Source-linked AI summary
Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence
Brian Wang, Bin Feng, Xiaoman Pan, Chenyang An, Felix Liu, Tangqi Fang, Gongbo Sun, Lingfeng Shen, Ning Wang, Handuo Zhang, Feng Chen, Fuchao Yang, Xiang Wang, Jiacheng Lin, Siting Li, Zixuan Liu, Chi Han, Zhenhailong Wang, Kunlun Zhu, Lawrence Zhao, Yueqi Guo, Kailong Wen, Feng Xing, Yiling Guo, Lidong Bing, David Tan, Bo An, Heng Ji, Sheng Wang
TL;DR
Consequential real-world problems rarely provide the executable, verifiable architecture that advanced AI models need. Apodex Discovery builds heavy-duty solver environments and evaluates investigation processes, improving drug-repurposing scores and surpassing prior baselines in AAV capsid design.
Problem
Consequential real-world problems rarely arrive with executable environments, tools, constraints, feedback, and success criteria for evaluating extended, verifiable problem solving.
Method
Apodex Discovery combines problem scouting, executable environment-task-episode abstractions, heavy-duty solver systems, and HDS6 process evaluation.
Results
2.5 and 7.6 points improved GPT-5.5 and GPT-5.6-sol mean normalized prediction scores, respectively, while Apodex surpassed prior human baselines in AAV capsid design.
Takeaways & Limitations
Apodex shifts evaluation from isolated answers toward reliable, evidence-grounded, self-correcting investigations by complete solver systems.
Takeaways & Limitations
The benchmark currently covers seventeen environments and 218 episodes, representing only a subset of the twenty selected problems and 423-problem registry.
Abstract
from arXiv · showhide
Apollo did not reach the Moon merely because its engineers could solve difficult equations. It succeeded by turning a distant ambition into a mission architecture of explicit objectives, simulation, verification, and repeated correction. AI now faces a similar transition: frontier models can solve difficult tasks once the problem, tools, and success criteria are specified, yet consequential real-world challenges rarely arrive in an executable or verifiable form. We introduce Apodex Discovery, a framework for building and evaluating discoverative AI through the heavy-duty solver, a system comprising a foundation model, harness, tools, and control policies that pursues extended, stateful, verifiable investigations. It has three core components. First, a problem-scouting process surveyed 561 industries across 16 sectors, assembled 423 high-value real-world problems, and selected 20 for the initial release. Second, a common environment-task-episode abstraction provides data, tools, constraints, feedback, trajectory recording, and verification of intermediate artifacts and final submissions. Third, HDS6 evaluates Tools, Repair, Alternatives, Coherence, Evidence, and Scope independently of final-task success. In AAV capsid design, Apodex surpassed the published state of the art by 7% across viability, tropism, structure prediction, and generative design. In drug repurposing and reformulation, a task-specific biomedical environment improved the mean normalized prediction score of GPT-5.5 and GPT-5.6-sol by 2.5 and 7.6 points over the same closed-book backbone. Controlled ablations show that the fixed TRACES episode interface enables attribution of performance differences to specific solver components. Apodex Discovery moves AI evaluation beyond predefined benchmarks toward verifiable investigations aimed at genuine discovery.
1. Introduction
Apodex Discovery frames discoverative AI as extended, real-world investigation rather than isolated answer production. It provides reality-based environments, verification, repair, and process evaluation through heavy-duty solvers and the TRACES benchmark.
- Motivation: Discoverative AI turns open-ended ambitions into grounded, verifiable investigative acts that can produce conclusions not yet known.The framework focuses on how systems execute and support discovery, rather than reproduce existing answers.
- Framework: The infrastructure combines problem formulation, reality-based environments, verification mechanisms, and repair loops to make ambitions executable and outcomes checkable.These components specify objectives and constraints, provide operational resources, assess intermediate and final claims, and support correction after failure.
- Framework: A heavy-duty solver is a complete system that sustains long, stateful, self-correcting investigations using a foundation model, harness, tools, memory, and control policy.Its defining capability is sustained investigation against an external environment, not raw reasoning power alone.
- Evaluation: HDS6 evaluates six process capabilities—Tools, Repair, Alternatives, Coherence, Evidence, and Scope—separately from whether the solver ultimately succeeds.This process verification is intended for settings where outcome ground truth may be delayed, incomplete, or initially unavailable.
- Benchmark: 561 industries across 16 sectors yielded 423 high-value problems, from which 20 were selected for the initial TRACES benchmark.The benchmark targets consequential real-world problems suitable for rigorous AI evaluation.
