Apodex 1.1: Scaling Agentic Intelligence for Complex Work

• Working capability turns reasoning into sustained work: finding evidence, operating on files and code, maintaining and revising plans, recovering from failures, and delivering inspectable artifacts.Its unit is completed work, including progress through tools and stateful environments, adaptation to changing observations, and satisfaction of a delivery contract.Source: §1 · p. 3
• Environment Scaling expands diverse, faithful, and verifiable file, search, and code worlds so executable trajectories become part of the learning distribution.These environments specify state, tools, transitions, interaction budgets, failure conditions, and completion checks.Source: §1 · pp. 3–4
• Agentic Coordination Scaling trains agents to decompose objectives, delegate work, integrate asynchronous results, revise shared plans, and reorganize unfinished branches.Agent Team realizes these behaviors through adaptive parallel effort across evidence gathering, file analysis, implementation, verification, and counteranalysis.Source: §1 · pp. 3–4
• A common execution harness connects the model to File, Search, and Code environments, while AgentOS preserves workspace, artifact, provenance, and branch state across tools and agents.The harness also defines observations and completion contracts and provides replay and verification hooks.Source: §1 · pp. 3–4
• Unified SFT and agentic RL use executable environment trajectories and coordination traces to improve task execution, recovery, delivery, and coordination within one policy.Real tasks, benchmark errors, and runtime failures feed a managed loop in which Task Pipeline generates tasks, environments supply worlds and verifiers, and evaluation, expert review, and user feedback guide training effort.Source: §1 · pp. 3–4
• The full Apodex 1.1 Agent Team system reaches the strongest values shown in FrontierFinance and FrontierScience-Research comparisons and remains competitive across professional work, reasoning, search, mathematics, and coding.Apodex 1.1 mini, with 35B parameters, reaches the performance band of selected frontier systems and improves markedly over Apodex 1.0 mini on overlapping tasks; evaluation also covers science, finance, professional file work, mathematics, coding, internal long-horizon tasks, and end-to-end cases.Source: §1 · p. 4

2.1 Completed Work Is the Unit of Agentic Intelligence

• Success requires a useful, verified delivery under a fixed objective and contract, with a defensible path that preserves valid work, revises plans, and recovers from failures.Clarifications that do not change the requested outcome remain within the same task contract; material changes to the objective or deliverables create a new contract.Source: §2.1 · pp. 5–6

2.2 Executable Environments Are a Scaling Surface

• Environment Scaling expands executable file, search, and code task contracts, balancing diversity with fidelity and verification while training workflows that connect evidence, code, and authoritative artifacts.File environments teach artifact handling, search environments teach evidence acquisition and reconciliation, and code environments teach transformation, testing, and recovery.Source: §2.2 · p. 6

2.3 Agentic Coordination Is a Scaling Surface

• Agentic Coordination Scaling trains decomposition, delegation, result integration, and replanning, with staged returns, shared state, and termination enabling parallel work to adapt as evidence or priorities change.Its value lies in moving results, decisions, and user feedback through an active task, not merely increasing the number of agents or samples.Source: §2.3 · pp. 6–7

2.4 A Common Harness Connects Both Scaling Dimensions

• A common harness aligns environment and coordination execution, while AgentOS maintains persistent task state so interventions, dependency changes, and recovery preserve unaffected artifacts and progress.When dependencies change, descendants can be invalidated and pending work revised without discarding causally independent progress.Source: §2.4 · p. 7

2.5 Training Exploits Both Scaling Dimensions

• Training converts executable environment and coordination trajectories into capability through unified SFT and agentic RL; runtime failures, benchmark errors, and human feedback guide subsequent development.Training is how the model learns from the first two scaling dimensions, rather than a third independent scaling axis.Source: §2.5 · p. 7

2.6 Evaluation Mirrors the Scaling Design

• Evaluation measures working capability across correctness, artifact completeness, provenance, recovery, intervention, time, and compute, distinguishing underlying policy, coordinated execution, and end-to-end completed work.Minimal-scaffold ReAct runs expose the model policy, while Agent Team runs measure trained coordination with additional organized computation.Source: §2.6 · p. 8

