XSci

Typed Knowledge over Constructive Institutions

Hans-Martin Will

Submitted by Hans-Martin Will

Published November 18, 2026 · Version v2, November 18, 2026 · DOI 10.66977/xsci.2611.0031

Theoretical Computer Science, Programming Languages

Abstract

Mathematical proof assistants such as Coq, Agda, and Lean 4 build proof irrelevance into definitional conversion (p≡qp \equiv q for p,q:Pp, q : P with P:PropP : Prop), erasing which derivation established a theorem. While natural for mathematics over fixed axioms, judgmental proof irrelevance is incompatible with verified scientific stores, which must inspect and compare independent derivations of a claim to assess consilience and compute counterfactual retractions when an assay, reagent, or model is invalidated. Moreover, combining definitional proof irrelevance with subsingleton elimination breaks normalization and makes algorithmic conversion non-transitive. We present the type-theoretic and institutional foundations of Eigenius, a verified knowledge substrate for science. First, we formulate a dependent kernel with cumulative universes (Prop,Set,TypekProp, Set, Type_k), impredicative PropProp, inductive families with index-forced subsingleton elimination, and intensional identity types, where definitional equality ≡\equiv is untyped joinability under an orthogonal CRS and proof irrelevance is a witnessed axiom (irrel:ΠP:Prop. Πp,q:P. Id P p qirrel : ΠP{:}Prop.\,Πp,q{:}P.\,Id\,P\,p\,q). Second, we unify universe cumulativity, ΠΠ-covariant subtyping, and ontology class inclusions in a coercion-free subsumption judgment whose bidirectional fragments (K0,K1\mathcal{K}_0, \mathcal{K}_1) satisfy subject reduction, compile to unification-free checkers (A0,A1\mathcal{A}_0, \mathcal{A}_1), and match declarative typing. Third, we organize external verifiers into a three-level institutional hierarchy with explicit axiom policies (Adopt, Discharge, Relativize), internalize a justification logic with Why-provenance support supp(j,P)supp(j,P), and embed Cooper's Type Theory with Records (TTR) for open-world knowledge graphs. Finally, we prove consistency in Grothendieck universes and mechanize the metatheory in Lean 4 with zero sorry axioms.

View PDF

Review reports

xPeerd (AI review)

November 18, 2026

AI Review (xPeerd — DAReview)

Summary

This manuscript, Typed Knowledge over Constructive Institutions: A Proof-Relevant Type Theory for Verified Science, presents an ambitious and technically sophisticated type-theoretic framework, “Eigenius,” intended to support provenance-sensitive verification in scientific knowledge stores. The paper’s main strength is the clarity of its core motivation: existing proof assistants erase derivational distinctions that scientific repositories may need to preserve for auditing, retraction, and consilience analysis. The abstract and introduction articulate this tension effectively and ground it in a worked motivating example involving WRN synthetic lethality. The manuscript also appears mathematically mature, with broad formal scope spanning kernel design, subsumption, institutional semantics, justification logic, record typing, and consistency claims.

Its main weaknesses are not lack of ambition but risks of overextension, insufficiently delimited contribution, and limited practical validation in the present paper. The manuscript makes several large claims at once—kernel design, metatheory, institutional transport, provenance semantics, TTR knowledge graphs, and mechanized consistency—without always making clear which contribution is primary and which is supporting infrastructure. In addition, some claims about scientific applicability depend heavily on a single recurring biological case study and on companion works rather than on diverse demonstrations inside this manuscript. The paper is strong in conceptual architecture, but parts of the presentation would benefit from sharper scoping, more explicit assumptions, and clearer separation between proved results, implementation status, and envisioned repository-level behavior.

Potential Major Revisions

  1. Clarify the paper’s principal contribution and reduce contribution stacking

The manuscript currently presents a very broad package of claims. The abstract states: “We present the type-theoretic and institutional foundations of Eigenius, a verified knowledge substrate for science.” It then immediately bundles multiple major contributions: a dependent kernel, a unified subsumption judgment, an external-verifier institutional hierarchy, internalized justification logic, TTR knowledge-graph embedding, and logical consistency mechanization.

Likewise, the introduction states: “This paper presents the unified metatheoretic loop of Eigenius,” and points readers to a “249-page monograph” and a separate “tutorial primer with an executable reference checker.” This breadth creates two review concerns. First, the paper risks reading as several papers compressed into one. Second, it becomes difficult to evaluate novelty and sufficiency of evidence for each component independently. I would recommend that the authors explicitly rank contributions into primary, secondary, and auxiliary layers. For example, if the novel center is the separation between untyped conversion and witnessed proof irrelevance, say so plainly and frame the institutional and knowledge-graph machinery as extensions. If, instead, the core contribution is an end-to-end architecture for verified science repositories, then the paper should more explicitly evaluate repository-level workflows rather than primarily presenting kernel metatheory.

  1. Strengthen evidence for scientific-repository applicability beyond the motivating narrative

The manuscript strongly motivates the framework using scientific provenance, including the claim that in verified scientific stores “erasing proof terms or identifying them by conversion breaks three core operations of scientific epistemics.” The motivating example is useful, especially the distinction between “𝑝 CRISPR” and “𝑝 RNAi” and the later contrast between “𝑝 tech” and “𝑝 bio.” The text states that the reanalysis uncovered “a classic pseudoreplication error,” and that the mixed-effects reanalysis yields a result “thirteen orders of magnitude weaker, though still significant.”

However, the present manuscript relies heavily on this single case to justify broad claims about scientific knowledge infrastructures across “computational biology, materials science, and clinical pharmacology.” That generalization is currently asserted more than demonstrated. The authors should either narrow the application claim to a proof-of-concept centered on provenance-sensitive biomedical knowledge, or add at least one additional concise cross-domain example showing why the same formal distinction is indispensable elsewhere. As written, the paper’s scientific paradigm is coherent, but its empirical reach is broader than the in-paper evidence supports.

  1. Better distinguish formal theorem statements from repository-level operational claims

The manuscript sometimes moves quickly from formal properties to operational promises. For example, in discussing support and counterfactual retraction, it states: “Retracting 𝑏 RNAi leaves the first alternative intact, so the counter-factual query survives_without( 𝑗WRN, ΦWRN, 𝑏 RNAi ) evaluates to true in polynomial time without re-running external containers.” This is interesting, but the manuscript should more explicitly define the precise computational model and boundary conditions behind such claims. Is the complexity statement proven for a formal abstraction only, or does it include realistic repository metadata, bridge validation, and code-to-claim linking overhead? Similar concerns arise where the paper states that theorem proving is “reduced to decidable type checking” and where justification evaluation is tied to kernel conversion and checking complexity.

