Some tumors cheat. When a drug cuts off their blood supply, the most malignant cells do not simply die; a subpopulation of stem-like tumor cells loses its epithelial packing, becomes migratory, and physically rewires itself into hollow, fluid-conducting channels — a bypass circulation called vasculogenic mimicry (VM) that needs no host blood vessels at all. The result is a therapy that looks like a success on every conventional readout while the tumor quietly acquires a direct pipeline for metastasis (Danfang Zhang et al.). This is, almost literally, the failure mode that a parallel line of work in machine learning was built to catch: programmatic system-state verification, which evaluates software agents by inspecting the persistent state they leave behind — database contents, file systems, application records — rather than their self-reports, screenshots, or action sequences, because those proxies routinely pass agents whose real effects are broken or damaging (Echoverse, AppWorld).
The claim of this proposal is that these two problems are the same problem, and that nobody has yet connected them. A reinforcement learning agent that schedules cancer therapy against a simulated tumor, scored on tumor volume, is exactly the kind of agent that AppWorld showed can complete its primary task while committing unnoticed "collateral damage" in surrounding state tables. We propose to build the first in silico environment in which therapy-induced VM is a first-class, auditable persistent state, and the first verification layer that programmatically distinguishes a functional, perfused escape network from the superficial cellular clumping that marker-based or imaging-based checks mistake for it. The biological analogue of "the agent clicked the button but the function never fired" (Automotive-ENV) is "the tumor stains positive for channel markers but no fluid moves."
Core Claim
A drug-scheduling policy trained and evaluated on tumor-volume reward in an existing tumor-microenvironment simulator will pass every conventional check while leaving the simulated tumor in a high-connectivity VM state — and a programmatic state audit over the simulator's persistent fields (phenotypic composition, channel topology, and solved fluid flux) will catch this failure that the reward function cannot see. Adding this audit to training changes the learned policies toward schedules that do not trigger the hypoxic escape.
The claim has two halves. The first is a diagnosis: the hypoxic escape rate, the fraction of volume-reward policies that achieve tumor control but leave a percolating, perfused VM network, is nonzero and material. The second is a fix: policies trained or verified against the audit avoid it. Both halves are testable in simulation, on top of an existing open platform.
Research Program
The environment. We extend PhysiGym, which wraps the open-source PhysiCell agent-based cell simulator in the standard Gymnasium interface and already trains RL agents to modulate drug dose in a 2D tumor microenvironment of 512 tumor cells, 128 immune cells, and four diffusing chemicals. PhysiGym is the right substrate but deliberately minimal: its cells only proliferate, die, and get killed, and its policies are scored by discounted scalar return. We add three biological mechanisms, each grounded in the cited literature rather than invented. First, a cancer-stem-cell subpopulation with the plasticity to adopt an endothelial-like phenotype, motivated by spheroid studies showing tumor-initiating cells enriched for exactly this differentiation capacity (SN12C spheres). Second, a hypoxia-driven plasticity switch: when the local oxygen field falls below a threshold, cells activate a HIF-1α-like state that raises a Twist1/LOXL2-like EMT propensity, matching the documented chains from hypoxia through LOXL2 to channel assembly in hepatocellular carcinoma (HIF-1α/LOXL2 in HCC, LOXL2 and tumour aggressiveness) and the sunitinib–hypoxia–Twist1–CSC expansion loop in triple-negative breast cancer (Zhang et al.). Third, channel assembly: mesenchymal-phenotype cells under sustained hypoxia rearrange into hollow cords whose conductivity depends on matrix remodeling, so that destroying host vessels with an anti-angiogenic action worsens hypoxia and accelerates the bypass — the paradox documented clinically and experimentally (Zhang et al., Wang et al.). We calibrate the switch qualitatively against these papers' reported directions of effect; none of them reports a dynamical rate we can copy, so parameter sweeps over the plausible range are part of the study, not a nuisance.