3 Apodex 1.1

• Apodex 1.1 trains one policy to reason, search, manipulate files, execute code, recover from failures, coordinate parallel work, and deliver verifiable artifacts.Its workspace accumulates evidence, executable analyses, intermediate artifacts, and final deliverables under a fixed task objective and delivery contract.Source: §3 · p. 8
• Environment Scaling expands executable file, search, and code worlds, while Agentic Coordination Scaling develops decomposition, delegation, integration, verification, and revision across agents and time.These are the two scaling dimensions introduced in Section 2.Source: §3 · pp. 8–9
• A common harness binds both dimensions to persistent state, typed observations, replay, and completion checks, with AgentOS as the execution substrate.The following sections cover the scaling dimensions, runtime support, and training that converts trajectories into model capability.Source: §3 · pp. 8–9

3.1 Environment Scaling

• Environment Scaling broadens initial workspaces, objectives, actions, transitions, budgets, and delivery-verification pairs across file, search, and code worlds under one task contract.The goal is to expand coverage, fidelity, and structural depth, not merely add tools or prompts.Source: §3.1 · p. 9

3.1.1 Environment Families

• File worlds distribute authoritative information across heterogeneous, versioned workspaces, requiring agents to reconstruct relationships and produce professional artifacts.Coverage spans 33 domains, 318 occupations, and 1,208 deliverable clusters; difficulty also varies with contributing systems, business-logic depth, and inferred delivery requirements, not file count alone. Graded quantities are re-derived from authoritative sources or linked to documents through provenance.Source: §3.1.1 · pp. 9–10
• Search worlds train discovery, source inspection, evidence reconciliation, and stopping across distributed evidence, intermediate entities, and plausible non-authoritative sources.Their gold objects include relevant sources, claim-to-evidence alignment, and explicit uncertainty where sources conflict; acquisition uses structured search and fetch, with a constrained shell path when needed.Source: §3.1.1 · p. 10
• Code worlds combine reusable execution infrastructure with harvested pull-request tasks and synthesized variants, using sandboxed execution to verify task completion.For harvested tasks, fail-to-pass tests must fail before and pass after the reference change, while pass-to-pass tests succeed in both states. Synthesized tasks also probe reward bypasses; scoring is isolated from the solver, and failed executions help diagnose construction errors.Source: §3.1.1 · p. 10

3.1.2 Scaling Coverage and Difficulty

• Environment scaling targets recurring capability gaps, while difficulty calibration reflects the behavior each world is intended to teach.Acquisition pressure depends on plausible candidates, load-bearing evidence transitions, and tool-call budget, but is not a universal difficulty measure; code worlds use coordinates such as dependency depth, state transitions, test observability, and distance to executable feedback. Generated tasks return through family-specific verification before training.Source: §3.1.2 · pp. 10–11

3.1.3 Verification, Isolation, and Replay

• Trajectories contribute training signal only when worlds are reconstructible, execution is isolated, and verification can be replayed.Immutable manifests and fresh sandboxes separate solver observations from hidden verifier state; replay records seeds, versions, actions, observations, file deltas, and termination reasons. Independent checks and blind solver probes help detect verifier errors, while generators make construction cheap and inverse recovery by the agent constrained.Source: §3.1.3 · p. 11

3.2 Apodex Agent Team 1.1: Interactive Self-Organizing Teams

• Apodex 1.1 retains problem-driven delegation and adds asynchronous intervention, asymmetric verification, adaptive effort, and evidence-grounded synthesis.A lead agent decomposes the problem and constructs specialized subagents on demand rather than selecting from a fixed role catalogue.Source: §3.2 · p. 12

3.2.1 From Latent Decomposition to an Explicit Task Board

• The task board makes decomposition an inspectable, persistent coordination record that scopes subagents, stores returned evidence, and carries plan revisions.Entries record bounded objectives, dependencies, resolution state, assigned agents, and returned evidence or artifacts; results can unlock dependent work while slow branches need not stall independent tasks.Source: §3.2.1 · p. 12

3.2.2 Asynchronous Human Intervention

• During execution, user messages can update the live task board while preserving valid work, and task-preserving queries can be answered as independent executions continue.Clarifications or method changes can revise relevant entries, invalidate affected descendants, and notify active subagents; material changes to objectives or acceptance constraints begin a new task contract while retaining valid workspace state.Source: §3.2.2 · p. 12

3.2.3 Asymmetric Verification

• Asymmetric verification targets a specific claim and its evidence, producing focused checks and actionable corrections instead of reproducing the full solution.Verifiers can search for counterexamples, triangulate with an independent source class, check atomic details, or test delivery constraints; feedback lets the lead reopen a board item and integrate a repair.Source: §3.2.3 · p. 13

