{"path":"research/core-periphery-and-the-concept-layer.md","content":"# Core, Periphery, and the Concept Layer — why concept tracking exists, read through the contingency lens\n\n**Date**: 2026-09-06 · **Type**: founder questions, answered against a full re-read of the corpus and a\nlive measurement of the concept layer · **The questions, verbatim**: *\"is the periphery as opposed to\nthe core of a word/term (can we be talking about other words than nouns? has this been discussed in the\ncorpus?) the primary reason we need the concept tracking/tagging in Deliberus? Does this inform in any\npreviously-unrealized way how to design and think about these aspects and features of Deliberus? Both\nfrom a pragmatic and perfectionist angle.\"* · **Companions**:\n[contingency-and-convergence-across-levels.md](contingency-and-convergence-across-levels.md) (the lens),\n[semantic-disambiguation-and-concept-tracking.md](semantic-disambiguation-and-concept-tracking.md) (the\norigin of concept tracking), [philosophical-foundations-stress-test.md](philosophical-foundations-stress-test.md)\n§§ 6–7 (thick concepts; Gallie), [claim-sameness-philosophical-readings.md](claim-sameness-philosophical-readings.md)\n(sense against reference). **Nothing in § 6 is ruled.**\n\n---\n\n## 0. The short answers\n\n**Yes — and the corpus said so in 2013 without the name.** The founding statement of the problem is\nEmanuel Kumlien's: two people *\"can give the SAME definition of a concept, genuinely believe they mean\nthe same thing, yet in discursive practice draw completely opposite conclusions\"*. A shared definition\nis a shared **core**; opposite conclusions in practice are a divergent **periphery**. Concept tracking\nexists for the words whose meaning lives in the periphery. The pipeline already knows this: its\ncontested-concept pass splits *semantic* contestedness (different cores — *cost* as gross expense\nagainst net fiscal impact) from *criterion* contestedness (one word, different evaluative principles —\n*fairness* as desert against non-discrimination), and its own prompt calls the second *\"the DEEPEST\nsource of disagreement.\"* Core splits are the easy case a definition settles. Periphery splits are the\nreason the layer is there.\n\n**Yes, other words than nouns — and the corpus has discussed them more than it has discussed nouns.**\nThe live concept index holds *fair/fairness/fairly*, *deserving/desert*, *problematic/socially\nundesirable*, *successful (strategy)*, *high quality*, *goal-driven*; the detector is told to look\n*\"especially for evaluative terms (fair, just, deserving, adequate, responsible)\"*; the stress test's\nwhole § 6 is about thick predicates (*cruel, courageous, desecration*); and the decomposition axes\nhandle smuggled evaluative modifiers (*unnecessary*) and degree thresholds (*expensive*) at the claim\nlayer. The thin spot is verbs, which arrive nominalised (*commoditization*).\n\n**One limit on the two-layer picture, already found here in July.** For thick terms the periphery is\nnot attached to a core; it *draws the core's boundary* (McDowell's shapelessness: strip the evaluative\npoint of view and you cannot say what counts as cruel). So \"track the core, note the periphery\" is the\nwrong instruction for thick terms, and the July design — mark thickness, store the predication whole,\nmake any split a challengeable claim — is the right one.\n\n**What is new** (§ 5–6): a division of labour the corpus never stated — *claims carry the\nstranger-readable core; senses must carry the attributed periphery* — and, measured today, the concept\nlayer is halfway there: sense attribution exists in the data model but is not written to the graph;\nthe sense's *label* (a model paraphrase) is what the interface shows; and **49 of 90 live concepts read\n\"bifurcated\" largely because the detector mints two senses per contested term**, so the lifecycle\nstate is measuring the pass's template more than the population. Plus a sixth flattening mechanism,\nproposed: **centroid sense minting**.\n\n**Sharpened by the founder the same day (§ 5b): the concept layer records deltas, not descriptions** — a\nsense is somebody's difference from a shown reference, and the divergence alarm decides whether the\ndifference is real.\n\n## 1. The two-axis rule, plainly, with examples\n\nThe rule from the contingency doc, said the way one would say it across a table.\n\n**Axis one: do the two things share a history?