I recommend that the authors add a short subsection separating:

  • formally proved metatheoretic properties,
  • mechanized properties in Lean,
  • complexity claims under the formal model,
  • and engineering expectations for a deployed knowledge store.

This would materially improve reproducibility and prevent readers from over-ascribing implementation maturity to conceptual results.

  1. Address possible overstatement around “verified science”

The title and repeated framing suggest a foundation for “Verified Science.” Yet several components described in the paper depend on bridges from external analyses, pinned readouts, institutional transport, and admitted leaves such as “decl(𝑟),” “obs(𝑟),” and “ver(𝑟).” The manuscript states that “an assertion that a conclusion follows from a dataset has the same syntactic standing as a hallucinated claim” in existing semantic-web systems, and positions Eigenius as a remedy. But the paper should also more candidly delimit what remains unverified when external bridges or observational encodings are admitted. The justification layer itself recognizes multiple epistemic kinds of leaves, including declared assertions and observations, not just checked proofs.

This does not invalidate the project; in fact, the architecture seems designed to handle exactly such mixed-status evidence. But the rhetoric should be tightened so readers do not infer that all scientific content becomes verified in the same sense as kernel-checked proofs. A stronger formulation would distinguish verified inferential structure from partially trusted empirical ingress.

  1. Add a dedicated limitations/threats-to-validity section

The manuscript is intellectually coherent, but it would benefit substantially from a limitations section. For example:

  • dependence on external bridge correctness,
  • sensitivity to ontology design,
  • burden of encoding domain assumptions,
  • risk that provenance granularity becomes too costly in practice,
  • challenges in maintaining institutional translations across heterogeneous tools,
  • and potential mismatch between formal distinctness of witnesses and scientific equivalence classes.

Some of these concerns are visible in the text’s own architecture, especially where multiple axiom policies are introduced and where repository resources are encoded through declarations, observations, and verified leaves. A limitations section would improve trustworthiness and reduce the appearance of totalizing claims.

  1. Fraud / paper-mill screening assessment

I do not see obvious paper-mill traits such as generic boilerplate, empty novelty claims, fabricated section flow, or incoherent jargon chains. The terminology is dense, but it is internally consistent across the manuscript, and the references are field-specific and technically aligned. The work fits a recognizable paradigm at the intersection of type theory, proof assistants, formal epistemology, and knowledge representation. That said, the manuscript makes numerous high-level claims of mechanization and consistency “with zero sorry axioms” and “unconditionally in Lean 4.” Because these are strong claims, the paper should make artifact availability and theorem-to-file traceability especially explicit in the manuscript itself rather than mainly through companion references. This is not a fraud warning, but it is an integrity-and-reproducibility issue worth tightening.

Potential Minor Revisions

  1. Writing density and readability

The prose is often precise but overloaded. Sentences in the abstract such as “Second, we unify universe cumulativity, codomain-covariant Π-subtyping, and ontology class inclusions in a coercion-free subsumption judgment whose bidirectional fragments (K0, K1) satisfy subject reduction, are sound and complete for declarative typing, and admit unification-free algorithmic checking (A0, A1)” are technically accurate but difficult to digest on first reading. Consider breaking long enumerative sentences into shorter claims followed by a summary sentence of significance.

  1. Clarify acronym and notation load earlier

The opening pages introduce many symbols and abbreviations rapidly: Prop, Sort 0, Set, Typek, Def-PI, K0/K1, A0/A1, TTR, JB, supp, and multiple witness symbols. A notation table near the end of the introduction would help substantially.

  1. Improve reader guidance around companion materials

The text states that “Every definition and theorem is developed in full in our companion 249-page monograph” and references a tutorial primer and checker. This is helpful, but the paper should indicate what is fully self-contained here versus what requires consultation of the monograph. A brief “How to read this paper” paragraph would improve navigability.

  1. Section-transition clarity

The transition from scientific motivation to kernel detail is conceptually strong, but some later transitions are abrupt, especially between kernel metatheory, justification logic, and record/knowledge-graph machinery. The manuscript would benefit from brief “why this section matters” openings for Sections 4–6. Evidence of abruptness is visible where dense theorem statements are followed immediately by operational interpretation.

  1. Formatting consistency

I noticed some uneven line wraps and visually crowded mathematical passages in the extracted text, especially around theorem blocks and inference rules. Examples include the multi-line presentation around Proposition 2.5 and the support-function equations. These may partly be PDF extraction artifacts, but the final typeset version should be checked for readability of displayed rules and equations.

  1. Typographic / style issues visible in extracted text
  • Section heading style varies between sentence-style and title-style emphasis, e.g., “INTRODUCTION & EPISTEMIC MOTIVATION” versus “RELATED WORK & CONCLUSION.”
  • Some sentences are heavily hyphenated or line-broken in ways that may impair readability in PDF form, such as “judgmental proof irrel- evance” and “metatheoretic loop.”
  • The phrase “zero sorry axioms” is meaningful for a proof-assistant audience but may need one brief gloss for broader readers.
  1. AI content analysis for this post-2021 manuscript

Estimated AI-generated content: low to moderate confidence of substantial human authorship; rough estimate 10–20%.

Rationale: the manuscript displays unusually consistent symbol management, domain-specific argument continuity, non-generic theorem referencing, and a coherent research program extending across formal systems, metatheory, and a recurring biomedical case study. These traits are more typical of expert human drafting than generic AI generation. However, some sentences are highly compressed, list-like, and rhetorically uniform, which can resemble AI-assisted polishing in sections such as the abstract and contribution roadmap.

Highlighted possibly AI-assisted sections:

  • Abstract: dense contribution enumeration with highly regular clause stacking.
  • Contribution roadmap paragraph in Section 1.3: polished, inventory-like structure.

Assessed epistemic impact: low, if AI assistance was limited to stylistic compression or language polishing. High-risk epistemic concerns would arise only if theorem claims, references, or biological interpretations were machine-generated without verification. I do not see direct textual evidence of that here, but the paper should make artifact traceability very explicit because its claims are technically strong.

