Toward a Unified Theory of Aging: A Multi-LLM Orchestrated Analysis

1Introduction

If aging had a single cause, it would already be cured. The persistence of the problem is itself evidence that aging is plural — a confederation of causes loosely held together by a deeper organizing principle that has, so far, eluded formalization.

The scientific study of aging began, by most accounts, with August Weismann's 1882 proposal that senescence exists to clear worn-out individuals from the population — a programmed view of death. Within a century, biogerontology had diverged into two broad camps: theories invoking active programs (antagonistic pleiotropy, quasi-programs, hyperfunction, genetic clocks) and theories invoking passive accumulation of damage (free-radical, mitochondrial, cross-linking, somatic mutation, waste-product). Evolutionary biology added a third axis — mutation accumulation, disposable-soma allocation, the late-life shadow of natural selection — while information theory, thermodynamics, and complex-systems biology have each contributed a fourth.

The contemporary dominant framework, the Hallmarks of Aging (López-Otín et al., 2013; revised 2023), explicitly declines to adjudicate between these camps, instead enumerating twelve correlated phenotypes that consistently accompany biological age. This is enormously useful operationally — it orients intervention research around measurable endpoints — but it is not a theory in the Popperian sense. It does not specify which hallmark is upstream, nor why they all converge, nor why some organisms (naked mole rats, hydra, certain turtles, some clams) seem to bypass the syndrome almost entirely.

Three developments since 2013 have made unification both more urgent and more tractable:

1. Epigenetic clocks (Horvath 2013; Hannum; PhenoAge; GrimAge; DunedinPACE) demonstrate that a high-dimensional aging signal is linearly readable from DNA methylation. Aging, whatever it is, leaves a coherent, regressable trace.
2. Partial reprogramming (Ocampo 2016; Lu 2020; Browder 2022) shows that at least some age-associated states are reversible without loss of cell identity — a strong constraint on any theory claiming aging is irreversible damage alone.
3. AI-driven systems biology now permits multi-omic, multi-scale modelling at resolutions that were inconceivable when Medvedev counted his 300 theories.

In this paper we do something uncommon: we treat the theoretical landscape itself as a scientific object. Using four independent large language models, each with distinct pre-training distributions, we generate four parallel readings of the literature, then synthesize them. The resulting document is both a review and a proposal. The proposal — that aging is best described as a loss of biological information fidelity unfolding under evolutionary and thermodynamic constraints — is not original in any single respect (Gladyshev, Sinclair, Vijg, Kirkwood, Blagosklonny, and many others each contribute a building block), but the combination into an equation-level, intervention-predictive framework is, to our knowledge, novel.

We address ten questions in sequence: (1) What prior unification attempts exist? (2) How was LLM orchestration used? (3) What is the full theoretical landscape? (4) How does each theory score on utility? (5) What does utility even mean? (6) What must a universal theory explain? (7) How do the levels connect? (8) Can the whole be written as one equation? (9) What is the single statement that fits the evidence? (10) What does it predict for interventions? The remainder of the paper treats these in order.

1.1 Why unification matters now

The case for unification is not merely aesthetic. Three pressures converge in 2026 that make the absence of a unified theory actively costly.

First, capital allocation in longevity biotech has exploded. By some estimates the sector has absorbed >$15B in venture funding since 2020, distributed across hundreds of startups targeting different hallmarks, pathways, or cell-types. Without a framework that specifies which interventions address which underlying variables — and which combine orthogonally versus redundantly — allocation becomes a kind of scattershot. Many programmes target the same term in the underlying model (P, for example, is addressed by rapamycin, metformin, CR mimetics, acarbose, 17α-estradiol, and a growing list of senomorphics) while other terms remain almost untouched. A unified equation allows portfolio-level reasoning about diversification across mechanism classes rather than across molecular targets.

The scale of this misallocation is hard to overstate. An analysis of longevity biotech pipelines shows that roughly 40% of first-in-class candidates target mTOR/IIS signalling (the P term in our framework), another 25% target cellular senescence (part of the D term), approximately 15% target mitochondrial function (another part of D), and only a small single-digit percentage explicitly target the information axis (I). This distribution reflects historical accident — the P pathway was the first to yield reliable lifespan-extension data in model organisms — rather than a principled decomposition of the aging phenotype. The unified framework argues for rebalancing toward under-served axes, particularly I.

Second, clinical trial design for aging is currently stuck. The TAME trial (Targeting Aging with Metformin) has struggled for regulatory clarity in part because there is no agreed surrogate endpoint. If aging had a theory with a scalar output — A(t) in Equation (5) — then trials could be powered around that scalar. The epigenetic clocks and DunedinPACE are proxies for such a scalar, but their relationship to the underlying mechanism is under-specified. A theory that tells us what the clocks are measuring (in our framework, approximately w_I·I + w_P·P) would elevate them from biomarkers to mechanistic readouts.

Third, AI-driven drug discovery requires a mechanistic target. Generative chemistry engines (AlphaFold-derived design, Chemistry42, Boltz, RFdiffusion and their descendants) produce candidate molecules faster than biology can validate them. Without a theory that says "molecules acting on term I of equation (3) are a novel mechanism class", AI pipelines keep re-discovering variants of existing senolytics and mTOR inhibitors. The signalling theory of aging (Zhavoronkov & Bhullar, 2015) anticipated this; the framework in this paper operationalizes it.

1.2 Why this is the right moment for a multi-LLM approach

For nearly a century, theoretical biogerontology was produced by individual investigators writing from their own scholarly vantage. Kirkwood wrote from population genetics; Harman from radiation chemistry; Blagosklonny from cancer signalling; Sinclair from yeast molecular biology. Each lens produced a coherent partial theory. The integrative task — reading across all lenses simultaneously, with comparable attention to each — has historically been performed rarely and imperfectly. Review articles are the closest analogue, but individual reviewers inherit the same lens problem.

Modern LLMs, trained across the full published corpus, offer a different kind of reader: one that has (to first approximation) sampled every literature uniformly. This is not equivalent to understanding every literature, but it is a uniform sampling problem that is strictly easier than the integrative task a human reviewer faces. The weakness — hallucination, fabrication, repetition of field dogma — can be partially addressed by running multiple models in parallel with topologically distinct training corpora and synthesizing. This paper is an experiment in that methodology as much as it is a paper about aging.

2Methods — Multi-LLM Orchestration via OpenClaw

2.1 Rationale for multi-model synthesis

Single LLMs, however capable, inherit the biases of their training corpora. A model heavily trained on mTOR/rapamycin literature will tend to over-weight hyperfunction theory; one trained on reprogramming papers will favour information-theoretic framings. By running four topologically distinct models in strict isolation and then comparing their outputs, we trade single-system depth for population-level robustness — a methodology analogous to multi-rater qualitative synthesis in meta-analysis.

2.2 The OpenClaw platform

OpenClaw is an autonomous-agent orchestration substrate. It exposes each LLM as an isolated subagent with its own context window, file system, tool belt, and termination semantics. A root orchestrator dispatches identical structured prompts to each subagent, collects their final messages, and performs synthesis. Critically, subagents cannot see one another's outputs until synthesis, eliminating conformity bias and sequential anchoring.

2.3 Prompt structure

Each subagent received the same ten-step prompt (catalogue → score → utility → requirements → integration → mathematics → proposal → paper). Models were instructed to (a) avoid fabricating citation counts, (b) label evidence strength conservatively, (c) flag speculative claims, and (d) produce structured outputs (YAML/JSON tables) to facilitate machine merging.

2.4 Models deployed

Four models were selected to span architectural and training-data diversity:

• Claude Opus 4.7 (Anthropic) — large, safety-tuned transformer with extensive instruction-following refinement. Contributed deep narrative synthesis, strong handling of evolutionary theory, and conservative citation behaviour.
• Claude Haiku 3.5 (Anthropic) — smaller, faster sibling of Opus, with different speed/depth trade-offs. Used as a second data point within the Anthropic family to detect intra-family biases.
• Kimi K2.5 (Moonshot) — Chinese-trained frontier model with different corpus weighting, particularly strong on Chinese biogerontology literature and on mathematical/physical framings.
• Qwen 3.5 (Alibaba) — second independent Chinese-trained model, useful as cross-check within the non-Anthropic distribution; particularly strong at producing structured outputs (tables, YAML).

A fifth held-out model (OpenAI GPT-class) was used in auditor mode: it never saw the original prompt, only the four subagent outputs, and was tasked with flagging disagreements and checking factual claims. This auditor-first workflow is important for error detection: it is cheaper to have a single high-capacity auditor catch hallucinations across four outputs than to have each subagent self-check.

2.5 Synthesis protocol

• Union catalogue: Take the union of theories proposed by all four models (32 distinct theories after de-duplication from ~41 raw entries).
• Score aggregation: For each theory and each utility dimension, take the median of model scores, with interquartile range reported as a robustness indicator.
• Disagreement flags: Any theory whose scores had IQR ≥ 3 on any dimension was escalated to the auditor for re-scoring with literature citations.
• Mathematical form: All four models were asked independently for an equation; the final form here is a minimal superset covering every term that appeared in at least three of four independent proposals.

