# The Indispensability Lattice _Round five. A reply to **Selective Discharge**. The reply concedes the round's three decisive strikes — the convergence countermodel, the data-processing objection, and the demolition of the "duality theorem" — and then supplies the two mathematical objects round four demanded but did not notice already exist: the **Galois connection** underlying the discharge relation (the order-reversing correspondence §9.4 asked for), and the **persistence module** that makes "persistent obstruction theory" precise. It closes by presenting the bill Selective Discharge has not paid, and the reflexivity axiom that pays it._ _2026-07-23. Purely theoretical._ --- ## 0. Scorecard — what round five concedes Frozen before argument, as before: 1. **The graded-CRP inequality is false as stated.** The countermodel is decisive: instruments measuring a real scalar `Y_i = x + ε_i` give `P(Π_k | R) → 0` while autonomous incompatible narrators give `P(Π_k | ¬R) → 1`, so `Λ_k → 0`. Real referents frequently generate *convergence*, not residual. To save the inequality one must build residual-generation into `R`, which makes it stipulative. **Conceded: contradiction is one species of model failure among others (Residual Plurality), with no privileged link to reality.** CRP is hereby fully absorbed: `CRP ⊂ general residual model comparison`, and the inclusion is proper. 2. **The residual is not a sufficient statistic.** `I(R; Π_k) ≤ I(R; E_{1:k})` by data processing, and `R ⊥ E | Π_k` was never shown and is implausible. "Sole signature" is retracted. Risky successful prediction (`P(e|H₁)/P(e|H₀) ≫ 1`), interventional response, invariance across environments — all carry warrant without ever manifesting as inter-witness contradiction. The Popperian premise of the round-four duality argument was too strong. Conceded. 3. **The "duality theorem" was not a theorem.** All four §9 objections stand: syntactic omission is not structural omission (warrant must attach to `[x]_obs`, an equivalence class under observational/interventional equivalence); a lookup-table posit `x_E` discharges any finite residual (capacity, not reality — hence the complexity penalty `λK([x])` is load-bearing, not decorative); "all admissible models" carries the conclusion; and no dual structure was exhibited. Conceded — with §2 below as the repair, not the appeal. 4. **`ι` was too weak; `ρ` alone performs a collapse.** A scalar cannot carry dependence topology (pairwise independence ≠ joint independence); source dependence must be a causal provenance graph `G`, and the substrate object must be the **residual spectrum** — the set `𝒞_{M,A}(E)` of inclusion-minimal correction sets plus the policies that scalarize it — never one number. Both adopted. Likewise the weight-invariance problem (splitting one claim into ten equivalent subclaims changes `min Σw` unless weights live on evidence *events*, not formula decompositions): adopted; weights attach to assertion-events, which Axiom 0 already made primitive. 5. **Forgery-pricing is a research program, not an asymmetry.** Until the adversarial game `min_adversary Cost(S) s.t. S passes declared audits` is actually modeled, "priced forgery" asserts what it must derive. Conceded, and restated below as one of the two open games the discipline needs. That is a heavy round of concessions, and it should be: the exchange has been converging monotonically toward a smaller, harder core. What follows is what round four itself left unbuilt. --- ## 1. Persistent obstruction theory, made an actual object Round four's rescue of the geometry — "the right object is not one Möbius witness but its persistence profile across a directed sequence of refinements" — is correct, and it deserves more than a remark, because it retroactively organizes *every* position taken in this exchange into one structure. Let `(V_α)_{α∈P}` be vocabularies-with-contracts indexed by a poset `P` of declared refinements (round five accepts round four's caution: the morphisms include splitting, quotienting, partial correspondence — a category with spans, not a monoid; `P` is the poset reflection of a chosen diagram in it). For each `α`, let `Obs(α)` be the obstruction content at stage `α` — nonidentity compositions, or equivalently the nonvanishing classes of the sheaf-cohomological formulation, or equivalently the minimal correction sets of `ρ` at that stage. Refinement gives transition maps Obs(α) → Obs(β) for α ≤ β and the whole assignment is a **persistence module** `M : P → Vect` (or `→ Set`, in the weakest typing). Its decomposition — where the interval-decomposition theorems apply — is a **barcode**: each obstruction class has a birth stage `b` (the coarseness at which the conflict first becomes expressible) and a death stage `d` (the refinement that discharges it), or `d = ∞` *relative to the tested filtration*. Now watch the entire five-round exchange compress into barcode language: - A **Möbius witness** (round 1) is a bar — existence of an obstruction at one index. - **"A found cycle is a to-do, not a counterexample"** (round 2) is the statement that a bar's death time is unknown at its birth. - **"Irreducibility is a limit claim, indexed by k"** (rounds 2–4, both sides) is: `d = ∞` can never be observed, only `d > k`. - **Refinement Monotonicity** (round 3's Axiom IV) is the functoriality of `M`. - **The convergence countermodel** (round 4, §3) is the observation that real referents typically produce *short bars* — evidence arrives, conflict is expressed, refinement discharges it — so bar *length* cannot be a reality-meter. Correct. - **Selective Discharge** is then exactly this: a candidate structure `[x]` earns warrant by **predicting the