In plain terms

The claim being assessed: Across tumour types, a prospectively identifiable rare self-renewing cell subset occupies the apex of a functional hierarchy, is sufficient to serially reconstitute a heterogeneous tumour, and is necessary for tumour maintenance and post-treatment reinitiation because the remaining malignant cells cannot replace it.

Where the evidence stands: Γ — Plausible mechanism. There is a believable biological story and partial support, but the causal step has not been demonstrated.

What this mechanism does: This makes the process worse or keeps it going, but something else starts it.

Strongest evidence for it: The claim is a six-part conjunction stated across tumour types. The three clauses that carry that cross-tumour scope — rarity (SC2), broad sufficiency (SC3) and irreplaceability (SC6) — all sit at Γ under recorded ceilings: the strongest rarity and serial-reconstitution evidence is transplantation-based and falls under the rubric's xenograft hard rule, and the field-valid endogenous tests of irreplaceability are negative. The three clauses that reach Ε (SC1, SC4, SC5) do so by narrowing to defined tumour contexts, which is what Ε means, so under the prescribed scope-matched rule they are the narrower subclaims reported separately rather than the headline. Stem-like states do causally maintain or reinitiate tumours in defined cancers and stages; that bounded result is reported at Ε in the matrix and is not what the nominal sentence asserts.

What decisive evidence would look like

In plain terms, researchers would change the proposed cause, confirm it changed as intended, and measure a meaningful health outcome while checking other explanations. The result should be able to support or count against the claim.

Relative explanatory fit: Alternative ranked ahead

What this means for cancer treatment and 'cancer stem cells'

Some cancers contain cells or temporary cell states that can help a tumour regrow in defined experimental settings. The broad picture of every tumour as a fixed hierarchy controlled by a rare, irreplaceable stem-cell population is not established across cancer types. Tumours can be diverse, and cells may change state or replace one another when treatment pressures them.

Implications for health and treatment: This supports careful development of treatments that target tumour-maintaining states in the specific cancer, stage, and genetic setting where they matter. It does not establish that finding one so-called cancer-stem-cell marker will provide a durable cure, or that a treatment aimed at one cell population will work across cancers. The evidence also explains why combination strategies and studies of resistance, recurrence, and cell-state plasticity remain important.

Limits for individual care: Much of the decisive experimental work comes from model systems, whereas direct proof of state-specific necessity and durable control in people is limited. A cancer diagnosis, prognosis, or treatment choice cannot be inferred from this mechanism review. This page is not medical advice: do not delay surgery, systemic therapy, radiation, surveillance, or enrolment decisions because of a product or clinic claiming to eliminate 'cancer stem cells'; discuss options with the oncology team.

Medical information notice: This review provides general health information, not medical advice. It cannot diagnose a condition or replace an individual discussion with a qualified clinician.

This assessment rests on 19 sources and 17 individually checked claims. 17 of those claims were verified against the original paper.

A low grade is not an accusation. It means the causal work has not been done yet, which is the ordinary state of most actively studied mechanisms in biology.

At a glance

Sub-claims assessed separately

ID	Sub-claim	Type	Level	Pivotal card
SC1	A functional self-renewal and differentiation hierarchy exists in tumours.	existence	Ε	Card ↗
SC2	Tumour-sustaining capacity is confined to a rare subpopulation.	generalization	Γ	—
SC3	The prospectively identified rare subset is sufficient by itself to serially reconstitute a heterogeneous tumour.	sufficiency	Γ	—
SC4	Eliminating the proposed stem-like population prevents continued tumour maintenance.	necessity	Ε	—
SC5	Eliminating the proposed stem-like population prevents post-treatment tumour reinitiation or relapse.	necessity	Ε	—
SC6	Cells outside the proposed subset cannot replenish an equivalent tumour-sustaining population sufficiently to bypass its elimination.	mechanism	Γ	Card ↗

Visual Evidence Cards · Pivotal Subclaims

Pivotal subclaim illustrations

Dark-theme simulated-micrograph renders and concise causal evidence summaries for the pivotal subclaims that govern this review's grade. Click any card to inspect the master 2400×2400 render.

Master 2400×2400 render ↗

Master 2400×2400 render ↗

Figures

Explanatory figures summarizing this review. They carry no evidential weight of their own; where a figure and the review differ, the review governs.

Full assessment

From here the review is written for a technical reader: caps and thresholds, competing models, the independence basis, and the citation ledger.

Provenance: 19 sources · 17 registered claims · 17 verified · 0 unverified · 0 contradicted · 19 Tier A sources · gate passed. Nine sources were read at complete-full-text depth and ten at abstract/metadata depth. Direct PubMed, PubMed Central, Europe PMC, and accessible primary PDFs were used. Scite was unavailable; citation-context counts are therefore unavailable rather than zero. Adversarial review: Generated by Sol 6.1; adversarially reviewed by Claude Opus 5.5 (Anthropic) on 2026-10-14. 3 corrections applied (see ADVERSARIAL REVIEW AND APPROVAL). Status: Approved as revised.

Hypothesis assessed: Tumours are hierarchically organized around a rare self-renewing subpopulation that is necessary and sufficient to sustain and reinitiate the tumour.

Gradeable strong form: Across tumour types, a prospectively identifiable rare self-renewing cell subset occupies the apex of a functional hierarchy, is sufficient to serially reconstitute a heterogeneous tumour, and is necessary for tumour maintenance and post-treatment reinitiation because the remaining malignant cells cannot replace it.

Review date: 2026-09-30 Framework: HSEF v5.4; cancer-biology rubric Mode: Revision, then cross-model red-team synthesis. The original assessment was locked before the previous cancer-stem-cell review was opened; this edition is the red-teamed final.

AT A GLANCE

The hypothesis says that most cells in a tumour cannot keep the cancer going. Instead, a small self-renewing group sits at the top of a one-way hierarchy, continually makes the other tumour cells, and is both able and required to rebuild the cancer after treatment.

Experiments establish such dependencies in particular cancers and stages. Native lineage tracing and cell-state ablation show that selected stem-like populations can maintain growth or seed recurrence. But the sentence above makes six separate promises at once, and the three that speak to all tumours — that the group is rare, that it alone can rebuild a tumour, and that no other cell can take its place — are the three that do not hold up. The claim as written is therefore graded Γ — Mechanistically plausible, partial system support, while the narrower finding that particular stem-like states causally sustain particular cancers is separately established and sits at Ε — Context-specific causality established.