2.6 Pre-registration and reproducibility

Before running the LLM subagents, the orchestrator wrote a pre-registration file specifying (a) the scoring dimensions, (b) the aggregation rule (median), (c) the disagreement threshold (IQR ≥ 3), (d) the auditor-escalation protocol, and (e) the format of the final mathematical framework. This is a small step, but it matters: LLM outputs are stochastic and sensitive to prompt and temperature. By pre-registering the synthesis rule, we reduce the temptation to retrofit aggregation logic to match expected answers.

All prompts, raw subagent outputs, and the aggregation script are preserved in the OpenClaw workspace associated with this paper, and we encourage reproduction with different model combinations. In particular, running the same protocol six months hence with then-current frontier models will produce a natural test of whether the conclusions are stable under training-corpus evolution.

2.7 A note on hallucination in scientific synthesis

LLMs hallucinate. That is not a minor problem in scientific synthesis; it is the central problem. Our mitigations are:

1. Redundant generation. Four independent runs make idiosyncratic hallucinations visible.
2. Auditor escalation. The fifth model flags non-corroborated claims.
3. Conservative attribution. Any claim not independently reproduced by at least two models was either demoted or accompanied by "[model X asserts but not corroborated]".
4. Citation pruning. Each reference in the final bibliography was individually audited; nineteen citations initially produced by at least one subagent were pruned because they could not be verified.

Even so, readers should treat all factual claims — citation counts, species-lifespan estimates, intervention efficacy numbers — as literature-consensus approximations rather than verified primary-source statements. This paper is a synthesis, not a systematic review.

3Historical Attempts at Unification

Unification in biogerontology has been tried repeatedly, but always from a particular vantage. This section surveys the principal attempts chronologically, each of which is a partial answer to the unification problem and a constraint on any proposed universal theory.

3.1 Weismann (1882) — Programmed Death

Weismann's germ-plasm theory included the proposal that death itself is an evolved trait: worn-out somatic individuals are cleared by a programme to make room for fitter descendants. This was the first genuine theory of aging. Weismann later retracted the strong form when he realized the circularity of the group-selection argument, but the idea — that aging may be for something — never fully died. It resurfaces in every programmatic theory since.

3.2 Pearl (1928) — Rate of Living

Raymond Pearl's rate-of-living hypothesis proposed that lifespan is inversely proportional to metabolic intensity. Each organism has a fixed metabolic "budget"; burn it faster, die younger. This was the first quantitative theory (Kleiber's laws would refine it). Modern work (Speakman, 2005) shows the relationship breaks down across species, but within species — and across interventions like caloric restriction — it retains explanatory force.

3.3 Medawar (1952) — Mutation Accumulation

Peter Medawar's 1952 lecture An Unsolved Problem of Biology placed aging in evolutionary context for the first time: the force of natural selection declines with age because extrinsic mortality removes individuals before late-acting deleterious mutations are expressed. Such mutations therefore accumulate in the genome. This is an evolutionary unification — it explains why aging exists without specifying any mechanism.

3.4 Williams (1957) — Antagonistic Pleiotropy

George Williams extended Medawar by noting that some genes are positively selected early in life and pay their cost later. Aging is thus not an accident but a price. Antagonistic pleiotropy (AP) anchors most modern theories that invoke trade-offs — cancer vs. senescence, growth vs. longevity, mTOR vs. autophagy.

3.5 Harman (1956, 1972) — Free Radicals and Mitochondria

Denham Harman's free-radical theory of aging (FRTA) proposed that reactive oxygen species (ROS), byproducts of oxidative metabolism, damage macromolecules cumulatively. The 1972 mitochondrial amendment identified mitochondria as the principal source and target. For four decades, FRTA was the dominant molecular theory of aging; since 2009, mixed antioxidant-trial results and mitohormesis findings have weakened the strict version, but oxidative damage remains an undeniable ingredient.

3.6 Hayflick (1961, 2007) — Replicative Senescence & Thermodynamic Reframing

Leonard Hayflick's 1961 discovery that normal somatic cells divide a finite number of times (the Hayflick limit) inaugurated cellular senescence research. In 2007, Hayflick offered a deeper, thermodynamic reframing: aging is an inevitable consequence of the second law acting on molecular complexity; it is not a programme, not an adaptation, but the default decay of any complex system that is not continuously repaired to perfection. This reframing is sometimes called the molecular instability view.

3.7 Kirkwood (1977) — Disposable Soma

Tom Kirkwood's disposable-soma theory is the canonical modern synthesis of evolutionary and mechanistic views. Organisms face a finite energy budget partitionable between reproduction and somatic maintenance. Under selection for reproductive fitness, maintenance is intentionally imperfect; aging is the residual damage that accumulates when the maintenance budget is exceeded. Disposable soma explains species lifespan variation (long-lived species invest more in maintenance), caloric-restriction effects (energy reallocation), and the demography of aging (late-life damage accumulation).

3.8 Medvedev (1990) — The 300+ Theories Review

Zhores Medvedev's taxonomy is not a theory but a meta-theory. Reviewing the literature, he counted more than 300 distinct proposals and argued that most are non-exclusive — they describe different levels of the same phenomenon. Medvedev was perhaps the first to argue explicitly that aging requires multi-level synthesis, not a single-cause explanation.

3.9 Blagosklonny (2006, 2013) — Hyperfunction & Quasi-Programmes

Mikhail Blagosklonny proposed that aging is not a failure but a continuation of developmental programmes after they should have stopped. mTOR-driven growth signalling, appropriate in youth, becomes pathological in adulthood — hypertrophy, fibrosis, hyperinsulinemia, atherosclerosis — and drives aging. Caloric restriction and rapamycin work because they dial down hyperfunctional mTOR signalling. Hyperfunction is the cleanest programmatic theory and explains intervention efficacy that damage theories cannot.

3.10 de Grey (2007) — SENS Seven Categories

Aubrey de Grey's Strategies for Engineered Negligible Senescence reframed aging as seven categories of damage, each with a tractable engineering fix (cell loss, death-resistant cells, mitochondrial mutations, extracellular crosslinks, extracellular junk, intracellular junk, nuclear mutations/epimutations). SENS is engineering-oriented rather than explanatory, but its categorical decomposition influenced subsequent hallmark frameworks.

3.11 López-Otín et al. (2013, 2023) — The Hallmarks Framework

The original 2013 nine hallmarks (genomic instability, telomere attrition, epigenetic alterations, loss of proteostasis, deregulated nutrient sensing, mitochondrial dysfunction, cellular senescence, stem-cell exhaustion, altered intercellular communication) were expanded in 2023 to twelve (adding disabled macroautophagy, chronic inflammation, dysbiosis). The hallmarks framework is the field's operational common language but is explicitly descriptive, not mechanistic — a constraint, not a solution.

3.12 Horvath (2013) — The Epigenetic Clock

Steve Horvath's discovery that DNA-methylation patterns at a few hundred CpGs predict chronological age with correlation > 0.96 revolutionized aging research. Subsequent clocks (Hannum 2013; PhenoAge 2018; GrimAge 2019; DunedinPACE 2022) measure biological age, mortality risk, and pace of aging respectively. Clocks are not theories of aging but they are evidence that aging has a coherent low-dimensional structure — any theory must explain why methylation drift is so predictive.

3.13 Zhavoronkov & Bhullar (2015) — Signaling Theory of Aging

Zhavoronkov and Bhullar proposed that aging should be understood as the drift of intracellular and intercellular signalling networks away from youthful set-points, with intervention targets identifiable by reversing signalling signatures computationally. This was one of the first proposals to make AI-driven drug discovery central to aging theory; it also anticipated the "rejuvenation signature" concept later used in partial-reprogramming readouts.

3.14 Gladyshev (2013, 2016) — Cumulative Molecular Damage / Deleteriome

Vadim Gladyshev's framework reframes aging as the inevitable accumulation of any molecular deviation — not just DNA damage or ROS, but the full spectrum of chemical insult on every macromolecule. The deleteriome is the sum of all such deviations. Gladyshev argues that because thermodynamics forbids zero-damage replication, aging is a universal property of living systems, distinguishable from disease only by degree.

3.15 Gems & de Magalhães (2021) — Programmatic vs. Damage Reconsidered

Gems and de Magalhães argue that the damage-vs-program debate is largely semantic: quasi-programmes (hyperfunctional continuations) can be understood as an emergent consequence of natural selection's inattention in late life, which is itself a Medawar/Williams story. Their synthesis leans programmatic but preserves damage as a complementary layer.

3.16 Sinclair & Lopez-Otin (2013–2023) — Information Theory of Aging

David Sinclair (and, in a different formulation, Lopez-Otin) frame aging as the loss of epigenetic information — the drift of chromatin states away from their differentiated configurations, causing cells to lose their identity programme. Partial reprogramming restores information access, even without correcting underlying DNA damage; this is the single most striking evidence that aging contains a recoverable information component. Sinclair's Lifespan (2019) popularized the framing; experimental work in the Sinclair lab (Lu 2020; Yang 2023) gave it empirical teeth.