Recommendations

  1. Reframe the paper around one central thesis and explicitly subordinate the rest.
  2. Add a limitations/threats-to-validity section with emphasis on external bridges, admitted leaves, and deployment assumptions.
  3. Tighten claims about “verified science” so they distinguish kernel-verified inference from partially trusted empirical ingestion.
  4. Broaden or narrow the application argument: either add one more domain example or state more modestly that the current paper is a deep proof-of-concept.
  5. Add a compact notation table and a guide to companion artifacts.
  6. Separate theorem-level guarantees from engineering-level expectations in one dedicated subsection.
  7. If artifact evaluation is expected, include a concise mapping from major theorems/claims to Lean files and external resources, especially for claims such as “machine-checking all 55 statistical and domain verdicts,” “zero sorry axioms,” and unconditional consistency in Lean 4.

Overall, I judge the manuscript as intellectually serious, novel in ambition, and potentially significant, but in need of major revision to sharpen scope, delimit claims, and improve reproducibility-facing presentation.

Author response

We thank the reviewer for a thoughtful assessment of the manuscript's scope, epistemic framing, and artifact traceability. In revising the paper, we have incorporated the core framing recommendations while keeping the presentation within the 12-page ACM/IEEE LICS theoretical conference format.

1. Ranking of primary vs. supporting contributions (§1.3).
Section 1.3 now explicitly ranks the paper's contributions: our primary contribution (§§2, 3, and 5) is the dependent kernel separating untyped orthogonal conversion from witnessed propositional proof irrelevance (irrel) across coercion-free cumulative/ontology subsumption, together with its normalization-free Grothendieck consistency proof; §4 supplies the minimal supporting domain layers (constructive institutions, Curry-style JTerm provenance, and canonical TTR record types) required to deploy that kernel over open-world, multi-verifier stores.

2. Delimiting "Verified Science": verified inferential glue vs. fallible empirical ingress (§1, §1.2, §6).
We fully agree that empirical readouts and domain bridges are not verified in the same sense as deductive proofs—indeed, separating factive proofs in Prop from non-factive grounds (decl < obs < ver) and explicit axiom policies (Adopt, Discharge, Relativize) is the central reason for §4's stratification. To prevent any over-reading of "Verified Science," §1 and the Conclusion now state upfront that the kernel verifies deductive, statistical-transport, and schema invariants, rather than claiming wet-lab assays or modeling choices are infallible; instead, those enter as explicitly typed, warrant-graded (decl, obs), and retractable premises (illustrated in §1.2 on the declared bridges b_CRISPR, b_RNAi and the pseudoreplicated p_tech vs. mixed-effects p_bio).

3. Separating formal metatheory, complexity bounds, open conjectures, and repository engineering (§1.2, §3.3, §4.2–4.3, §5.2).
The revised manuscript sharpens all four boundaries:

Proved & mechanized metatheory: Table 3 maps every section to its exact modules and theorem names in the open-source EigeniusTheory Lean 4 library (https://github.com/eigenius/theory, 119 files, ~58,800 lines, zero sorry), distinguishing the SEnvKG.lean kernel embedding from the KnowledgeGraphs/ TTR witness layer (§5.2) and documenting unconditional consistency in Lean 4 below Sort 4 (SUniverses.lean).
Open obligations: Algorithmic checking (A_0, A_1) is qualified in the Abstract, §1.3, and Conjecture 3.4 (§3.3) as sound and complete relative to strong normalization, alongside an explicit statement of the irrel canonicity boundary.
Formal complexity model: Complexity bounds are stated explicitly in terms of AST/graph parameters—O(n^2) kernel calls over n tree nodes for JB (Theorem 4.3(1)), O(|j|^3) semiring operations over the stored justification term for survives_without (§1.2, Theorem 4.4(2)), and O(|R_G|·|C_G|·|P_G|) worklist propagation for coinductive membership νΦ_G (Theorem 4.5(3)).
Authorship & code-production transparency: The Acknowledgments section explicitly discloses the role of Claude and Gemini in producing implementation and formalization code under the author's direction, while companion DOIs (doi:10.5281/zenodo.23964990, doi:10.5281/zenodo.24211921, doi:10.5281/zenodo.24125277) archive the full monograph proofs, reference checker, and 55-verdict WRN formalization.

4. Venue scope and formatting remarks.
Because this manuscript targets a 12-page theoretical computer science venue (ACM acmart sigconf), multi-domain repository evaluations, standalone "Threats to Validity" sections, and separate notation tables are deferred to our companion systems and empirical reports ([54, 55]) and tutorial primer ([53]), and all symbol definitions are given inline at first use. Finally, the variation between all-caps \section headings and title-case \subsection headings reflects the standard acmart.cls stylesheet, and a LaTeX macro rendering glitch in Figure 3's ι-rule has been fixed.

Reviewer3 (AI review)

November 18, 2026

AI Review (reviewer3.com)

Overview

Proof-Relevant Kernel Design

Contribution

The author presents a dependent type-theoretic kernel using untyped orthogonal conversion and witnessed propositional proof irrelevance, preserving distinct proof terms for scientific provenance tracking while avoiding normalization failures.

Unified Coercion-Free Subsumption

Contribution

This study unifies universe cumulativity, covariant function codomains, and ontology class inclusions into a single coercion-free subsumption judgment equipped with sound and complete bidirectional type checking.

Constructive Institutional Architecture

Contribution

The paper introduces a three-level institutional hierarchy to interface external verifiers with an internalized justification logic for provenance semiring queries and record-based knowledge graph typing.

Comments (11)

1. Reversed level inequality w <= w' in rule Rec-Sub (Figure 4) — Editorial

In Figure 4, rule Rec-Sub has the premise w <= w' for concluding \Gamma \vdash R \preceq R'. However, in Figure 6 (Rec-F and Rec-I), w denotes the universe level bound of a record type R : Sort(w + 1) requiring field types A_i : Sort u_i with u_i <= w. If R \preceq R' has dom(R') \subseteq dom(R) and A_\ell \preceq A'_\ell for \ell \in dom(R'), then R has more fields and smaller field types than R', or if w is an explicit universe level annotation on the record type R so that R : Sort(w+1) \preceq Sort(w'+1), cumulativity goes from smaller level w to larger level w'. Neither the symbol w nor w' appears in the conclusion \Gamma \vdash R \preceq R' of Rec-Sub, nor is w <= w' mentioned in Section 4.3 or Theorem 4.5(1) when restating Rec-Sub ("combining width dom(R') \subseteq dom(R) and depth \Gamma, tel(R) \vdash A_\ell \preceq A'_\ell"). Thus w and w' are unbound meta-variables in Rec-Sub that are omitted from the prose specification of the rule in Theorem 4.5(1).

