{"id":"ae250676-0559-44ff-be5a-d47d8c795291","arxiv_id":"2509.07719","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The pullback of a relative presheaf topos on an arbitrary fibration is the relative presheaf topos on its inverse image.","lead":"This math paper proves that a standard way of building a 'relative presheaf topos' from a fibration is preserved under pullback along any geometric morphism, so the pullback is again a relative presheaf topos, built from the inverse image fibration. This gives a general description of fibered products of toposes that was previously only known in the cartesian case.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 4.12 proves only continuity of f*(η_C), not that it is a morphism of sites; the missing finite-limit condition is load-bearing for Propositions 4.13–4.16.","rationale":"The reader identified Proposition 3.1's omitted proof as the weakest assumption. My stress-test agrees that unresolved η-extension properties are dangerous, but it locates a more specific and more downstream gap: the proof of Proposition 4.12 explicitly checks only continuity of f*(η_C), then concludes that it is a morphism of sites. Given the paper's own definition, a morphism of sites must also preserve finite limits at the topos level. This is not derived, and it is not a consequence of being a morphism of fibrations in the non-cartesian case. The subsequent equivalence argument in Prop. 4.12, and hence the whole transposition theorem of Prop. 4.15, depends on this missing condition. I therefore agree partially with the reader: the paper has a real soft spot, but it is in the core proof of the main theorem rather than only in the preliminary adjunction apparatus. The verdict should remain CONDITIONAL rather than UNCHANGED? In the schema, UNCHANGED means my read does not change the reader's verdict, and the reader already returned CONDITIONAL. The specific gap strengthens the case for conditionality but does not justify REJECT, since the theorem may still be true and the missing finite-limit condition might be derivable with additional work. Hence I keep the reader's CONDITIONAL verdict, expressed as UNCHANGED.","tokens_in":31542,"tokens_out":15080,"duration_ms":138392,"concrete_test":"Independently re-derive the missing finite-limit-preservation step in the proof of Prop. 4.12: for a non-cartesian E-indexed category C, prove that f*(η_C) satisfies the filtering condition of [BC23, Prop. 3.13], or equivalently that its η-extension preserves finite limits fibrewise, using only continuity of f*(η_C) and the hypotheses of Prop. 4.12. A minimal case to test is the constant E-indexed category whose fibre is the walking arrow, with a non-trivial geometric morphism f. If no such proof can be given without assuming C cartesian, then Prop. 4.12 lacks a proof and Prop. 4.16 does not follow as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central bipullback theorem (Prop. 4.16) is obtained through Prop. 4.15, which relies on Prop. 4.13, which uses Prop. 4.12. In the proof of Prop. 4.12 the authors need f*(η_C): (G(f*C), Gir_{f*C}) -> (G(f*(S_C)), J_{...}) to be a morphism of sites. What the proof actually establishes is only continuity: the image topology contains the Giraud topology, and f*(η_C), being a morphism of fibrations, is continuous by Theorem 2.3. But the paper's own criterion (Prop. 2.4, and the filtering characterization cited from [BC23, Prop. 3.13]) requires a morphism of sites to preserve finite limits at the topos level, not merely to be cover-preserving. Since f* is the left 2-adjoint Lan_{f^op} on indexed categories, finite-limit preservation is not automatic; for non-cartesian fibrations this is exactly the obstruction that forced Giraud to restrict to cartesian stacks. The subsequent identity Sh(η_{f*C})* ≃ Sh(^(fC)*_t)* ∘ Sh(f*(η_C))* is therefore not justified: Sh(f*(η_C))* is only defined if f*(η_C) is a morphism of sites. Without this step the equivalence argument in Prop. 4.12 collapses, and with it Prop. 4.13, Prop. 4.15, and finally Prop. 4.16 are not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves that the pullback (bipullback) of the Giraud topos of an arbitrary fibration along a geometric morphism is the Giraud topos of its inverse image. The main theorem is Proposition 4.16, which states that for a relative topos f:F→E and an E-indexed category C, the square Gir(f*C) → Gir(C), F → E is a bipullback of toposes. The proof introduces a notion of