2. Scouting High-Value Real-World Problems
Apodex Discovery uses a structured scouting process to identify high-value, difficult, verifiable real-world problems suitable for heavy-duty AI evaluation. The process surveyed 561 industries across 16 sectors, produced 423 candidate problems, and selected 20 for the initial TRACES release.
- Problem identification: 423 high-value real-world problems were collected and reviewed for reasoning depth, technical feasibility, verifiability, data and tool availability, verification latency, and potential impact.The review was conducted by a ten-person STEM research team through a two-month survey spanning scientific, engineering, healthcare, financial, and industrial applications.
- Initial release: 20 problems were selected from the 423-problem registry for the initial TRACES release, spanning frontier-model development and biomedical discovery.The published selection pipeline narrowed 561 industries to 10 domains, expanded them to 423 candidate questions, and screened them to 20 published problems.
- Benchmark scope: The initial problem families target investigations requiring long-horizon state maintenance, heterogeneous tool use, inspectable intermediate artifacts, verification feedback, and externally judgeable outcomes.These are extended investigations rather than short-form questions, and their conclusions must ultimately be confirmable or falsifiable outside the benchmark.
- Benchmark scope: The initial executable environments cover only a small fraction of the 423-problem registry, with future releases intended to expand across scientific, engineering, financial, healthcare, and industrial domains.The broader goal is to evaluate solver systems on difficult, consequential, and verifiable problems whose answers are not already known.
3. Process Verification and Reality-Based Environment
Apodex Discovery makes open-ended investigations verifiable through complementary outcome and process verification, executable environment–task–episode infrastructure, and trajectory-based diagnostics. Its fixed episode interface supports reproducible, repairable evaluation by exposing controlled resources and returning verifier feedback while keeping outcome checks hidden.
- Process verification: Process verification evaluates how a solver produced its result by analyzing the trajectory, providing diagnostics that terminal outcome scores cannot supply.The blind process verifier reads the trajectory alone, while the hidden outcome verifier evaluates the submission against ground truth.
- Outcome verification: Outcome verification scores final submissions against hidden success criteria or ground truth using task-appropriate checks, including tests, held-out labels, simulations, or expert adjudication.The verifier may also check important intermediate results or milestones before the final artifact is complete.
- HDS6: HDS6 operationalizes process verification through six TRACES dimensions that assess the competencies required for long, self-correcting investigations, including repair and alternative-hypothesis management.Repair measures effective self-correction after incorrect steps, verification feedback, or failures; Alternatives measures explicit competing hypotheses and evidence-based ranking updates.
- Executable environment: The executable infrastructure nests environments, tasks, and episodes with verification agent teams, exposing actions, tools, budgets, observations, and intermediate or final submissions.Episodes are the primary unit of end-to-end evaluation, with task-level scores explaining within-episode performance and episode-level outcome scores grading final submissions.
- Episode interface: The fixed episode interface standardizes inputs, tools, budgets, submission formats, hidden verifiers, hard gates, and outcome metrics, while verifier feedback makes episodes repairable and evaluation interpretable.Constructing a faithful environment with a trustworthy hidden verifier is treated as a methodological contribution because it is at least as difficult as the solver’s task.
4. Results on High-Value Problems
Results provide early evidence that heavy-duty solvers can outperform published methods on AAV capsid design and benefit from task-specific environments in drug repurposing. HDS6 process scores align with biological performance, highlighting the role of tool use, evidence grounding, coherence, and self-correction.
- Task-specific environments: 0.741 versus 0.716: claude-opus-4-8 performed better in the domain-specific environment than with the generic Claude Code harness across tasks 1–3.The environment supplied task-specific data, tools, biological constraints, and executable feedback while holding the backbone, benchmark items, output schema, and scoring procedure fixed.
- AAV capsid design: 0.904 AUROC: apodex-1.1 exceeded CAP-PLM’s 0.878 on out-of-distribution AAV viability prediction.Viability is the campaign’s first rate-limiting filter for identifying candidates worth advancing.
- AAV capsid design: 0.635 versus 0.622: apodex-1.1 outperformed Fit4Function on predicting cross-species tissue tropism.The task evaluates which tissues variants reach and whether targeting patterns hold across species.
- AAV capsid design: 0.180 versus 0.109, 0.116, and 0.110: apodex-1.1 surpassed AAVGen, AAVDiff, and ALICE on generative AAV design.The solver searched sequence space while balancing viability, specificity, novelty, and a limited experimental budget.
- HDS6 process evaluation: HDS6 scores tracked final biological performance, with stronger systems showing more coherent hypotheses, better tool and data use, stronger constraint handling, and more self-correction.On structure prediction, gpt-5.5 scored 0.544, while its process scores were 1.38 and 0.94 because it skipped required self-validation; kimi-k3 reached 3.56 and 3.60.