3.2.4 Adaptive Max Team Effort

• Adaptive Max Team Effort assigns independently scoped investigations to weak, contested, or load-bearing claims and revises allocation as results return.Later investigation waves target unresolved or disconfirmed branches, and load-bearing conclusions require focused verification before finalization.Source: §3.2.4 · p. 13

3.2.5 Evidence-Grounded Synthesis

• A dedicated synthesis stage uses the terminal task state and verification findings to build an evidence graph before a writer agent produces the constrained deliverable.The graph marks claims as load-bearing, corroborated, disputed, or unresolved; untraceable claims are qualified or omitted, missing load-bearing dependencies return to the lead, and verifier judgments determine what survives into the final evidence state.Source: §3.2.5 · pp. 13–14

3.3 AgentOS: An Execution Substrate for Long-Horizon Work

• AgentOS maintains persistent workspace and tool state, turns runtime events into model-visible observations, preserves useful progress, and governs the transition from intermediate artifacts to final deliverables.It retains the task-agnostic kernel–plugin architecture and generic ReAct loop introduced with AgentOS 1.0, extending the runtime for persistent execution, asynchronous control, multi-agent coordination, and controlled artifact delivery.Source: §3.3 · p. 14

3.3.1 Persistent Workspace and Tool Execution

• The workspace tracks files, retrieved evidence, executable state, artifacts, and provenance, while /inputs, /workspace, and /outputs distinguish supplied files, intermediate work, and candidate deliverables.Deployments may mount a read-only, user-owned /shares library; workspace visibility varies by isolation backend, and file tools are capability-scoped by execution profile.Source: §3.3.1 · p. 14

3.3.2 Coordination State for Agent Teams

• The Task Board stores coordinator-owned work-item state separately from runtime execution status, and finalization requires active items to be resolved or cancelled.Items track owners, dependencies, evidence or artifact references, and resolution; a reported subagent execution does not itself resolve an item, and planning mode additionally requires a delegated branch and independent verifier before finalization.Source: §3.3.2 · p. 15

3.3.3 Execution Continuity over Long Horizons

• An opt-in follow-up channel delivers client interventions into model-visible history, allowing runs to revise future actions without discarding workspace state, including while the coordinator awaits subagent reports.Typed directives can gate delegation or publication, and accepting a follow-up renews the run’s wall-time budget.Source: §3.3.3 · p. 16
• AgentOS handles context and time exhaustion separately: measured token pressure triggers tiered compaction, while soft deadlines prompt consolidation and bounded finalization can recover a partial result.Pauses occur between completed turns; completed observations, evidence, and files remain stored after pause or abnormal termination, with Agent Team also retaining the latest board projection and completed reports.Source: §3.3.3 · pp. 16–17

3.3.4 Controlled Artifact Delivery

• Final delivery requires a publisher to declare an exact /outputs manifest under a single active lease; writes are restricted to declared paths, while shared /outputs/scratch/ remains excluded from collection.The runtime checks manifest entries against a baseline and flags incomplete or unverified delivery; its contract does not guarantee source, computation, method, or conclusion correctness, and coordination state is run-scoped rather than atomically checkpointed with workspace files.Source: §3.3.4 · pp. 17–18
• The publication guard mediates built-in file and shell tools rather than monitoring every filesystem syscall, refusing writes whose output paths cannot be determined.Where supported, a mount-level restriction backs the tool policy; process-restart recovery and historical filesystem rewind are outside the current contract.Source: §3.3.4 · p. 18

3.4 Training

• Training trajectories represent delegation, staged return, synthesis, branch revision, and recovery as behaviors to strengthen, not merely as inference-time wrapper properties.Source: §3.4 · p. 18

3.4.1 Supervised Fine-Tuning

• SFT combines diverse reasoning and tool-use data in a shared stateful schema, filters for behavioral validity, and merges domain variants into a unified initialization.Filtering removes invalid interactions, inconsistent state, ignored observations, and incomplete delivery; rubric-scored tasks use task-relevant dimensions and thresholds. General, agentic, and coding variants are combined through model-soup merging.Source: §3.4.1 · p. 18