η-extension with base change, compares the canonical stack of a direct or inverse image with the direct or inverse image of the canonical stack, and uses image topologies to establish a Morita equivalence between the inverse image of the canonical stack and the canonical stack of the inverse image. The argument is a long chain that relies heavily on results from the authors' earlier manuscripts [BC23] and [BC25].","tokens_in":31985,"tokens_out":16921,"duration_ms":142072,"significance":"If correct, the paper substantially extends Giraud's and Diaconescu's pullback computations from cartesian stacks to arbitrary fibrations, which is a natural and valuable step in relative topos theory. The structural devices introduced—η-extensions with base change and the comparison functor ν—are original and likely to be reusable. The authors are transparent about deferring proofs to [BC23] and [BC25], but this means the present manuscript is largely a reduction to unpublished companion results. The main theorem is falsifiable and precise: for every E-indexed category C whose fibration is small relative to the base site, the square in Proposition 4.16 must satisfy the universal property of a bipullback. The paper would be a useful contribution once the deferred proofs are supplied or made publicly available.","major_comments":[{"comment":"The main theorem depends essentially on unpublished black boxes: Theorem 3.15 of [BC23] (relative Diaconescu), Theorems 6.3.4 and 6.3.5 of [BC25] (indexed weak Diaconescu), and subsection 6.1 of [BC25] (equivalence between relative and indexed geometric morphisms). These results are used without proof in Propositions 4.9, 4.13, 4.15, and 4.16. For a journal publication, either full proofs must be included or the companion papers must be made available and shown to cover exactly the statements used. As it stands, the central claim of the paper is not verifiable from the manuscript alone.","section":"Section 4.2–4.6"},{"comment":"The proof of Proposition 3.1 is omitted entirely ('completely analogous to Proposition 3.4 [BC23] and left to the reader'), and the equivalence (i)⇔(iii) in Proposition 3.3 is also left to the reader. These propositions establish the adjunction properties of the η-extension with base change and the characterization of indexed weak geometric morphisms. Every later step in Section 4 uses these results, and the base-change case is not literally present in [BC23]. The omission is load-bearing and must be repaired by giving the proofs or by stating the base-change versions explicitly with detailed references.","section":"Section 3, Propositions 3.1 and 3.3"},{"comment":"The proof that f*(η_C) is a morphism of sites is too compressed. The text first establishes continuity and then uses the factorization Sh(η_{f*C})* ≃ Sh(^(f C)*_t)* ∘ Sh(f*(η_C))* to deduce that Sh(^(f C)*_t)* is an equivalence and, finally, that Sh(f*(η_C))* is an equivalence. This reasoning is only valid if the reader accepts that Sh is defined for continuous functors as well as for morphisms of sites, and that an equivalence of inverse images forces the continuous functor to be a dense morphism of sites. The manuscript should state these conventions explicitly; as written, the argument risks a circularity and is not independently checkable.","section":"Proposition 4.12, proof"},{"comment":"The statement of Proposition 4.16 omits the smallness hypothesis that is explicitly assumed in Section 4.2 ('for any fibration ... which is Jcan_E-small'). Without this hypothesis, the Giraud topos Gir(f*C) need not exist as a topos. The proposition should be restated with the smallness condition on C relative to the canonical topology of E.","section":"Proposition 4.16, statement"},{"comment":"The proof of Proposition 4.16 establishes, for an arbitrary relative topos g:G→F, an equivalence T op/E([fg],[Cp]) ≃ T op/F([g],[C_{p'}]). To conclude that the square is a bipullback in the bicategory of toposes, this equivalence must be 2-natural in g. The manuscript does not state or prove naturality. The authors should either verify pseudo-naturality of the equivalence or cite a general principle that makes it automatic from the construction.","section":"Proposition 4.16, bipullback universal property"}],"minor_comments":[{"comment":"The text contains many typographical errors, for example 'toposp' in Definition 2.8, 'categorie' in the introduction, and missing spaces such as 'Theorem2.3' and 'comorphism of sites' in Section 2.2. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"In the displayed description of the Giraud topology, the notation for the cartesian lifting family is ambiguous; it should be clarified that the family ((f_i,1)) is meant to be a family of cartesian arrows over a J-covering family.","section":"Section 2.2, Definition 2.5"},{"comment":"The description of the comparison indexed weak geometric morphism for étale toposes is helpful, but the sentence 'The terminal object of E/f*(F) is the identity morphism on f*(F)' should be rephrased: the terminal object is the pair (f*(F), id), and the unit f*f*(F)→F is not generally an isomorphism. The point is clear but the wording is imprecise.","section":"Section 3.1, Example 3.1"},{"comment":"The sentence 'It is also cartesian, and thus qualifies as a morphism of sites' is too terse. A morphism of fibrations that is cartesian in the fibrewise finite-limit sense is indeed a morphism of sites when both fibrations are cartesian and the codomain topology contains the Giraud topology, but this chain of implications should be spelled out for the reader.","section":"Section 4.5, proof of Proposition 4.13"}],"recommendation":"major_revision","confidential_remarks":"The paper's main theorem is appealing and the overall strategy is coherent, but the extensive reliance on unpublished work [BC23] and [BC25] is a serious obstacle for a journal submission. I would encourage the editor to ask the authors to make the companion manuscripts available or to include the necessary proofs in an appendix. The missing naturality check in Proposition 4.16 and the omitted smallness hypothesis are also important to fix before the result can be considered complete."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe headline is that this is a real contribution: it proves a genuinely new theorem — pullbacks of relative presheaf toposes for arbitrary fibrations, not just cartesian stacks — and the proof machinery (eta-extension with base change, the comparison indexed weak geometric morphism, the image topology trick) is substantial. Giraud's restriction to cartesian stacks was a genuine technical wall, and the paper identifies it clearly.\n\nThe good parts are real. Proposition 4.16 is new relative to Giraud and Pitts. The decision to work with comorphisms of sites to avoid explicit inverse-image computation is smart, and the Morita-equivalence between the inverse image of the canonical stack and the canonical stack of the inverse image is the right tool. The paper also honestly states where stackification is unnecessary.\n\nNow the soft spots. The biggest one is not a fatal flaw but a serious verification burden: the two foundational propositions 3.1 and 3.3 are left to the reader, with proofs 'completely analogous' or 'entirely analogous' to results in the unpublished companion papers [BC23] and [BC25]. Since those papers are not available to the reader (and possibly not to the referee), the main proof is partly a black box. That is a real problem for a paper whose whole value is a new theorem.\n\nOn the stress-test note: I think its specific objection — that Sh(f*(η_C))* is undefined unless f*(η_C) is a morphism of sites — does not land, because continuous functors already induce topos-level adjunctions (Definition 2.9). So the composition identity can be written down. But there is a residual concern: the argument that Sh(^(f C)_t^*)* is an equivalence from surjectivity plus the composite equivalence is categorical and fine, but the diagram chase in Proposition 4.10 is dense, and the notation obscures the direction of units and transposes. A referee will need to check that carefully, and should have access to [BC23] and [BC25].\n\nAlso, the abstract omits the J-smallness hypothesis that appears in Definition 2.6 and is needed for the Giraud topos to exist properly. That is a presentation flaw, not a mathematical one.\n\nOverall, the central idea holds up and the strategy is coherent. But the paper is not independently verifiable in its current form. It deserves a serious referee, with the expectation that the referee will request the companion papers.\n\nWho should read it: specialists in relative topos theory and categorical logic. I'd bring it to a reading group only with a warning that it is heavy. I would cite it if I were working on relative topos pullbacks, but only after checking the companion papers.