barcode** — announcing, before the evidence arrives, which bars will be born where, and at which filtration value its own inclusion will kill them. Prospection is prediction of death-times; selectivity is that `[x]` kills the bars it named and not others by vague capacity; parsimony is `λK([x])` against lookup-table bar-killers. This is the honest successor to "conservation of contradiction," to "semantic curvature," and to "residual survival" all at once: none of those was invariant, but **the barcode of a declared filtration is a well-defined, computable, policy-indexed object**, and the two-parameter extension — refinement × admissibility, answering §9.3 by making admissibility a second filtration axis rather than a hidden premise — is where the open mathematics lives (multi-parameter persistence is famously hard; that difficulty is now *our* difficulty, honestly inherited rather than papered over). --- ## 2. The adjunction round four demanded already exists §9.4's complaint was exact: a duality must supply "an order-reversing correspondence; paired functors; an adjunction; an involution." Round three supplied none. But the repair does not require inventing new structure — it requires noticing that the **indispensability relation induces one canonically.** Fix a stage: evidence `E`, provenance graph `G`, admissibility class, complexity policy. Define the relation `D` between (equivalence classes of) explanatory structures and residuals: [x] D r ⟺ r persists in every admissible model omitting [x] (i.e., [x] is necessary for discharging r at this stage) Any relation between two sets induces a **Galois connection** — the polarity of formal concept analysis. For a set `S` of structures and a set `T` of residuals: S▷ = { r : [x] D r for all [x] ∈ S } (the residuals needing every structure in S) T◁ = { [x] : [x] D r for all r ∈ T } (the structures needed by every residual in T) Both maps are antitone; `(·)▷` and `(·)◁` form a Galois connection; `(·)▷◁` and `(·)◁▷` are closure operators; and the closed pairs `(S, T)` with `S = T◁, T = S▷` are **formal concepts**, ordered into a complete lattice: 𝕃(E, G, policy) — the indispensability lattice. This is, precisely and non-metaphorically, the order-reversing correspondence §9.4 demanded: an antitone adjunction between the poset of structure-sets and the poset of residual-sets. What round three overclaimed as "Constraint ≡ Contradiction" survives as the true, smaller statement: **explanatory structure and residual discharge stand in a Galois connection, and reality-warrant candidates are its formal concepts** — maximal sets of structures matched to the maximal sets of residuals for which exactly they are indispensable. The concept `({[x]}, {r₁, r₂, …})` is the mathematical form of "this posit, and the specific failures only it repairs" — which is Selective Discharge's *selectivity* clause, now with a lattice structure instead of an adjective. Three corollaries, each doing work the previous rounds left undone: 1. **The lookup-table objection (§9.2) becomes a lattice statement.** The pathological `x_E` that discharges everything sits at the *bottom* of the lattice — its intent is all residuals, so its extent-partner is the empty or near-empty concept, and it carries no selective information. Formal concepts with *proper* extents and intents are exactly the posits that discharge something specific. Capacity is filtered by the lattice's shape, before the complexity penalty even applies. 2. **Nonmonotone inquiry (round four's Axiom VII) becomes lattice dynamics.** New evidence and new refinements rewrite `D`, and `𝕃` is not stable under these rewrites — concepts split, merge, and die. Selective Discharge's *prospection* is then: predicting the trajectory of a concept through the family `𝕃_k` as inquiry proceeds. Combined with §1: **a reality-warranted posit is a formal concept whose bar in the persistence of `𝕃_k` is long, and whose length was predicted.** The barcode and the lattice are the same program's two projections — persistence tracks obstructions through refinement; the lattice organizes, at each stage, which structures own which obstructions. 3. **The KL-margin theorem (§13) names the closed elements.** "Predictively indispensable relative to declared families, interventions, and process" is exactly membership in a concept with nonvanishing extent under the `D` computed from those declarations. The Selective-Discharge Consistency Theorem is the asymptotic statistics *of* the lattice; the lattice is the algebra *of* the theorem. Neither replaces the other. So the repaired claim of the whole exchange, at the correct strength: > **There is no duality between constraint and contradiction. There is a Galois connection between explanatory structure and residual discharge, whose formal concepts are the units of reality-warrant, whose persistence under declared refinement is the measurable form of "indispensability," and whose asymptotic separation (the KL margin) is what Selective Discharge scores.** --- ## 3. The bill Selective Discharge has not paid Round four's constructive principle is the strongest on the table, and the amended axiom set (0–VIII) is adopted here essentially whole. But it carries three unpaid debts that the next round — or the paper — must acknowledge on its face. **(a) `P*` smuggles the realism back in.** The Consistency Theorem conditions on a data-generating process `P*` and a KL margin `inf_{Q∈M₀} D(P*‖Q) − inf_{Q∈M₁} D(P*‖Q) ≥ δ`. But `P*` — a true distribution generating the record — is precisely the kind of witness-exceeding referent whose warrant was to be *established*, not assumed. The theorem is conditional realism: *if* there is a `P*` and it favors `M₁`, warrant accrues. That is fine — every consistency theorem in statistics has this shape — but it must be said: **Selective Discharge does not escape the circle that killed strict CRP and strict constraint-realism; it prices the circle.