3.4.2 Reinforcement Learning

• PIVOT-RL localizes credit to consequential decisions, retaining useful prefixes and using temporary corrective hints before mixing localized continuations with unhinted full tasks.Asynchronous optimization processes completed trajectories without waiting for slower episodes. Figure 4 shows search, knowledge, and science evaluations improving as RL compute increases.Source: §3.4.2 · pp. 18–19, Figure 4 · p. 19

4 Evaluation

• Evaluation spans public benchmarks for model breadth and Agent Team gains, internal structured search and end-to-end research delivery, and HDS6 assessment of execution-process quality.The progression covers outcomes, capability-specific evidence, and reliability of the work used to produce them.Source: §4 · p. 19
• Agent Team matches or exceeds strong reported reference points on several professional-work, finance, and scientific-research evaluations, while remaining broadly competitive in reasoning, search, mathematics, and coding.Across complex work, Apodex 1.1 operates in the leading performance band of current agentic systems.Source: §4 · p. 19
• The 35B-parameter Apodex 1.1 mini reaches selected frontier systems’ performance band across representative work, finance, and scientific-research tasks, and substantially improves over Apodex 1.0 mini on overlapping evaluations.Source: §4 · p. 19

4.1 Empirical Setup

• Apodex is evaluated in ReAct and Agent Team modes; scores are reported to one decimal place unless native units differ, semantic evaluations use stated judges or rubrics, and captions identify internally reproduced comparisons.Source: §4.1 · p. 19
• ReAct alternates reasoning, tool interaction, and observations, minimizing orchestration so performance primarily reflects the model’s reasoning and tool-use policy.Source: §4.1 · p. 20
• Agent Team can spawn subagents to parallelize or specialize work and allocate focused verification, testing decomposition, delegation, integration, and replanning under organized computation; online-product Section 3.2.5 is excluded from offline evaluations.Source: §4.1 · p. 20

4.2 Main Results

• Agent Team adds gains over ReAct, including the strongest shown FrontierFinance and FrontierScience-Research results and competitive performance on GDPVal, APEX-Agents, HLE, and DeepSearchQA.The comparison spans professional work, finance, business simulation, scientific research, general reasoning, and deep search.Source: Table 3 · p. 20, Table 4 · p. 21, §4.2 · p. 21
• Broader executable environments, agentic training, and organized coordination can produce frontier-band complex-work capability at a compact 35B model scale.The comparison is stated as a performance-band result because parameter counts for several proprietary reference systems are not public.Source: §4.2 · pp. 21–22
• On FrontierFinance, FrontierScience-Research, and APEX-Agents, Apodex 1.1 mini Agent Team scores 50.2, 51.7, and 27.7, up from ReAct’s 40.0, 45.0, and 24.2.These are gains of 10.2, 6.7, and 3.5 points, respectively.Source: §4.2 · p. 21, Table 5 · p. 22
• On overlapping FrontierFinance and APEX-Agents evaluations, ReAct improves Apodex 1.1 mini over Apodex 1.0 mini from 33.2 to 40.0 and from 15.4 to 24.2, respectively.With ReAct alone, the 35B model also reaches 45.0 on FrontierScience-Research.Source: §4.2 · p. 21, Table 5 · p. 22

4.3.1 Professional Work

• GDPVal scores 78.8 with Agent Team versus 69.5 with ReAct, while APEX-Agents scores 38.5 with Agent Team, 34.4 with ReAct, and 16.5 for Apodex 1.0.GDPVal evaluates completed work products across 44 occupations; APEX-Agents tests sustained execution across 480 cross-application tasks in 33 data-rich worlds.Source: §4.3.1 · pp. 22–23
• FrontierFinance scores 54.3 with Agent Team versus 48.7 with ReAct, a 5.6-point gain on its macro-averaged rubric qualification rate.The benchmark covers 220 open-ended queries graded against 11,543 expert-written rubric items.Source: §4.3.1 · p. 23

4.3.2 Scientific Research

• On FrontierScience-Research, Apodex 1.1 scores 63.3% with Agent Team versus 55.0% with ReAct and 28.3% for Apodex 1.0.On BioMysteryBench’s revised 17-task Human-difficult Set, it scores 35.3% (6/17) with Agent Team and 23.5% (4/17) with ReAct; Claude Opus 5 reports 49.4%. The Claude 4.x Table 4 values use the original 23-task set as historical references. Agent Team improves over ReAct across both benchmarks, particularly on the biomedical tasks.Source: §4.3.2 · p. 23