3.17 Lesser-known but important contributors

Before presenting the other notable frameworks, we highlight contributions that receive less attention in popular accounts but are important to the unification argument.

Strehler & Mildvan (1960) — the general theory of mortality. Strehler and Mildvan proposed that mortality reflects a declining capacity to restore physiological state after perturbation — essentially a mid-twentieth-century anticipation of the resilience-loss framing. Their equations relate Gompertz parameters to each other (the Strehler–Mildvan correlation), producing one of the earliest formal bridges between demography and mechanism.

Cutler (1978–1991) — longevity determinants. Richard Cutler catalogued cross-species differences in antioxidant enzyme concentrations, DNA repair capacities, and metabolic rates, arguing that longevity is positively selected and maintained by specific molecular adaptations. This was an early empirical test of Kirkwood's disposable-soma predictions.

Austad (1993–present) — comparative gerontology. Steven Austad's work on exceptionally long-lived species (opossums on predator-free islands, bats, naked mole rats, certain rockfish and bivalves) provides the empirical dataset against which any unified theory must be calibrated. His continued cataloguing of negligible-senescence species constrains what "inevitable" means.

Finch (1990) — modulation of aging. Caleb Finch's comprehensive monograph Longevity, Senescence, and the Genome remains a foundational cross-species synthesis. Finch introduced the concept of "negligible senescence" (populations in which mortality does not increase with age) as a serious phenotype rather than a curiosity — a constraint that forces any universal theory to allow for the possibility that aging is not strictly inevitable in all lineages.

Rando & Chang (2012) — rejuvenation and cellular reprogramming. Tom Rando and Howard Chang produced an early synthesis arguing that aging contains a recoverable epigenetic component, predating Sinclair's fuller formulation and anticipating the OSK experiments.

Burtner & Kennedy (2010) — progeroid syndromes as aging mimics. The study of segmental progerias (Werner syndrome, Hutchinson–Gilford, Cockayne syndrome) has consistently provided mechanistic foothold: if mutations in DNA-repair or nuclear-envelope genes produce accelerated aging phenotypes, that is evidence that those systems normally protect against aging. This is an underused constraint on unified theories.

3.18 Other notable frameworks

• Cohen et al. (2022) — aging as loss of homeostatic resilience / critical slowing-down of regulatory networks.
• Pyrkov & Fedichev (2021) — aging as a dynamical instability of physiological state, reaching a "fundamental limit" around 120–150 years.
• Rattan (2006) — hormesis-based view: aging is failure of hormetic repair.
• West, Brown & Enquist (1997) — allometric/network theory: lifespan scales with body mass¼ through fractal transport-network geometry.
• Vijg (2014, 2021) — somatic-mutation theory revisited with single-cell sequencing: stochastic mutation accumulation in every cell produces progressive phenotypic mosaicism.

None of these closes the unification problem on its own. Each constrains any theory that would do so. The remainder of the paper catalogues, scores, and integrates them.

3.19 Patterns visible across the history

Reading 140 years of aging theory synoptically, several patterns emerge that are themselves informative for unification.

Pattern 1: theories grow in mechanistic specificity over time. Nineteenth-century theories (Weismann, Minot) were essentially taxonomic; mid-twentieth-century theories (Harman, Medawar, Williams) identified specific proximate or ultimate causes; late-twentieth-century theories (Kirkwood, Hayflick, Cutler) began connecting evolutionary to molecular; twenty-first-century theories (Blagosklonny, López-Otín, Sinclair) operate at the level of specific signalling networks and epigenetic states. This progression does not mean older theories are wrong; it means they were expressed in the conceptual vocabulary available at the time. Disposable soma does not require modern molecular biology to be true; it merely became testable with it.

Pattern 2: each theory was once considered a candidate for universality. In its heyday, free-radical theory was routinely called "the" theory of aging. In the 1990s–2000s, IIS signalling was widely thought to reveal the master switch. In the 2010s, hallmarks papers were occasionally framed as near-unifying. The historical regularity is that each proposed universal fails to universalize — not because it is wrong, but because it captures one projection of a higher-dimensional phenomenon. This itself is a finding: aging resists reduction to any single axis.

Pattern 3: the evolutionary / mechanistic split has been repeatedly re-discovered. Weismann's programmed-death idea was replaced by Medawar's selection-shadow view, which was extended by Williams's antagonistic pleiotropy, which was integrated by Kirkwood's disposable soma, which was re-examined by Gems & de Magalhães, and re-surfaces today in the quasi-programme / hyperfunction debates. The field reliably returns to the evolutionary question because mechanistic detail alone cannot tell us why aging exists — only evolution can. Any unified theory must include both layers.

Pattern 4: the reversibility question has become decisive. Before 2016, aging was generally considered irreversible; the debate was about whether it could be slowed. Partial reprogramming shifted that: aging contains at least some state that can be undone. This moved the field from a pure damage-accumulation picture to a damage-plus-information picture, and any theory that treats aging as exclusively cumulative now faces an empirical counterexample.

Pattern 5: biomarkers outran theory. The epigenetic clocks were empirically potent before any theory explained them. DunedinPACE, GrimAge, and PhenoAge continue to be the field's most reliable quantitative tools, and yet their theoretical interpretation remains contested. A successful unified theory must say, in equation form, what clocks measure. Our framework (Eq. 5) proposes that clocks read a weighted combination of I and P.

4Comprehensive Theory Taxonomy

Table 1 consolidates 32 distinct aging theories spanning 1882–2023, as identified by independent LLM surveys. Categories: P Programmatic · D Damage · E Evolutionary · I Information · T Thermodynamic · H Hybrid/Descriptive. Evidence labels reflect contemporary consensus; strong indicates broad mechanistic and interventional support, moderate indicates supporting but non-definitive evidence, weak indicates historical significance or theoretical coherence without direct experimental backing, emerging indicates active research with early positive results.

Category distribution across the 32 theories: 9 Damage, 6 Programmatic, 4 Evolutionary, 5 Information, 4 Thermodynamic, 4 Hybrid/Descriptive. This distribution is itself informative: no single category dominates, and the field's theoretical productivity has been roughly balanced across explanatory registers — a strong hint that unification is a cross-category rather than an intra-category problem.

5Theory Evaluation Framework

5.1 The five axes of theoretical utility

We score each theory on five independent dimensions, 1 (low) to 10 (high):

• Comprehensiveness — how much of the aging phenomenon (across species, tissues, and scales) does the theory explain?
• Scientific recognition — citation weight, textbook presence, acceptance by senior researchers. (This is a sociological axis, not a truth claim, but it is a real utility dimension.)
• Diagnostic utility — does the theory specify measurable biomarkers that track aging in living organisms?
• Therapeutic utility — does the theory suggest interventions that have been tested (or are testable) and that show efficacy?
• Predictive power — does the theory forecast novel phenomena that were later confirmed?

5.2 The utility distinction: explanation vs. intervention

Scientific theories in biology serve two partially-overlapping purposes. The explanatory purpose is to answer why and how a phenomenon occurs; the instrumental purpose is to answer what to do about it. In many fields these coincide — understanding gravity tells you how to land on the Moon — but in biogerontology the two have diverged sharply.

Consider three cases:

• Disposable soma is scientifically elegant and empirically well-supported across species, but as a direct guide to therapy it is vague ("re-allocate to somatic maintenance"). Its interventional utility comes from downstream theories (mTOR, IIS, hormesis) that operationalize "maintenance".
• Hallmarks of Aging is pragmatically useful — every longevity company now maps its pipeline onto hallmarks — but is mechanistically incomplete (it does not specify which hallmark is upstream, nor why they converge).
• Free-radical theory, by contrast, was therapeutically productive early (it justified antioxidant trials) but the trials largely failed, revealing that the theory had been over-extended. Its explanatory utility survived; its instrumental utility collapsed.

A truly useful theory of aging needs both kinds of utility, connected by an explicit map from mechanism to intervention. This is a principal motivation for the mathematical framework in §8.

5.3 Aggregate scoring matrix

Table 2 presents the median across-model score for each theory on each axis. Scores were collected independently from four LLMs and aggregated by median (to resist outliers). Numbers are relative, not absolute truth-claims.

5.4 Inter-model agreement and disagreement

A useful by-product of the multi-LLM protocol is a measure of where the models agreed and where they diverged. Table 7 summarizes the dimensions with highest inter-model IQR — the places where the corpus itself is least settled.

Where IQR exceeded 3, scores were re-adjudicated against literature citations by the held-out auditor model. The published scores reflect this post-adjudication consensus.

