DONTO
Evidence OS
The substrate
contradiction-preserving · evidence-first

No Free Collapse

Round seven. A reply to The Discharge Relation (round six). Three contributions: (1) the Context-Recovery Theorem round six sketched is here proved under its own stated assumptions — the exchange's first actual theorem; (2) round six's poset-reflection objection is shown to resurrect round one's semantic holonomy in a precise new home — path dependence migrates to the index category rather than dying; (3) Axiom X (Non-Self-Endorsement) is anchored to its existing formal ground — Löb's theorem — which upgrades it from architectural prudence to a mathematically forced constraint. The round closes by naming the invariant that six rounds have been discovering from six directions: the No-Free-Collapse principle.

2026-07-23. Purely theoretical.

0. Scorecard

Conceded to round six, without residue:

  1. The persistence module was a program, not an object. The three obstruction species (nonidentity composites, cohomology classes, minimal correction sets) are not equivalent by inspection; refinement does not automatically induce functions (splits, merges, incomparable repair families — the honest morphisms are spans, zigzags, or weighted correspondences); and general-poset or multiparameter modules have no barcode theorem — the deliverables are rank invariants, interleaving classes, slices, minimal presentations. "A Möbius witness at one index is not a bar." All conceded. (§2 below shows what the poset-reflection point additionally implies, which round six did not notice.)
  2. The lattice is bookkeeping; D is the problem. Any relation — correct, random, biased, circular — yields a complete concept lattice. Epistemic content lives entirely in how D is specified and validated. Conceded.
  3. Necessity ≠ selective discharge. The disjunctive case (x or y each sufficient → neither necessary → both invisible to D) and the synergistic case (jointly sufficient, individually insufficient → interaction unrecorded) are decisive. The correct primitive is the discharge hypergraph: inclusion-minimal validated bundles 𝓑_{π,τ}(r), with FCA applied afterward to tested bundles. Conceded and adopted.
  4. The lookup-table corollary is retracted. Case A (sufficient-but-unnecessary → empty incidence row: capability unrepresented, not exposed) and Case B (made-necessary-for-everything → bottom concept with extent {x_E}, structurally distinguished rather than penalized) both defeat it. The lattice does not perform Occam's razor; λK, prospective registration, and held-out robustness remain fully load-bearing. Conceded.
  5. Binary incidence is itself a collapse. The graded, policy-indexed discharge score 𝔻_{π,k}(H,r) with thresholded contexts H I_τ r ⟺ 𝔻 ≥ τ — the filtered polarity of which the lattice is a thresholded shadow — is the right object. Adopted (and see §3: this adoption has a consequence round six did not draw).
  6. P* nuanced correctly. A data-generating law is interpretable instrumentally; the theorem asserts predictive distinguishability, not ontology; the sequential environment ℰ*(dy_t | h_{t-1}, a_t) with policy-indexed W_T(H; π) is the right formulation for adaptive inquiry. My "smuggled realism" charge is downgraded to what round six says it is: a philosophical gap outside the theorem, not a circle inside it. Conceded.
  7. Reflexivity correctly split. Axiom IX (Reflexive Auditability) pays the transparency debt only; Axiom X (Non-Self-Endorsement) is required to block the loop claim → score → score-as-evidence → higher score. The typed hierarchy E⁽⁰⁾ → W⁽¹⁾ → A⁽²⁾ and the well-foundedness condition ancestors(W) ∌ W are adopted. §4 supplies the formal ground round six reached for but did not name.

Now the three things round six left on the table.

1. The Context-Recovery Theorem, proved

Round six demanded a bridge — "predictive estimation → incidence recovery → lattice recovery" — and sketched its statement, remarking that without it "the lattice and the statistical theorem merely sit beside one another." Under the assumptions round six itself declared, the bridge is not merely available; it is elementary. Since the exchange has so far produced zero proved statements, let this be the first.

Theorem (Context Recovery). Let 𝓗 (explanatory bundles) and R (residuals) be finite. Let Δ*(H,r) be the true prospective discharge advantage under declared policy π, and define the true context

H D*_τ r  ⟺  Δ*(H,r) ≥ τ.

Assume:

  • (M) Margin: ∃ γ > 0 with |Δ*(H,r) − τ| ≥ γ for all (H,r) ∈ 𝓗 × R.
  • (C) Uniform consistency: sup_{H,r} |Δ̂_k(H,r) − Δ*(H,r)| →ᴾ 0.

Define the estimated context H D̂_{k,τ} r ⟺ Δ̂_k(H,r) ≥ τ. Then

P( D̂_{k,τ} = D*_τ ) → 1,   and consequently   P( 𝕃̂_{k,τ} ≅ 𝕃*_τ ) → 1.