4.3.3 General Reasoning and Search

• Agent Team scores 56.1 on Humanity’s Last Exam and 92.4 F1 on DeepSearchQA, versus ReAct’s 53.2 and 88.2, respectively.Humanity’s Last Exam tests broad expert-level reasoning with tools, while DeepSearchQA emphasizes multi-step evidence seeking and synthesis.Source: §4.3.3 · p. 23

4.3.4 Mathematical Reasoning

• On IMO 2025, IMO 2026, and USAMO 2026, Apodex 1.1 Agent Team scores 36.5, 30.5, and 26.5, exceeding the stated reference thresholds of 35, 29, and 25.Against Apodex 1.0, its scores rise from 12.5 to 36.5 on IMO 2025, 13.0 to 30.5 on IMO 2026, and 5.8 to 26.5 on USAMO 2026; it also exceeds the stated IMO gold-medal cutoffs. On IMO-ProofBench, Agent Team scores 96.7% on Basic and 63.3% on Advanced, versus ReAct’s 80.0% and 46.4%.Source: §4.3.4 · pp. 23–24, Table 6 · p. 24

4.3.5 Coding and Software Engineering

• Apodex 1.1 scores 70.8 on Terminal-Bench 2.1 and 77.7 on SWE-bench Verified.Terminal-Bench 2.1 measures sustained execution in realistic command-line environments; SWE-bench Verified measures repository-level issue resolution against executable tests.Source: Table 7 · p. 24, §4.3.5 · p. 24

4.4.1 FrontierSearchBench

• FrontierSearchBench comprises 41 verifiable deep-search tasks requiring structured deliverables reconciled across sources, with task construction, ground truth, and scoring completed independently of evaluated runs.Tasks use frozen ground-truth dimensions rather than a single answer string, and their correct answers do not change with retrieval date.Source: §4.4.1 · p. 25
• Each FrontierSearchBench task is scored through claim extraction, alignment to frozen ground truth by fixed judge models, and a deterministic rubric that penalizes incorrect assertions.The aggregate is an unweighted mean of 41 normalized task scores on a 100-point scale, with positive, zero, and negative task-score fractions also reported.Source: §4.4.1 · p. 25
• Agent Team achieves the best FrontierSearchBench result, surpassing the strongest proprietary reference; it roughly halves zero-credit tasks versus ReAct while keeping incorrect assertions rare.Table 8 reports mean normalized task scores over 41 tasks and the proportions scored positive, zero, and negative; the supplied passages do not give the systems’ numeric scores.Source: Table 8 · p. 25, §4.4.1 · p. 25

4.4.2 FrontierResearchBench

• FrontierChallenge contains 97 executable scientific tasks across ten domains, requiring analyses and mutually consistent research artifacts in task-specific Docker environments.Deliverables can include executable code, structured data, figures, domain-specific files, and a written report.Source: §4.4.2 · p. 25
• Custom graders verify required files, numerical results, formats, executable outputs, and cross-artifact consistency; any unmet requirement makes a task a non-pass.Deterministic rules combine with rubric-defined semantic judgments when qualitative assessment is needed, but the task Grader computes the final outcome.Source: §4.4.2 · pp. 25–26
• 12.4% Pass Rate for Apodex 1.1 with Agent Team contrasts with only 20.6% of tasks receiving full credit for GPT-5.6-Sol with Codex and Grok-4.6 with Claude Code.Apodex 1.1 remains behind frontier systems, and reliably completing the specified workflow and delivering every required artifact remains challenging for all evaluated systems.Source: Table 9 · p. 26, §4.4.2 · p. 26

Framework

• HDS6 scores six process capabilities through 24 rubric items and applies a separate integrity gate to test whether delivery rests on a defensible execution process.The capabilities are state coherence, evidence fidelity, hypothesis management, boundary and failure reasoning, tool use and execution-state management, and self-correction under verification.Source: §4.5 · p. 26, Figure 5 · p. 27

Process-level judging protocol

• The process judge stitches asynchronous multi-agent activity into an ordered log, then uses specialized mapping, judging, review, and arbitration stages to assign cited 0/1/2 scores.Grading is outcome-blind, excludes private reasoning, and rechecks cited events against the recorded trajectory; fabricated results or narrated actions that never occurred trigger an outright zero.Source: §4.5 · p. 26

