens), each of which relocates rather than removes the problem, and bottoms out on the **hard problem of consciousness**: why is there something it is like to be in the painful state at all. That terminus is **none of the platform's four residue types**.\n\n## The taxonomy is OUGHT-biased (the load-bearing finding)\n\nThe four residue types (fittingness, structural, axiom-choice, permissive-zone) were induced from value debates: death penalty, minimum wage, assisted dying, and now the white lie. They are all **ought-shaped**. The moment a descent goes down a deep *is*-claim, it hits a terminus none of them fits: an **open, brute-or-reducible question about the world**, currently unanswered and possibly fundamental. Pain's badness is the specimen; \"is the felt awfulness ontologically fundamental or will it reduce\" is the leaf, and it is neither empirical-dissolving, nor fittingness, nor axiom-choice, nor permissive-zone.\n\nTwo honest responses, a founder call:\n1. **Add an is-bedrock / open-empirical-metaphysical residue type** (Deutsch-flavored: a soluble problem we do not yet have the explanation for, held as a challengeable open node with its competing readings). This widens the map to factual-metaphysical descents.\n2. **State explicitly that the residue map covers normative disagreement only**, and that factual-metaphysical descents (consciousness, fundamental physics) are out of scope by design. Cleaner, narrower, honest.\n\nEither way, the current silent behavior is a mis-typing: without a decision, the classifier will force \"pain is bad\" into `fittingness` or collapse it to `undecided`, and describe it wrongly. This connects to the philosophical stress test's separate demand to split the permissive-zone four ways (parity / indeterminacy / incomparability / permissivism); both are the same underlying message, that the taxonomy induced from a handful of value debates is too coarse for the space of real descents.\n\n## The brake question got a worked answer\n\nThe July 10 live thread was: is a residue genuine bedrock, or arrested inquiry wearing a robe? Descent 2 contains the perfect specimen, the extracted claim \"the badness is not a further fact beyond how it feels.\" Said early it is a thought-terminating cliche; said after the functional, representational, and physicalist decompositions have each been tried and failed, it is a hard-won terminus. **The only thing that distinguishes the two is the record of moves tried.** This is the effort-certificate proposal from the brake deliberation, now with a concrete case that makes the argument concrete: a bedrock verdict is only honest if it carries the record of the decompositions attempted and found wanting. Moves-tried metadata is not bookkeeping; it is what separates a terminus from a refusal to think, which is exactly the founder's worry about a religiously-adhered brake inheriting humanity's premature stopping.\n\n## Method notes for the pipeline comparison\n\nBoth roots are live (`claim_a876a3f2302e`, `claim_c14f1e4dc6b2` / `claim_605b1fd10966`). When the balance allows the large decompose calls, the registered predictions to score are: descent 1 bottoms out fittingness-plus-parity with a small residue fraction and an empirical trust sub-terminus; descent 2 bottoms out on an is-bedrock the four-type classifier cannot name, and if the classifier forces it into `fittingness` or `undecided` without holding the two readings, that is the predicted mis-typing. A useful extraction cue fell out of descent 2: when a `normative` root's descent keeps generating `empirical` and mechanism sub-questions, suspect an is-claim in ought's clothing and re-type before descending further.\n\n## Operational note (the reason this ran by hand)\n\nThe project Gemini balance is razor-thin rather than depleted: a one-token probe returns HTTP 200, but a full extraction or decompose estimates above the remaining balance and Google rejects it as \"prepayment credits depleted.\" Small entry-extractions slipped through; the large decompose calls did not. A larger top-up, not a config change, is the fix. Separately, a real bug was fixed in the dogfood driver: the fleet-wide shell `GEMINI_API_KEY` was shadowing the project `.env` key (dotenv never overrides existing env vars), so the driver now forces the project key. That one cost an outage-diagnosis detour and is worth keeping.\n\n\n**Update (2026-08-22)**: the ought-bias is now one of three worked instances that the ontology encodes theory it never declared — alongside the terminus enum's is/ought partition and the additivity assumption. The sorting test that separates a commitment which must live in the schema from one the graph can hold as an ordinary claim: [what-belongs-in-the-ontology.md](what-belongs-in-the-ontology.md).\n\n**Update (2026-08-20)**: the is-bedrock terminus that \"pain is bad\" reached now has a naturalistic reading — Heylighen grounds valence directly in autopoiesis (pain IS the self-maintenance signal), i.e. the descent bottomed out where argument hands over to biology. [heylighen-value-complexity-and-deliberus.md](heylighen-value-complexity-and-deliberus.md) §2.\n"}