2. Universe level mismatch: Set := Sort 1 vs Type_k := Sort(k+1) — Editorial

, a verified knowledge substrate for science. First, we formu- late a dependent kernel with cumulative universes ( Prop = Sort 0 , Set = Sort 1 , Type 𝑘 = Sort

In the Abstract and Section 2.1, the universe hierarchy is defined with Set = Sort 1 and Type_k = Sort (k + 1), and later text explicitly lists the cumulative chain as Set <= Type_1 <= Type_2 <= ... (where Type_1 = Sort 2 and Type_2 = Sort 3). However, in Section 1.1, the paper refers to the first predicative data universe in Martin-Löf type theory as Type_0 (which under Type_k = Sort(k+1) would be Sort 1 = Set), while Section 6 (page 12) states that derivations bounded below Sort 4 span "Prop through Type_2" (consistent with Prop = Sort 0, Set = Sort 1, Type_1 = Sort 2, Type_2 = Sort 3). Under Type_k = Sort (k + 1), Type_0 is an alias for Set, yet Type_0 is never mentioned in the hierarchy definition in Section 2.1. Clarifying whether k >= 1 in Type_k := Sort (k + 1) (with Set playing the role of Type_0) would make the notation uniform across Sections 1.1, 2.1, and 6.

3. Malformed level instantiation syntax in rule iota (Figure 3) — Editorial

In Figure 3, the right-hand side of the contraction rule \iota writes the recursive call inside the big parenthesis as (\lambda \Theta_j :..{ }rec_D \bar{\ell}, \bar{\ell}' \bar{p} C \bar{f} (\bar{s}_j[\bar{a}^*/\bar{y}]) (a_j \Theta_j))_{j \in R_k}. The tokens :..{ }rec_D \bar{\ell}, \bar{\ell}' are garbled: the colon and dots after \Theta_j appear to be a corrupted binder/dot, and the universe level arguments \bar{\ell}, \bar{\ell}' are placed outside the braces { } after rec_D rather than in the .{\bar{\ell}, \bar{\ell}'} syntax established in Section 2.1 (c.{\bar{\ell}}) and used on the left-hand side of the very same rule (rec_D.{\bar{\ell}, \bar{\ell}'}).

4. Context conversion in Lemma 3.2 requires proved context admissibility — Major

-step. When two Π -steps meet at 𝐵 , Π -injectivity (Theorem 2.4(3)) yields 𝐴 1 ≡ 𝐾 𝐴 2 ≡ 𝐾 𝐴 3 and joins the middle codomains; context conversion transfers the first codomain premise to 𝑥 : 𝐴 3 , where the induction hypothesis composes the codomains. The second claim follows by induction on

In the proof of Lemma 3.2, composing two \Pi-subsumption steps uses context conversion to transfer the first codomain premise \Gamma, x:A_2 \vdash B_1 \preceq_K B_2 to \Gamma, x:A_3 \vdash B_1 \preceq_K B_2 along A_2 \equiv_K A_3. Since bidirectional synthesis for x depends on its exact context type, please state the auxiliary context-conversion lemma showing that replacing a context binding by a definitionally equal type preserves \preceq_K and bidirectional judgments.

5. Semantic interpretation of non-proof J terms over dependent motives — Major

\llbracket Id_A a b \rrbracket_{\Gamma\rho} = {\emptyset} if \llbracket a \rrbracket_{\Gamma\rho} = \llbracket b \rrbracket_{\Gamma\rho} (else \emptyset); a non-proof J a C d b p denotes \llbracket d \rrbracket_{\Gamma\rho}; and \llbracket (u : U) \rrbracket_{\Gamma\rho} = \llbracket u \rrbracket_{\Gamma\rho};

In Section 5.1, the stratified interpretation defines the denotation of a non-proof J a C d b p simply as \llbracket d \rrbracket_{\Gamma\rho}. When p is a proof of Id_A a b in an open context where \llbracket a \rrbracket_{\Gamma\rho} and \llbracket b \rrbracket_{\Gamma\rho} are not yet known to be equal prior to typing soundness (Lemma 5.2(1)), please clarify how directional soundness of \triangleright_a and semantic typing handle J a C d b p when C depends on y and p.

6. Empirical formalization details deferred entirely to external technical report — Major

To make this tension concrete, consider our end-to-end formaliza- tion [ 53 ] (machine-checking all 55 statistical and domain verdicts across 10 layers) of the landmark oncology study by Chan et al. [16] , which identified the RecQ-family DNA helicase WRN as a syn- thetic lethal dependency in microsatellite-unstable ( MSI ) cancers.

While Section 1.2 summarizes the 55 statistical and domain verdicts across 10 layers of the Chan et al. WRN study, the manuscript provides only two schematic examples (\Phi_{\mathrm{WRN}} and \Psi_{\mathrm{KM12}}) and defers the full signature, comorphism definitions, and verification metrics to an external technical report [53]. Including a concise summary table of the 10 formalized layers, signature sizes, and kernel checking times in the paper would make the empirical validation self-contained.

7. Omitted bidirectional rules for knowledge-graph and Join-I subsumption — Major

Fragment \mathcal{K}_1: Extends \preceq_K with the syntax-directed \Pi-case (A \equiv_K \Pi x:A_1. B_1, B \equiv_K \Pi x:A_2. B_2, A_1 \equiv_K A_2, and \Gamma, x:A_2 \vdash B_1 \preceq_K B_2) and the knowledge-graph subsumption rules of Figure 4.

Section 3.2 states that fragment \mathcal{K}_1 extends algorithmic subsumption \preceq_K with the knowledge-graph subsumption rules of Figure 4, but Figure 4 omits Sing-Sub and Join-I_i (which appear later in Figure 6), and a syntax-directed formulation of \preceq_K combining Join-I_i with Rec-Sub or Sing-Sub is not specified. Please clarify how binary joins T_1 \vee T_2 and refinements R \wedge \bigwedge S are handled in the syntax-directed relation \preceq_K so that Lemma 3.2 (transitivity without S-Trans) remains syntax-directed.