Findings

• HDS6 comparisons are intended within each Figure 6 panel, not across panels, and Deep Discover v1.0 aggregates eight runs while v1.1 uses a single run.The panels differ in task family, model scale, execution mode, and—in Deep Discover—the aggregation protocol, so that comparison is not a strictly matched single-run ablation.Source: §4.5 · p. 27, Figure 6 · p. 27
• Initial Decomposition improves by +1.3 and Final Verification by +0.8; evidence fidelity and hypothesis management also improve across rubric items.These are the two largest single-item deltas; HDS6 localizes gains from combined model, environment, coordination, and harness changes but does not isolate any component’s causal contribution.Source: §4.5 · pp. 27–28, Figure 6 · p. 27

Qualitative cases

• Three end-to-end cases cover clinical survival analysis, molecular docking, and electrochemical corrosion analysis, with appendix documentation of inputs, work division, tool execution, checks, and artifacts.The cases complement aggregate scores.Source: §4.5 · p. 28

General-purpose agentic models

• Apodex follows frontier models’ shift toward sustained tool- and environment-mediated work, while organizing development around Environment Scaling and Agentic Coordination Scaling.Recent releases combine long-context reasoning, coding, tool use, information seeking, and sustained execution (Team et al., 2026; OpenAI, 2026d; Anthropic, 2026a; Zeng et al., 2026; DeepSeek-AI, 2026).Source: §5 · p. 28

Executable environments and real work

• Apodex’s File, Search, and Code environments extend execution-grounded benchmarks by serving jointly in task synthesis, training, replay, and evaluation.Prior work includes WebArena (Zhou et al., 2023), OSWorld (Xie et al., 2024), SWE-bench, and APEX-Agents (Vidgen et al., 2026).Source: §5 · p. 28

Search and deep-research agents

• Apodex treats search as one stage in a broader working process, where retrieved evidence can feed code, file transformations, parallel investigation, and delivery verification.Deep-research work includes autonomous search and reasoning (Li et al., 2025b), real-world web training (Zheng et al., 2025), and multi-turn retrieval (Jin et al., 2025).Source: §5 · p. 28

Agents for scientific discovery

• Unlike domain-specific laboratory agents, Apodex develops a general-purpose model and runtime whose file–search–code policy serves scientific, software, and professional workflows.Scientific-agent examples include chemistry tools (Bran et al., 2023), automated research loops (Lu et al., 2024; Yamada et al., 2025), and asynchronous scientific-agent coalitions (Gottweis et al., 2025).Source: §5 · p. 29

Multi-agent coordination and test-time scaling

• Apodex’s Agent Team uses staged result return, shared task state, subagent execution control, and user intervention rather than fixed dialogue graphs or majority-vote ensembles.Prior approaches include programmable conversations (Wu et al., 2023), structured software roles (Hong et al., 2023; Qian et al., 2023), and debate (Du et al., 2023; Liang et al., 2023); collaboration is useful when decomposition and integration are reliable.Source: §5 · p. 29

Verification and process supervision

• Apodex combines environment checks, artifact and provenance checks, and independent Statement Review to verify executable outcomes, completed workspaces, and consequential claims.Prior verification approaches include self-feedback, reflection, targeted questions, and process supervision (Madaan et al., 2023; Shinn et al., 2023; Dhuliawala et al., 2023; Lightman et al., 2023).Source: §5 · p. 29

Environment scaling and agentic reinforcement learning

• Apodex integrates environment scaling and agentic reinforcement learning through shared delivery contracts, provenance state, replay, independent verification, and failure-driven task construction.Prior work scales synthesized or real-world environments (Fang et al., 2025; Tu et al., 2026; Dong et al., 2026) and develops interaction optimization and scalable RL infrastructure (Jin et al., 2025; Wang et al., 2025b; Yao et al., 2025).Source: §5 · p. 29

