{"path":"research/evidence-division-and-the-foundation.md","content":"# Evidence Division Under a Mother Claim: Why It Feels Foundation-Threatening, and What the Corpus Already Answers\n\n**Date**: 2026-08-21 · **Occasion**: founder root-question, verbatim: *\"how to divide up the relevant evidence between different subclaims under a mother-claim, and aggregate the evidence in a way that's epistemologically honest and well-calibrated, while honoring the fuzzy boundaries between some semantic domains that the different subclaims might be talking about — a central one that in my mind might challenge the foundation of what I imagine Deliberus could become.\"* · **Grounding**: full reads this session of [decomposition-axes.md](decomposition-axes.md), `deliberus/evidence_routing.py`, [strength-layer-audit.md](strength-layer-audit.md), the Heylighen additivity warning ([heylighen-value-complexity-and-deliberus.md](heylighen-value-complexity-and-deliberus.md) §3), plus the standing evidence-as-subgraph and self-similar-decomposition designs.\n\n## §1 Why the instinct is right — naming what it detects\n\nThe unease has a 120-year-old name: **confirmation holism** (Duhem, later Quine). Evidence never confirms an isolated claim; it confirms a claim *together with* a web of auxiliaries. Strictly speaking, \"divide the evidence among the subclaims\" is not merely hard — it is **metaphysically the wrong verb**, because evidential bearing is a property of claim-webs, not claim-atoms. The fuzzy boundaries between subclaims' semantic domains are not an engineering nuisance to be tuned away; they are holism showing up in the ontology. The founder's intuition is detecting that.\n\nThe Bayesian statement of the same point: the epistemologically honest aggregation is conditioning a **joint** model in which subclaim correlations are first-class. A tree-shaped sum has no correlation term — verified in our own code (`subclaim_energy` sums with none) — so QBAF-style propagation is a *tractable approximation* that diverges from honest aggregation exactly where the founder's question lives: where subclaims share domains and evidence smears across them.\n\nAnd the reason it feels **foundation-threatening** rather than merely hard: the wager says decomposition clarifies. If evidence cannot be honestly apportioned, decomposition instead *manufactures precision* — four children each looking independently evaluable while the evidence smears across them, double-counting inflating the mother, omission starving her — and the badge, the system's one public signal, inherits the arbitrariness at its most load-bearing layer. A map that is confidently wrong precisely where it decomposed deepest would be the failure mode the confession principle exists to prevent, produced by the flagship operation.\n\n## §2 Why it does not sink the foundation\n\n**The foundation never promised a correct division. It promised visible, contestable judgments.** The claims-all-the-way-down ontology is the rescue: there is no primitive \"evidence\" type — evidence is claims whose own support bottoms out empirically — so a routing judgment is a *link*, links carry provenance, and links can be attacked. Holism would sink a system that needed silent, authoritative division; nothing forces ours to be either. The threat is real only for the naive implementation, and the corpus has already, piece by piece, declined to build the naive implementation:\n\n1. **Link, never divide** (`evidence_routing.py`, from the evidence-as-subgraph design): evidence is a node in a DAG with as many parents as it genuinely bears on. Sharing is *visible structure*, not invisible copies. The web that holism insists on is representable — division was the wrong metaphor and the architecture already refuses it.\n2. **Abstention is first-class**: `undecided` is distinct from `irrelevant`, the unrouted fraction is reported, and nothing routes by similarity (the forty-seven-papers example: maximally similar to the domain-child, bearing on it not at all). Mis-routing manufactures support; abstention only admits ignorance — the literature confirms the asymmetry.\n3. **Fair-share discounting** (built, deliberately unadopted): evidence shared by *k* siblings counts once across them, identical to today when nothing is shared, only ever deflating. The correlation term the sum lacks, added as a correction that cannot flatter.\n4. **Decompose the evidence too** (the founder's own 2026-08-15 correction): most *apparent* smear is coarse evidence — \"she published 47 papers\" is a bundle, and at the right granularity the publication-record child draws on all of them, the domain child on a subset, the standing child on citations. Much of the fuzziness dissolves when evidence granularity matches claim granularity; the discount then handles only the genuine residual sharing.\n5. **Non-additive marks where contested** (`support_semantics.necessary`, shipped): the first escape from additivity, with Heylighen's warning on record that interaction is the norm in value-space — the vocabulary grows only where someone contests, per the standing decomposition rule.