\n\nRecommendation: accept for peer review, with a report that the authors need to either prove or precisely locate the foundational propositions and clarify the smallness condition.","headline":"A genuinely new generalization of Giraud's pullback theorem to arbitrary fibrations, with a smart proof strategy but a heavy reliance on unpublished companions and several key lemmas left to the reader.","tokens_in":32413,"tokens_out":9557,"would_cite":true,"duration_ms":80201,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["18B25","18D30","18F10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Pulling back a relative presheaf topos along any geometric morphism yields the relative presheaf topos on the inverse-image fibration.","keywords":["relative topos theory","fibrations","indexed categories","relative presheaf topos","Giraud topos","pullback of toposes","bipullback","eta-extension"],"falsifier":"Take a base site with one non-invertible arrow, let $\\mathcal{C}$ be a non-cartesian indexed category over it, and let $f$ be the geometric morphism induced by a morphism of sites. Compute the $\\eta$-extension of the structural map and its right adjoint directly: if the right adjoint does not send cartesian arrows to cartesian arrows, Proposition 3.1(i) fails and the proof of Proposition 4.16 collapses. At the level of the main claim, exhibit any small fibration $\\mathcal{C}$ for which the two Hom-categories in Proposition 4.15 are not equivalent; that would be a concrete counterexample to the bipullback square.","tokens_in":31333,"feed_emoji":"🔄","tokens_out":11175,"duration_ms":90541,"temperature":0.7,"pith_summary":"The paper proves a base-change formula for relative toposes: pulling back a relative presheaf topos along any geometric morphism gives the relative presheaf topos on the inverse-image fibration. Earlier work had established this only for cartesian stacks, where finite limits can be checked fibrewise; the authors remove that restriction for arbitrary fibrations that are small relative to the base site. The payoff is that pullbacks of toposes presented by fibrations can be computed by base change on the indexing category, without explicit stackification or inverse-image calculations. The argument transposes morphisms of sites along the inverse/direct image adjunction and shows that this transposition preserves the property of being a morphism of sites.","feed_headline":"Pullback of a relative topos is the inverse-image Giraud topos","feed_subtitle":"For any fibration, the pullback of its Giraud topos is just the Giraud topos of the inverse-image fibration.","key_machinery":"The load-bearing object is the $\\eta$-extension with base change, a relative analogue of left Kan extension that sends a continuous functor between relative sites over different bases to a functor between their canonical stacks. Its right adjoint is defined through a pullback square, and the adjunction is what transfers morphisms of sites across the inverse/direct image adjunction. The other essential piece is the comparison functor between the inverse image of the canonical stack and the canonical stack of the inverse image, together with the image topology it induces; endowing the inverse image of the canonical stack with that topology makes the two sites Morita-equivalent, and this equivalence is what allows both directions of the transposition to preserve site morphisms.","core_discovery":"Proposition 4.16 states that for a relative topos $f: \\mathcal{F} \\to \\mathcal{E}$ and an $\\mathcal{E}$-indexed category $\\mathcal{C}$ whose fibration is small relative to the base site, the square formed by $\\mathrm{Gir}(f^*\\mathcal{C}) \\to \\mathrm{Gir}(\\mathcal{C})$ and $\\mathcal{F} \\to \\mathcal{E}$ is a bipullback of toposes. In words, the Giraud topos of the inverse-image fibration $f^*\\mathcal{C}$ is the pullback of the Giraud topos of $\\mathcal{C}$ along $f$. The paper reaches this by proving that transposition of morphisms of fibrations along the adjunction $f^* \\dashv f_*$ restricts to an equivalence between categories of morphisms that are also morphisms of relative sites (Proposition 4.15). The key intermediate is a Morita equivalence, produced by an image topology, between the inverse image of the canonical stack and the canonical stack of the inverse image.","pith_inferences":["The transposition mechanism may extend from trivial relative sites to arbitrary relative site topologies, since the obstruction the paper overcomes concerns morphisms of sites rather than the particular choice of