6 Conclusion

• Apodex 1.1 frames agentic intelligence as the ability to sustain progress through changing tasks, use tools and evidence effectively, recover from failure, and deliver verifiable results.Source: §6 · pp. 29–30
• Environment Scaling broadens executable file, search, and code worlds, while Agentic Coordination Scaling improves decomposition, delegation, integration, and reorganization across agents and time.A common harness and AgentOS maintain persistent execution state, and unified training converts environment and coordination trajectories into a stronger working policy.Source: §6 · p. 30
• Evaluations span scientific research, professional work, mathematics, search, coding, and internal long-horizon tasks, while HDS6 analysis and end-to-end cases examine tool use, evidence, repair, alternatives, and reviewable artifacts.ReAct versus Agent Team estimates the system-level lift of trained coordination under additional organized computation; together, evaluations position Apodex 1.1 as a common model and execution stack.Source: §6 · p. 30
• The 35B Apodex 1.1 Mini reaches the performance band of selected frontier systems on representative professional, financial, and scientific tasks, improves over Apodex 1.0 Mini on overlapping evaluations, and retains substantial Agent Team gains.This compact-model result indicates that both scaling dimensions improve capability at a smaller model scale, as well as peak system performance.Source: §6 · p. 30
• The long-term goal is a Heavy-Duty Solver for ambitious, long-running, verifiable work, pursued by scaling executable-environment coverage and fidelity, longer-horizon coordination, hierarchical-trace training and credit assignment, and failure-driven task construction, training, and evaluation.Apodex 1.1 establishes a development path centered on completed work rather than isolated responses.Source: §6 · p. 30

A Case Studies

• Three recorded runs transform raw inputs into checkable deliverables: a protein simulation package, cell-death measurements from six microscope images, and candidate genes linked to traits across 50 pigs.Cases 1 and 2 report agent task boards, tool calls, actions, and returns; Case 3 uses pipeline-stage blocks because its agent-level trace was not retained.Source: A Case Studies · p. 35

Case 1: Coarse-Grained Structure Preparation and Energy Minimisation of a Protein Hexamer

• The repaired 22-file Martini 3 package minimized in 255 steps and was verified as one run, but retained a disclosed 131-position residue-sequence mismatch and is only an equilibration starting point.The minimization input coordinates matched ionized.gro for all 63,484 particles, with maximum absolute difference 0.Source: Case 1: Coarse-Grained Structure… · pp. 37–38

Case 2: TUNEL/DAPI Apoptosis Quantification from Fluorescence Micrographs

• Two defensible fluorescence-intensity definitions produced opposite group conclusions; arbitration selected shared-ROI field means because per-mask means were threshold-sensitive and discarded the roughly 2.7× TUNEL-positive-area difference.The shared-ROI ratio was stable across plausible tissue ROIs, but not all ROIs, because background haze shifted C1 from 83.6 to 103.6.Source: Case 2: TUNEL/DAPI Apoptosis… · pp. 40–41
• Control TUNEL/DAPI ratio was 113.80 ± 31.67 versus Experimental 47.64 ± 8.92; Cohen’s d = 2.84, but Welch two-sample P = 0.0593 at n = 3 each.The reported ratio is 2.4× higher in Control, a large effect that does not reach significance.Source: Case 2: TUNEL/DAPI Apoptosis… · pp. 41–42
• Two agents analyzed the same fluorescence images using per-mask and shared-ROI/k-means definitions, and a local verifier arbitrated their outputs.The run took 24.7 minutes and recorded 324 steps, with all seven board items resolved.Source: Case 2: TUNEL/DAPI Apoptosis… · p. 42, Figure 8 · p. 42

Case 3: WGCNA of a Pig RNA-seq Cohort with Twenty Specified Deliverables

• The WGCNA workflow audited 6,000 genes and 50 samples, selected 5,000 variable genes, and produced ten co-expression modules.It retained 5,983 genes after low-expression filtering, transformed counts as log2(CPM + 1), and used power 20 as a maximum-fit fallback because no tested power reached R2 ≥ 0.85.Source: Case 3: WGCNA of a Pig RNA-seq Cohort… · pp. 43–44
• Five phenotypes mapped to different non-grey modules, all at FDR < 10^-16; hub screening yielded 437 candidate records.The records were screened in the strongest non-grey module per phenotype using FDR < 0.05, |MM| > 0.8 and |GS| > 0.2; they are candidates for validation, not causal or regulatory findings.Source: Case 3: WGCNA of a Pig RNA-seq Cohort… · p. 44
• The network results are correlational and qualified by absent batch information, limited covariate modelling at n = 50, and sensitivity of module partitions to analysis parameters.The report says analysis.R was not run end to end because R was unavailable; numerical results came from an equivalent implementation with file-level consistency checks, leaving possible R version or package-API differences on rerun.Source: Case 3: WGCNA of a Pig RNA-seq Cohort… · p. 44