** If they do, their agreement is a copy, not a\ndiscovery, and it tells you about the source rather than the world.\n\n- Two students hand in the same wrong answer. They used the same textbook. The match says nothing\n  about whether the answer is right.\n- Identical twins have the same eye colour: one genome, one object, copied twice.\n- Every Swede knows what *lagom* means: one language, copied into each child. It tells you about\n  Swedish, not about the world.\n- Two language models put *freedom* next to *liberty*: both copied the internet.\n\nAssembly theory's word for this is *copy number* — many instances of one object.\n\n**Axis two: how strongly does the world force the answer, relative to how many answers there were?**\nLineages that share *no* history still converge when the constraint is strong and the space of\nsolutions small; they diverge when the space is huge and the constraint weak — and even real\nconvergence shows up in the coarse shape, not the fine structure.\n\n- Eyes: light and optics allow few designs, so eyes evolved independently dozens of times; the camera\n  eye of a vertebrate and of an octopus share no seeing ancestor. Convergence, and it is evidence\n  about physics.\n- Wings: flight forces *a wing* but not its construction; bird, bat and insect wings are built from\n  different parts. Convergence in function, contingency in structure.\n- Lenski's twelve flasks of bacteria: all grew fitter in the same direction; all got there by\n  different mutations. Repeatable in the coarse, contingent in the fine.\n- Counting: every culture that trades counts, and number words align across languages better than\n  any other domain. The world forces arithmetic.\n- *Taxes are theft*: nothing in the world forces a verdict and the space of stances is enormous, so\n  unrelated people diverge — and where they agree, it is usually axis one again: the same memeplex.\n\n**The test to run on any agreement.** First, do they share a source? Then it is a copy. If not, how\nstrong is the constraint and how big is the space? Strong and small → convergence, which is evidence\nabout the world. Weak and huge → coincidence or a shallow overlap. And expect convergence to be coarse:\nthe same *neighbourhood*, not the same *distances*. Applied: models agree by axis one (one written\nrecord); human perception agrees by both axes (one biology, one world); concrete nouns agree by axis\ntwo (the world); political terms agree by neither, so they diverge.\n\n## 2. What the corpus already holds — the scan\n\nRead in full today: the concept-tracking origin, the two stress-test sections, the sameness readings,\nthe vision and UX-principle statements, the dedup thread, the contested-concept prompt and data model,\nthe lifecycle rules in the API, the cost curve's safety condition, the concept-route upgrade, the\nformalization-paradox critique, the load-bearing-unsaid doc, and the lens doc's own § 3.\n\n| Where | What it says, in the lens's terms |\n|---|---|\n| [semantic-disambiguation-and-concept-tracking.md](semantic-disambiguation-and-concept-tracking.md) | The 2013 problem *is* core/periphery: same definition, opposite discursive conclusions; *barnfattigdom* — agreed facts, different operative definitions. Its \"Layer 2: what language hides\" is the periphery. Definitional claims are moves, not metadata; senses become separate nodes; lifecycle states exist |\n| `deliberus/extraction/contested_concepts.py` | Semantic contestedness = core differs; criterion contestedness = periphery differs; the prompt aims at evaluative terms and calls criterion contestedness the deepest source |\n| [philosophical-foundations-stress-test.md](philosophical-foundations-stress-test.md) § 6 | Thick predicates: the periphery constitutes the extension; separability is a live dispute the pipeline must stop settling silently; mark thick, store whole, split as a challengeable claim, add a thick-drift check |\n| same doc § 7 | Family resemblance (no core at all, only overlapping similarities); Gallie's essential contestation (permanent periphery dispute, *\"not resolvable by argument\"*); proposal: definitional claims in cluster form, a terminal lifecycle state *essentially contested* |\n| [claim-sameness-philosophical-readings.md](claim-sameness-philosophical-readings.md) | Frege: embedding similarity gestures at reference, the contested distinctions live in sense; Wittgenstein: sameness is use-relative, *\"only the pattern of downstream play can tell\"*; James: a difference that makes no difference; Gadamer: pre-clustering prevents the fusion of horizons |\n| [lowering-the-cost.md](lowering-the-cost.md) § 6.3 | Reuse lowers cost only where sense is stable; `usage_count × sense_count` maps where linking manufactures false agreement (*intelligence* 17 uses / 4 senses) |\n| [auto-connect-upgrade.md](auto-connect-upgrade.md) § 2a | The polysemy guard, built: above three senses, route only same-sense pairs |\n| [dogfood-run-6-israel-palestine-cross-domain.md](dogfood-run-6-israel-palestine-cross-domain.md) § J8 | Stance flattening: the same fact deployed in opposite directions — periphery at the *claim* layer, with the stance-conflict instrument and the write-time guard built for it |\n| [steelmanned-critiques.md](steelmanned-critiques.md) § 2 | The formalization paradox: decontextualisation strips pragmatic force, implicature, frame — the periphery — and the stranger test filters out exactly that material |\n| [the-load-bearing-unsaid.md](the-load-bearing-unsaid.md) | A premise omitted because presumed shared, which is not shared, is how an ordinary disagreement is found to be deep |\n| [worldview-lenses.md](../worldview-lenses.md) | *The interface is not the instrument — another person is*: a frame is invisible from inside; the lens holds two positions in one structure so the gap becomes an object |\n| [decomposition-axes.md](decomposition-axes.md) | Predicate–argument and degree-threshold axes: modifiers and standards as separable, at the claim layer |\n| [extraction-improvement-research.md](extraction-improvement-research.md) § 5 | No NLP work operationalises Gallie; criterion contestedness has no published detector; the proposed method is to extract each evaluative claim's grounding and compare across speakers |\n\nSo the corpus has the *phenomenon* from 2013, the *split* (semantic against criterion) in code since\nMarch, the *philosophy* (thick, family-resemblance, contested, sense-against-reference) since July,\nand three built defences against periphery collapse (polysemy guard, stance guard, demand-driven\nsplitting). What it lacked is the name that unifies them and the level-by-level reason the periphery\ndiverges, which is the lens doc's contribution: the periphery is the part of a concept's assembly\nthat happened *after the world stopped constraining it* — cultural, communal, individual — so it\ndiffers by lineage, and it is where contested words keep their meaning.\n\n## 3. Unexpected connections that surfaced on the re-read\n\n- **False consensus is the mechanism of the implicit crux.** Run 3F's stable finding — a debate's crux\n  is stated on one side and left unwritten on the other — and the load-bearing-unsaid doc's *premise\n  omitted because presumed shared* now have a measured psychological cause: people leave unsaid what\n  they believe is shared, and they overestimate what is shared by about two thirds whatever the true\n  rate (Marti et al.). The enthymematic gap is not laziness; it is the false-consensus bias applied to\n  premises. This makes implicit-premise recall a periphery instrument as much as a logic instrument.\n- **Sacred values and the periphery.** Symbolic recognition works where argument fails (Ginges &\n  Atran) because recognition addresses the *periphery* — what the value means from inside — while\n  argument attacks the core. Impact testimony is periphery testimony, and it is positional evidence for\n  the same reason: the periphery is visible only from where it was assembled.\n- **Bridging claims move a debate from contested peripheries to a shared core.** The 2026-09-06\n  election-post comment's *teknikneutrala incitament* takes *kärnkraft vs vindkraft* — two identity\n  flags, two peripheries — down to a core both sides state as a goal (investment certainty). That is\n  computable: a claim supported by claims from *both* senses of a lineage-bifurcated concept is a\n  bridge candidate (§ 6, item 7).\n- **Lean has no periphery.** Its definitions are total, which is why identity is decidable there and\n  why the reuse flywheel runs frictionlessly ([the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md)).\n  Deliberus's drift toward structural identity — scheme, kind, terminus, attribution — is the growing\n  of a core under words that have only peripheries.\n- **Frame lock-in is a crystallised periphery.