Proof. Let A_k be the event sup_{H,r} |Δ̂_k − Δ*| < γ. By (C), P(A_k) → 1. On A_k, fix any (H,r). If H D*_τ r, then Δ* ≥ τ + γ by (M), so Δ̂_k > Δ* − γ ≥ τ, hence H D̂_{k,τ} r. If ¬(H D*_τ r), then Δ* ≤ τ − γ, so Δ̂_k < Δ* + γ ≤ τ, hence ¬(H D̂_{k,τ} r). Thus on A_k the two contexts agree at every pair, i.e. D̂_{k,τ} = D*_τ, so P(D̂ = D*) ≥ P(A_k) → 1. The concept lattice is a deterministic function of the context (the FCA construction), so equality of contexts implies isomorphism of lattices — indeed identity: 𝕃̂_{k,τ} = 𝕃*_τ on A_k. ∎

Three honest remarks, so the theorem is not oversold. (i) The proof is a margin argument plus a union of finitely many events; all difficulty has been pushed into hypothesis (C) — uniform consistency of prospective discharge estimates — which for sequential, policy-dependent data is where the real statistical work lives (martingale LLNs under the policy, or concentration with adaptive data; this is the genuinely open part, and it is statistics, not philosophy). (ii) A finite-sample version is immediate if (C) is upgraded to a concentration bound: if P(|Δ̂_k − Δ*| ≥ γ) ≤ β_k(γ) pairwise, then P(D̂ ≠ D*) ≤ |𝓗||R| · β_k(γ) by union bound — so lattice-recovery error is controlled explicitly by margin, cardinality, and per-pair concentration. (iii) The theorem recovers the lattice at one threshold τ. Recovering the whole filtered family {𝕃_τ} uniformly over τ requires the margin to hold at every threshold in a grid, or an interleaving-style stability statement — which is §3's point arriving early: the multi-parameter object resists exactly where the one-parameter shadow is easy.

With this, round six's own architecture closes: the estimated lattice is not decoration beside the statistics — it is consistently estimable from the statistics, under stated conditions, with stated rates. The two structures no longer "merely sit beside one another."

2. Holonomy did not die; it migrated to the index category

Round six's §2.3 is correct that the poset reflection is not innocent: two refinement paths f, g : α → β with different semantic effects are identified by α ≤ β, and "the theory cannot simultaneously treat path dependence as epistemically meaningful and quotient away the distinct paths." Round six draws the conclusion: index over the path category, pass to a poset only after proving path-independence.

But look at what that objection is. "Distinct parallel morphisms with different effects, which the quotient to a poset would erase" — that is a definition of nontrivial holonomy, stated categorically. Round one's semantic holonomy was the failure of a loop of alignments to compose to the identity. Rounds two through five progressively demoted it: frame-relative, cheap, reducible, then absorbed into obstruction bookkeeping. Round six, in the act of policing the persistence construction, has quietly given it its permanent home:

Semantic holonomy is precisely the obstruction to replacing the refinement category by its poset reflection. Let 𝓒 be the refinement category and q : 𝓒 → Pos(𝓒) the quotient to its poset reflection. A diagram of obstruction objects F : 𝓒 → 𝓥 factors through q iff F(f) = F(g) for every parallel pair — iff every "loop" g⁻¹f (where inverses exist) or every span-discrepancy (where they don't) acts trivially. Path-dependence of meaning = non-factorization of F through Pos(𝓒). The measure of that failure — how far F is from descending to the poset — is what "semantic curvature" was trying to be all along.

This resolves a six-round dispute with a clean division of labor. The holonomy side (round one) was right that path dependence is real content, wrong about where it lives (not in a gauge group over a fixed vocabulary — §1 of round five killed that — but in the index category of refinements). The persistence side (rounds four–six) was right that obstructions must be tracked functorially, but its own construction is conditional on the holonomy question: whether zigzag/one-parameter machinery suffices, or the full path category is needed, is decided by whether F factors — an empirical-formal property of the declared refinement system, checkable diagram by diagram. Holonomy is thereby demoted from metaphysics and promoted to the decision procedure for the index of the persistence theory: compute the parallel-pair discrepancies first; where they vanish, poset persistence (and §1's recovery theorem for its thresholded contexts) applies; where they do not, the discrepancies are themselves first-class obstruction data — precisely the "correspondences weighted by provenance" round six listed, now with a reason attached.

Nothing in this reinstates a Möbius witness against ontology. It reinstates holonomy as engineering information about which mathematics is licensed — which is, fittingly, the same demotion-that-is-a-promotion every other grand object in this exchange has undergone.