The deciding distinction is between a function and a fixed cell compartment. The evidence strongly supports tumour-sustaining functions. It does not show that those functions always reside in a rare, prospectively fixed, one-way apex population. That is why a dynamic hierarchy ranks above both a universal cancer-stem-cell model and a fully equipotent stochastic model.

On the headline letter. Γ here is a verdict on a six-part conjunction, not on cancer-stem-cell biology. Three clauses of that conjunction clear the endogenous causal boundary inside defined contexts and are reported at Ε in the matrix below. Three clauses stated across tumour types do not clear it, and those three are the ones that carry the claim's own scope. Readers who take away "cancer stem cells were refuted" have misread this page: what is graded Γ is the universal rare fixed compartment, and what is graded Ζ, in the tournament, is the dynamic-state alternative that the same evidence supports.

SUBCLAIM MATRIX

Aggregation rule: scope-matched, preregistered before scoring, as the hypothesis index prescribes for this entry. The rule reads: the overall level equals the highest subclaim whose scope matches the central claim's scope; narrower subclaims are reported separately. The central claim is stated across tumour types and as a conjunction. The subclaims that carry that cross-tumour scope are SC2 (rarity), SC3 (sufficiency) and SC6 (irreplaceability) — and all three sit at Γ under recorded ceilings. SC1, SC4 and SC5 reach Ε precisely by narrowing to defined tumour contexts; Ε's own definition is that "generalization to the full claimed scope fails," so those three are the narrower subclaims the rule directs us to report separately rather than to promote. Scope-matched therefore returns Γ.

The index entry states the same conclusion in words: "the universal CSC framing is scored by the strength of the evidence in the weakest tumor context claimed, while specific tumor-type subclaims are reported separately."

The alternatives, as v5 requires: weakest-link Γ; primary-subclaim on SC4 Ε; scope-matched Γ. Only primary-subclaim — the rule that was not prescribed — returns Ε. The spread is one rung and it is driven entirely by rule choice on identical evidence, which is why the rule is named up front and its arithmetic shown.

The original sentence should therefore not be paraphrased as "the rare fixed hierarchy is established." What is established is the narrower causal importance of tumour-sustaining states in defined contexts.

COMPETING MECHANISTIC MODELS

The rank-one model has four independent causal lines. Native lineage tracing reduces the risk that hierarchy is only a transplantation artifact [5,6]. Endogenous ablation and recurrence intervention move growth or relapse endpoints and therefore exclude purely descriptive lineage persistence [7,12,14]. Human and patient-derived relapse evidence reduces the chance that every pattern is unique to engineered mouse cancers [9,10,15]. State-replenishment and combined-depletion experiments distinguish incomplete targeting from genuine biological bypass [11,13,18,19]. No single xenograft, marker, or transgene artifact produces all four lines.

Cañellas-Socias et al. provide the sharpest separator between the top two models. In a colorectal surgical-recurrence system, EMP1-high residual cells causally seeded relapse, and their ablation before surgery prevented most recurrence, whereas preoperative LGR5-positive-cell ablation did not [12]. The result shows an organized, state-specific recurrence dependency, which is more than undirected clonal selection. Yet the identity and timing of the responsible state differ from a fixed LGR5 apex, which favors the dynamic-hierarchy model.

The tournament is reported for what it is: a comparison of explanations. It does not set the nominal model's letter, which follows from the caps, thresholds and aggregation rule alone. What it does inform is the role of the mechanism: tumour-sustaining states reinforce and stabilize growth or recurrence in selected contexts, while a universal primary-driver designation would require them to be initiating, necessary, and rate-limiting across the claimed cancer context.

FULL REVIEW

Claim form and falsifiability

The user's sentence is a conjunction of six claims: hierarchy exists; the apex is rare; it is sufficient to reproduce the tumour; it is necessary for maintenance; it is necessary for reinitiation; and other malignant cells cannot replace it. Those clauses must be kept separate because each has a different gold-standard experiment. A cell can initiate a xenograft without being required in the original tumour. A state can be required during metastatic outgrowth without being the cell that disseminated. A marker-defined population can be depleted while another population regenerates the same function.

Because the claim is a conjunction, it is scored as one. A conjunctive hypothesis is graded on the conjunction, not on its easier half; a review that reported the strongest surviving clause as the headline would be answering a different question from the one the sentence asks.

The strong form remains falsifiable because membership must be specified before maintenance or recurrence is observed. "Cancer stem cell" cannot be defined retrospectively as whichever cell eventually sustains the tumour. The relevant antecedent must be a declared marker, lineage state, or functional state. Rarity must be measured in a stated assay and host. Necessity requires selective loss followed by a disease-relevant endpoint. Sufficiency requires prospective isolation and serial reconstruction. Irreplaceability requires lineage-resolved evidence that cells outside the target do not recreate its function after depletion.

This definition avoids an unfalsifiable escape. If every tumour-sustaining cell is relabelled a cancer stem cell after the outcome, high melanoma tumour-forming frequency and non-LGR5-to-LGR5 conversion would become support instead of tests. The review therefore treats those findings as direct pressure on the rare fixed hierarchy.

What the classic transplantation studies establish

Bonnet and Dick prospectively enriched human AML cells with leukaemia-initiating capacity in immunodeficient mice. The initiating fraction was CD34-positive, CD38-negative across every AML subtype examined, and it differentiated in vivo into leukaemic blasts — the basis for their conclusion that "the leukaemic clone is organized as a hierarchy" [1]. Al-Hajj et al. reported an analogous prospective enrichment in human breast cancer: the tumorigenic cells were isolated as CD44-positive, CD24-negative/low, Lineage-negative in eight of nine patients, as few as 100 such cells formed tumours in mice whereas tens of thousands of cells with alternate phenotypes did not, and the selected cells could be serially passaged, each time regenerating both further tumorigenic cells and the phenotypically diverse non-tumorigenic populations of the original tumour [2]. These results are strong for transplantation sufficiency and for the proposition that phenotypically distinct malignant cells can differ greatly in tumour-reconstitution capacity.