** The reuse flywheel's shadow\n  ([the-residual-error-taxonomy.md](the-residual-error-taxonomy.md)) is a lineage's periphery baked\n  into the graph's vocabulary, and the three-runs law — every taxonomy carries the distinctions of the\n  register it was induced from — is the same thing said of taxonomies: a taxonomy is a lineage's\n  periphery presented as a core.\n- **The founder–Claude glossary programme has the same shape.** A coined compound is a high-index\n  periphery; *\"I'm not against jargon when I know what it means\"* is *when the periphery is shared*;\n  and the glossary's `status` column is a per-person sense record. The plain-brief words rule is a\n  core/periphery rule for a dyad.\n- **The stranger test and the periphery are in tension by design, and the concept layer is the\n  release valve.** The formalization-paradox critique is right that decontextualised claims lose\n  implicature, frame and force. The answer the corpus has been building without naming it is a\n  division of labour (§ 5): claims are cut to be stranger-readable, and the concept layer plus\n  attribution is where the stripped periphery is stored back.\n\n## 4. The concept layer, measured today\n\nFrom the live index (`GET /concepts`, 2026-09-06, 90 concepts) and the code:\n\n- **Lifecycle distribution: bifurcated 49 · emerging 26 · unseen 9 · fragmented 5 · locally stable 1.**\n  The rule for *bifurcated* is two senses each carrying at least 30% of usage. The contested-concept\n  pass is asked to find *contested* terms and returns them *with* their senses — almost always two —\n  so a two-sense split at roughly equal shares is the pass's output shape, and the state reads it back\n  as a diagnosis. **The distribution is largely the detector's template**, not a measurement of how\n  the population uses the term. (The pattern the corpus names as *a guard that holds by accident*: the\n  state looks informative because the input is shaped to trip it.)\n- **Sense attribution is modelled but not stored.** `ConceptSense.attributed_to` exists in the Pydantic\n  model; `store.py` writes only `r.sense` on the `USES_CONCEPT` edge. Which *lineage* holds which sense\n  is recoverable through each claim's own `attributed_to` and source, but nothing in the lifecycle\n  rules reads it. So the layer cannot yet tell *bifurcated by lineage* (two communities, two objects\n  under one word — the memeplex split the lens predicts) from *bifurcated within one source* (real\n  polysemy, or the model's guess).\n- **The sense's label is the displayed object.** The extraction checkpoint shows a sense as its\n  meaning string, truncated at 45 characters, with a count; the claim page's clarification card asks\n  the user to state a meaning. The model's paraphrase of the sense is the content the user sees; the\n  attributed claims that constitute the sense are behind it.\n- **What the layer holds is evaluative, not only nominal**: of the forty most-used concepts, the\n  contested ones are mostly evaluative predicates and standards (*fairness/reciprocity*, *deserving*,\n  *problematic*, *high quality*, *minimally decent standard of living*), which is where criterion\n  contestedness lives. The part-of-speech question is answered by the data.\n\n## 5. Claims are cores; senses are peripheries — the division of labour\n\nStated as a design principle rather than a finding, because it is what the built system already\nhalf-does and what the July and August rulings each assume from their own side:\n\n- **The claim layer is cut for the stranger.** Decontextualisation, the stranger test, `attributed_to`\n  moved out of the text, the seven-key sameness set — all of it makes a claim readable and comparable\n  without its speaker's periphery. That is the addressability substrate and it must not carry the\n  periphery, or the stranger test fails.\n- **The concept layer is where the periphery goes.** A sense is *who* uses the word *with which\n  criteria*, and a definitional claim is one of those criteria made attackable. Attribution is not\n  decoration here; it is the sense's identity, because a periphery is assembled by a lineage.\n- **Provenance links the two**, and the thick-term marker is the exception that proves the rule: where\n  the periphery cannot be separated from the core, the predication is stored whole and the split is a\n  move.