The verification layer. This is the contribution that has no precedent in the biological simulation literature. Following the design logic of Copy-on-Write Scoring, which decomposes agent evaluation into a structural score over the shape of database changes and a content score over their values, we define two programmatic scores over the simulator's persistent state at any step. The structural score $s^{struct}$ inspects the graph of channel-forming cells: node degrees, connected components, and whether the VM network percolates between the simulated host-vessel boundaries. The content score $s^{content}$ inspects the field values: local oxygen, drug concentration, and the fraction of tumor mass in the stem-like or mesenchymal phenotype. The critical and novel piece is that neither score accepts morphology alone. To count as a functional channel rather than a passive cord, a conduit must pass a flow check: we solve a pressure-driven transport problem on the channel graph and require the resulting flux to exceed a physiological threshold. This is the direct analogue of Echoverse scoring agents on the raw SQL diff instead of the interface, and of Automotive-ENV catching actions that completed visually while the underlying function never triggered. The biophysical verification gap — that VM channels stain positive for PAS and matrix proteins but nobody can tell perfused conduits from debris in standard histology — is the open problem this mechanism transfers wholesale into a domain where it becomes computable.
The audit itself is a conjunction of programmatic unit tests, in the spirit of AppWorld's state-based tests that explicitly flag unintended modifications to unrelated state. A policy passes a trial only if
$$\text{volume reduction} \geq \tau_v \;\;\text{and}\;\; \frac{J_{\text{VM}}}{J_{\text{baseline}}} < \tau_c,$$
where $J_{\text{VM}}$ is the total solved fluid flux through tumor-derived channels at trial end and $J_{\text{baseline}}$ is the flux through an untreated control, so the second condition fails exactly when the tumor has built a perfused bypass. The primary new metric is the hypoxic escape rate: the fraction of otherwise-successful policies that fail the second condition.
Policy training under audit. We train dose-scheduling agents twice: once with the stock volume reward (the condition under which we predict escape), and once with the audit expressed as a constraint or shaped penalty, and compare the learned schedules. The interesting scientific output is qualitative: we expect audit-aware policies to avoid the maximally aggressive anti-angiogenic bursts that induce the deepest hypoxia, echoing how PhysiGym's RL agent already discovered a non-intuitive two-phase regime (aggressive burst, then maintenance) that static schedules missed.
Counterfactual branching, decided empirically. Copy-on-Write Scoring achieves non-destructive, per-operation evaluation by redirecting writes to a changes table instead of cloning sandboxes, at a cost of roughly 250 lines of core code in Plane. We will evaluate whether an analogous transactional branching of simulator state pays off here — it would let one tumor trajectory serve many counterfactual dose perturbations without full state copies — but we treat this as a measured engineering question rather than a claim, because at PhysiGym's current scale (hundreds of cells, 2D) plain resets may be cheap enough. We report the crossover point and adopt whichever wins; this honesty is deliberate, and expanded below.
Baselines
The naive reference point is an untreated simulation and a fixed maximum-tolerated-dose schedule, evaluated by tumor volume alone. This is the clinical default and the "do nothing" of evaluation methodology: it cannot, even in principle, detect VM, because no quantity it reads depends on the bypass network. Its role is to fix the floor and to demonstrate the blindness concretely.
The strongest existing competitor is the PhysiGym image-based RL agent itself (PhysiGym): the only published trained therapy policy in this simulator family, which observes the microenvironment as an image and modulates drug dose, reporting a discounted cumulative return of about 46 versus about 44 for a scalar-observation agent after 500k training steps. It is the strongest competitor because it is the current best method for exactly our task — controlling the same simulated tumor under the same physics — and because our claim is that it succeeds on its own metric while failing a state audit it was never given. For a fair comparison we run it in our extended environment unchanged: same observation space, same action space, same 500k-step training budget, same hyperparameter search, and score it both on its native return and on our audit. Beating it means achieving comparable tumor-volume control while reducing the hypoxic escape rate from whatever level the volume-reward agent exhibits to near zero under the audit-trained agent, at matched training compute.
The ablation that isolates our key ingredient replaces the flow-based functional verification with a marker-count check: count PAS-analogue-positive, channel-marker-positive structures, exactly as standard histology does. This ablation keeps the biology, the environment, the training budget, and even the audit interface identical, changing only whether the verifier solves for fluid transport or counts morphology. If the marker-count verifier passes policies that the flow verifier fails — and the biology strongly suggests it will, since the entire open problem is that stained structures are not necessarily perfused ones — then the functional flow check is shown to be the load-bearing component, not the added biology or the added training signal.
A second ablation adds a scalar VM penalty directly to the reward without any programmatic verifier, to test whether the audit contributes anything beyond richer reward shaping. If reward shaping alone matches the audit's escape-rate reduction, the verification layer's value shrinks to evaluation hygiene rather than training signal, and we will report that plainly.
Why not just…?