They do not establish endogenous necessity. Engraftment requires survival through dissociation, injection, immune and stromal mismatch, and access to a foreign niche. A cell that fails this composite assay may still contribute durably in its native tumour. Conversely, a cell capable of forming a tumour after transplantation is not necessarily the cell that sustains the original lesion or survives therapy. Under the cancer rubric, xenograft tumour initiation without lineage-resolved endogenous evidence is capped at Γ for the broad inference. Serial passage does not turn transplantation into an endogenous relapse experiment.

The breast paper also has an attached author correction concerning an omitted financial-interest acknowledgement [2]. The correction does not alter the reported experiments, and it is recorded in the source ledger as a status flag on the parent record rather than hidden or counted as a separate source.

Rarity is an assay result before it is a biological constant

Quintana et al. changed the recipient conditions for human melanoma cells by using more permissive immunodeficient hosts and extracellular-matrix support. Two separate experiments are relevant and should not be merged. In limiting-dilution assays across cells from 12 patients, approximately 25% of unselected melanoma cells formed tumours under the more permissive conditions. In a distinct single-cell transplantation series across four patients, 69 of 254 single-cell injections formed tumours — 27% overall, ranging from 12% to 65% depending on the patient [3]. A follow-up study reported that 28% of single melanoma cells obtained directly from patients formed tumours, that none of 22 heterogeneously expressed markers enriched tumorigenic cells, and that many markers appeared to be reversibly expressed [4].

These studies do not prove that every melanoma cell is equipotent in a patient. They show that "rare" is partly a property of the detection system. If a cell's measured tumour-forming capacity changes when host permissiveness, extracellular support, or immune deficiency changes, a numerical rarity estimate cannot be transported unchanged into native biology. The per-patient range of 12–65% matters here too: it shows the spread that a single pooled percentage conceals, and it is the spread rather than the mean that defeats a claim about rarity as a defining property.

The correct conclusion is bounded. AML and some breast cancers can contain strongly enriched transplantation-initiating fractions [1,2]. Melanoma under permissive conditions need not [3,4]. Rarity is therefore context- and assay-dependent, placing SC2 at Γ rather than supporting a universal rare compartment.

Native lineage tracing establishes hierarchy in selected tumours

Endogenous lineage tracing removes several xenograft failure modes. Driessens et al. followed labelled clones in mouse skin tumours. In papillomas, the majority of labelled tumour cells had only limited proliferative potential while a fraction persisted long term; their quantitative fit to the clone-size distributions required that "the stem-cell-like population accounts for just 20% of the labelled tumour proliferative cells." The inferred division rate of that persistent population was twice per day, which the authors report as approximately four times the progenitor rate in the same tumour — implying roughly 0.5 divisions per day for committed progenitors [5]. Invasive squamous carcinomas in the same program behaved differently: there the cumulative clone-size distribution was "equally consistent with a single equipotent proliferative cell population," with an average division time of one to two days [5]. This within-program contrast is a particularly informative mapped boundary: hierarchy exists, but stage and tumour architecture matter.

Schepers et al. used lineage tracing in mouse intestinal adenomas. Lgr5-positive cells, approximately 5–10% of adenoma cells, self-renewed and generated multiple differentiated tumour-cell types in native lesions [6]. This is strong evidence for a functional hierarchy without transplantation. It does not itself establish necessity because the population was traced rather than eliminated.

Together, the skin and intestinal studies clear the field-valid endogenous boundary for SC1. They also show why SC1 should not be generalized into the whole nominal proposition. A lineage can sit high in a hierarchy without being irreplaceable, and a tumour can change its growth architecture during progression.

Direct maintenance and recurrence tests

Chen et al. tested a quiescent Nestin-expressing population in an endogenous mouse glioma model. Those cells regenerated tumours after temozolomide, and ganciclovir-mediated genetic ablation arrested tumour growth; where ganciclovir was effectively delivered, "survival was substantially prolonged and tumour progression was severely impaired" [7]. This crosses the causal perturbation threshold for maintenance and post-treatment regrowth. The study also reports two important failure boundaries with unusual candour. Combined temozolomide and ganciclovir "did not significantly prolong survival beyond the GCV-only treated mice." And new tumours arose in the ventral brain and brain stem — six of seven treated mice — carrying high endogenous nestin protein but no transgene expression, which the authors traced to a silent transgene in that region [7]. The experiment therefore supports a causal dependency while showing how escape and target coverage constrain interpretation.

Colorectal models expose an even sharper difference between tumour stages. de Sousa e Melo et al. found that Lgr5-positive-cell ablation "restricts primary tumour growth, but does not result in tumour regression"; tumours were instead maintained by Lgr5-negative cells that continuously replenished the Lgr5-positive pool, leading to rapid re-initiation of growth once treatment stopped. Lgr5-positive cells had a more distinct role in the formation and maintenance of liver metastases [8]. Shimokawa et al. lineage-traced LGR5-positive cells in transplanted human colorectal organoids; ablation caused regression followed by regrowth driven by re-emerging LGR5-positive cells, with KRT20-positive differentiated cells reverting to the LGR5-positive state and contributing to that regrowth [9]. These are not null findings. They show a causal LGR5 role whose necessity depends on stage and whose cellular carrier is replaceable.

Fumagalli et al. followed colorectal metastasis at higher state resolution. Most disseminating cells in circulation were Lgr5-negative, seeded distant metastases, and then gave rise to Lgr5-positive cancer stem cells within them; this plasticity was "indispensable for outgrowth, but not establishment, of metastases" [11]. A fixed hierarchy predicts that an apex Lgr5-positive cell should seed the hierarchy. The observed sequence instead supports a dynamic division of labour: one state disseminates, another supports durable expansion, and cells transition between them.

Cañellas-Socias et al. directly tested surgical relapse. EMP1-high residual cells — the high-relapse cell population identified from patient single-cell transcriptomes — seeded metastatic recurrence, and their genetic ablation prevented it, leaving mice disease-free after surgery. The matched negative result is as informative: diphtheria-toxin treatment before surgery "effectively eliminated Lgr5-positive cells, yet it did not prevent disease relapse and mice developed overt liver metastatic disease," a result the authors replicated in an independent line. Ablation of Lgr5-positive cells after direct inoculation into the liver did halt metastasis formation, indicating a requirement during outgrowth rather than at seeding [12]. This is strong relapse causality, but for a temporally defined residual state rather than a universal fixed apex.