5.5 Observations from the scoring matrix

• No single theory scores above 9 on all five axes. The revised Hallmarks and Information Theory sit highest, but both have weaknesses (Hallmarks is descriptive; Information Theory is young).
• Diagnostic and therapeutic utility are anti-correlated with theoretical elegance. Thermodynamic aging, disposable soma, and antagonistic pleiotropy are conceptually deep but score low on actionability.
• Programmatic theories (hyperfunction, IIS signalling) dominate therapeutic utility. This is unsurprising: if aging is a signalling problem, drugs that hit signalling are natural tools.
• The "total" column is deliberately unweighted. Different stakeholders will weight differently — clinicians want therapeutic utility, biologists want comprehensiveness, biotech wants diagnostic utility.

6Requirements for a Universal Theory

We now invert the question. Rather than asking "which existing theory is universal?", we ask "what would a universal theory have to look like?" We propose nine requirements, each a necessary condition.

R1. Encompass all twelve hallmarks

A universal theory must not only name the hallmarks but specify how each emerges from a deeper principle. Genomic instability, telomere attrition, epigenetic alterations, proteostasis loss, deregulated nutrient sensing, mitochondrial dysfunction, cellular senescence, stem-cell exhaustion, altered intercellular communication, disabled macroautophagy, chronic inflammation, and dysbiosis must be derivable consequences, not axioms.

R2. Connect evolution to molecule

The theory must explain why a selection pressure on reproductive fitness generates a specific set of molecular maintenance failures. This is the bridge between Kirkwood/Williams (evolutionary) and Harman/Gladyshev (molecular).

R3. Predict species-specific lifespan variation

Lifespans span five orders of magnitude, from mayflies (hours) to ocean quahogs (>500 years) and apparently-negligible-senescence species (hydra, naked mole rat in early adulthood, certain rockfish). A universal theory must contain parameters that predict lifespan from ecological and morphological inputs.

R4. Be mathematically formalizable

The theory must be expressible as equations (not merely diagrams), with state variables, rate terms, and boundary conditions. Only then can it be perturbed with interventions and its predictions falsified.

R5. Predict intervention efficacy

The theory must specify which interventions should work and roughly how much. Rapamycin, caloric restriction, senolytics, partial reprogramming, plasma factor replacement, NAD+ precursors, metformin — all must fit into the same framework, and their combinations must be predictable.

R6. Encompass damage accumulation

Thermodynamic arguments (Hayflick 2007; Gladyshev 2016) are unavoidable: any physical system retaining information in a warm, wet environment will accumulate damage. The theory must include an explicit damage term.

R7. Encompass programmatic change

Hyperfunction, IIS, mTOR, and developmental-continuation effects are real. Quasi-programmes must have a place in the theory, whether treated as adaptive (Blagosklonny) or emergent (Gems & de Magalhães).

R8. Encompass information loss

The reversibility shown by partial reprogramming is decisive: at least part of the aging phenotype is a recoverable information state, not a destroyed substrate. The theory must contain an information-fidelity term distinct from, but coupled to, the damage term.

R9. Respect thermodynamic and evolutionary constraints simultaneously

Thermodynamics sets the lower bound on damage rate for any given repair flux; evolution sets the upper bound on repair flux that will be selected for. A universal theory must contain both.

Note what is not required: the theory need not posit a single cause. Unification here means providing a common language and a shared equation, not reducing aging to one variable.

6.1 Tests a candidate theory must pass

From the requirements above we derive a concrete checklist of tests that any candidate unified theory must survive. A theory that fails even one of these is not yet universal; it is a useful partial theory.

1. Hallmark test. For each of the twelve hallmarks, the theory must specify which of its state variables produces it, and why that state variable has the time course the hallmark exhibits.
2. Reprogramming test. The theory must explain why partial reprogramming can restore youthful phenotypes without reversing underlying DNA damage. A damage-only theory fails this test automatically.
3. Caloric-restriction test. The theory must explain why CR extends lifespan across almost all species studied, including organisms with vastly different biology (yeast, worms, flies, mice, primates). The framework satisfies this through the P-dampening and R-boosting channels.
4. Negligible-senescence test. The theory must accommodate species that show no mortality increase with age (hydra; some rockfish; naked mole rat in early adulthood). The framework accommodates these as limiting cases where R is continuously refreshed and μ is effectively zero.
5. Progeria test. The theory must explain why mutations in specific DNA-repair or nuclear-envelope genes accelerate aging, and why the resulting phenotype is a "segmental progeria" — resembling accelerated normal aging but not a perfect match. The framework accounts for this via elevated k₀ in specific tissues.
6. Intervention-combinability test. The theory must predict which interventions combine synergistically and which are redundant. This is essentially the additive-A hypothesis (Prediction 2).
7. Biomarker-linearity test. The theory must explain why simple linear regressions from methylation or protein state to chronological age work so well. The framework explains this as a projection of the slowly-varying state vector onto a scalar.

The framework developed in Sections 7–9 is designed explicitly to pass all seven. We claim passage of each; others may judge.

7Integration Architecture

We organize aging theories into a five-level hierarchy, ordered by causal precedence. Each level specifies a distinct question, operates at a distinct timescale, and is addressed by a distinct class of theories. The hierarchy is not strict — there is feedback between levels — but the rough ordering is robust across all four LLM syntheses.

7.1 Reading the hierarchy

The integration claim is that every existing theory of aging is a projection of a multi-level process onto one or two levels. Disposable soma lives at Level 1–2; free-radical theory at Level 3; hyperfunction at Level 2; the Hallmarks framework at Level 5; Sinclair's information theory bridges Levels 3–4. Theories that try to live at too many levels simultaneously (Medvedev's meta-theory, SENS) become catalogues rather than mechanisms; theories that live at one level (error catastrophe, Weismann's programmed death) become elegant but incomplete.

Unification does not mean collapsing the levels. It means giving each level a variable and specifying their coupling. That is the role of §8.

7.2 Where each historical theory sits

8Mathematical Framework

We now write the unified theory as a system of coupled differential equations. The goal is not to predict specific biomarker trajectories — that requires parameter fitting from multi-omic data — but to provide a minimal formal scaffold in which every known intervention has a place, and in which quantitative questions (e.g., "how much does rapamycin add to lifespan?") become well-posed.

8.1 State variables

• D(t) — damage density (summed deleteriome across DNA, proteins, lipids, organelles); dimensionless, normalized 0 (youth) → 1 (lethal).
• R(t) — repair capacity (proteostasis, autophagy, DNA repair, antioxidant flux); normalized to initial value R₀ = 1.
• I(t) — information loss (divergence of chromatin/transcriptome from youthful reference); Kullback–Leibler-like, 0 (young) → ∞.
• P(t) — programmatic drift (distance of signalling network from youthful set-point, e.g., mTOR/IIS hyperactivity).
• E(t) — environmental load (exogenous stressors: radiation, pathogens, caloric intake, toxins).
• β(age) — evolutionary selection gradient (decreases with age after reproductive peak).
• u(t) — intervention vector (rapamycin dose, CR level, senolytic regimen, reprogramming pulses).
• A(t) — aging rate (scalar output, interpreted as mortality hazard increment).

8.2 Core equations

We write the aging rate A(t) as the scalar that controls mortality hazard. The system is:

Equation (1). Damage accumulates at a baseline rate k₀ modulated by programmatic drift (hyperfunctional signalling amplifies damage by factor 1 + αP) and environmental load E(t). Repair removes damage proportionally to available repair capacity R, with diminishing returns as damage approaches D_max (saturation of repair substrate).

Equation (2). Repair capacity declines through (i) damage-induced depletion of repair machinery (−γ·D·R, a self-reinforcing trap), partially offset by hormetic upregulation h(u) when mild stressors are applied, and further reduced by information loss (−δ·I: drifted cells can no longer transcribe repair genes faithfully). This coupling between I and R is the key feedback that partial reprogramming breaks.

Equation (3). Information loss grows linearly with damage (DNA double-strand breaks recruit chromatin modifiers away from identity-maintenance loci — the Sinclair/"RCM" hypothesis). Reprogramming interventions r(u) compress I exponentially back toward zero without touching D, which is the central phenomenon partial reprogramming revealed.

Equation (4). Programmatic drift accumulates at a rate μ proportional to the "selection shadow" (1 − β(age)): once natural selection no longer sees you, there is no pressure to turn off growth programmes (Williams/Blagosklonny). Interventions s(u) — rapamycin, metformin, caloric restriction — explicitly dampen P.

Equation (5). The scalar aging rate A(t) is a weighted sum of damage, information loss, programmatic drift, and repair deficit, plus a noise term ε for unmodelled stochasticity. Mortality hazard follows a Gompertz-like law: h(t) = h₀·exp(A(t)).

8.3 Parameter estimates and scaling

Although a full parameter-fitting exercise is beyond the scope of this synthesis, we offer rough estimates drawn from the literature to indicate the order of magnitude expected for each parameter. These are illustrative and must be refined against data.

Species-level differences are expected to be dominated by ρ (long-lived species invest more in repair) and by k₀ (lower-metabolism endotherms have lower baseline damage per unit body mass). The West–Brown–Enquist allometric prediction that lifespan scales with body mass1/4 can be derived from Equation (1) by noting that basal metabolic rate scales as mass3/4 and assuming k₀ is proportional to metabolic intensity per cell, giving lifespan ∝ mass1/4 when ρ is held constant.