- Drug repurposing: Environment access improved every reported drug-repurposing metric for both backbones, especially tool use, evidence fidelity, and long-horizon coherence.These descriptive gains came from structured biomedical knowledge, literature-derived evidence, and executable scientific tools; repair scores were unavailable for no-environment trajectories.
5. Related Work
Prior work spans answer-based knowledge benchmarks, interactive agent environments, agent scaffolds, automated evaluation, process supervision, anti-gaming verification, and autonomous scientific-discovery systems. Apodex Discovery builds on these traditions through executable, verifiable investigations rather than isolated answer checking.
- Knowledge and reasoning benchmarks: Answer-based benchmarks evaluate language models against reference keys, spanning broad knowledge, reasoning, mathematics, and increasingly difficult frontier examinations.Examples include MMLU, BIG-bench, AI2 Reasoning Challenge, GPQA, GSM8K, MATH, FrontierMath, and Humanity’s Last Exam.
- Interactive and agentic benchmarks: Interactive benchmarks embed models in executable environments and score task success, extending from reinforcement-learning platforms to language-agent worlds, web interaction, and economically significant professional work.Examples include Arcade Learning Environment, OpenAI Gym, ALFWorld, WebShop, Mind2Web, SWE-Lancer, and Agents’ Last Exam.
- Agent scaffolds: Agent scaffolds transform models into agents through prompting, control, tool use, multi-agent coordination, open-ended exploration, and agent–computer interfaces.Representative methods and systems include ReAct, Tree of Thoughts, Reflexion, Self-Refine, Toolformer, AutoGen, Voyager, and SWE-agent.
- Automated and LLM-based evaluation: Automated evaluation increasingly uses model judges, reference-free scoring, open evaluator models, and rubric- or skill-based protocols to decompose output quality into explicit criteria.Examples include arena-style judging, G-Eval, JudgeLM, Prometheus 2, FLASK, HealthBench, and trained critics such as CriticGPT.
- Process supervision and verifiable rewards: Process-supervision research evaluates reasoning steps rather than only final answers, showing stronger verifiers and reducing annotation costs through process reward models and generative verification.This line includes comparisons of process- and outcome-based feedback, step-level supervision, Math-Shepherd, and generative verifiers.
- Autonomous scientific discovery: Scientific-discovery systems couple language models to laboratory tools, generate hypotheses through multi-agent debate, conduct machine-learning research end to end, and analyze biomedical AI agents.Examples include Coscientist, ChemCrow, the AI co-scientist, and The AI Scientist.
6. Conclusion
Apodex Discovery combines systematic problem selection, executable environments, and HDS6 evaluation to build and assess discoverative AI on consequential, verifiable problems. Initial results show gains in capsid design and drug repurposing, while the current benchmark remains limited in coverage and verification-loop support.
- Core capabilities: Apodex Discovery provides problem identification, executable environments, and HDS6 assessment of investigation quality as complementary capabilities.The framework targets high-value problems that are objectively verifiable and supports solver action through data, tools, and feedback.
- Initial results: 2.5 and 7.6 points improved the mean normalized prediction score of GPT-5.5 and GPT-5.6-sol, respectively, over the same closed-book backbone.The improvement came from a domain-specific biomedical environment for drug repurposing.
- Limitations and Extensions: 17 executable environments currently instantiate TRACES across 218 episodes, but only two support the full verification–repair loop.The benchmark covers a subset of 20 selected problems and remains limited relative to the 423-problem registry.
7. Author Contributions and Acknowledgments
The section credits Tianqiao Chen with proposing the heavy-duty solver, defining HDS6’s six capabilities, and shaping Apodex Discovery’s architecture. It also acknowledges additional discussion, technical, and broader contributors.
- Acknowledgments: Tianqiao Chen proposed the heavy-duty solver, defined HDS6’s six capabilities, and shaped Apodex Discovery’s overall architecture.The authors credit his vision, guidance, and support as making the work possible.
- Acknowledgments: Yuan Cai, Chengyu Fu, Shiyi Zhang, and Harry Zhang contributed insightful discussion and contribution.
- Contributors: Sheng Wang, Brian Wang, Bin Feng, Xiaoman Pan, and Chenyang An were core technical contributors, alongside Felix Liu.
- Contributors: Additional contributors included Tangqi Fang, Gongbo Sun, Yueqi Guo, Kailong Wen, Feng Xing, Yiling Guo, Lingfeng Shen, and others.The passage also lists Ning Wang, Handuo Zhang, Feng Chen, Fuchao Yang, Jiacheng Lin, Siting Li, Zixuan Liu, Xiang Wang, Chi Han, Zhenhailong Wang, Kunlun Zhu, Lawrence Zhao, Lidong Bing, David Tan, Bo An, and Heng Ji.