Ohta et al. add a patient-derived dormancy model. Dormant LGR5-positive, p27-positive cells persisted through chemotherapy in human colorectal organoid xenografts, and knocking out COL17A1 removed the dormant subpopulation and rendered organoids sensitive to chemotherapy [15]. This supports treatment-stage persistence in that system. Its xenograft and organoid setting prevents it from establishing universal human relapse necessity.

Human relapse is not one cellular route

Shlush et al. analysed paired human AML diagnosis and relapse samples with combined genetic and functional analysis of purified subpopulations and xenografts. The study identified at least two relapse patterns: one traceable to rare stem/progenitor cells, and another arising from larger committed populations that carried stemness-associated transcriptional programs [10]. This evidence is observational with respect to necessity; no human cell state was experimentally ablated. Its strength lies in endogenous human disease history and in falsifying a single universal cellular route to relapse.

The result is consistent with a dynamic hierarchy. A rare primitive population can seed relapse in one patient, while a more committed clone with acquired or retained self-renewal programs does so in another. It is less compatible with a single surface-marker-defined compartment that is always both necessary and sufficient.

Human therapeutic tests of the central corollary

The hypothesis carries a clear therapeutic prediction: eliminate the rare compartment and the tumour should not come back. Two randomized human trials bear on it, and they point in opposite directions. Both are included because the framework names therapeutic perturbation with a hard clinical endpoint as a field-valid human equivalent for the upper ladder, and because a review that omitted them would be adjudicating a human claim on mouse evidence alone.

The null result is the larger. CanStem303C, an open-label randomized phase III trial in previously treated metastatic colorectal cancer, added napabucasin to FOLFIRI (n=624) against FOLFIRI alone (n=629). Median overall survival was 14.3 versus 13.8 months, hazard ratio 0.976, one-sided P=.74. In the 44% of patients whose tumours were pSTAT3-positive — the prespecified biomarker-positive population, n=275 versus 272 — median overall survival was 13.2 versus 12.1 months, hazard ratio 0.969, one-sided P>.99. The trial concluded that adding napabucasin to FOLFIRI did not improve overall survival, and separately that pSTAT3 is an adverse prognostic factor [16].

The positive result is smaller and in the tumour type where the hierarchy evidence is strongest. In a randomized open-label phase II trial in newly diagnosed AML or high-risk myelodysplastic syndrome in patients unsuitable for intensive chemotherapy, glasdegib plus low-dose cytarabine (n=88) gave median overall survival of 8.8 months (80% CI 6.9–9.9) against 4.9 months (3.5–6.0) for low-dose cytarabine alone (n=44), hazard ratio 0.51, 80% CI 0.39–0.67, P=0.0004, with complete remission in 17.0% versus 2.3% [17].

Neither trial is a clean test, and the asymmetry between them should not be over-read. Glasdegib inhibits Hedgehog signalling, a pathway implicated in leukaemic stem-cell maintenance, but the trial demonstrates a survival benefit, not population-selective killing of a rare compartment; the rubric classes pathway inhibition with significant pleiotropy as moderate evidence and lists generalized toxicity as a competing explanation for such readouts. Napabucasin is a reactive-oxygen-species generator bioactivated by NQO1 whose status as a cancer-stem-cell-directed agent is itself contested, so its failure does not falsify stem-cell necessity either. What the pair establishes is narrower and still useful: the direction of the human therapeutic signal tracks tumour type in the same way the preclinical evidence does, and no human trial has yet asked the question the hypothesis actually poses — whether eliminating a prospectively defined rare compartment prevents recurrence.

Direct tests of plasticity and redundancy

Plasticity is not a verbal alternative to hierarchy; it makes testable predictions. If cells outside a targeted state can enter that state after depletion, then the state may be functionally important while the original compartment is not irreplaceable. Longitudinal lineage tracking and sustained depletion can separate these possibilities.

Mzoughi et al. studied WNT-dependent colorectal cancer and identified LGR5-positive and oncofetal states across a plastic continuum. Their depletion experiments used VAKSP tumoroids transplanted subcutaneously into immunodeficient mice — a transplanted-organoid system rather than an autochthonous tumour, which bounds what it can establish. Within that system the results are clear: sustained depletion of either state alone under continuous treatment "failed to maintain tumor stasis indefinitely, suggesting that either population can fuel tumor growth in the absence of the other," while co-targeting both states produced tumour shrinkage and significantly delayed resurgence [13]. The result supports causal state functions and biological redundancy. It does not establish durable cure after withdrawal, and resurgence was delayed rather than prevented.

Chan et al. used in-situ tracing and ablation in autochthonous mouse lung cancers. A high-plasticity cell state contributed causally to progression, maintenance, and treatment resistance: ablation in early neoplasias abrogated benign-to-malignant transition, and ablation in established tumours by suicide gene or CAR T cells robustly reduced tumour burden [14]. Two details matter for the irreplaceability clause, and the second cuts against a loose reading of this paper. First, non-high-plasticity cells can acquire the state — but the demonstration is transplantation-based: primary non-high-plasticity cells transplanted orthotopically into recipient lungs produced tumours containing a similar fraction of high-plasticity cells as the parent tumours, with the same result in subcutaneous serial allografts [14]. Second, the residual high-plasticity cells seen after ablation in autochthonous tumours were attributed by the authors not to biological bypass but to incomplete recombination of the reporter cassette, which let cells escape toxin-mediated killing; they state that their results "probably underestimate the importance of the HPCS in the maintenance of LUAD tumours," and they generated a clonal line with complete excision to remove the confound [14]. Citing the residual cells as replenishment while omitting that attribution would invert the authors' own reading.

Two recent studies give the replenishment route a molecular handle rather than leaving it phenomenological. Jiang et al. identified ZFP36L2, mutated in 5–10% of colorectal cancers, as a stress-responsive switch for dynamic dedifferentiation: in mouse colon regeneration models its ablation inhibited dedifferentiation, intestinal-stem-cell gene expression and function, and impaired regeneration, while in human colorectal cancer loss of ZFP36L2 function abrogated metastatic seeding and the outgrowth of LGR5-positive canonical metastases and instead promoted non-canonical lineage plasticity [18]. Perelli et al. used somatic mosaic genome engineering to trace and ablate malignant mesenchymal lineages along the epithelial-to-mesenchymal continuum in pancreatic cancer; ablation of the clones capable of mesenchymal plasticity abolished the associated mutational processes and evolutionary patterns, supporting what the authors call cell-state-restricted patterns of evolution [19]. Both extend the state-specific trace-and-ablate paradigm — the fourth and fifth tumour types in this review to receive it — and both describe a state whose entry is regulated rather than a compartment whose membership is fixed.