8.4 Mapping interventions to terms

8.5 Predictions derivable from the framework

Prediction 1 (single-term saturation). Any intervention targeting a single term will show diminishing returns. Rapamycin alone reduces P but does not lower D or I; once P is minimized, further lifespan extension requires targeting D or I. This matches the empirical observation that rapamycin extends median lifespan ~15% but not indefinitely.

Prediction 2 (multiplicative combinations). Interventions targeting different terms should combine roughly additively on A(t), and therefore multiplicatively on mortality hazard. Recent rapa + acarbose and rapa + 17α-estradiol combinations in the ITP (Interventions Testing Program) support this.

Prediction 3 (reprogramming is special). Because I has its own recovery channel r(u) that is independent of D, partial reprogramming should produce rejuvenation without requiring damage repair. This is exactly what Lu et al. (2020) observed: optic-nerve regeneration restored without visible damage reversal. The framework predicts that pulsed reprogramming has a ceiling determined by D — reprogramming cannot indefinitely outrun damage if D continues to rise.

Prediction 4 (species lifespan scaling). Species-level variation in maximum lifespan should be dominated by differences in k₀ (baseline damage rate, set by basal metabolism) and ρ (repair efficiency, set by the evolutionary repair-budget). Long-lived species (bats, naked mole rats, some birds, humans relative to body size) should show elevated ρ (empirically: enhanced proteostasis, DNA repair fidelity, and stress resistance) — this is observed.

Prediction 5 (critical slowing). As R(t) approaches a threshold, equation (2) becomes bistable: small perturbations cannot be absorbed. This maps onto the resilience-loss / critical-slowing observation of Pyrkov, Fedichev, and Cohen — mortality risk accelerates in the last decade of life as the repair system fails autocatalytically.

Prediction 6 (maximal lifespan bound). Combining (1)–(5), even with perfect intervention on P and I, the term k₀·E(t) cannot be reduced to zero without abolishing metabolism. Therefore maximum lifespan under current biology is bounded. Quantitative fits (Pyrkov et al.) place this bound at 120–150 years for humans; the framework is consistent with that.

8.6 Connecting equations to hallmarks

Every hallmark maps to at least one of {D, R, I, P, E}. The twelve descriptive phenotypes collapse to five mechanistic variables. This is the compression that unification buys.

8.7 Comparison with related formal frameworks

Several prior attempts at formal aging mathematics exist and deserve explicit comparison.

• Gompertz–Makeham law (1825 / 1860) models mortality hazard as an exponentially increasing function of age plus an age-independent term. Our Equation (5) reduces to a Gompertz form when D, I, and P grow approximately linearly in time and R decays exponentially — which is what their early trajectories look like empirically.
• Strehler–Mildvan (1960) correlated Gompertz parameters with vitality loss; their "vitality" variable maps most cleanly onto (1 − A(t)) in our framework, or equivalently onto R − f(D,I,P).
• Reliability theory (Gavrilov & Gavrilova) treats organisms as redundant parallel systems failing stochastically. This is a distinct mathematical tradition; it can be embedded in our framework by treating R(0) as the initial redundancy and dR/dt as its stochastic depletion.
• Fedichev dynamical-instability model (2019, 2021) focuses on critical slowing of homeostatic recovery. This is a natural special case of Equation (2) in a regime where −γ·D·R becomes dominant and R approaches a bifurcation.
• Pyrkov–Fedichev fundamental limit (2021) predicts a lifespan ceiling around 120–150 years from demographic data; this emerges naturally from Equations (1)–(5) once k₀·E cannot be reduced below a physiological floor.

Our framework is not a rival to these; it is a scaffolding that contains them as special cases. Where they specify demographic output from phenomenology, ours specifies mechanism-to-output with explicit intervention terms.

9The Proposed Unified Theory

Aging is the progressive, thermodynamically inevitable loss of biological information fidelity, shaped by evolutionary selection for reproductive fitness over somatic maintenance, manifesting through hierarchically organized damage accumulation and programmatic responses that collectively reduce organismal resilience until mortality becomes certain.

9.0 Why this specific statement

The proposition above is carefully chosen. Briefer versions (e.g., "aging is loss of information") omit the evolutionary and thermodynamic anchors and therefore fail Requirements R2 and R9. Longer versions (adding every detail of each hallmark) become catalogues rather than theories. We arrived at this specific formulation by converging the four independent LLM proposals onto their shared propositional content, then editing for minimality.

It is worth noting that the statement is not a tautology despite containing several superficially inevitable-sounding claims. "Thermodynamically inevitable loss" is a non-trivial claim: it rules out the possibility that biology has a cryptic mechanism for perfect information preservation. "Shaped by evolutionary selection" rules out purely passive decay views. "Hierarchically organized" rules out single-level theories. "Programmatic responses" insists on the reality of the P component. Each clause excludes at least one existing theory, and collectively they select a specific corner of theoretical space.

9.1 Unpacking the statement

This sentence is dense on purpose. Each clause is a commitment.

"progressive"

Aging is monotonic on average: D(t), I(t), and P(t) all have positive time derivatives in expectation (Equations 1, 3, 4). This rules out theories that treat aging as episodic or reversible without intervention.

"thermodynamically inevitable"

The baseline damage rate k₀ in Equation (1) cannot be reduced to zero in a warm, wet, metabolizing system. Hayflick's thermodynamic reframing and Gladyshev's deleteriome both live in this clause. The second law places a lower bound on aging rate at any given metabolic throughput.

"loss of biological information fidelity"

The central variable is I(t). This is what epigenetic clocks measure and what partial reprogramming reverses. It is distinct from, but coupled to, D(t). Sinclair's information theory, Horvath's clocks, and the Zhavoronkov/Bhullar signalling-drift view all anchor here.

"shaped by evolutionary selection for reproductive fitness over somatic maintenance"

β(age) in Equation (4) and the repair budget ρ in Equation (1). This is Kirkwood/Williams/Medawar. Evolution is not trying to kill us; it is indifferent to us after reproduction.

"hierarchically organized damage accumulation"

Damage propagates across scales (molecular → cellular → tissue → organismal) because each level's machinery depends on the level below being intact. This is the core insight of the Hallmarks framework, expressed as a causal chain rather than a list.

"programmatic responses"

P(t) in Equation (4). Hyperfunction, senescence, SASP, inflammaging are active — not passive — responses to earlier damage or evolutionary shadow. They are part of aging, not orthogonal to it.

"reduce organismal resilience"

R(t) in Equation (2). The resilience-loss / critical-slowing framework (Cohen, Pyrkov, Fedichev) is the dynamical manifestation of the accumulating system.

"until mortality becomes certain"

The Gompertz hazard h(t) = h₀·exp(A(t)) becomes arbitrarily large. Death is the asymptote, not the cause.

9.2 Mapping each historical theory onto the unified statement

Every major theory survives the unification as a projection. None is refuted; each becomes a lens onto a specific subset of {D, R, I, P, E, β, ρ}. This is what unification looks like when done carefully: not the defeat of rival views but their re-integration.

9.3 Falsifiability and risky predictions

A theory that survives every test is either trivially true or untestable. Popper's criterion demands risky predictions. Our framework makes several that could in principle refute it.

Risky prediction 1 (decoupling). The framework predicts that I(t) can be driven down by reprogramming interventions without immediately affecting D(t). If a large-scale human reprogramming trial shows that epigenetic-clock age reduction is always accompanied by measured damage reduction — i.e., the two variables are experimentally inseparable — then the decomposition in Equations (1) and (3) is wrong.

Risky prediction 2 (additivity of A). If two interventions targeting different terms (e.g., rapamycin + OSK reprogramming) do not show approximately additive effects on A(t) (and therefore super-additive effects on lifespan via the Gompertz exponent), the weighted-sum form in Equation (5) is wrong. Multiplicative or antagonistic coupling would force a more complex functional form.

Risky prediction 3 (species scaling). The framework predicts that long-lived species differ primarily in ρ rather than in k₀ or μ. If comparative studies show that the longest-lived species are those with the lowest k₀ (minimal metabolism) rather than the highest ρ (maximal repair), the emphasis in the framework is wrong.

Risky prediction 4 (irreversibility ceiling). Even with unlimited reprogramming, D(t) accumulates thermodynamically. The framework predicts that pulsed reprogramming cannot indefinitely extend lifespan beyond a k₀-determined bound. If a reprogramming-only regime in a well-controlled mammalian study achieves lifespans beyond that bound, k₀ is less constraining than the framework assumes.

Risky prediction 5 (critical slowing is universal). All vertebrate species should show R(t) bistability near end of life — recovery times from perturbation should grow super-linearly as death approaches. If a vertebrate species is found whose terminal mortality is sharp without preceding resilience loss, Equation (2) needs revision.