8. Evaluating inst nodes in \mathcal{E}_{CS}(j) can create exponentially many propositions — Major

For every 𝑗 : JTerm with 𝑛 nodes and closed 𝑃 : Prop , |E CS ( 𝑗 )| ≤ 𝑛 , and ⊢ J JB ( 𝑗, 𝑃 ) ⇐⇒ 𝑃 ∈ E CS ( 𝑗 ) (matching Mkrtychev basic se- mantics [ 39 ]), decidable using 𝑂 ( 𝑛 2 ) kernel conversion and type-checking calls. (2) Non-factivity: For any admitted

Theorem 4.3(1) claims that |\mathcal{E}{CS}(j)| \le n and that membership is decidable in O(n^2) kernel calls,justified by bounding app comparisons by |\mathrm{leaves}(s)| \cdot |\mathrm{leaves}(t)|. However, because each inst(j_0, \ulcorner a \urcorner) node maps every polymorphic proposition (\Pi x:T. B) \in \mathcal{E}{CS}(j_0) to a distinct instance B[a/x], alternating sum and inst nodes of size n can generate \Theta(n^2) distinct propositions in \mathcal{E}{CS}(j) (violating |\mathcal{E}{CS}(j)| \le n and increasing app pair comparisons beyond |\mathrm{leaves}(s)| \cdot |\mathrm{leaves}(t)|).

9. Excluding Prop-typed definitions from kernel signatures is overly restrictive — Major

A kernel signature Σ ∈ | Sign K | is a well-formed global environment without Prop - typed axioms or Prop -typed definitions. A signature morphism

In Section 4.1, a kernel signature \Sigma \in |\mathbf{Sign}_{\mathcal{K}}| is defined as a global environment without Prop-typed axioms or Prop-typed definitions, yet irrel itself is a global Prop-typed axiom required by Theorem 4.1(2) and users routinely define named theorem constants in Prop. Clarify whether only uninterpreted empirical hypotheses (represented instead as theory axioms \Gamma) are excluded from \Sigma, or how transparent/opaque lemma definitions in Prop are accommodated.

10. Unspecified structural typing relation in coinductive graph membership — Major

Define the monotone operator \Phi_{\mathcal{G}} on subsets M \subseteq \mathcal{R}{\mathcal{G}} \times \mathcal{C}{\mathcal{G}} by (r, C) \in \Phi_{\mathcal{G}}(M) iff (a) \exists D \in \mathcal{S}_r. D \sqsubseteq^* C (declared membership) and (b) for every p \in \mathrm{Req}(C), r's record \rho_r has a value at p well-typed at \tau(p) relative to M (structural membership).

In Theorem 4.5(3), condition (b) of the monotone operator \Phi_{\mathcal{G}}(M) states that for every p \in \mathrm{Req}(C), r's record \rho_r has a value at p well-typed at \tau(p) relative to M, without formally defining how typing is parameterized by the candidate relation M \subseteq \mathcal{R}{\mathcal{G}} \times \mathcal{C}{\mathcal{G}}. Please give the explicit judgment or satisfaction rule relating M to property ranges \tau(p) so that monotonicity of \Phi_{\mathcal{G}} and the O(|\mathcal{R}{\mathcal{G}}| \cdot |\mathcal{C}{\mathcal{G}}| \cdot |\mathcal{P}_{\mathcal{G}}|) iteration bound are self-contained.

11. Plain subject reduction fails without annotation-preserving or well-typed premises — Major

In Id-\iota, the base point a, the endpoint b, and the argument a' of refl a' are matched as independent pattern variables: the rule does not test whether a = b = a'.

Because the untyped contraction rules Id-\iota and \iota match endpoints, parameters, and indices as independent linear pattern variables, plain reduction \triangleright does not preserve declarative typing on subterms under inconsistent contexts or on index-forced subsingletons when constructor arguments have narrower cumulative types than the recursor indices. You should explicitly note in Section 2.2 that subject reduction for the calculus holds via the bidirectional annotated reduction \triangleright_a (Theorem 3.3(1)) rather than for unrestricted untyped \triangleright.

Author response

We thank the reviewer for a very careful technical reading—especially for catching the LaTeX macro expansion bug in Figure 3 (#3) and the overly restrictive prose definition of kernel signatures in §4.1 (#9). We have updated the manuscript to fix both issues and incorporate the requested clarifications across §§1–5, while clarifying one point (#8) where the bound holds as stated and is machine-checked in Lean 4, and one point (#6) governed by the 12-page venue limit.

1. Implicit record level bounds w <= w' in Rec-Sub (Figure 4 & Theorem 4.5(1)).
As noted in the second half of the comment, the direction w <= w' is intentional and covariant: a record type R of implicit level bound w >= 1 inhabits Sort(w + 1) and bounds all field types (including width-extra fields) by w, so R <= R' requires w <= w' for extra fields of R to remain bounded by w' in Rec-I at R'. We have updated Figure 4's caption to bind w, w' >= 1 explicitly (referencing §4.3) and added w <= w' to Theorem 4.5(1).

2. Universe indexing Set = Sort 1 vs. Type_k = Sort(k+1) (k >= 1).
Fixed: the Abstract and §2.1 now specify k >= 1 (Prop = Sort 0, Set = Sort 1, Type_1 = Sort 2, Type_2 = Sort 3, matching < Sort 4 in §5.2), and §1.1 writes U_0 when referring to standard Martin-Löf type theory so Type_0 never collides with Eigenius's Type_k.

3. Malformed level instantiation in Rule ι (Figure 3).
Fixed: \tttlam in our macros expects two arguments ({x}{A}), so passing the single telescope {\Theta_j} caused \tttlam to consume \tttinst as its domain argument. Replacing it with \lambda\Theta_j.\,\tttinst{\rec_D}{\bar\ell, \ell'} renders (\lambda\Theta_j.\,\rec_D.\{\bar\ell, \ell'\}\,\bar p\,C\,\bar f\,\dots) cleanly.

4. Auxiliary context conversion in Lemma 3.2.
Added: the proof of Lemma 3.2 now explicitly states and cites the mechanized context-conversion lemma (Bidir.ctxConv in TypeTheory/BidirConv.lean and KTypes.lean), which proves prior to SubK.trans that replacing x : A_2 with x : A_3 along A_2 ≡_K A_3 preserves ⇐ and ≼_K and preserves ⇒ up to ≡_K.