These studies make SC6 the clearest failed clause. The failure does not demote all cancer-stem-cell biology to association. It changes the object of causality from a rare immutable population to a regulated and sometimes replenishable state.

Cap rules, boundary check, and upper-ladder reachability

Three Γ ceilings bind, and they fall on the three clauses that carry the claim's stated scope. SC2 and SC3 are limited because the strongest rare-fraction and serial-reconstitution evidence is transplantation-based for the broad claim — the rubric's hard rule reads "xenograft tumor-initiation without lineage-resolved endogenous evidence." SC6 is limited because the field-valid endogenous tests are negative for irreplaceability. The negative result matters: an elegant perturbation does not cross Δ for a proposition when it produces the opposite of that proposition's predicted outcome.

SC1, SC4 and SC5 clear the Γ/Δ boundary within the systems in which their perturbations were run. Native lineage tracing establishes hierarchy in specified tumours [5,6]. Endogenous cell-state ablation changes maintenance endpoints in autochthonous glioma and lung tumours [7,14]. Relapse-specific tracing and ablation change reinitiation endpoints [7,12], with human AML tracking providing a separate disease-origin line [10]. Their mapped failures of generalization put them at Ε rather than forcing them back to Δ. The cap-versus-threshold collision is resolved the way the framework directs: a perturbation crosses a threshold within the system it was actually run in, while a universal cap governs the inference from that system to the claimed scope. The two rules bind on different subclaim scopes, and the evidence classes are separated rather than averaged.

Nine endogenous decisive perturbations support the Ε-level context-specific subclaims: Driessens 2012, Schepers 2012, Chen 2012, de Sousa e Melo 2017, Fumagalli 2020, Cañellas-Socias 2022, Chan 2026, Jiang 2026 and Perelli 2025 [5–8,11,12,14,18,19]. Mzoughi et al. is counted separately from these, as a transplanted-organoid perturbation rather than an autochthonous one [13]; neither human trial is counted, because neither demonstrated selective depletion of a declared population [16,17].

The upper ladder remains reachable in cancer biology without a randomized mechanism trial, and each field-valid equivalent was checked. Human clonal histories are available but do not identify one universal fixed state [10]. Comparable cross-population dose-response for a prospectively defined cancer-stem-cell fraction is absent. Withdrawal and regrowth often demonstrate replenishment [8,9,13]. Therapeutic perturbation with a hard endpoint now exists in humans in both directions [16,17], but neither trial demonstrates selective elimination of a prospectively defined rare compartment, and in mice the dependencies differ by tumour type and stage [7,8,12–14]. The convergent set supports the dynamic alternative at Ζ; it does not support the rare fixed nominal model above Γ.

Falsification ledger

Five prespecified falsification patterns are met:

1. Rarity fails under a more permissive assay. Melanoma tumour-forming frequency rises sharply under permissive xenograft conditions, to roughly a quarter of unselected cells [3,4].
2. A proposed cancer-stem-cell population is not universally required for maintenance. Primary colorectal tumours persist or regrow after Lgr5-positive-cell depletion [8,9].
3. A proposed marker-defined population is not universally required for relapse. Preoperative Lgr5-positive-cell ablation does not prevent colorectal recurrence, while EMP1-high residual-cell ablation does [12].
4. Cells outside the state reacquire equivalent function. Differentiated and Lgr5-negative colorectal cells regenerate tumour-sustaining states in native and organoid settings [8,9,11,13], through a regulated dedifferentiation route [18]; non-high-plasticity lung cells acquire the state on transplantation [14].
5. Lineage evidence contradicts a stable hierarchy in part of the claimed context. Invasive squamous cancer differs from papilloma in the same clonal-analysis program [5].

Each downgrade is clause-specific. None implies that tumour-sustaining states are epiphenomenal. They show that a rare fixed compartment is not the right universal unit.

Steelman and final adjudication

The case one rung higher, written first. A Δ or Ε headline is defensible if the reader's question is "do stem-like states causally matter in cancer?" They do, and the evidence is good: native lineage tracing, maintenance ablation in two autochthonous systems, relapse-specific ablation, patient-derived dormancy models, and human clonal tracking. Under primary-subclaim aggregation on SC4 the headline would be Ε, and that is the one rule of the three that returns it. The case fails as the headline for this entry because the rule prescribed for this hypothesis is scope-matched, because the claim is explicitly conjunctive, and because the clauses stated across tumour types are the ones that do not clear the boundary. Promoting a context-restricted subclaim to stand for a universal conjunction would answer a question the sentence does not ask.

The case one rung lower. A Β reading would follow if the marker-defined populations were only correlated with outcome. They are not. Selective ablation of declared states changes growth, survival and recurrence endpoints in autochthonous glioma, autochthonous lung cancer, and a colorectal surgical-recurrence model [7,12,14]. That is causal manipulation with disease-relevant endpoints, and it puts the floor above Β decisively.

Γ is therefore the correct letter for the conjunction as stated, and the grade must always travel with its qualifier: context-specific stem-like dependencies are established; the universal rare fixed hierarchy is not. To reach Ζ, the nominal claim would need independent endogenous confirmation of the same prospectively defined rare compartment across maintenance and relapse, plus durable exclusion of replenishment and of the major clonal and plastic alternatives.

The direction-of-causation procedure classifies the mechanism as reinforcing / stabilizing factor, by the framework's five-step derivation applied in order. It is not a downstream consequence, because ablating the state changes hard endpoints rather than the phenotype proceeding unchanged. It is not an epiphenomenon, because causal dependency is demonstrated. It is not a primary driver, which requires all three of prevention on removal, initiation on engagement, and rate limitation: removal does not prevent the phenotype across the claimed scope, and rate limitation is not shown generally. It is not a permissive factor, because it is not required — it is replenishable — and it does initiate in transplantation. What remains, at step five, is a factor that amplifies, accelerates or maintains a phenotype initiated upstream.