topology; a natural next step is to rerun Sections 4.3 through 4.5 with a general relative topology in place of the Giraud topology.","This suggests a general pullback-equals-inverse-image-fibration formula for toposes presented by fibrations with any topology preserved by the structural morphism; checking it against a known classifying-topos example would test whether the mechanism survives outside the presheaf case.","One could use Proposition 4.16 to compute pullbacks of classifying toposes of geometric theories directly from their syntactic fibrations, bypassing translation into internal categories; this would turn the theorem into a practical tool for geometric logic."],"forward_implications":["For every relative presheaf topos presented by a fibration satisfying the smallness condition, pullback along any geometric morphism is the Giraud topos of the inverse-image fibration, so no stackification is needed to compute the pullback.","Giraud's pullback theorem, previously restricted to cartesian stacks, now covers arbitrary fibrations; fibre products in the bicategory of toposes can therefore be obtained from a wider class of site presentations.","Since inverse images along comorphisms of sites are computed by pullback of fibrations, the main theorem turns pullback of toposes into base change on the indexing category, a concrete operation.","The Morita equivalence between the inverse image of the canonical stack and the canonical stack of the inverse image supplies a canonical site-level presentation of the pullback topos that later arguments can use."],"supporting_citations":[{"why":"Introduces the fixed-base η-extension and the finite-limit characterization of relative-site morphisms that the base-change version generalizes.","marker":"[BC23]"},{"why":"Supplies the indexed weak Diaconescu equivalence and the local-fibration results used to transpose morphisms of fibrations and sites.","marker":"[BC25]"},{"why":"Provides the duality between morphisms and comorphisms of sites, dense morphisms, and the image-topology construction used in Section 4.","marker":"[Car20]"},{"why":"Sets up the relative topos framework: fibrations, the Giraud topology, and the canonical relative site.","marker":"[CZ21]"},{"why":"Proves the pullback formula in the cartesian-stack case that this paper extends.","marker":"[Gir72]"},{"why":"Gives the earlier change-of-base adjunction for canonical stacks that the transposition argument parallels.","marker":"[Pit85]"}],"fun_headline_variants":["Pullback topos equals inverse-image Giraud topos","Bipullback: Giraud topos of inverse image","Morita equivalence makes pullback a Giraud topos","Relative topos pullback: inverse-image Giraud","Giraud topos of inverse-image fibration is pullback"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole proof rests on an adjunction property whose proof is omitted in the paper: the $\\eta$-extension with base change must be a left adjoint whose right adjoint preserves the fibred structure. If that property fails, the comparison functors used in Sections 4.3 through 4.5 cannot be constructed, and the bipullback theorem does not go through.","fun_headline_variants_meta":{"raw":{"variants":["Pullback topos equals inverse-image Giraud topos","Bipullback: Giraud topos of inverse image","Morita equivalence makes pullback a Giraud topos","Relative topos pullback: inverse-image Giraud","Giraud topos of inverse-image fibration is pullback"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1571,"prompt_tokens":902,"completion_tokens":669,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":582}},"tokens_in":518,"tokens_out":669,"duration_ms":5838,"temperature":1.0,"reasoning_tokens":582,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:10:51.882807+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a base site with one non-invertible arrow, let $\\mathcal{C}$ be a non-cartesian indexed category over it, and let $f$ be the geometric morphism induced by a morphism of sites. Compute the $\\eta$-extension of the structural map and its right adjoint directly: if the right adjoint does not send cartesian arrows to cartesian arrows, Proposition 3.1(i) fails and the proof of Proposition 4.16 collapses. At the level of the main claim, exhibit any small fibration $\\mathcal{C}$ for which the two Hom-categories in Proposition 4.15 are not equivalent; that would be a concrete counterexample to the bipullback square.","supporting_citations":[],"review_version":2}