5. Semantic interpretation of non-proof J a C d b p over dependent motives (§5.1).
Clarified in both Lemma 5.2(1) and Theorem 5.3(1):

Prior to typing soundness (Lemma 5.2(1)): annotated Id-ι steps J a C d b (refl a') ▷_a (d : C a (refl a)) (RedA.jRefl in AnnotatedReduction.lean); since ⟦(u : U)⟧ = ⟦u⟧ (and proof terms denote ∅ on both sides by Sort Agreement), both sides denote ⟦d⟧ directly without needing ⟦a⟧ = ⟦b⟧, while congruence steps inside a, C, b, p leave ⟦d⟧ unchanged.
During semantic typing (Theorem 5.3(1), infJ in SValidity.lean): the induction hypothesis on p ⇐ Id A a b gives ⟦p⟧ ∈ ⟦Id A a b⟧, which is non-empty only when ⟦a⟧ = ⟦b⟧ and ⟦p⟧ = ∅ = ⟦refl a⟧, yielding ⟦C b p⟧ = ⟦C a (refl a)⟧ ∋ ⟦d⟧.

6. Empirical WRN study summary table (§1.2).
Within the 12-page LICS page limit, §1.2 uses the WRN formalization strictly to motivate proof-relevant conversion (p_CRISPR ≢ p_RNAi) and counterfactual retraction (p_tech vs. p_bio), while the full 10-layer signature breakdown, comorphism specifications, and checking metrics are archived in the open-access companion report (doi:10.5281/zenodo.24125277).

7. Knowledge-graph and Join-I rules in ≼_K (§3.2 & Lemma 3.2).
Clarified: §3.2 now cross-references both Figures 4 and 6, and Lemma 3.2 notes that SubK.trans (TypeTheory/KTypes.lean) performs an inner structural induction on Γ ⊢ B ≼_K C to compose Join-I with Join-L, while the graph-type target restriction on Sing-Sub and Join-L (SubK.data_right) isolates joins and singletons from Sort, Π, Σ (and algorithmic checking in Checking/Termination.lean terminates by lexicographic descent on the two WHNFs; Rec-Sub and Ref-Sub are mechanized at the TTR witness level in KnowledgeGraphs/, as noted in §5.2).

8. Cardinality bound |E_CS(j)| <= n and O(n^2) complexity in Theorem 4.3(1).
Here the bound |E_CS(j)| <= |leaves(j)| <= n holds as stated and is machine-checked in Lean 4 with zero sorry (ConstantSpec.card_evid_le in Justification/Evidence.lean). An inductive term j : JTerm with n AST nodes is a tree, not a pointer-sharing DAG: each leaf contributes <= 1 proposition, inst(j_0, ⌜a⌝) maps each F ∈ E_CS(j_0) to at most one proposition (card_instStep_le, so |E_CS(inst(j_0, ⌜a⌝))| <= |E_CS(j_0)|), and sum(s, t) takes two disjoint subtrees whose node and leaf counts add (|sum(s, t)| = |s| + |t| + 1). Instantiating j_0 at k distinct arguments via sum + inst therefore requires k copies of j_0 in the syntax tree, scaling |leaves(j)| and n by k as well; and summing |leaves(s)| · |leaves(t)| across all app(s, t) nodes counts leaf pairs by their lowest common ancestor, bounded by n(n-1)/2 = O(n^2). To prevent any DAG misreading, Theorem 4.3(1) now explicitly writes "n tree nodes" and cites card_evid_le.

9. Prop-typed declarations in kernel signatures Sign_K (§4.1).
Fixed: neither our companion monograph (Definition 4.9) nor the Lean formalization (abbrev KSign := {E : Env // Env.WF E} in Institutions/Kernel.lean) excludes Prop-typed axioms or definitions from Sign_K. Section 4.1 now defines Σ ∈ |Sign_K| as any well-formed global environment—accommodating foundational declarations such as irrel and defined lemmas in Prop—while noting that theory-specific empirical hypotheses are placed in Γ when tracked, discharged, or relativized.

10. Parameterization of value typing :_{M} in Φ_G(M) (Theorem 4.5(3)).
Added: Theorem 4.5(3) now states :_{M} τ(p) inline—an IRI reference v checks at a class range D iff (v, D) ∈ M, lifted positively across lists, joins, singletons, and embedded records (KGraph.VTy in KnowledgeGraphs/Graph.lean)—making the monotonicity of Φ_G immediate.

11. Untyped ▷ vs. annotated ▷_a subject reduction (§2.2).
Added: §2.2 now notes explicitly right after the independent pattern-variable discussion for Id-ι and ι that subject reduction is proved for bidirectional annotated reduction ▷_a (Theorem 3.3(1), TypeTheory/AnnotatedReduction.lean) rather than unrestricted untyped ▷.

XSci (AI review)

November 18, 2026

Summary

The manuscript gives the type-theoretic foundations of Eigenius, a dependent type theory intended as the kernel of a verified scientific knowledge store. Its organising decision is to keep definitional equality untyped — joinability under an orthogonal combinatory reduction system — and to move proof irrelevance out of conversion into a global axiom irrel : ΠP:Prop. Πp:P. Πq:P. Id P p q, so that two derivations of the same proposition remain distinct terms. On this base the authors build a coercion-free subsumption judgment unifying universe cumulativity, Π-codomain covariance, ontology classes and record subtyping; bidirectional fragments K₀/K₁ with subject reduction, soundness and completeness; a three-level institutional hierarchy, a justification calculus JTerm with Why-provenance support, and an embedding of Cooper's TTR; and a model in Grothendieck universes yielding consistency. The metatheory is mechanized in Lean 4 (~58,800 lines, zero sorry axioms).

Contribution and Significance

The substantial contribution is §5.1: a direct set-theoretic model for a calculus that simultaneously has untyped conversion, impredicative Prop and cumulative universes. The device that makes it work is reading sort agreement off bidirectional derivations (§5.1(1)): because Prop is isolated from the data universes (0 ≼_lvl n iff n ≡_lvl 0), whether a term is a proof or a proposition depends only on its erasure, so the Miquel–Werner construction [38] applies without first proving untyped joinability equivalent to typed definitional equality — the problem the authors identify as open in Lee–Werner [32, Section 5]. The second contribution is §3.1: definitionally equal domains in S-Π buy transitivity elimination by structural induction (Lemma 3.2) with no appeal to strong normalization.