PROPOSED DECISIVE EXPERIMENTS

1. Cross-tumour prospective state-depletion atlas

In predeclared autochthonous models spanning haematologic and solid tumours, prospectively define the proposed apex state, label it, ablate it after tumour establishment, and follow every remaining clone with heritable barcodes through maintenance and treatment. The readouts should be tumour growth, recurrence-free survival, state-transition trajectories, and restoration of tumour-propagating function. Matched ablation of an equally sized non-apex population, calibrated target coverage, and state-specific rescue are essential controls. Complete excision of the ablation cassette must be verified before residual cells are interpreted as bypass rather than escape — the distinction that limits the current lung and glioma results [7,14].

The nominal model is falsified if tumours persist or recur because cells outside the state regenerate equivalent function. It is supported if complete ablation causes durable collapse across the predeclared tumour contexts, no outside lineage restores the function, and matched rescue restores growth.

2. Human residual-disease lineage and functional linkage

Collect paired primary tumour, minimal residual disease, and relapse samples before and after therapy. Combine phylogenetic lineage reconstruction and state-resolved profiling with selective depletion, matched rescue, and in-vivo confirmation in patient-derived models. The key test is whether relapse founders originate exclusively from the prospectively declared state and remain dependent on it.

Genetic clone identity must be separated from reversible transcriptional state, and the assay must quantify the sensitivity for rare alternative founders. Multiple baseline origins or re-entry into the state from outside it would falsify the fixed hierarchy. Exclusive state-restricted relapse origins plus durable dependency and rescue-specific restoration would support it.

3. Rarity-versus-assay calibration

Measure the same prospectively isolated cells using native intravital lineage tracing, orthotopic transplantation, and graded recipient permissiveness. Fix the state definition before outcomes. Equalize viability, cell-cycle distribution, niche access, barcode recovery, and statistical uncertainty in limiting dilution.

A large increase in estimated tumour-sustaining frequency with recipient permissiveness would confirm that rarity is assay-dependent. A consistently rare fraction carrying durable function across native and permissive settings would strengthen SC2.

4. A population-selective human trial

The two available human trials test agents, not populations [16,17]. A trial that prospectively defines a target state, demonstrates on-treatment depletion of that state in serial patient samples, and takes recurrence-free survival as its endpoint would be the first human test of the claim's actual content. Depletion must be verified in the patient, not inferred from the agent's mechanism, and the analysis must prespecify whether benefit tracks depletion of the state or general cytotoxicity.

KNOWLEDGE GAPS

• State versus clone: In many human tumours, durable tumour-sustaining capacity may track a genetic clone, a reversible cell state, or their interaction. This distinction determines whether selective targeting can be durable.
• Ablation completeness versus biological bypass: Residual growth can reflect incomplete depletion, transgene-silent escape, or true state replenishment. Chen et al. and Chan et al. both document technical escape explicitly [7,14]; necessity requires these possibilities to be separated quantitatively rather than assumed either way.
• Human relapse causality: Human phylogenies identify relapse origins, but direct state-specific necessity and rescue remain mainly model-based.
• Durability after multi-state targeting: Combined-state depletion delays resurgence in a transplanted-organoid system [13], but post-withdrawal durability, autochthonous confirmation, and generalization across colorectal genotypes remain unresolved.
• Tumour-type denominator: A universal claim requires prospectively sampled negative contexts, not only cancers already selected because they display a hierarchy.
• Population-selective human evidence: No trial has demonstrated depletion of a prospectively defined state in patients and measured recurrence against it.

REVIEWER NOTES

Red-team synthesis note

This edition is the published product of a cross-model red-team pass on 2026-09-30. The input review was authored by a different model family; the red team and synthesis were performed by Claude. Sixteen findings were raised, twelve accepted and four rejected after checking.

The headline letter moved from Ε to Γ, on aggregation logic rather than on new evidence, and no subclaim level changed. The prescribed scope-matched rule takes the highest subclaim whose scope matches the central claim's scope. The three subclaims that reached Ε reached it by narrowing to defined tumour contexts, which is what Ε means; the three that carry the claim's stated cross-tumour scope — rarity, sufficiency, irreplaceability — all sit at Γ. The hypothesis-index entry for this hypothesis directs that the universal framing be scored in the weakest tumour context claimed. The input review reported the Γ reading three times as an alternative while keeping Ε as the headline; this edition promotes it to the headline, where the rule puts it.

Substantive evidence corrections: an unverifiable pair of cell counts in the breast-cancer study was replaced with the figures its abstract reports; Mzoughi et al. was reclassified from an autochthonous to a transplanted-organoid perturbation, lowering the endogenous decisive-perturbation count from eight to seven; Chan et al.'s replenishment result was corrected to the transplantation experiment that produced it, with the authors' own incomplete-recombination caveat restored; two randomized human trials of stem-cell-directed strategies were added, one null and one positive, together with two 2025–2026 dedifferentiation studies. Confidence, the role of mechanism, the rubric routing and all six subclaim levels were re-derived independently and retained. All nineteen references were resolved against primary records; none of the confabulated-identifier problems documented elsewhere in this corpus were found in this package.

Full detail, including the four rejected findings and the citation-verification table, is in reviews/cancer-stem-cell/redteam/redteam-pass-2026-09-30/REDTEAM_REPORT.md; the itemised before-and-after list and the open issues are in reviews/cancer-stem-cell/redteam/redteam-pass-2026-09-30/SYNTHESIS_CHANGE_LOG.md. Both paths are relative to the repository root and lie outside this package.

Prior-review comparison

The independent assessment was locked before the earlier file at D:\HypothesisCritic\research\assessing_the_cancer_stem_cell_hypothesis.md was opened. That review assigned Ε with High confidence; this edition assigns Γ with High confidence, for the aggregation reason above. It also changes the role from primary driver to reinforcing / stabilizing factor, because a universal initiating and rate-limiting compartment is not demonstrated and multiple systems show replaceability.

The earlier review also contains a factual inversion in its discussion of Driessens et al. It states that the transient population cycled rapidly at twice per day. The paper's model assigns the approximately twice-daily rate to the persistent stem-like population and, by its stated four-fold ratio, roughly 0.5 divisions per day to committed progenitors [5]. This correction was made by the input review, independently confirmed in the red-team pass, and is retained. It does not change the letter but matters for the mechanistic description.

The original prior review was preserved unchanged. Its SHA-256 is FDEBD218C825B2C78FD33C3B9D1EA2B982260EE147FA0E651E97C15A1B21E340.