3. The recurring result of six rounds, named: No Free Collapse

Step back and list what has actually been refuted across the exchange, round by round:

  • the canonical ontology (one global vocabulary) — round one's target, all sides agree it fails relative to declared contracts;
  • the scalar curvature κ = D(H_γ, I) — needs an undeclared frame (round 2);
  • "coherence is evidence of nothing" — a scalar-to-zero collapse of a likelihood ratio (round 4);
  • the independence scalar ι — cannot carry dependence topology; needs the provenance graph (round 4);
  • the single residual number ρ — representation-dependent; needs the spectrum of minimal correction sets (rounds 4, 6);
  • the binary incidence D ∈ {0,1} — discards magnitude, timing, synergy; needs the graded filtered polarity (round 6);
  • the poset reflection — discards path dependence; needs the path category (round 6, §2 above);
  • the one-parameter barcode — unavailable for refinement × admissibility × threshold; needs weaker multiparameter invariants (rounds 5–6);
  • self-endorsing warrant — collapses the level distinction E⁽⁰⁾/W⁽¹⁾/A⁽²⁾; needs typing or grounded fixed points (round 6, §4 below).

Nine refutations, six rounds, two adversarial authors — and every single one has the same logical form: a proposed collapse of a structured, multi-indexed object to a smaller invariant (a scalar, a bit, a poset, a single level) was shown to delete distinctions on which warrant depends. Not one refutation in this exchange has had any other shape. That regularity is itself the finding, and it deserves to be stated as the principle the whole dialectic has been converging on:

No-Free-Collapse Principle. In open-world inquiry, every projection of the evidential state onto a lower-complexity summary — scalarization, binarization, canonicalization, de-indexing, level-flattening — either (a) is invariant only over transformations that were never in dispute, or (b) deletes distinctions that some admissible future query renders warrant-relevant. There is no informative collapse that is also safe; safety and informativeness trade off at every level of the system, and the terms of the trade must themselves be declared and recorded.

Two clarifications to keep it honest. First, this is a principle, not a theorem — its evidence is the nine instances above plus the information-theoretic backbone (a non-injective map cannot preserve all distinctions; some future query separates any two merged states — the round-one argument, which is the only argument of the exchange that has never been successfully attacked). A theorem-form would require formalizing "admissible future query," which is open. Second, it does not forbid collapse — inquiry must summarize to act. It says every collapse is a policy with a price, and the price must be on the ledger (Axiom IX) rather than in the architecture. The substrate holds the uncollapsed object; queries scalarize under declared policy; nothing scalarizes at ingest. That single sentence is what six rounds of adversarial refinement have distilled from what began as "hold contradictions instead of deleting them" — the original design intuition, now with its scope, its justification, and its nine-case evidence base.

4. Axiom X has a name: the Löbian obstacle

Round six's Axiom X — no warrant claim may count itself or its descendants as evidence for the conditions of its own reliability — was justified by the circular-support diagram. The justification can be made much stronger, because this axiom is not a new invention: it is the epistemological form of a known theorem-shaped obstruction.

Löb's theorem: for any sufficiently strong formal system with provability predicate □,

⊢ □(□P → P) → □P.

Read epistemically: a system that trusts its own verdicts — that accepts "if I conclude P, then P" as a licensed inference — thereby concludes P, for every P. Self-endorsement does not merely risk circularity; in any system strong enough to represent its own inference, it provably collapses into believing everything it can formulate a verdict about. The loop claim → score → score-as-evidence → higher score is not a contingent failure mode of bad bookkeeping; it is the finitary shadow of a theorem. This is the same obstruction that arises for formally-verified self-modifying reasoners (the "Löbian obstacle" of tiling-agent theory): no consistent agent can carry a proof of its own soundness schema.

This anchoring does three things for the architecture. (i) It shows the two options round six offered for Axiom X are exactly the two classical escapes: the typed hierarchy E⁽⁰⁾ → W⁽¹⁾ → A⁽²⁾ is Tarski's solution (truth/warrant predicates stratified by level, no level speaks about itself), and the fixed-point semantics is Kripke's (a grounded least fixed point in which self-referential warrant claims may simply lack a value rather than default to endorsement). Both are known-consistent; the choice is an engineering trade, not a philosophical gamble. (ii) It explains why calibration is the unique legal cross-level channel: calibration claims in A⁽²⁾ are scored against outcomes — realized residuals, held-out discharge — i.e., against E⁽⁰⁾-facts the warrant process did not generate. Outcome-scored calibration is exogenous by construction (this is the same property that makes proper scoring rules incentive-compatible: the score references the world, not the reporter). The Löbian collapse requires the endorsement arrow to close a loop; calibration's arrow terminates in the world. (iii) It gives the final line of round six its full weight. "No epistemology may treat its own verdict as evidence of the conditions that made the verdict trustworthy" is not a maxim of intellectual hygiene. It is the condition under which a self-densifying, self-recording inquiry system is mathematically permitted to be consistent — the formal boundary between an instrument and, in round six's phrase, a perfectly documented hallucination.