\n\n**And the irreducibly fuzzy residual has a formalism that turns it into an instrument.** Sheaf gluing (decomposition-axes §5): when local evidence-assignments are each coherent but refuse to glue into a global one, that is *contextuality* — a discovered incoherence with a name and a literature, not a bug. The shipped discursive-dilemma flag is already its judgment-aggregation instance. The honest system does not pretend the boundaries are crisp; it **measures where they refuse to be**, and publishes that as a finding.\n\n## §2b The conjunction problem is the same problem, seen from above (added 2026-08-21 on founder follow-up)\n\nThe strength audit's \"conjunction problem\" — the aggregation cannot tell **jointly required** parts (Dr. Smith's credibility needs publications AND domain-match AND standing) from **independently corroborating** ones (three separate studies), so a mother scores well while a required part is weak — is not a neighbor of the evidence-division question. It is the SAME question at the other joint. Evidence-division asks how support lands **downward** across children whose domains overlap; the conjunction problem asks how children's strengths combine **upward** through a join whose logic varies. Both are failures of one silent assumption: **independence** — of evidential bearing in the first case, of the parts' contribution-logic in the second. The honest name for the unified gap is the **join-semantics problem**: DECOMPOSES_INTO currently carries one implicit meaning, and reality has several.\n\nProbability theory grounds the design space exactly, and it turns out the shipped choice sits at a principled point in it:\n\n- **Sum** (the pre-`necessary` default): inflates conjunctions, rewards splitting (more children, more energy), and *amplifies* shared-evidence double-counting — every failure in this doc at once.\n- **Product** (honest for a conjunction of *independent* parts: ten conjuncts at 0.9 give ≈0.35): punishes decomposition itself — finer slicing of the same claim lowers its score — which would build an **anti-clarification incentive** into the flagship operation, and it double-punishes correlated conjuncts, which after evidence-decomposition is the common case.\n- **Ceiling/min** (`support_semantics.necessary`, shipped): implements the conjunction *bound* P(A∧B) ≤ min(P(A), P(B)) — the tightest thing that is always true regardless of correlation. It is splitting-neutral (the measured 3-parts-vs-12-parts both 0.502), and it *dampens* rather than amplifies shared-evidence inflation, since a uniformly inflated set of children moves a min far less than a sum.\n\nSo `necessary` is not just a patch — it is the correlation-agnostic upper bound, which is exactly the right default posture for a system that (per §1) cannot know the correlations. The remaining honesty gap is typed, not mysterious: corroborative joins still sum (correct only when evidence is genuinely independent — which is what fair-share discounting repairs), and genuinely independent conjunctions would deserve something between min and product (a founder-decision refinement, not urgent, because the anti-clarification hazard of product outweighs its honesty gain at current corpus maturity). The graph-scale face of the same mathematics is the **preface paradox**: every claim individually believable, their conjunction not — which is one more reason the residue map reports fractions rather than the corpus asserting its own grand conjunction.\n\n## §2c The assembly-theory boundary conundrum is the third face (added 2026-08-21 on founder connection)\n\nThe founder's connection holds, and it cuts in both directions. Walker names identifying **the boundaries of the selected units** as assembly theory's hardest problem — *\"the copy number is the hardest part — it is actually identifying the boundaries of the selected units\"* — and that difficulty stands even in chemistry, where part-identity is nearly given by bonds. Three strands tie it to the join-semantics problem:\n\n1. **Boundary choice → number inheritance.** AT's index depends on where the unit boundaries are drawn; our badge depends on claim boundaries, evidence granularity, and join types. Both formalisms compute over a decomposition the underlying material does not uniquely determine, and **every downstream number inherits the choice**. The LZ-compression equivalence attack that AT has not overturned exploits exactly this arbitrariness of the parts-alphabet — and the same attack shape aims at us (\"your strengths are artifacts of decomposition choices\"), which is why §5's stability-under-reasonable-routing falsifier is not optional: it is the test AT never ran on itself.\n2. **The join half AT has not even named.