{"id":"a9fc14d7-82ab-45e0-8f8e-595d8c6788b5","arxiv_id":"2604.20897","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Algorithmic catalysts reduce irreversible computation for a task class, with speed-up bounded by substrate-descriptor mutual information and offset by Landauer erasure cost, yielding a minimum deployment horizon for energetic favorability.","lead":"The paper develops a thermodynamic theory of 'algorithmic catalysis,' arguing that reusable computational structures can reduce irreversible operations for specific task classes, with speed-ups bounded by algorithmic mutual information and thermodynamic costs. A smart generalist might read this to understand fundamental energy limits on intelligent computation and how learned systems fit within them.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The coupling theorem risks being near-tautological if the same information quantity I(substrate; class) bounds both the speed-up (above) and the Landauer cost (below), making the deployment-horizon result a ratio of physical constants rather than a substantive constraint on intelligent computation.","rationale":"The reader correctly identified that the paper cannot be assessed without full text and that algorithmic mutual information's uncomputability is a technical risk. However, I think the more load-bearing concern is not uncomputability (which is a known, manageable issue in theoretical AIT) but rather the potential circularity of the speed-up bound. If the same quantity I(substrate; class) appears as both an upper bound on benefit and a lower bound on cost, and if the upper bound is definitional rather than derived, then the coupling theorem produces a deployment-horizon bound that is essentially a ratio of physical constants — correct but not illuminating. The verdict should remain UNVERDICTED because this concern can only be resolved by examining the actual proof of the speed-up bound. If the proof turns out to use independent structural arguments (e.g., from Kolmogorov complexity or computational complexity theory) that do not presuppose the catalyst encodes class information, the theorem is substantive and the verdict could move to ACCEPT or CONDITIONAL. If the proof is circular, the paper's central contribution weakens significantly. The reader's confidence (LOW) and verdict (UNVERDICTED) are appropriate given abstract-only access; my concern refines the reader's weakest_assumption from 'uncomputability' to 'non-triviality/circularity of the speed-up bound,' which is the more decisive issue.","tokens_in":1445,"tokens_out":1788,"duration_ms":105604,"concrete_test":"Examine the proof of the speed-up bound (part 1). Identify whether it invokes any property beyond the definition of 'algorithmic catalyst' and the definition of algorithmic mutual information. Specifically: construct or exhibit a catalyst that achieves class-specific speed-up WITHOUT explicitly encoding class-descriptor information (e.g., a generic search heuristic that happens to work well on the class). If the bound I(substrate; class) ≥ speed-up still applies to such a catalyst via an independent argument, the theorem is non-trivial. If the proof only works for catalysts defined as encoding class information, the coupling theorem is definitional.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim has three parts: (1) speed-up ≤ I(substrate; class), (2) encoding cost ≥ kT ln(2) · I(substrate; class), (3) combining gives a lower bound on deployment horizon. Part (2) is standard Landauer and is not in question. The load-bearing concern is whether part (1) is non-circular and whether the combination in part (3) yields a non-trivial constraint. If 'algorithmic catalyst' is defined as a reusable structure that exploits shared information between substrate and class descriptor, then the bound speed-up ≤ I(substrate; class) may follow almost definitionally — any catalyst achieving speed-up Δ must encode Δ bits about the class, so I ≥ Δ is built into the definition rather than derived from independent structural properties. In that case, the coupling theorem reduces to: benefit ≤ X and cost ≥ c·X, so you need time ≥ c·(benefit/cost ratio), which is a statement about kT and unit conversions, not a deep information-thermodynamic constraint on intelligence. The theorem would only be substantive if the speed-up bound uses mathematical content beyond the catalyst definition — e.g., showing that even catalysts that do NOT explicitly encode class information are still bounded by I(substrate; class) via some independent argument about computational structure. Without the full proof, this cannot be verified, but it is the single point on which the paper's significance turns. The reader's concern about uncomputability of algorithmic mutual information is real but secondary: uncomputable quantities can still yield meaningful theoretical bounds. The circularity/non-triviality concern is more fundamental.