{"path":"research/what-the-necessary-default-changes.md","content":"# What the Necessary-Default Changes — in plain words\n\n**Date**: 2026-08-31 · **Type**: plain-language companion to the padding-defense ruling\n(founder: *\"Go with option C, both\"*). Written because the change was executed in one turn and its\nconsequences deserve unhurried reading. No jargon; the technical version lives in\n[strength-layer-audit.md](strength-layer-audit.md) and `deliberus/support_semantics.py`.\n\n---\n\n## 0. What changed, in one paragraph\n\nWhen a claim is broken into parts, the system asks how the parts' strengths should combine into the\nwhole's strength. Before this ruling, the answer for untagged parts was **add them up**. Now it is\n**the weakest part decides**. That single switch is the whole change. Everything below is what\nfollows from it.\n\n*Precision added 2026-09-17: for untagged parts the weakest part decides the **energy** (so splitting\ncannot raise a parent, and a weak part pulls it down a little), while the hard **cap** (a whole can be\nno stronger than its weakest required part) still applies only to parts explicitly tagged\n`necessary`. That gap was noticed 2026-09-02 and kept on purpose: a bare cap on untagged parts would let a\nspurious weak part drag its parent down 1:1, and would let any fresh decomposition cap an evidenced\nclaim at 0.5 until every part is examined. **The designed completion was built 2026-09-17**: a\nconfirmed untagged part now caps its whole at a modest, rebuttable weight (half of the full\nweakest-link cap), an unconfirmed machine proposal caps nothing, and anyone can tighten the cap by\narguing the requirement as an ordinary claim (\"W requires X: …\") whose own strength the cap then\nfollows — support it and the cap approaches the full reading, attack it and the cap closes\n(`weakest-link-arithmetic-and-the-merge-hunch.md` § 9d).*\n\n## 1. The attack it closes\n\nTake *\"Dr. Smith is credible as an expert in labor economics.\"* Honestly split, that is three\nparts: she has published work, she holds a real position, and her field matches the topic.\n\nUnder the old adding rule, someone could split it into twelve trivial parts instead — *\"she has\npublications,\" \"the publications exist,\" \"they appear in journals,\" \"the journals are real,\"* and\nso on. Each is easy to support. Each added a little. The measured effect: three parts at strength\n0.70 gave the parent **0.516**; twelve identical parts gave **0.671**. **The claim looked\nbetter-evidenced purely because it had been sliced thinner** — no new knowledge, just more slices.\n\nUnder the weakest-part rule, three parts and twelve parts score **the same**. Slicing earns nothing.\nNobody needs to detect the trick, because the trick stopped paying — which is the shape this\nproject prefers for any defense: remove the reward rather than police the behaviour.\n\n## 2. First consequence: what protects a small good claim changed\n\n**The old property.** A small, well-supported claim sitting among nine siblings that nobody had\nevidenced was not dragged down by them. The neutral siblings contributed exactly zero — under\naddition, zero-plus-zero-plus-zero is still zero, so the little claim kept its standing. The project\nhad a name for this: *pebbles are not buried*.\n\n**What is true now.** For untagged parts, that protection is gone — and gone *correctly*. If a claim\ngenuinely requires three things and only one of them is established, then the claim genuinely is not\nestablished. A chain with one unverified link is an unverified chain. That is the honest reading of\na whole made of required parts.\n\n**Where the protection still lives**: on parts explicitly marked *corroborative* — meaning parts\nthat each raise the case on their own rather than being required together, like three independent\nstudies pointing the same way. Mark them, and the additive behaviour returns, including the\npebble protection.\n\n**So the change is really this**: the protection moved from being automatic to being **claimed**.\nWhere independence is real, you say so and get it back. Where you say nothing, the system assumes\nthe parts belong together — which, under the separability floor blessed the same day (split a claim\nuntil nothing separable is still bundled), is what a decomposition child *is*.\n\n**The felt difference**: a parent claim with several unevidenced parts will now show weaker than it\ndid yesterday. Nothing broke; the display became more conservative and more truthful.\n\n## 3. Second consequence: the \"which part matters most\" pointer got sharper\n\nThere is an instrument that answers *if you changed your mind about one part of this argument, how\nmuch would the conclusion move?* — the sensitivity ranking that decides which question is worth\nasking first.\n\n**Before**, with parts adding up, that sensitivity was spread thinly: each part contributed its own\nslice, so no single part dominated, and the ranking was flat-ish.\n\n**Now**, with the weakest part deciding, sensitivity concentrates on the part the whole actually\nhangs on. Ask about *that*, and the conclusion swings. Ask about a part that is already solid, and\nlittle happens — correctly, because it is not what the argument depends on.\n\n**Why this is an improvement**: the instrument's purpose is to find the crux. Under the old rule it\nwas averaging across a conjunction, which blurred the crux; under the new one it points at it.\nA note in the code records the earlier version of this problem: denying a jointly required part used\nto move the conclusion by about 0.02 when it should have cratered it — *the instrument whispered\nexactly where it should have shouted*.