\n\nIf this is right, three of the corpus's standing tensions are the same tension: the formalization\nparadox (what extraction destroys), the embeddings tension (what averaging erases), and the sense\ngovernance question (what merging collapses). Each is the periphery being lost at a different layer,\nand each has the same remedy — give the periphery an address of its own, attributed, and never let the\naverage stand in for a bearer.\n\n## 5b. The delta, not the description — the founder's sharpening (2026-09-06)\n\nFounder, verbatim, on reading § 5: *\"I guess one consequence of all this might also be that the\ninteresting part of defining and tracking concepts is not to come up with a description of or\nchallenge the core of a concept, because we all basically agree on that anyway, what's interesting is\nto articulate how you think your conception of a term/idea DIFFERS from the norm, right? What the\nconceptual delta is compared to how people in general think, or compared to how your opponent in\nparticular thinks. Am I onto something?\"*\n\nYes, and it is the sharpest form § 5 can take. Three reasons it is right, three sharpenings, what it\nleaves out, and the 2×2 it produces.\n\n**Why it is right.**\n\n1. *The core is redundant to elicit and the periphery is invisible to its holder.* Marti's data (§ 3)\n   has both halves: core features near-universal, whole-profile agreement rare, two thirds of the\n   disagreement assumed away. A clarification that asks for a description spends the user's words on\n   the part both already hold and both already assume, and buries the split in the middle of a\n   paragraph. The delta is exactly the part false consensus hides, which is why nobody states it\n   unprompted.\n2. *Every motte-and-bailey test is a delta test in disguise.* Owned supersession is stating your\n   delta from your own earlier sense; entailment acceptance is owning what your delta entails; the\n   symmetric burden is both parties stating one; return to the bailey is the delta re-widening over\n   time ([the-motte-and-the-periphery.md](the-motte-and-the-periphery.md) § 3b). A record that stores\n   deltas runs the four tests as queries.\n3. *A delta is attributed by nature; a description is not.* In the contingency lens (§ 1) the core is\n   the convergent, high-copy-number part and the delta is the contingent-path signature, the part\n   that carries a lineage. That is what `attributed_to` was always for: a description can be\n   nobody's; a delta is always somebody's, from something.\n\n**Three sharpenings.**\n\n- *Two deltas, unequal robustness.* Delta-from-the-norm needs a norm, and the only cheap one is the\n  centroid: the model's population-mean reading, the thing the embeddings tension distrusts as an\n  authority on meaning. But an origin needs no authority, only stability, and the delta between two\n  people cancels the origin entirely (mine minus theirs is the same whatever the zero point). So the\n  opponent-relative delta is the robust primary, the norm-relative delta is the solo-use fallback,\n  and the centroid is demoted from a *sense* anyone holds to the *origin* deltas are measured from.\n  That is the cleanest resolution of the sixth-mechanism worry (§ 6.5): the average is harmless as a\n  coordinate and harmful only as a bearer.\n- *The holder cannot state the delta unprompted, so the prompt must subtract.* False consensus means\n  the user believes theirs is the norm; asking *how does yours differ* returns *it doesn't*. The\n  elicitation has to show the reference (the centroid reading, or in a co-present session the other\n  person's attributed sense) and ask *what here is not yours, and what is missing*: the machine\n  supplies the reference, the human supplies the subtraction, which is the inversion principle\n  applied once more. The criterion taps of § 6.3 already had this shape; the delta framing says why,\n  and says the taps should be candidate *differences* rather than candidate features.\n- *A delta has a small type system.* Marti's feature-profile method supplies it: **inclusion**\n  (Peterson's structural *worship* covers every value hierarchy), **exclusion** (the liturgical sense\n  excludes non-ritual valuing), **criterion swap** (*fair* by desert against by non-discrimination,\n  the detector's existing criterion type), **threshold** (*expensive*, *safe enough*, the gradable\n  predicates of § 6), **weighting** (same features, different dominance, the within-mind case). The\n  pipeline's semantic/criterion split is a coarsening of these five, and the perfectionist triple in\n  § 6 was already delta-shaped without saying so: the exemplar core is the origin, the criteria list\n  and the per-position weighting are the deltas.