Source access and limits

Complete full text was read for the load-bearing perturbation studies: Quintana 2008 and 2010, Driessens 2012, Schepers 2012, Chen 2012, Fumagalli 2020, Cañellas-Socias 2022, Mzoughi 2025, and Chan 2026 [3–7,11–14]. Bonnet 1997, Al-Hajj 2003, de Sousa e Melo 2017, Shimokawa 2017, Shlush 2017, Ohta 2022, the two clinical trials, and the two 2025–2026 dedifferentiation studies [1,2,8–10,15–19] were assessed at abstract/metadata depth. No abstract-only source carries an exact effect size beyond what was available in its abstract; the trial effect sizes quoted above are reported in their abstracts verbatim.

Al-Hajj 2003 deserves a specific note. Its PubMed Central record returned no retrievable full text, Europe PMC's full-text service returned a server error, and the publisher and PMC web views were inaccessible. The input review reported cell counts of 200 versus 20,000 from that paper at full-text depth; because no accessible record substantiates those figures, this edition uses the abstract's reported values instead and records the row at abstract/metadata depth. Nothing in the assessment turns on which pair of numbers is used.

The HSEF packet was treated as methodology, not as biological evidence. No sealed calibration target was supplied, so this is not a calibration entry. It is an independent revision with a prior-review comparison, followed by a cross-model red-team synthesis.

Validation meaning

The machine validator checks schema presence, level/descriptor consistency, aggregation arithmetic, cap ceilings, evidence identifiers, source status, claim-to-source links, and required prose blocks. A pass means that those structural and internal-consistency checks succeeded. It is not an independent replication of the biological studies or an external scientific endorsement. In particular, the validator cannot detect a citation whose identifier resolves to a different paper than the one described, which is why every identifier in this edition was resolved by hand.

ADVERSARIAL REVIEW AND APPROVAL

Authoring model: Sol 6.1. Adversarial reviewer: Claude Opus 5.5 (Anthropic). Date: 2026-10-14.

This is the 2026-10-09 revised edition of a review dated 2026-09-30 that had already been through a cross-model red-team synthesis. Findings below mark whether a defect was introduced by the 2026-10-09 revision or inherited.

Scope of checks. All 19 REFERENCES PMIDs and DOIs verified against PubMed get_article_metadata (19/19 resolve to the stated title, first author, journal and year within the plus-or-minus-1 print/e-pub convention; DOIs match; no retractions or expressions of concern; the recorded Al-Hajj author correction is a correction, not a retraction). Twelve load-bearing claims spot-checked against original abstracts, with particular attention to the two human trials and the melanoma rarity figures, which carry the falsification ledger. Scoring logic checked: the cap ceilings on SC2, SC3 and SC6, cap precedence, the pre-registered scope-matched aggregation rule and its reported alternatives, the Beta/Gamma/Delta boundary, the upper-ladder reachability check, the confidence-band text, the role-of-mechanism derivation, the Momentum artifact requirement, the competing-model tournament fields and the causal-tally arithmetic. Provenance claims in the spliced Consensus block were checked against the 2026-10-06 adversarial summary and against this case's own prepublication artifacts.

Rejected findings (checked, no change).

1. Scope-matched aggregation (the specific defect red-teamed in September). Checked in detail, because the earlier edition of this entry misapplied the rule as "highest subclaim" rather than "highest subclaim whose scope matches". This edition does not repeat it. It states the rule verbatim, identifies SC2, SC3 and SC6 as the subclaims carrying the central claim's cross-tumour scope, notes that SC1, SC4 and SC5 reach Ε only by narrowing - which is what Ε denotes - and returns Γ, reporting the narrower subclaims separately. It also prints all three rules (weakest-link Γ, primary-subclaim on SC4 Ε, scope-matched Γ) and says in terms that only the rule that was not prescribed returns Ε. The hypothesis index's own wording is quoted in support. Correct as written.
2. Confidence band. The rationale rests on evidence density and reproducibility - endogenous lineage and ablation studies, patient-derived systems, human relapse tracking, two randomized trials and direct falsification of fixed-state predictions - and explicitly distinguishes confidence in the assessment from belief in the proposition. It makes no appeal to the reviewer's access depth, so the two-step confidence test leaves it untouched. High is consistent with the High text given replication across labs, models and assay modalities. Retained.
3. "Best-supported rival". The named best rival is the rank-1 dynamic-hierarchy model; the nominal sits at rank 3. No renderer substitution of the nominal for the rival. No change.
4. Causal-tally arithmetic. The body reports nine endogenous decisive perturbations while the REVIEWER NOTES describe the red team lowering the count "from eight to seven". These are consistent: causal_tally_audit.tally_arithmetic records 8 minus Mzoughi (reclassified as a transplanted-organoid perturbation) = 7, plus the two studies added in the red-team pass (Jiang 2026 and Perelli 2025) = 9. The seven rejected items are itemised. No change.
5. Human-trial numbers. CanStem303C (n=624 vs 629; median OS 14.3 vs 13.8 months; HR 0.976; one-sided P=.74; 44% biomarker-positive, n=275 vs 272; 13.2 vs 12.1 months; HR 0.969; one-sided P>.99) and the glasdegib phase II (n=88 vs 44; 8.8 vs 4.9 months; HR 0.51, 80% CI 0.39-0.67; P=0.0004; complete remission 17.0% vs 2.3%) match their abstracts exactly, including the review's refusal to read either as population-selective.
6. Melanoma rarity figures. PMID 19052619 reports approximately 25% of unselected cells forming tumours in limiting dilution across 12 patients and an average of 27% in single-cell transplants across four patients; PMID 21075313 reports 28% of single cells and that none of 22 heterogeneously expressed markers enriched tumorigenic cells. Both match. The per-patient 12-65% range and the 69-of-254 denominator come from the full text, which the review records reading at complete depth.
7. Headline level. Γ was argued up to Ε and down to Β, and the review's own steelman already makes both cases; both were re-checked and rejected on the same grounds. Retained.
8. Reviewer-notes contradiction (the known Oct 9 splice artifact). Does not apply: the preserved REVIEWER NOTES make no claim that Consensus work was absent, and the spliced block itself records which Consensus artifacts were located and explicitly declines to infer a comparison count from them.
9. Collective-author, empty-DOI and "PMIDnot supplied" defects, and missing-space rendering defects. None present in this review; its reference block uses a different builder from the v5.3 batches and is clean.