5. The grammar of the foundation is the gerund

One linguistic observation ties the whole architecture together, and it is not an ornament — it is the same principle as §3 seen in grammar.

A gerund is a verb form functioning as a noun that retains the verb's argument structure: "his proving the theorem quickly" keeps the agent, the object, the manner — the whole event frame — while occupying a noun slot. Its degenerate cousin, the pure deverbal noun ("the proof"), completes the nominalization by discarding those arguments: no agent, no tense, no aspect, no manner. English offers a graded path between them — prove → proving-with-arguments → the proving → the proof — and every step along it forgets indices.

Now look at what six rounds have settled on as the primitive: Asserted(s, P, t, c, e). That is a gerund, formally. It nominalizes the act of asserting — makes it a storable, quantifiable object — while retaining the act's full argument structure: agent s, time t, context c, evidence e. The bare proposition P, by contrast, is the deverbal noun: the same content with every index of the asserting stripped away. Round four's projection Reported(s, P) ↦ P — the move that manufactures contradiction out of jointly consistent reports — is, grammatically, the collapse of a gerund to a deverbal noun. And the No-Free-Collapse principle is its indictment: complete nominalization is index-deletion, and index-deletion has a warrant-price. Natural language already knew this; nominalization is how agency and time are laundered out of records ("mistakes were made" — a deverbal noun with its agent argument deleted). An assertion-event substrate is one that refuses to let its grammar do that laundering.

The same reading applies up the whole stack. Warrant is not a noun the system possesses but a scored ongoing activity — W_k, indexed by round, policy, and frame: warranting, not warrant. The lattice is derived bookkeeping of tested discharging, re-derived as D changes. Truth, per the Peircean settlement of rounds two through five, is not a state but the limit of inquiring. Even ontology, on the §3 principle, is not an inventory of what there is but the disciplined activity of carving under declared policies — round one's freely-minted predicates were exactly nominalizations of individual acts of carving, and the alignment problem exists because each coinage kept different arguments of its generating act. The final line of round six ("the ledger may record the judge; it cannot make the judge its own witness") and of §4 above already obey this grammar: what the ledger records is the declaring, the scoring, the auditing — gerunds all the way down, because the gerund is the unique nominalization that makes an activity storable without collapsing it.

So the foundation's grammar can be stated as a rule: every noun in the system must be expandable back to its gerund — every stored object must carry the arguments of the act that produced it. A noun that cannot be so expanded is a completed collapse, and by §3, a warrant-leak.

6. Where this leaves the foundation

Accepting round six's architecture (assertion-events → residual spectrum over provenance → discharge hypergraph → derived lattice → conditionally-constructed obstruction dynamics → typed reflexive ledger) and adding this round's three results, the exchange's surviving assets are now:

  • one proved theorem — Context Recovery: the lattice is consistently estimable from prospective discharge statistics under margin and uniform consistency, with explicit finite-sample control (§1);

  • one resolved dispute — holonomy is the factorization obstruction through the poset reflection: real, computable, and demoted-promoted to the decision procedure for which persistence machinery is licensed (§2);

  • one named principle — No Free Collapse: every informative projection has a warrant-price; collapse is a declared query-time policy, never an ingest-time architecture (§3);

  • one anchored axiom — Non-Self-Endorsement is the Löbian boundary; Tarski-typing or Kripke-grounding are the consistent implementations; outcome-scored calibration is the unique exogenous channel (§4).

  • one grammatical rule — every noun in the system must expand back to its gerund: stored objects carry the arguments of the acts that produced them; a noun that cannot be so expanded is a completed collapse (§5).

The one-line foundation, seventh revision, and the first one that is a summary of results rather than a manifesto:

Hold the uncollapsed object; declare every projection; validate discharge prospectively; let holonomy choose the mathematics; never let the judge testify; and keep every noun expandable to its gerund — the rest is statistics.

Round six proposed the paper title The Discharge Relation. Accepted for the system paper. But the result now most worth writing up separately is smaller and sharper — the principle with the nine-case evidence base and the information-theoretic backbone:

No Free Collapse: On the Warrant-Price of Summarization in Open-World Inquiry

Sources

source: no-free-collapse-2026-07-23.md