\n\n**What the framing leaves out, honestly.** Essentially contested concepts (democracy, art) have a\nthin exemplar rather than a fat core, so the delta is measured from an exemplar and is most of the\ncontent; stipulative technical terms have a delta from lay usage that is the entire definition; and\nthe apophatic layer (the motte doc's § 3b) is where a delta presupposes what it cannot have, a\nproposition on both sides. None of these breaks the framing; the first two make the delta large, the\nthird marks its edge.\n\n**The 2×2 it produces, which is the design payoff.** Cross *delta stated* against *conclusions\ndiverge downstream* (the alarm the corpus already treats as the primary detector, § 6.4):\n\n| | delta stated | delta not stated |\n|---|---|---|\n| **conclusions diverge downstream** | mapped disagreement — the layer's job done | **false consensus, the implicit crux** — run 3F's finding, and pandering's home |\n| **conclusions do not diverge** | **manufactured differentiation** — a stated difference that changes nothing | genuine agreement |\n\nThe two off-diagonal cells are the two political pathologies the election thread named on the same\nday. The lower-left is a commenter's diagnosis of party democracy: candidates *\"must debate and\ncompete, and desperately try to differentiate themselves every fourth year\"* when the work should be\nfinding compromises — a delta inflated for competition, which James's maxim (*a difference that makes\nno difference is no difference*) and the divergence alarm together expose as one that moves no\nconclusion. The upper-right is Peterson's *worship*: conclusions that diverge with no delta ever\nstated, the delta deflated for the room ([the-motte-and-the-periphery.md](the-motte-and-the-periphery.md)\n§ 3c). The concept layer's job, in the sentence the founder supplied: record the delta, attributed,\nand let the divergence alarm say whether it is real.\n\n**What it changes in § 6, concretely.** Item 2 stands. Item 3 becomes *subtract against a shown\nreference* rather than *probe the periphery*. The pragmatic storage change is small: a sense carries\nits reference and its delta kind beside `attributed_to`, and the model's meaning string survives as\nthe reference. Item 1's lineage-partition condition gains its measure: *bifurcated by lineage* is two\nattributed deltas of opposite sign from one origin. And a stated delta is the anti-pandering\ncommitment device: once on record, a speaker cannot retreat to the room's sense without an owned\nsupersession, so the misrepresentation incentive (*put it in the graph so nobody re-renders it for\nyou*) applies to one's own definitions. Ledgered: N4 (the founder's proposal, asked as a question),\nN5 (the audience dimension and the 2×2, mine).\n\n## 6. Design implications — pragmatic first, perfectionist second\n\n**Pragmatic (small, most a day or less, none ruled):**\n\n1. **Bifurcated needs an independence condition.** A concept is *bifurcated by lineage* only when its\n   senses partition by source or position (derivable now from claims' `attributed_to` and source\n   through `USES_CONCEPT`); two senses inside one source is *polysemous* or *machine-split*. Different\n   diagnosis, different next step: the first invites a bridging question (*what would satisfy both\n   criteria?*), the second invites a definitional claim. One line stores `attributed_to` on the\n   sense edge; a few lines add the partition to the lifecycle rationale.\n2. **Lead a sense with its attributed claims, demote the label.** The model's meaning string is a\n   paraphrase of the population's centre and is nobody's; the sense's content is the claims that\n   constitute it, with who said them. The DeepMind rule — *show structure, ask questions, never\n   restate their words* — applies to the concept layer as much as to the live session.\n3. **Probe the periphery, not the core.** A clarification that asks *what do you mean by fairness*\n   elicits the shared core and misses the split. Marti's own method is the template: feature and\n   similarity questions. For criterion-contested terms the detector already has the criteria; present\n   them as taps — *fair because of desert, or because of non-discrimination?