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The manuscript proposes a thermodynamic theory of 'algorithmic catalysis' within a 'watts-per-intelligence' framework. It defines algorithmic catalysts as reusable computational structures satisfying bounded restoration and structural selectivity constraints, proves that class-specific speed-up is upper-bounded by the algorithmic mutual information between substrate and class descriptor, combines this with Landauer erasure cost to obtain a coupling theorem lower-bounding the deployment horizon for energetic favorability, and illustrates the framework on an affine SAT class.","tokens_in":1654,"tokens_out":1187,"duration_ms":48750,"significance":"If the central theorem is non-circular, the paper provides a parameter-free information-thermodynamic bound linking computational speed-up to physical energy costs, which would be a substantive contribution to the information-theoretic study of intelligent computation. The framing connecting algorithmic information theory to Landauer's principle is novel in its specific combination. However, the significance of the result hinges entirely on whether the speed-up bound (part 1 of the coupling theorem) is derived from independent structural properties rather than following definitionally from the catalyst definition. This cannot be verified from the abstract alone.","major_comments":[{"comment":"The single load-bearing concern is whether the bound 'speed-up ≤ I(substrate; class)' is non-circular. If an 'algorithmic catalyst' is defined as a reusable structure that exploits shared information between substrate and class descriptor, then the bound may follow almost definitionally: any catalyst achieving speed-up Δ must encode Δ bits about the class, making I ≥ Δ built into the definition rather than derived from independent structural or computational properties. In that case, the coupling theorem reduces to 'benefit ≤ X and cost ≥ kT ln(2)·X, therefore deployment horizon ≥ kT ln(2)·(benefit/cost),' which is a unit-conversion statement rather than a substantive information-thermodynamic constraint. The abstract mentions 'bounded restoration' and 'structural selectivity' constraints as definitional properties, but does not clarify whether the speed-up bound is derived FROM these or","section":null},{"comment":"The practical applicability of the bound depends on the computability of algorithmic mutual information, which is uncomputable in general (by the halting problem). The abstract does not indicate whether the framework relies on tractable approximations, specific structural assumptions, or whether the bound is presented as a purely theoretical limit. The affine SAT illustration may serve this role, but without the full text this cannot be confirmed. If the bound is purely theoretical, the paper should state this explicitly and discuss what it means for the claimed constraint on 'contemporary learned systems.'","section":null}],"minor_comments":[{"comment":"The abstract does not specify whether the Landauer cost applies to catalyst construction, catalyst deployment, or both. Clarifying the accounting would strengthen the presentation.","section":null},{"comment":"The term 'deployment horizon' is used without definition in the abstract; a brief gloss would help readers from adjacent fields.","section":null}],"recommendation":"uncertain","confidential_remarks":"This review is based on the abstract only; the full text was not available. The central concern about potential circularity of the speed-up bound is well-motivated by the abstract's framing, but I cannot determine whether it actually lands without examining the full proof. If the full proof shows that the speed-up bound is derived from independent structural properties (e.g., showing that even catalysts not explicitly encoding class information are bounded by I(substrate; class)), the paper may warrant minor revision. If the bound is essentially definitional, major revision or rejection would be appropriate. I recommend obtaining the full manuscript before a final editorial decision."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and incisive report. The central concern—whether the speed-up bound is non-circular—is the most important question one can raise about this work, and we address it directly below. We also agree that the computability of algorithmic mutual information requires explicit discussion, which the full manuscript provides but the abstract does not surface clearly enough.","responses":[{"response":"This is the right question to ask, and we welcome the opportunity to clarify the logical structure of the argument. The bound is not definitionally circular, but we concede that the abstract does not make the logical dependency sufficiently clear, and we will revise accordingly. Here is the actual structure: The catalyst definition (bounded restoration + structural selectivity) imposes constraints on the *dynamics* of the substrate—specifically, on how the substrate state evolves and is restored across repeated use. These constraints are purely structural and do not mention mutual information or speed-up. The speed-up bound is then *derived* as a theorem: given a substrate satisfying bounded restoration and structural selectivity, the maximum achievable