\n\n**The felt difference**: the \"what should I look at first\" suggestions will more often point at one\nspecific weak part rather than spreading attention across a claim's children.\n\n## 4. Both consequences share one root\n\nThey are the same fact seen from two sides. **Adding treats parts as independent; taking the\nminimum treats them as required together.** Independence is what protected the pebble and what\nspread the sensitivity; requirement is what removes both — and what makes padding worthless.\n\nYou cannot have the padding defense and the old pebble behaviour at once for the same parts, because\nthey are consequences of opposite arithmetic. What you *can* do, and what the design now does, is\nlet each set of parts declare which it is. The default just moved to the truthful answer for\ndecomposition: parts of a whole are required.\n\n## 5. If either consequence feels wrong\n\nThe flip is a one-line change and could be narrowed to apply only to edges created from now on,\nleaving today's 132 untagged edges under the old adding rule. That would keep the corpus's current\ndisplay exactly as it was while still closing the attack for everything new. Not proposed — the\nconstitutive reading is the correct one — but the escape hatch is cheap and worth knowing about.\n\n**Ruled 2026-08-31, provisionally.** Founder verbatim: *\"The crux-pointer concentrating I guess\nsounded rational. The pebble protection requiring a tag seems to me too nitty-gritty complex for\nmy brain rn and unclear what the ACTUAL consequences are, but lemme just bless it for now and we\ncan always come back to revisit it in the future, doesn't feel super-central right now, we have\nbigger fish to fry, right?\"* Both consequences stand blessed; the blessing is explicitly\nprovisional, the revisit flag is registered in TODO's decision stack, and this escape hatch stays\navailable.\n\n---\n\n# Deriving the tag from the content\n\n*The founder's question after ruling: could the system work out for itself whether parts are\nrequired-together or independently-corroborating, instead of defaulting?*\n\n## 6. Short answer: yes, and the ruling makes it easier than expected\n\nThe naive version of this is a three-way judgment call: look at the parts, decide *required*,\n*independent*, or *unclear*. But the separability floor changes what the question even is.\n\n**Under that floor, a genuine part is required by construction.** You split a bundle into what it is\nmade of; a whole is not the whole without its parts. So the interesting case is not \"which flavour\nof decomposition is this?\" — it is:\n\n> **A part that looks independently-corroborating is usually not a part at all. It is a piece of\n> evidence wearing the wrong label.**\n\nThree studies are not three parts of a claim. They are three witnesses *for* it. In the graph's\nvocabulary, they should be attached as evidence, not as pieces of a decomposition.\n\n## 7. So the pass has three outcomes, not two\n\nFor each existing decomposed piece, a review pass proposes one of:\n\n1. **Required part** — the whole needs it. Confirms the default; no action.\n2. **Independent support** — genuinely reinforces without being required. Tag it, and the additive\n   behaviour (with its pebble protection) returns.\n3. **Mislabelled** — this was never a part; it is evidence. Propose changing the relationship type\n   itself.\n\n**The third bucket is the unexpected prize.** It turns a tagging chore into a repair list for the\ngraph's structure — and it is the honest answer to *why would a decomposition child ever be\ncorroborative?* Usually because somebody attached a supporting fact using the wrong kind of link.\n\n## 8. What the machine can read from the words\n\nSome of it needs no model at all:\n\n- A piece that **shares the parent's subject and carves off a fragment of what is said about it**\n  is a part. *\"Dr. Smith is credible\" → \"Dr. Smith has published work\"*: same subject, one strand of\n  the same assertion.\n- A piece that is **its own complete statement about something else** — a study, a citation, a\n  worked example, a source — is evidence. *\"A 2019 survey found X\"* is not a fragment of anything;\n  it is a separate fact brought in to help.\n\nThose two signals are countable without judgment. A model handles what is left, which is the genuine\ngrey zone, and everything it proposes goes through the founder's own review flow: a list of\nbefore-and-after, agree or disagree, keyboard-fast.\n\n## 9. Where this ends up\n\nOnce such a pass has run, nearly every relationship carries a reviewed, explicit label. The default\nthen matters only for the brief moment between a new piece being created and its label arriving —\nwhich is the quiet, good end-state: **the question \"what should the default be\" mostly stops mattering,\nbecause almost nothing is left on the default.**\n\n---\n\nDeep companion (founder-commissioned research pass, 2026-08-31):\n[weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md) —\nthe academic ground for every questionable pillar, the Bayes analysis, the auto-merge hunch\nverdict, four worked trees, and the surfaced up-transmission question. Interactive presentation:\nthe **Weakest Link** artifact.\n\nCross-references: [strength-layer-audit.md](strength-layer-audit.md) (the measured numbers behind\n§ 1) · [decomposition-axes.md](decomposition-axes.md) (the separability floor; the eight axes) ·\n[the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) (the house-choice class this\nbelongs to) · [fractal-scales-and-temporal-frame.md](fractal-scales-and-temporal-frame.md)\n(payoff-removal before detection; the adversary classes) · `deliberus/support_semantics.py` and\n`deliberus/graph/qbaf.py` (the implementation, with the ruling recorded in their docstrings)\n"}