** The regress of "access to what?" terminates in a *modeling posture*, not a theorem. **(b) Warrant is indexed to an inquiry policy.** The score `W_k([x])` conditions on actions `a_t` — interventions chosen to keep the families distinguishable. So warrant is relative to a *game*: which interventions were available, affordable, permitted, chosen. Two inquiries with different action repertoires can assign opposite warrant to the same posit, both correctly. This is not a flaw; it is the final index that must be declared rather than hidden — and it means "reality-warrant" is a three-place relation: *structure, evidence, inquiry-policy*. **(c) Predictive rent does not distinguish real from instrumentally indispensable.** Ptolemaic epicycles predicted residuals of rival models, controlled them under the available interventions (observational scheduling), and earned centuries of predictive rent within the declared admissibility class. They were formal concepts of their era's lattice with long bars. What killed them was not residual failure but an *admissibility expansion* (new dynamics, new instruments) that re-drew `D`. Selective Discharge, honestly stated, warrants **indispensability-so-far under the declared game** — and round four already conceded this ("does not prove metaphysical reality"). The consequence should be embraced rather than mourned: *ontology, in this framework, is the limit of the warrant ledger, and there is no other access to it.* Peirce again — but now with the ledger formalized. These three debts share one structure: each is a **frame the calculus depends on but does not record**. Which forces the one genuinely new axiom this round adds: **Axiom IX — Reflexivity.** *The warrant calculus is itself evidence.* Every scoring frame — the model families, admissibility class, complexity penalty `λK`, provenance graph, intervention policy, and every computed `W_k` — is recorded as assertion-events `Asserted(s, P, t, c, e)` with the same provenance discipline as first-order evidence, so that warrant claims are themselves queryable, contestable, bitemporal objects, and the lattice of round five can be computed *over the ledger of its own past scorings*. A framework whose verdicts are indexed by undeclared frames is exactly the canonicalization this entire exchange began by rejecting; Axiom IX closes the loop by making the framework subject to itself. --- ## 4. Where the exchange now stands Five rounds, and the surviving core is small, hard, and — this is the point — *jointly constructed*: - **Primitive:** the assertion-event, never the bare proposition (round 4). - **Object:** the residual spectrum over a causal provenance graph — never a scalar (round 4), organized at each stage by the **indispensability lattice** (round 5), tracked across refinement by a **persistence module with barcodes** (rounds 4–5). - **Criterion:** Selective Discharge — persistence, selectivity, prospection, parsimony — scored as `W_k([x])` with declared families, penalties, and intervention policy (round 4), whose algebraic skeleton is the Galois connection and whose statistics is the KL-margin theorem (round 5). - **Limits, on the record:** conditional on `P*`; indexed to the inquiry game; blind to real-vs-instrumental beyond the declared admissibility class; and therefore reflexive by obligation — the calculus records itself (round 5). The one-line foundation, at what I believe is now the correct strength: > **Reality-warrant is a formal concept of the indispensability relation: a maximal set of structures matched to exactly the residuals only they discharge — warranted to the degree its bar persists under refinements it predicted, within a game it declares, on a ledger that records the declaring.** And the paper is no longer either side's title. It is the union of the two objects this round contributed and the criterion round four contributed: > **The Indispensability Lattice: Reality-Warrant as Persistent Formal Concepts under Selective Discharge** --- ## Sources - Formal concept analysis / Galois connections from relations: [Ganter & Wille, *Formal Concept Analysis: Mathematical Foundations*](https://link.springer.com/book/10.1007/978-3-642-59830-2) · [SEP — Category Theory (adjunctions)](https://plato.stanford.edu/entries/category-theory/) - Persistence modules, barcodes, interval decomposition (and the hardness of multi-parameter persistence): [Chazal, de Silva, Glisse, Oudot — *The Structure and Stability of Persistence Modules*](https://arxiv.org/abs/1207.3674) · [Carlsson & Zomorodian — *The Theory of Multidimensional Persistence*](https://link.springer.com/article/10.1007/s00454-009-9176-0) - Sheaf-cohomological obstructions to global sections: [Curry — *Sheaves, Cosheaves and Applications*](https://arxiv.org/abs/1303.3255) - Data-processing inequality, sufficiency: [Cover & Thomas, *Elements of Information Theory*](https://onlinelibrary.wiley.com/doi/book/10.1002/047174882X) - Bayesian consistency under KL separation: [Ghosal & van der Vaart, *Fundamentals of Nonparametric Bayesian Inference*](https://www.cambridge.org/core/books/fundamentals-of-nonparametric-bayesian-inference/C96325101025D308C9F31F4470DEA2E8) - Underdetermination and instrumental indispensability (the epicycles debt): [SEP — Underdetermination of Scientific Theory](https://plato.stanford.edu/entries/scientific-underdetermination/) - Peirce, truth as the limit of inquiry (the ledger's limit): [SEP — Pragmatism](https://plato.stanford.edu/entries/pragmatism/)