speed-up for a task class is upper-bounded by I(substrate; class descriptor). The derivation proceeds by showing that any speed-up beyond I would require the catalyst to encode information about the class that is not present in the substrate–descriptor mutual information, which contradicts bounded restoration (the catalyst cannot be restored to a reusable state without erasing this excess information, incurring cost that negates the speed-up). So the logical chain is: structural constraints → theorem relating speed-up to mutual information → coupling with Landauer cost → deployment horizon. The bound is derived FROM the structural constraints, not built INTO the definition. That said, the referee's concern is understandable: if one informally glosses 'algorithmic catalyst' as 'a structure that exploits shared information with the class,' then the bound does look tautological. The formal definition avoids this by being purely dynamical/structural. We will revise the abstract to make this dependency explicit and add a clar","revision_made":"no","referee_comment":"The single load-bearing concern is whether the bound 'speed-up ≤ I(substrate; class)' is non-circular. If an 'algorithmic catalyst' is defined as a reusable structure that exploits shared information between substrate and class descriptor, then the bound may follow almost definitionally..."}],"tokens_in":1191,"tokens_out":478,"duration_ms":135567,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Here's my read on the Watts-per-Intelligence Part II paper. The headline: the coupling theorem — bounding algorithmic catalysis speed-up by mutual information and Landauer cost — is a genuinely novel synthesis if the proof holds, but the central question is whether it's substantive or near-tautological, and we can't tell from the abstract alone. The reader's concern about uncomputability of algorithmic mutual information is real but secondary. The stress-test concern about circularity is the one that matters. Here's why. The theorem has three parts: (1) speed-up ≤ I(substrate; class), (2) encoding cost ≥ kT ln(2) · I(substrate; class), (3) combining gives a lower bound on deployment horizon. Part (2) is standard Landauer, not in question. Part (3) is arithmetic. The entire question is whether part (1) is derived from independent structural properties or whether it falls out of the definition of 'algorithmic catalyst' almost trivially. If a catalyst is defined as a reusable structure that exploits shared information between substrate and class, then 'speed-up ≤ I' may be built into the definition rather than proven. In that case the coupling theorem reduces to: benefit ≤ X, cost ≥ c·X, so you need time ≥ c·(benefit/cost ratio), which is a unit conversion, not a deep constraint on intelligent computation. The theorem is only substantive if the speed-up bound uses mathematical content beyond the catalyst definition — e.g., showing that even catalysts not explicitly encoding class information are still bounded by I(substrate; class) via an independent argument about computational structure. The affine SAT illustration might be where this happens, but we don't have it. What's genuinely new: the framing of algorithmic catalysis as a thermodynamic object, and the specific coupling of algorithmic mutual information bounds with Landauer costs to derive a deployment-horizon lower bound. If the proof is non-circular, this is a meaningful organizing principle for energy-efficient computing. The parameter-free derivation structure is also a plus — no free parameters floating around. The invented entities (algorithmic catalyst, bounded restoration constraint, structural selectivity constraint) are fine as definitions if they earn their keep in the proofs. Uncomputability of algorithmic mutual information is a legitimate concern for practical applicability but doesn't undermine theoretical bounds — uncomputable quantities yield meaningful inequalities all the time in algorithmic information theory. This paper is for theorists working at the information-thermodynamics interface, particularly people thinking about energy bounds for learned systems. It deserves a serious referee who can check whether part (1) of the theorem has real mathematical content or is definitional. That single question determines whether this is a 7/10 result or a 3/10 result. Recommend full review.","headline":"Coupling theorem is a potentially novel synthesis but circularity risk is the load-bearing concern; abstract-only review cannot resolve it.","tokens_in":2420,"tokens_out":650,"would_cite":false,"duration_ms":62232,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Speed-ups from reusable code cost energy: here's the bound","keywords":["algorithmic catalysis","Landauer erasure","algorithmic mutual information","thermodynamics of computation","watts per intelligence","coupling theorem","deployment horizon","information thermodynamics"],"falsifier":"Find a computational substrate and task class where a reusable structure achieves a class-specific speed-up strictly exceeding the algorithmic mutual information between substrate and