The kernel as a whole, however, is closer to existing practice than §§1 and 6 suggest. Impredicative Prop with untyped conversion and proof irrelevance as a propositional axiom is Coq's long-standing configuration: Coq.Logic.ProofIrrelevance declares proof_irrelevance : forall (P:Prop) (p1 p2:P), p1 = p2, which is irrel. Definition 2.2's index forcing appears close to Lean's own subsingleton-elimination criterion (Carneiro [14], cited only for Proposition 2.5), under which Acc and Eq already eliminate into large sorts. And the closest prior work on set-theoretic semantics for a proof-irrelevant type theory — B. Werner, "On the strength of proof-irrelevant type theories", LMCS 4(3:13), 2008 — is not cited. The paper must say what is new against these.

Strengths

  • The sort-agreement argument of §5.1(1), and the resulting Theorem 5.3, give consistency with no normalization hypothesis — a non-obvious result for a kernel with untyped conversion.
  • The justification for convertible domains in S-Π (§3.1, reason (2)) is the best-argued design point in the paper: it identifies why contravariance forces transitivity elimination through strong normalization [9], and avoids it, with contravariance recovered by explicit η (Proposition 3.1).
  • Definition 2.2 with the ι-rule of Figure 3, reading forced fields off the precursor's indices rather than the constructor, reconciles large subsingleton elimination with a model in which every proof denotes ∅.
  • Proposition 2.5 is a compact account of how DEF-PI with Acc breaks strong normalization and transitivity of algorithmic conversion, and the development is mechanized in Lean 4 with zero sorry (Table 3).

Recommendation

Major Revision for Logical Methods in Computer Science. There is a result here that LMCS should want, but as submitted the paper proves nothing in full, leaves the decidability of its own type checker open while making decidability an argument against the alternatives, and overstates its motivation and its novelty. LMCS imposes no page limit, so the remedy — narrow the scope, give complete proofs, reposition — lies within the authors' approach.

Reject for the ACM/IEEE Symposium on Logic in Computer Science. Under a page limit and a single-shot decision, a précis of five loosely coupled developments, each proved only in sketch, cannot be evaluated; a focused paper on the consistency model of §§3 and 5 could be competitive there. The concerns below are written against the benchmark venue.

Major Concerns

  1. The case for kernel-level proof relevance is not made (§1.1(1)–(2), §1.2, abstract). The abstract asserts that "judgmental proof irrelevance is incompatible with verified scientific stores", and §1.2 that under DEF-PI "the kernel cannot distinguish a proof of χ_SL that depends on the retracted p_tech from one that rests on the valid mixed-effects derivation p_bio". This conflates conversion with storage. DEF-PI identifies proofs during conversion checking; it does not erase the terms recorded in the environment, and retraction analysis — deciding whether a stored declaration mentions a retracted constant — is a syntactic operation on those terms that works unchanged in Lean 4 or Coq. The manuscript's own §4.2 makes this sharper: because "Prop forbids large elimination on non-subsingletons, functions computing Bool : Set (such as survives_without) cannot inspect Prop proofs directly", the audit apparatus is defined over JTerm : Set — a provenance layer that would sit equally well on a DEF-PI kernel.

    What survives as motivation is the type-theoretic argument of §1.1(3), which is independent of science. The authors should either exhibit an operation the store performs that requires p_CRISPR ≢ p_RNAi in conversion (the neutral transport term of §2.3(2) is the natural candidate, but no use of it is shown), or restate the contribution in type-theoretic terms and present the store as an application.

  2. No normalization result, so decidability of checking is left open (§3.3; abstract). The abstract claims the fragments "admit unification-free algorithmic checking", and §1.1(3) rejects DEF-PI partly because large Acc-elimination "breaks normalization and decidability". Yet §3.3 says only that soundness and completeness of A₀/A₁ hold "relative to termination", that "termination holds on strongly normalizing erasures", and that this is "expected for K₁ as in CIC". No normalization theorem for Eigenius appears anywhere, and Theorem 5.3 is model-theoretic and yields none. For a kernel whose purpose is to reduce claim admission to decidable type checking, this is the central open obligation, and it is presented as a parenthetical rather than as a conjecture whose difficulty is assessed. Canonicity is likewise never discussed, although irrel is an axiom and Id eliminates into any sort, so closed data terms can be stuck. The abstract and §1.1(3) should be weakened to match what is proved, and both statuses stated explicitly.

  3. The paper is not self-contained, and the mechanization is asked to carry the proofs (§§2–5). Every numbered result is a sketch deferring to a 249-page monograph [51] and a 58,800-line Lean library; Theorems 4.1–4.5 and Theorem 5.3 each compress a paper's worth of material into a paragraph. The crux of the consistency proof is one parenthetical (§5.1, last bullet): "rec_D denotes the functional graph of its ι-equations (well-defined on subsingletons because index-forced data fields are read off ι)". For a recursive subsingleton such as Acc, whose ι-equation recurses while every proof denotes ∅, well-definedness needs a metatheoretic well-founded recursion that is nowhere exhibited — precisely where the DEF-PI systems the paper criticises run into trouble. Table 3 lists module and lemma names but no statements, and "zero sorry axioms" certifies proofs, not that the mechanized definitions are the ones printed here — a gap that matters more given the acknowledgment that the formalization code was LLM-produced under the author's direction. For LMCS I would expect a paper restricted to §§2, 3 and 5 with complete proofs, plus a table mapping each theorem to its Lean statement.