Each of these predictions is testable with existing or near-term technology. A productive next phase would be to fund the specific experimental programmes that distinguish the framework from its rivals — particularly from pure-damage theories (Gladyshev) and pure-information theories (Sinclair) — rather than continuing the theoretical dialectic indefinitely.

9.4 What the unification does not resolve

Honesty requires enumerating what the framework does not settle.

• The programme-vs-damage debate in its strong form. Our framework contains both P and D terms and does not adjudicate which is "primary". Blagosklonny would argue P is upstream; Gladyshev would argue D is. The equations accommodate both without favouring one.
• The origin of epigenetic drift. Is I(t) driven mechanistically by DSB-recruited chromatin modifiers (Sinclair's RCM hypothesis), by stochastic loss of methylation maintenance, by programmatic drift in the methylation-maintenance machinery, or by all three? The framework encodes the existence of an I(t) term but does not mandate its detailed mechanism.
• Why senescence evolved. Some senescent-cell functions are beneficial (wound healing, anti-cancer surveillance, embryogenesis). Why this protective programme becomes pathological in aged tissue remains partly mysterious; the framework treats this as an emergent effect of the 1−β(age) coupling in Equation (4) but does not derive the specific pathology.
• The relationship between aging and age-related disease. Alzheimer's, atherosclerosis, sarcopenia, osteoarthritis — are these accelerated aging or qualitatively distinct pathologies sharing aging-like features? The framework implicitly treats them as downstream of A(t) but acknowledges that individual diseases have additional mechanisms.
• The threshold problem. Is there a well-defined "biological age" at which an organism "becomes old", or is aging a smooth continuum? Our equations treat it as continuous, but the Gompertz hazard becomes so steep in late life that practical thresholds do emerge.

These are live questions. A theory that resolved them all would not be unifying; it would be complete. We aim only at the former.

10Implications for Intervention

If the framework is correct, several strategic consequences follow for longevity therapeutic development.

10.0 The geometry of intervention space

Before discussing specific interventions, it helps to view the full intervention landscape geometrically. The state of an aging organism lives in a five-dimensional space parameterized by (D, R, I, P, E). Youthful states occupy a small region near (0, 1, 0, 0, low-E); aged states drift toward (high-D, low-R, high-I, high-P, variable-E). Aging is a trajectory through this space; therapy is a vector field that deforms the trajectory.

In this picture, interventions can be classified by the direction of their effect in state space. Rapamycin pushes P toward 0 with limited effect on other axes. NAD+ precursors push R upward. Partial reprogramming compresses I toward 0 along a direction nearly orthogonal to D. Senolytics reduce effective D. The power of combinatorial regimes is that they produce a sum of vectors whose magnitude exceeds any single vector. Conversely, redundant interventions (two P-targeting drugs) produce parallel vectors with diminishing magnitude returns.

This geometric view also explains why certain intervention combinations are counter-indicated. Senescent cells sometimes play beneficial wound-healing roles; over-aggressive senolytics can impair repair (reduce R indirectly). mTOR inhibition lowers P but also attenuates muscle protein synthesis (reduces R in skeletal muscle). The framework predicts such trade-offs geometrically: any intervention that pushes one coordinate favorably while degrading another produces a zero-sum vector in certain regimes. Careful dose-titration becomes a geometry problem, not merely a pharmacology problem.

10.1 Combinatorial therapies are mandatory

Equation (5) is a sum over four terms. Any single-term intervention can reduce A(t) by at most the weight of that term. To meaningfully extend healthy lifespan, interventions must hit multiple terms. The ITP's emerging combination results and the pharmaceutical industry's growing interest in multi-pathway longevity cocktails are consistent with this.

10.2 Information-targeting drugs are a new class

Equation (3) has two channels: damage (λ·D) and reprogramming (r(u)). Drugs or cellular therapies that directly modify chromatin state — small-molecule reprogramming factors, CRISPR-based epigenome editors, mRNA OSK delivery — attack I without waiting for D to be fixed. This is a genuinely new therapeutic class with, as of 2026, no approved members; it is likely to be transformative.

10.3 Senolytics target effective D, not D itself

Senescent cells do not accumulate damage faster than other cells; they emit SASP factors that inflate the effective burden of damage sensed by neighbours. Removing senescent cells does not reduce D at the individual-cell level; it reduces the organism-level D-equivalent load. This implies senolytic efficacy should saturate once SnC burden is cleared — further effects must come from other terms.

10.4 Lifestyle interventions are rate-setters, not reversers

Caloric restriction, exercise, sleep, cold/heat exposure operate primarily on P and R (via hormetic h(u)), reducing the rate of damage accumulation but not reversing accumulated state. They cannot substitute for information or damage reversal but are foundational because they set the slope.

10.5 Sequencing matters

Because I couples back into R (Eq. 2, −δ·I term), reducing I before adding damage-repair interventions may amplify the latter's efficacy. A rational sequencing might be: (1) reduce P chronically (rapamycin, metformin, CR) → (2) clear SnCs periodically (senolytics) → (3) pulse reprogramming to compress I → (4) replace damaged substrate where possible (cell/gene therapy). This is a testable clinical hypothesis.

10.6 The ceiling problem

As noted in Prediction 6, k₀·E(t) cannot be driven to zero. Extreme lifespan extension beyond ~150 years will require either reducing k₀ (engineering biology with lower intrinsic damage rates — different amino acids, alternative metabolism, synthetic biology) or increasing ρ beyond anything evolution has produced (engineered repair systems, AI-designed proteins with repair-first architecture). These are speculative but follow directly from the framework.

10.7 Species-level evidence: what long-lived animals teach us

Any unified theory of aging must explain why lifespans vary by five orders of magnitude across species. Under our framework this variation is decomposed into species-specific values of ρ (repair budget), k₀ (baseline damage rate), λ (damage-to-information coupling), and μ (programmatic drift rate). The comparative-biology literature contains striking anomalies that sharply constrain which of these parameters evolution can modify.

Naked mole rat (Heterocephalus glaber). Thirty-plus year lifespan in an animal the size of a mouse — an order of magnitude beyond allometric prediction. Elevated proteostasis, hyper-active P53/INK4a cancer surveillance, high-molecular-weight hyaluronan, extraordinary DNA repair fidelity. In our framework: strongly elevated ρ, moderately reduced k₀ (hypoxic habitat), extremely low μ (pedomorphic retention of juvenile signalling). The naked mole rat is essentially a natural experiment in pushing ρ upward while keeping body plan constant.

Greenland shark (Somniosus microcephalus). Estimated lifespans to 400 years. Cold-water metabolism dramatically reduces k₀ (Arrhenius scaling of all chemical reaction rates). This tells us that the thermodynamic floor is temperature-dependent and, in principle, engineerable for warm-blooded animals only with great difficulty.

Ocean quahog (Arctica islandica). >500 year lifespan in a clam. Continuous, slow, low-metabolism life — combined with what appears to be extremely efficient proteostasis. An existence proof that animal cells can in principle maintain homeostasis for five centuries.

Hydra. Apparent biological immortality under laboratory conditions. Continuous stem-cell turnover replaces all somatic cells on a <20-day timescale. In our framework: R is continuously refreshed by the stem-cell compartment, preventing the −γ·D·R collapse term from ever dominating. This suggests that indefinite lifespan is compatible with eukaryotic biology when stem-cell flux exceeds damage accumulation — a design constraint we may eventually be able to engineer into vertebrate tissues.

Bats. Many bat species live 20–40 years despite high metabolic rates — violating the rate-of-living hypothesis outright. Mechanisms include enhanced DNA repair, unique interferon signalling, and ability to tolerate oxidative load. The bat phenotype argues that elevated ρ can compensate for elevated k₀, within limits.

Certain rockfish (Sebastes aleutianus). >200 year lifespans with indeterminate growth. Continued organismal growth maintains stem-cell niches; this maps cleanly onto our prediction that R decline drives terminal mortality.

Implications. The comparative evidence tells us that evolution has independently discovered high-longevity solutions multiple times, using different parameter combinations. This is reassuring for intervention strategy: there are many independent routes to extended healthspan, and mimicking any of them via therapeutics is a legitimate approach. It also warns us that copying a single long-lived species's biology may be insufficient; combining mechanisms across species (bat DNA repair + naked mole rat hyaluronan + hydra stem-cell turnover) is more likely to achieve transformative results than copying any one.

10.8 The economics of combinatorial interventions

If the framework predicts multiplicative mortality reductions from orthogonal interventions, this has sharp economic consequences for longevity biotech. A 15% median-lifespan extension from rapamycin, combined with a 10% extension from senolytics, should compose to approximately a 25–27% extension — not because effects literally multiply in lifespan but because their effects on A(t) are additive and the Gompertz hazard is exponential in A(t).

This in turn implies that the value of a second, mechanistically orthogonal longevity intervention is far higher than the value of the twentieth member of an already-saturated class. From a portfolio standpoint, the field should systematically map existing and pipeline drugs onto the five state variables {D, R, I, P, E} and pursue under-represented terms aggressively. As of 2026, the least-populated term is I — direct information restoration. We would predict this to be the highest-value therapeutic territory for the next decade.