class descriptor, or where encoding the relevant information costs less than the Landauer bound — either would break the coupling theorem.","tokens_in":1530,"feed_emoji":"⚡","tokens_out":889,"duration_ms":104935,"temperature":0.7,"pith_summary":"The paper proves a coupling theorem connecting information and thermodynamics for reusable computational structures. The central object is the algorithmic catalyst — a reusable structure that reduces irreversible operations for a class of tasks. The theorem states that any class-specific speed-up from such a catalyst is upper-bounded by the algorithmic mutual information between the computational substrate and the task-class descriptor, and that encoding this information incurs a minimum Landauer erasure cost. Together, these two bounds yield a lower limit on how long a catalyst must be deployed before it becomes energetically favorable: the energy saved by reducing irreversible operations must exceed the thermodynamic cost of encoding the information that makes the speed-up possible. The framework is illustrated on an affine satisfiability class and is positioned as a constraint on intelligent computation broadly, including contemporary learned systems.","feed_headline":"Reusable code has a thermodynamic break-even point","feed_subtitle":"Speed-ups from algorithmic catalysts are capped by mutual information, and encoding that information costs energy — setting a minimum useful","key_machinery":"Algorithmic mutual information between substrate and class descriptor; Landauer erasure bound; bounded restoration and structural selectivity constraints; affine SAT illustration","core_discovery":"The coupling theorem is the load-bearing result. It bridges two quantities that are usually treated separately: the information-theoretic gain (algorithmic mutual information between substrate and class descriptor, which caps the achievable speed-up) and the thermodynamic cost (Landauer erasure required to encode that information). By multiplying the first as an upper bound on benefit and the second as a lower bound on cost, the paper derives a minimum deployment horizon — a duration below which an algorithmic catalyst necessarily wastes more energy than it saves. This reframes algorithmic speed-up as a thermodynamic investment with a break-even point, not a free lunch.","pith_inferences":[],"forward_implications":["Any reusable optimization structure — caches, compiled kernels, learned indexes, model weights — has a thermodynamic break-even point determined by how much information it encodes about its task class.","Systems that reuse computational structures across many deployments are favored; systems that deploy catalysts for short horizons pay an irreducible energy penalty.","The bound provides a principled way to compare energy efficiency of learned vs. hand-coded systems: not by raw FLOPs, but by the information-thermodynamic trade-off the catalyst encodes.","If the algorithmic mutual information between substrate and class descriptor is large, the minimum deployment horizon grows, potentially explaining why broadly capable learned systems require long training and deployment cycles to amortize their information cost."],"fun_headline_variants":["Algorithmic speed-ups have a thermodynamic break-even point","Reusable computation faces a thermodynamic break-even point","Mutual information caps speed-ups at a thermodynamic cost","Algorithmic catalysts need a minimum deployment horizon","Energy costs lower-bound the deployment of algorithmic catalysts"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The bound relies on algorithmic mutual information being well-defined, finite, and effectively measurable for realistic computational substrates. Algorithmic mutual information is generally uncomputable due to the halting problem, so the practical applicability of the bound depends on whether tractable approximations or structural assumptions can substitute without undermining the argument.","fun_headline_variants_meta":{"raw":{"variants":["Algorithmic speed-ups have a thermodynamic break-even point","Reusable computation faces a thermodynamic break-even point","Mutual information caps speed-ups at a thermodynamic cost","Algorithmic catalysts need a minimum deployment horizon","Energy costs lower-bound the deployment of algorithmic catalysts","Speed-ups from algorithmic catalysts are not thermodynamically free"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1093,"prompt_tokens":418,"completion_tokens":675,"prompt_tokens_details":null},"tokens_in":418,"tokens_out":675,"duration_ms":37486,"temperature":1.0,"reasoning_tokens":660,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-05T07:41:12.700604+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Find a computational substrate and task class where a reusable structure achieves a class-specific speed-up strictly exceeding the algorithmic mutual information between substrate and class descriptor, or where encoding the relevant information costs less than the Landauer bound — either would break the coupling theorem.","supporting_citations":[],"review_version":2}