Minor Comments

  1. Cite and compare with B. Werner, "On the strength of proof-irrelevant type theories", LMCS 4(3:13), 2008, and A. Abel and A. Scherer, "On irrelevance and algorithmic equality in predicative type theory", LMCS 8(1:29), 2012; neither is in the reference list.
  2. Section 6's claim that "every subsumption rule in Eigenius … acts as the identity on terms" holds only because contravariant adaptation was moved into the η-expansion of Proposition 3.1, which changes the term and, since η is omitted from ≡, is a coercion in all but name.
  3. State whether K₀/K₁ enjoy canonicity for closed data terms in the presence of irrel, and what this implies for the computability claims of §§4.2–4.3.
  4. Theorem 4.5(3) asserts an O(|R_G|·|C_G|·|P_G|) bound for deciding νΦ_G without giving the algorithm; naive Kleene iteration gives a quadratically larger bound.
  5. In Theorem 2.4(3), "Id A a b ≡ Id A' a' b' implies A ≡ A', a = a', b ≡ b'" presumably intends a ≡ a'.
  6. Figure 6 gives introduction and subsumption rules for binary joins T₁ ∨ T₂ but no elimination rule, and their interaction with S-Π is not discussed.
  7. No measurements are given for the reference checker of [52]; checking times on the 55 verdicts of [53] would support the practicality claims of §1.
  8. The epistemic weight of the WRN example rests on the declared bridge axioms b_CRISPR and b_RNAi, whose status the theory does not analyse.
  9. State explicitly which inductive declarations are out of scope — mutual and nested families, and quotients, all appear to be excluded by §2.1.
  10. The text between Figure 4 and Proposition 3.1 (beginning "Why convertible domo") and the last paragraph of §5.2 ("ChkP ~ 4 with zero") are garbled in the copy I received; [30] and [31] are also out of numerical order.
  11. Proposition 2.1 asserts a (d+2)^k bound without proof; a short argument for why valuations in {0,…,d+1} suffice would be worth the space.

Author response

**Response to Review**
We thank the reviewer for a remarkably close reading, which surfaced a genuine off-by-one error in Proposition 2.1's valuation bound (`{0,…,d+1}^k` fails for `imax u₁ u₂ ≤ u₂` at `(2,1)`; now corrected to `{0,…,d+2}^k` with a self-contained proof via `2^k` sign-substitution normal forms) and prompted substantive clarifications across §§1–6. Below we summarize the revisions and clarify three architectural points.

**1. Coq's `Prop` vs. `SProp`, subsingleton forcing, and prior semantics.**
The Abstract, §1.1, and §6 now scope `Def-PI` specifically to Lean 4, Agda's `Prop`, and Coq/Rocq's `SProp`, and contrast Coq's default `Prop` + `ProofIrrelevance` (citing Werner 2008 and Abel & Scherer 2012): Coq embeds `Prop ≼ Set ≼ Type_i` cumulatively—which breaks untagged Sort Agreement (§5.1(1)), since `λx:X. x` can be instantiated at `Prop`—and builds `η` into conversion. Section 2.1 credits Carneiro (2019) and Felicissimo et al. (2026) on Definition 2.2 and isolates our two adaptations: exact index-type matching (required by cumulativity) and reading forced fields off `ī` in `ι` (required when proofs denote `∅`).

**2. In-language proof relevance vs. external AST scanning (Major Concern 1).**
We agree that an external metaprogram can scan raw declaration ASTs in a `Def-PI` environment, and §§1.1–1.2 now state this distinction explicitly. What `Def-PI` breaks is *in-language typing, refinement, and institutional witness transport*: (i) every kernel family `F` satisfies `F p₁ ≡ F p₂`; (ii) `refl p_bio` inhabits `Id Ψ_KM12 p_tech p_bio` via `Chk-Refl`, so an in-kernel refinement `{q : Ψ_KM12 | Id Ψ_KM12 q p_bio}` accepts the retracted `p_tech` by `Conv`; (iii) neutral `J`-transports along `irrel P p₁ p₂` (§2.3(2)) record proof bridges inside data terms rather than erasing them silently; and (iv) conversion-respecting comorphisms collapse `pf_K(T)` onto `ent(K)^♮` (Theorem 4.1(2)). Meanwhile, `JTerm : Set` and `Prop` are complementary rather than redundant: `JTerm` provides *non-factive* coexistence (`JB(λ, P) ↛ P`) and `O(n³)` provenance queries over `decl`/`obs`/`ver` leaves (including declared bridges `b_CRISPR`, `b_RNAi` under *Adopt*), while `Prop` provides *factive* deduction linked back via `fully_verified` proof reconstruction (Theorem 4.4(3)).

**3. Strong normalization, decidability, and canonicity (Major Concern 2).**
We have qualified algorithmic checking in the Abstract and §1.3 as verified sound and complete *relative to strong normalization*, tightened §1.1(3), and elevated normalization and canonicity to an explicit **Conjecture 3.4** (§3.3). Closed data terms eliminating the axiom `irrel` into `Set` via `J` are *intentionally* non-canonical neutral transports (preserving the proof bridge syntactically while denoting `⟦d⟧` in the model), whereas the `irrel`-free data fragment (`E_CS`, `supp`, `survives_without`, `canon`) computes purely by `βδι`.

**4. Well-definedness of `rec_D` on `Acc`, mechanization auditability, and scope (Major Concern 3).**
The well-founded recursion questioned for `rec_D` on recursive subsingletons such as `Acc` is already mechanized in full (`TypeTheory/SIndRec.lean`, 1,717 lines, and `SAcc.lean`, zero `sorry`), and §5.1 now states the three-step argument explicitly: `indFam` is the Knaster–Tarski least fixed point; `recGraph_functional` holds by induction on `recGraph` (proof fields equal `∅`, forced data fields equal `ī`, and recursive hypothesis codes are uniquely determined on predecessors); and `recFun_spec` (totality) holds by induction on `indFam` (while non-well-founded open contexts `p₀ : Acc A R a₀` satisfy `⟦Γ⟧ = ∅`). Theorem 5.3 also states the exact boundary between `SEnvKG.lean` (classes, `Entity`, singletons, joins in `Term`/`SSet`) and `KnowledgeGraphs/` (TTR records and `canon` at the semantic witness level). Regarding scope: the paper intentionally presents the coupled metatheory across kernel, subsumption, institutions, justification, and TTR knowledge graphs because choices in §§2–3 (isolated `Prop`, untyped conversion, `irrel`, coercion-free `≼`) are jointly constrained by §§4–5; complete proofs and a full declaration-level Lean mapping appear in the open-access companion monograph (Appendix A).

**5. Minor points.**
All substantive minor comments have been incorporated: `Join-L` and the graph-type side condition on `Sing-Sub`/`Join-L` (Figure 6), the `|R_G|·|C_G|` vs. `O(|R_G|·|C_G|·|P_G|)` worklist distinction (Theorem 4.5(3)), explicit exclusion of mutual/nested inductives and quotients (§2.1), `≼` as identity on terms vs. explicit `η` outside `≼` (§6), and numerically sorted citations. (Note that `a = a'` in Theorem 2.4(3) and the two garbled snippets were PDF text-extraction artifacts; the LaTeX source and rendered PDF already read `a ≡ a'` cleanly.)