** AT counts joining steps but treats all joins alike — a bond is a bond, though chemistry itself distinguishes covalent from van der Waals. Two objects with identical assembly indexes can differ in whether their parts are jointly required or merely aggregated, and the index cannot see it. The join-semantics problem is thus the boundary conundrum's missing half: boundaries say what the units ARE; join semantics say what combining them MEANS. A construction-counting formalism needs both, and each project holds one half more firmly than the other.\n3. **The escape move runs in our direction.** AT's life-detection claims need an observer-independent decomposition — which is precisely what its critics deny it has. Deliberus never needed observer-independence, only **visibility**: units and joins here are contestable, provenance-carrying, and challenge-relative (every verifier its own optimal grain), so the boundary conundrum converts from a foundational threat into governed judgment. The connection's yield both ways: Walker's naming certifies our claim-sameness and routing difficulties as the *known hard problem of a research field* (expect no clean answer from trying harder), and our propose-and-contest architecture is the move their formalism lacks.\n\n## §3 The two reframes that dissolve most of the remaining pressure\n\n**There is no THE division, and there does not need to be.** The atomicity-optimum result (ACL 2025) says every verifier has its own optimal grain — so the \"right\" apportionment is relative to the challenge being run, which is the contest-driven decomposition rule arriving from the evidence side. A division is an artifact of a question, not a fact about the world, and the graph can hold several without contradiction because routings are links with provenance.\n\n**Honesty means calibrated bookkeeping with published divergence conditions, not solved holism.** Science never solved holism either; statistics did not divide evidence between hypotheses — it modeled jointly and *confessed its assumptions*. The equivalent here is cheap and buildable: a **correlation-exposure figure per mother claim** — the fraction of its children's evidential support that is shared/multi-routed — displayed beside the badge, so a strength resting on largely-shared evidence says so. The approximation stays; the reader learns exactly when to trust it less. That is the confession principle applied to the aggregation itself, and it is the missing instrument this question points at.\n\n## §4 The approach, sequenced (mostly the corpus's own backlog, now seen as one program)\n\n1. **Measure the alignment question on the live corpus** — does decomposition today attach distinct evidence per child or repeat the mother's? (decomposition-axes §6.1; the degrading configuration is the measured danger.)\n2. **Build the recomposition check** (losslessness as a colimit: can a model rebuild the mother from the children alone?) — makes lossless a number.\n3. **Run the evidence-decomposition pass + routing with abstention** (both designed; routing built); then put the **fair-share adoption decision** to the founder with measured before/after strengths — it changes every number in the graph, which is why it waits.\n4. **Ship the correlation-exposure confession** beside the badge (new, small, this doc's concrete yield).\n5. **Generalize the gluing check** from the discursive-dilemma flag: ask any decomposition subgraph whether its local sections glue; surface the failures as findings.\n6. **Keep grain contest-driven** — molecular not atomic, axes not booleans, divisions as challengeable links.\n\n## §5 What would actually falsify the foundation here\n\nStated so the worry stays a wager rather than a mood: the foundation fails if, after evidence-decomposition + linked routing + fair-share + the confession figure, mother-claim strengths remain **dominated by routing choices reasonable people make differently** — i.e., if the contestable-judgment machinery produces strength distributions with no stable signal under reasonable routing variation. That is measurable (route the same corpus under k reasonable policies, compare badge distributions), and nobody has measured it. Until then, the honest position: holism forbids perfect division, the architecture never needed it, and every mechanism the problem demands is either built, designed, or named above.\n\n## Cross-references\n\n[decomposition-axes.md](decomposition-axes.md) (the warning, the axes, colimit + sheaf) · `deliberus/evidence_routing.py` (link/abstain/discount, built-not-adopted) · [strength-layer-audit.md](strength-layer-audit.md) (the missing correlation term; conjunction/corroboration) · [heylighen-value-complexity-and-deliberus.md](heylighen-value-complexity-and-deliberus.md) §3 (interaction as the norm) · [self-similar-decomposition-and-claim-ontology.md](self-similar-decomposition-and-claim-ontology.md) (claims all the way down; the founder's coarse-evidence correction) · [scheme-bounded-decomposition-and-evidence-as-subgraph.md](scheme-bounded-decomposition-and-evidence-as-subgraph.md) (evidence as subgraph) · [inductive-reasoning-research.md](../inductive-reasoning-research.md) (the Bayesian/QBAF landscape)\n"}