Why not just use standard RL reward on tumor volume, or an LLM judging the outcome? This is the most important objection and the proposal is built around it rather than around an answer to it. The software-agent literature already established that self-reporting and model-based judging fail on multi-step execution with hidden state (OS-Copilot) and that proxy metrics pass agents that damage unrelated persistent state (AppWorld). But that is analogy, not evidence about tumors. The decisive experiment is internal to this proposal: if the volume-reward agents in our extended environment show a negligible hypoxic escape rate, the core claim dies and the objection was right. If they escape at a material rate — which the documented sunitinib paradox (Zhang et al.) makes the expected outcome whenever the reward cannot see the channel state — then scalar reward is demonstrably insufficient in this domain, by direct measurement rather than by citation.
Why not just reset the environment between episodes instead of building copy-on-write branching? At the current scale, this objection is simply correct, and we have demoted CoW from a claim to a measured option accordingly. Gymnasium's standard reset is cheap for a 2D simulation of a few hundred cells, and Copy-on-Write Scoring's own motivation — sandbox replicas are resource-intensive and drift-prone — bites only when state is large. Our experiment will time both approaches across simulation scales and report the crossover; CoW-style branching earns a place only if it lets us run the counterfactual perturbation grid (many dose variants per base trajectory, needed for the audit's per-step attribution) at materially lower cost. If resets win everywhere we can reach, we say so and move on.
Why not just evaluate adherence to established clinical dosing protocols? Because sequence-matching is the exact evaluation style the verification literature discredits: MobileAgentBench and AndroidLab show that forcing agents onto a rigid reference trajectory penalizes valid alternatives and makes the metric brittle. The oncological version of this brittleness is worse than an inconvenience: fixed aggressive schedules are precisely what the cited work implics induces the escape — conventional chemotherapy promoted non-angiogenic blood-lake formation and resistant clonal expansion in telangiectatic osteosarcoma (Wang et al.), and sunitinib schedules expanded the channel-forming stem compartment (Zhang et al.). A protocol-adherence score would award full marks to the very schedules our audit is designed to flag. We still include a fixed-schedule baseline, but as a subject of the audit, not as the evaluator.
Decisive Checks
Three results would settle the proposal. First, the diagnosis: with the audit applied post hoc to volume-reward agents from PhysiGym-style training, the hypoxic escape rate should be material across the biological parameter sweep — if it is near zero everywhere, the premise is wrong and we report the negative result, which would itself be informative about how robust the simulated paradox is to parameterization. Second, the functional-versus-superficial gap: the marker-count ablation should pass a substantial fraction of the policies the flow-based audit fails, quantifying in silico the perfusion blind spot that histology has in vitro; the AppWorld collateral-damage audit is the template for reporting this as a table of passes, failures, and false acquittals. Third, the fix: audit-trained agents should achieve tumor control within a few percent of the volume-reward agents' volume reduction (which we can compare against PhysiGym's reported ~46 discounted return for its image-based agent at 500k steps as a sanity anchor) while driving the escape rate toward zero, and their discovered schedules should be systematically less hypoxia-inducing. A fourth, honest check we will run whether or not it flatters us: whether plain reward shaping matches the audit, which determines whether programmatic verification here is a training tool or only an evaluation discipline. Either answer closes a real question that, on the evidence assembled here, no one has yet asked.
References
- Twist1 expression induced by sunitinib accelerates tumor cell vasculogenic mimicry by increasing the population of CD133+ cells in triple-negative breast cancer
- AppWorld: A Controllable World of Apps and People for Benchmarking Interactive Coding Agents
- Automotive-ENV: Benchmarking Multimodal Agents in Vehicle Interface Systems
- PhysiGym: bridging the gap between the Gymnasium reinforcement learning application interface and the PhysiCell agent-based model software
- Spheres derived from the human SN12C renal cell carcinoma cell line are enriched in tumor initiating cells
- HIF-1α promoted vasculogenic mimicry formation in hepatocellular carcinoma through LOXL2 up-regulation in hypoxic tumor microenvironment
- <scp>LOXL</scp>2 promotes vasculogenic mimicry and tumour aggressiveness in hepatocellular carcinoma
- Telangiectatic osteosarcoma of the thoracic vertebra: a case report and literature review
- Copy-on-Write Scoring: Application-Specific Agent Evaluations
- OS-Copilot: Towards Generalist Computer Agents with Self-Improvement
- MobileAgentBench: An Efficient and User-Friendly Benchmark for Mobile LLM Agents
- AndroidLab: Training and Systematic Benchmarking of Android Autonomous Agents