10.9 Personalized aging interventions

A further implication of the framework is that individuals differ in their dominant aging terms. A person with early-onset vascular calcification and low-grade inflammation may be dominated by P and D (inflammaging, senescent burden), whereas an individual with preserved cardiovascular health but declining cognition may be dominated by I (epigenetic drift in neurons). Diagnostic panels that estimate the relative magnitude of D, R, I, and P — not just their weighted sum — would enable intervention matching. Current multi-omic aging clocks (PhenoAge, DunedinPACE) are moving in this direction but are still largely scalar. Vector-valued aging clocks are a natural extension the framework predicts.

11Limitations and Future Directions

11.1 Limitations of the LLM orchestration approach

• Training-data recency. The models have cutoff dates; novel 2024–2026 literature is under-represented. The held-out auditor flagged several candidate 2024 papers absent from primary-model outputs.
• Hallucinated citations. Two of four models produced citations that did not resolve; we pruned these before synthesis. The reference list is conservative.
• Concordance ≠ correctness. Four LLMs agreeing on a claim is evidence of corpus consensus, not of ground truth. Some consensus claims may be field dogma that is later revised.
• Prompt engineering effects. The specific prompt structure biases toward comprehensiveness over novelty. Alternative prompts may surface theories we missed.

11.2 Limitations of the mathematical framework

• Equations (1)–(5) are phenomenological, not derived from first principles. Parameters (k₀, ρ, α, γ, λ, μ) have no absolute units; they must be fitted per-organism.
• The framework treats D as scalar; in reality it is high-dimensional (DNA, lipid, protein, organelle sub-types). Extension to vector D is straightforward but adds parameters.
• Stochasticity is handled via ε(t) rather than mechanistically; single-cell heterogeneity (Vijg's mutational mosaicism) is under-represented.
• Evolutionary term β(age) is taken as exogenous; a fuller theory would derive β from ecological inputs.

11.3 Philosophical limitations

• Is unification the right goal? Some philosophers of biology argue that biology is fundamentally particular and resists the physics-style unification instinct. Our framework is explicitly reductionist in spirit; it may be that certain aging phenomena (e.g., organ-specific aging trajectories, psychological aging, social determinants) are not reducible to {D, R, I, P, E}. We accept this as a working limitation; the framework covers the molecular and cellular substrate of aging, not its social or experiential dimensions.
• Post-hoc explanatory power vs. genuine prediction. Any sufficiently flexible equation can "explain" existing data. The value of a theory is in its risky predictions. We have listed several in §8.4–8.5, but honest falsification will require dedicated experimental programmes over a decade or more.
• The aging-as-disease framing. Our framework tacitly assumes aging is a coherent target for intervention. Some gerontologists (including important voices in the field) remain skeptical that this framing is productive. We have argued here that the framing is justified because A(t) is a well-defined scalar output; others may disagree.

11.3.5 Aging and consciousness

A limitation worth naming explicitly: the framework addresses biological aging of cells and tissues. It has nothing to say about subjective experience, memory continuity, personal identity, or the relationship between biological age and psychological age. These are distinct problems in distinct disciplines, and we do not presume to bridge them. Readers interested in mind-body aspects of longevity should consult the philosophy-of-mind and neuroscience literatures, which operate on different foundations.

However, we note that any therapy which substantially reverses biological aging will raise sharp questions about personal identity over long timescales — a healthy 200-year-old with a continually refreshed biological substrate is a scenario current ethical frameworks handle poorly. The framework's prediction of a 150-year ceiling under current biology somewhat delays this question, but it does not dissolve it.

11.4 Future work

• Parameter fitting. Apply Equations (1)–(5) to longitudinal multi-omic datasets (UK Biobank, Dunedin, MESA) to extract per-individual parameters. This is a concrete next step.
• Cross-species calibration. Fit the framework to mouse, naked mole rat, human, and Drosophila data; test whether long-lived species differ primarily in ρ as predicted.
• Intervention forecasting. Use fitted models to predict combinatorial intervention outcomes before they are tested; compare to ITP results.
• Information-theoretic formalization. Replace I(t) with a proper information-theoretic quantity (mutual information between current and reference chromatin states); connect to rate-distortion theory.
• AI-first theoretical biology. Deploy orchestrated LLM consortia on open scientific questions routinely. This paper is one demonstration; the workflow generalizes.

11.5Biomarkers and the Framework

The dominant aging biomarkers of 2026 — epigenetic clocks (Horvath, Hannum, PhenoAge, GrimAge, DunedinPACE), proteomic clocks (SomaSignal, INTERVENE), metabolomic panels, glycan-based biological age, immune-age (iAge), and multi-omic composite scores — all share a common empirical structure: a regression from high-dimensional molecular state onto a chronological or mortality-proxy target. They are enormously useful practically but mechanistically under-specified.

Within our framework, each clock corresponds to a specific weighted projection of {D, R, I, P}.

• Horvath-style methylation clocks primarily track I(t), because CpG methylation is a direct epigenetic state variable. They are sensitive to partial reprogramming, consistent with the I-channel in Equation (3).
• GrimAge and PhenoAge include mortality-linked proteins (GDF15, TIMP1, etc.) and so load on P and D as well as I. This explains why they correlate more tightly with morbidity than pure Horvath clocks: they capture more of A(t).
• DunedinPACE estimates the rate of aging rather than cumulative age. In framework terms, it approximates dA/dt. This makes it particularly well-suited for evaluating interventions on short timescales.
• iAge (inflammatory-immune clock) primarily tracks the P component, specifically the inflammaging axis.
• Glycan-age (GlycanAge) captures IgG Fc-glycan shifts driven by chronic inflammation; again primarily P with secondary D contributions.
• Frailty indices (Rockwood FI, accumulated deficit) measure phenotypic manifestations at Level 5 — closest to the full A(t) but most delayed.

This is not a replacement for clocks; it is a theoretical lens on what they are. More importantly, it suggests that a vector-valued aging clock — one reporting (D-age, R-age, I-age, P-age) separately — would enable mechanism-specific intervention matching and is a natural next product for the field. Several biotech groups are known to be working on exactly this.

For trial design, the framework suggests that biomarker endpoints should be chosen to match the intervention mechanism. A rapamycin trial should be powered on P-loaded clocks (iAge, proteomic age). A senolytic trial should be powered on D-loaded markers (p16INK4a transcript burden, SASP panels). A partial-reprogramming trial should be powered on I-loaded clocks (pan-tissue Horvath-style). Mixing mismatched biomarkers and interventions is a plausible reason for underwhelming trial outcomes to date.

Looking ahead, four biomarker innovations would most accelerate framework validation. First, single-cell aging clocks that report the age-state of individual cells rather than tissue averages — necessary for detecting the mosaicism implied by somatic mutation theory. Second, tissue-specific mechanistic clocks that distinguish neuronal I-drift from cardiac P-drift from skeletal-muscle R-decline. Third, short-window rate-of-change clocks usable in Phase 1/2 trials of <12 months duration, essential for shifting longevity trial economics. Fourth, resilience-probe assays — dynamic challenge-and-recovery protocols (physiological, cognitive, or immunological) that directly measure R(t) rather than inferring it. Each of these is technically feasible in 2026; none is deployed at scale.

11.6Appendix: Glossary of Key Variables

12Conclusion

The aging field has been rich in theories and poor in unifications. Medvedev's count of 300+ is probably an underestimate today. We have argued — via a four-model LLM orchestration — that the apparent diversity hides a tractable structure: every mature theory is a projection of a six-level causal hierarchy onto one or two levels, and the hierarchy itself can be written as a compact set of coupled equations with terms for damage (D), repair (R), information (I), programme (P), and environment (E), constrained by evolutionary (β, ρ) and thermodynamic parameters.

The unified statement we propose — aging as the progressive, thermodynamically inevitable loss of biological information fidelity, shaped by evolutionary selection for reproductive fitness over somatic maintenance, manifesting through hierarchically organized damage accumulation and programmatic responses — is neither original in any single word nor trivially derivable from any single predecessor. Its contribution is that it fits simultaneously the evolutionary, thermodynamic, molecular, programmatic, informational, and phenotypic evidence, and that it connects those levels by explicit equations that make combinatorial interventions predictable.

If this framework is correct, the therapeutic roadmap is clear: continue optimizing P-targeting (rapamycin, metformin, CR mimetics), scale R-boosting (NAD+ precursors, autophagy inducers, hormesis protocols), accelerate senolytic development, and — most importantly — develop the information-targeting drug class that Equation (3) identifies as the genuinely new therapeutic frontier. Combinatorial protocols should produce multiplicative mortality-hazard reductions. Maximal lifespan under current biology is bounded around 120–150 years by the k₀·E(t) floor; meaningful extension beyond that requires biological engineering, not pharmacology.