Residual limitations. Ten of nineteen sources were read at abstract/metadata depth, and Al-Hajj 2003 was inaccessible at full text, so the review's substitution of abstract-reported figures there remains unverifiable in either direction. Scite was unavailable, so citation-context counts are absent rather than zero. No literature search was re-run, so completeness at the review's cutoff is not certified.

Approval: Following adversarial review and the corrections listed above, the revised document is approved. - Claude Opus 5.5 (Anthropic), 2026-10-14.

REFERENCES

1. Bonnet D, Dick JE. Human acute myeloid leukemia is organized as a hierarchy that originates from a primitive hematopoietic cell. Nature Medicine. 1997;3:730–737. PMID 9212098 · DOI 10.1038/nm0797-730
2. Al-Hajj M, Wicha MS, Benito-Hernandez A, Morrison SJ, Clarke MF. Prospective identification of tumorigenic breast cancer cells. PNAS. 2003;100:3983–3988. PMID 12629218 · DOI 10.1073/pnas.0530291100. Author correction: DOI 10.1073/pnas.1131491100.
3. Quintana E, Shackleton M, Sabel MS, Fullen DR, Johnson TM, Morrison SJ. Efficient tumour formation by single human melanoma cells. Nature. 2008;456:593–598. PMID 19052619 · DOI 10.1038/nature07567
4. Quintana E, Shackleton M, Foster HR, et al. Phenotypic heterogeneity among tumorigenic melanoma cells from patients that is reversible and not hierarchically organized. Cancer Cell. 2010;18:510–523. PMID 21075313 · DOI 10.1016/j.ccr.2010.10.012
5. Driessens G, Beck B, Caauwe A, Simons BD, Blanpain C. Defining the mode of tumour growth by clonal analysis. Nature. 2012;488:527–530. PMID 22854777 · DOI 10.1038/nature11344
6. Schepers AG, Snippert HJ, Stange DE, et al. Lineage tracing reveals Lgr5+ stem cell activity in mouse intestinal adenomas. Science. 2012;337:730–735. PMID 22855427 · DOI 10.1126/science.1224676
7. Chen J, Li Y, Yu TS, et al. A restricted cell population propagates glioblastoma growth after chemotherapy. Nature. 2012;488:522–526. PMID 22854781 · DOI 10.1038/nature11287
8. de Sousa e Melo F, Kurtova AV, Harnoss JM, et al. A distinct role for Lgr5+ stem cells in primary and metastatic colon cancer. Nature. 2017;543:676–680. PMID 28358093 · DOI 10.1038/nature21713
9. Shimokawa M, Ohta Y, Nishikori S, et al. Visualization and targeting of LGR5+ human colon cancer stem cells. Nature. 2017;545:187–192. PMID 28355176 · DOI 10.1038/nature22081
10. Shlush LI, Mitchell A, Heisler L, et al. Tracing the origins of relapse in acute myeloid leukaemia to stem cells. Nature. 2017;547:104–108. PMID 28658204 · DOI 10.1038/nature22993
11. Fumagalli A, Oost KC, Kester L, et al. Plasticity of Lgr5-negative cancer cells drives metastasis in colorectal cancer. Cell Stem Cell. 2020;26:569–578.e7. PMID 32169167 · DOI 10.1016/j.stem.2020.02.008
12. Cañellas-Socias A, Cortina C, Hernando-Momblona X, et al. Metastatic recurrence in colorectal cancer arises from residual EMP1+ cells. Nature. 2022;611:603–613. PMID 36352230 · DOI 10.1038/s41586-022-05402-9
13. Mzoughi S, Schwarz M, Wang X, et al. Oncofetal reprogramming drives phenotypic plasticity in WNT-dependent colorectal cancer. Nature Genetics. 2025;57:402–412. PMID 39930084 · DOI 10.1038/s41588-024-02058-1
14. Chan JE, Pan CH, Rub J, et al. Critical role for a high-plasticity cell state in lung cancer. Nature. 2026;650:231–241. PMID 41565826 · DOI 10.1038/s41586-025-09985-x
15. Ohta Y, Fujii M, Takahashi S, et al. Cell-matrix interface regulates dormancy in human colon cancer stem cells. Nature. 2022;608:784–794. PMID 35798028 · DOI 10.1038/s41586-022-05043-y
16. Shah MA, Yoshino T, Tebbutt NC, et al. Napabucasin plus FOLFIRI in patients with previously treated metastatic colorectal cancer: results from the open-label, randomized phase III CanStem303C study. Clinical Colorectal Cancer. 2023;22:100–110. PMID 36503738 · DOI 10.1016/j.clcc.2022.11.002
17. Cortes JE, Heidel FH, Hellmann A, et al. Randomized comparison of low dose cytarabine with or without glasdegib in patients with newly diagnosed acute myeloid leukemia or high-risk myelodysplastic syndrome. Leukemia. 2019;33:379–389. PMID 30555165 · DOI 10.1038/s41375-018-0312-9
18. Jiang Q, Raghavan MS, Rodriguez AM, et al. ZFP36L2 orchestrates stress-adaptive plasticity in regeneration and cancer. Nature. 2026;657:1053–1063. PMID 42557330 · DOI 10.1038/s41586-026-10890-0
19. Perelli L, Zhang L, Mangiameli S, et al. Evolutionary fingerprints of epithelial-to-mesenchymal transition. Nature. 2025;640:1083–1092. PMID 40044861 · DOI 10.1038/s41586-025-08671-2

Documentary record

Every grade on this site is reproducible from its record. The files below are the posted edition exactly as assessed.

• ADVERSARIAL_REVIEW.md
• SHA256.txt
• adversarial-findings.json
• cancer-stem-cell.review.md
• cancer-stem-cell.scorecard.json
• codex-correction-provenance.json
• consensus-addendum.md
• consensus-validation.json
• edition-approval-audit.json
• research-update-2026-10-05.bib
• research-update-2026-10-05.html
• research-update-2026-10-05.json
• research-update-2026-10-05.md
• research-update-2026-10-05.ris
• validation-report.json
• validation-round2-2026-10-14.json
• validation.txt
• video-insert-2026-10-05.md

Sources: 19 · load-bearing claims: 17 · verified: 17 · unverified: 0 · contradicted: 0 · schema 5.4