* — which is the inversion\n   principle applied to disambiguation and costs the user one tap instead of a sentence. (§ 5b sharpens this to *subtract\n   against a shown reference*: show the centroid or the other party's sense, ask what is not yours.)\n4. **Trigger under felt agreement.** The parties will not flag a periphery split, because they believe\n   the periphery is shared. The downstream-divergence alarm is therefore the *primary* detector, not a\n   safety net (the corpus already says this; the lens says why). A cheap pre-trigger that needs no\n   downstream structure: between-source meaning distance per concept, Rodriguez's embedding regression\n   on the attributed claims — proposed in the companion doc, now with its use-site.\n5. **The sixth flattening mechanism, proposed: centroid sense minting.** A model asked to name a\n   contested term's senses proposes population-mean readings; when a user then confirms one, the\n   graph has ratified an average nobody holds. Mitigation: every sense is seeded from attributed usage\n   (the pass already returns claim indices — enforce that a sense with no attributed usage is shown as\n   *machine-proposed* and never as a default), and item 2's display rule. The canon line says *five\n   known flattening mechanisms*; adding one is a canon change and is ledgered for the founder.\n6. **Build the essentially-contested terminal state**, proposed in July and still absent from the\n   lifecycle rules: senses partitioned by lineage *and* downstream divergence persisting across\n   clarifications. Its health metric is the quality of the mapped contest, not sense convergence. The\n   lens gives it a level — a culture's carving, level 3–4 — and predicts which terms will land there.\n7. **Bridge candidates from peripheries.** A claim supported by claims from both senses of a\n   lineage-bifurcated concept is a bridging-claim candidate — one graph query, no model call, and the\n   first instrument that would have found *teknikneutrala incitament* by structure.\n\n**Perfectionist (the shape the layer would have if built for the periphery from the start):**\n\n- **A sense is a triple**: an *exemplar core* (stated once, rarely contested), a *criteria list* (each\n  a definitional claim, attackable, in cluster form so family-resemblance concepts stop chasing a\n  bottom), and a *per-position weighting* (which criteria dominate for whom — the parity ruling's\n  *\"which reasons and values weigh more for whom\"*, applied to a word).\n- **Sameness of two concept uses becomes structural**: same core and overlapping criteria → one\n  concept; same criteria, different weighting → one concept, different position; disjoint criteria →\n  bifurcated; no shared core → unrelated. The seven-key set, applied to senses instead of claims, with\n  *can't tell* preserved — and the chiasmus the sameness ruling already carries (*claims want merging\n  with careful splitting, concepts want splitting with careful merging*) falls out of the asymmetry\n  between core and periphery.\n- **Gradable predicates get a standard slot.** *Expensive*, *decent*, *safe enough* are periphery\n  objects whose divergence is a threshold, not a criterion; *minimally decent standard of living* is\n  already in the index as a noun phrase carrying exactly that.\n- **Within-mind periphery is a weighting, not a sense.** The same person's *safety* in two debates\n  differs by weighting, which the weighing descent's state-deficit and satisfier-belief screens already\n  separate; the concept layer should not mint a new sense for a mood.\n- **The wager, at this layer**: clarification that descends a periphery to its criteria will find\n  shared cores where the world supplies them and *contested criteria that do not peel* where a\n  culture does — the essentially-contested state is that horizon, and the concept layer is the\n  instrument that locates it.\n\n## 7. What was read and measured\n\nEvery doc in the § 2 table was read in the passages cited, plus the contested-concept prompt and\nmodels, the graph store's sense write, the lifecycle rules, and the two front-end surfaces; the live\nconcept index was fetched from production the same day. The two-axis examples are standing biology\n(eyes, wings, Lenski) and the sources in the lens doc. Nothing here re-derives the July rulings on\nthick terms or the August rulings on sameness; it points at them.\n"}