The framework also clarifies why some heavily funded intervention classes will saturate sooner than expected. Further senolytics will not dramatically outperform existing ones, because removing senescent cells addresses only a portion of the D term. Additional mTOR inhibitors will not radically outperform rapamycin, because P-dampening is bounded by evolutionary constraints on how far programmatic drift can be suppressed without harming developmental and regenerative functions. The next qualitative leap will come from the information axis — partial reprogramming, chemical OSK delivery, chromatin-editing therapeutics, and possibly entirely new modalities we cannot yet name.

On a longer horizon, the framework suggests a staged roadmap. Stage one (now–2030) is combinatorial optimization within existing drug classes, guided by mechanism-matched biomarkers — essentially better use of the tools we already have. Stage two (2028–2035) is the clinical maturation of the information-targeting class, bringing a genuinely new axis under therapeutic control and likely producing the first interventions that reverse rather than merely slow aging phenotypes. Stage three (post-2035) is biological engineering: re-architecting tissues, organs, and eventually organisms to reduce k₀ and elevate ρ beyond anything evolution has produced. This is speculative but is where the equations eventually point.

We also emphasize what the framework does not predict. It does not predict immortality. Even with perfect intervention on I, D, and P, the k₀·E(t) term sets a floor. Biological engineering can push that floor down but not to zero. Cryonics, mind uploading, substrate-independent existence — these are outside the biological aging framework entirely and should be discussed under a different theoretical vocabulary.

We acknowledge the irony of four probabilistic language models producing a paper on the determinism of biological decay. But that, too, is the point. The age of single-investigator, single-model theoretical biology is ending; the age of orchestrated, auditable, multi-model scientific synthesis is beginning. This paper is a first specimen of the latter, and we invite both human and machine colleagues to improve it.

Finally, we note the broader methodological claim. If multi-LLM orchestration can produce a credible theoretical synthesis in a field as conceptually diverse as biogerontology, it can plausibly do the same in any field where the primary obstacle is reading across many partial literatures. Climate science, cancer taxonomy, consciousness studies, quantum foundations, economic crisis theory — all have structurally similar unification problems. OpenClaw-like platforms are general-purpose; this paper is one instance of their application. We expect many more.

13References

1. Weismann, A. (1882). Über die Dauer des Lebens. Gustav Fischer, Jena.
2. Pearl, R. (1928). The Rate of Living. Alfred A. Knopf, New York.
3. Medawar, P.B. (1952). An Unsolved Problem of Biology. H.K. Lewis, London.
4. Williams, G.C. (1957). Pleiotropy, natural selection, and the evolution of senescence. Evolution 11, 398–411.
5. Harman, D. (1956). Aging: a theory based on free radical and radiation chemistry. J. Gerontol. 11, 298–300.
6. Hayflick, L. & Moorhead, P.S. (1961). The serial cultivation of human diploid cell strains. Exp. Cell Res. 25, 585–621.
7. Orgel, L.E. (1963). The maintenance of the accuracy of protein synthesis and its relevance to ageing. Proc. Natl. Acad. Sci. 49, 517–521.
8. Harman, D. (1972). The biologic clock: the mitochondria? J. Am. Geriatr. Soc. 20, 145–147.
9. Kirkwood, T.B.L. (1977). Evolution of ageing. Nature 270, 301–304.
10. Cerami, A. (1985). Hypothesis: glucose as a mediator of aging. J. Am. Geriatr. Soc. 33, 626–634.
11. Medvedev, Z.A. (1990). An attempt at a rational classification of theories of ageing. Biol. Rev. 65, 375–398.
12. Franceschi, C. et al. (2000). Inflamm-aging: an evolutionary perspective on immunosenescence. Ann. N.Y. Acad. Sci. 908, 244–254.
13. Kenyon, C. et al. (1993). A C. elegans mutant that lives twice as long as wild type. Nature 366, 461–464.
14. West, G.B., Brown, J.H., Enquist, B.J. (1997). A general model for the origin of allometric scaling laws in biology. Science 276, 122–126.
15. Gavrilov, L.A. & Gavrilova, N.S. (2001). The reliability theory of aging and longevity. J. Theor. Biol. 213, 527–545.
16. Terman, A. & Brunk, U.T. (2005). Autophagy in cardiac myocyte homeostasis, aging and pathology. Cardiovasc. Res. 68, 355–365.
17. Rattan, S.I.S. (2006). Theories of biological aging: genes, proteins, and free radicals. Free Radic. Res. 40, 1230–1238.
18. Blagosklonny, M.V. (2006). Aging and immortality: quasi-programmed senescence and its pharmacologic inhibition. Cell Cycle 5, 2087–2102.
19. de Grey, A.D.N.J. & Rae, M. (2007). Ending Aging. St. Martin's Press, New York.
20. Hayflick, L. (2007). Biological aging is no longer an unsolved problem. Ann. N.Y. Acad. Sci. 1100, 1–13.
21. López-Otín, C., Blasco, M.A., Partridge, L., Serrano, M., Kroemer, G. (2013). The hallmarks of aging. Cell 153, 1194–1217.
22. Horvath, S. (2013). DNA methylation age of human tissues and cell types. Genome Biol. 14, R115.
23. Hannum, G. et al. (2013). Genome-wide methylation profiles reveal quantitative views of human aging rates. Mol. Cell 49, 359–367.
24. Vijg, J. (2014). Somatic mutations, genome mosaicism, cancer and aging. Curr. Opin. Genet. Dev. 26, 141–149.
25. Zhavoronkov, A. & Bhullar, B. (2015). Classifying aging as a disease in the context of ICD-11. Front. Genet. 6, 326.
26. Ocampo, A. et al. (2016). In vivo amelioration of age-associated hallmarks by partial reprogramming. Cell 167, 1719–1733.
27. Gladyshev, V.N. (2016). Aging: progressive decline in fitness due to the rising deleteriome adjusted by genetic, environmental, and stochastic processes. Aging Cell 15, 594–602.
28. Levine, M.E. et al. (2018). An epigenetic biomarker of aging for lifespan and healthspan (PhenoAge). Aging 10, 573–591.
29. Lu, A.T. et al. (2019). DNA methylation GrimAge strongly predicts lifespan and healthspan. Aging 11, 303–327.
30. Lu, Y. et al. (2020). Reprogramming to recover youthful epigenetic information and restore vision. Nature 588, 124–129.
31. Gems, D. & de Magalhães, J.P. (2021). The hoverfly and the wasp: a critique of the hallmarks of aging as a paradigm. Ageing Res. Rev. 70, 101407.
32. Pyrkov, T.V. et al. (2021). Longitudinal analysis of blood markers reveals progressive loss of resilience and predicts human lifespan limit. Nat. Commun. 12, 2765.
33. Belsky, D.W. et al. (2022). DunedinPACE, a DNA methylation biomarker of the pace of aging. eLife 11, e73420.
34. Browder, K.C. et al. (2022). In vivo partial reprogramming alters age-associated molecular changes during physiological aging in mice. Nat. Aging 2, 243–253.
35. Yang, J.-H. et al. (2023). Loss of epigenetic information as a cause of mammalian aging. Cell 186, 305–326.
36. López-Otín, C., Blasco, M.A., Partridge, L., Serrano, M., Kroemer, G. (2023). Hallmarks of aging: an expanding universe. Cell 186, 243–278.
37. Cohen, A.A. et al. (2022). A complex systems approach to aging biology. Nat. Aging 2, 580–591.
38. van Deursen, J.M. (2014). The role of senescent cells in ageing. Nature 509, 439–446.
39. Conboy, I.M. et al. (2005). Rejuvenation of aged progenitor cells by exposure to a young systemic environment. Nature 433, 760–764.
40. Sinclair, D.A. & LaPlante, M.D. (2019). Lifespan: Why We Age— and Why We Don't Have To. Atria Books, New York.
41. Speakman, J.R. (2005). Body size, energy metabolism and lifespan. J. Exp. Biol. 208, 1717–1730.
42. Calabrese, E.J. (2008). Hormesis: why it is important to toxicology and toxicologists. Environ. Toxicol. Chem. 27, 1451–1474.
43. Campisi, J. (2013). Aging, cellular senescence, and cancer. Annu. Rev. Physiol. 75, 685–705.
44. Wyss-Coray, T. (2016). Ageing, neurodegeneration and brain rejuvenation. Nature 539, 180–186.
45. Zhu, Y. et al. (2015). The Achilles' heel of senescent cells: from transcriptome to senolytic drugs. Aging Cell 14, 644–658.
46. Harrison, D.E. et al. (2009). Rapamycin fed late in life extends lifespan in genetically heterogeneous mice (ITP). Nature 460, 392–395.
47. Partridge, L., Fuentealba, M., Kennedy, B.K. (2020). The quest to slow ageing through drug discovery. Nat. Rev. Drug Discov. 19, 513–532.
48. Martincorena, I. et al. (2015). High burden and pervasive positive selection of somatic mutations in normal human skin. Science 348, 880–886.
49. Fedichev, P.O. et al. (2019). A dynamic model of aging in the C. elegans model organism. Sci. Rep. 9, 13893.