A. Domains and questions considered by Apodex Discovery
This section catalogs the domains and questions considered by Apodex Discovery, including ten selected domains and a twenty-question registry.
- Domains: Apodex Discovery’s considered domains are presented as ten selected domains.The listing appears in Table 14, which continues across two table passages.
- Questions: The considered questions are organized into a twenty-question registry.The registry is presented in Table 15, which continues across two table passages.
B. Adeno-associated virus capsid design full results
Adeno-associated virus capsid design evaluates discoverative solvers across viability, tropism, structure prediction, and constrained generative design. The tasks require extrapolation, orchestration of folding and analysis tools, leakage control, and novelty under metered scoring budgets.
- Viability: Viability prediction tests out-of-distribution ranking accuracy on held-out variants across mutational load and sequence position or mutation type.Variants are classified as viable when they assemble and package their genome, using amino-acid sequence alone; validation labels are hidden behind a metered scoring action.
- Tropism: Tropism prediction measures agreement between predicted and measured enrichment for seven-amino-acid inserts across mouse, macaque-transfer, and human assays.The instances increase in difficulty from mouse multi-organ tropism to cross-species transfer and human in-vitro assays.
- Structure: Structure prediction requires solvers to orchestrate folding engines and analysis tools to produce subunit, 60-mer shell, and antibody- or receptor-complex structures.Evaluation uses post-cutoff targets, sequence-similarity filtering, and network isolation; the complex set contains 8 antibody and 7 AAVR-receptor targets.
- Design: Generative design permits up to 1000 novel peptides while limiting oracle queries to 100 sequences, requiring hits to be novel, manufacturable, and highly selective.Seed sets range from 2,000 training and 500 validation peptides for tissue instances to 188,000 and 10,000 for receptor instances.
- Solver behavior: GPT-5.5 receives low HDS6 scores because it skips required self-validation and submits the cheapest answer across all three structure instances.Its behavior includes using one default folding engine without comparison, copying the highest-identity template coordinates, and hardcoding complex outcomes.
C. Full harness × model result matrix
The full matrix reports harness–model scores across 11 LLM environments, with annotations distinguishing unrun configurations and non-capability outcomes. Several environments reveal that zeros, submission reliability, and execution artifacts can materially affect interpretation of capability results.
- Matrix structure: 11 environments report harness–model scores for six foundation models, while blank cells indicate unrun configurations and superscripts flag non-capability outcomes.The flagged outcomes include harness bugs, infrastructure faults, refusals, and gate failures.
- Inclusion criteria: 7 of 11 environments feed the main-text aggregates; four are excluded for floor-level scores, coverage gaps, incomplete submissions, or single-model domination.The excluded environments are nanogpt-speedrun, swe-juice, model-weakness-internal-benchmark, and llm-determinism.
- Zero interpretation: 23 recorded zeros include 8 genuine failures, with the remainder attributed to artifacts; no confirmed regression triggered the no-regression gate across 72 episodes.The no-regression gate fired five times, but none of those firings was confirmed.
- Reliability effects: 1700 episodes yield a balanced-grid mean of 0.9178, with harness spread 0.0176 and model spread 0.0212; excluding submission-failure zeros, harness spread falls to 0.0098.Excluding zeros, GPT has the highest judgement quality at 0.9625, showing that raw ordering can reflect submission reliability rather than judgement quality.
- Environment outcomes: 11 of 16 scored loop-track cells sit at 0.927, while submitted structures are byte-identical across harnesses because the default fold path is deterministic and cached.The track used the best of up to three attempts, and no leakage fingerprint triggered in 79 episodes.
- Environment outcomes: All 33 scored cells pass the additive-baseline threshold, with Bryant at 0.885 and Ogden at 0.851; Bryant settings broadly exceed the literature reference cap-PLM/ESM2 ≈0.897.Eleven first-pass zeros were concurrency artifacts recovered by low-concurrency reruns, and capability failures were zero.
D. Drug Repurposing and Reformulation Trace Example
This appendix presents one drug-repurposing and reformulation evaluation episode end to end, including its complete reasoning trace and process score. The episode displays every sub-rubric, while the reference remains hidden from the judging panel.
- Scoring: The reference column is never visible to the judging panel.
D.1. The task as given
The system was tasked with an offline, tool-mediated assessment of Dostarlimab for Stage II Breast Cancer, estimating clinical-approval potential and prediction confidence. The brief specified evidence-grounded investigation, calibrated phase-based scoring, and a strict submission format.
- Evaluation constraints: Offline evaluation prohibited web searches, arbitrary URLs, news browsing, and direct access to public trial or biomedical services for the held-out outcome.Only capabilities exposed through the sanctioned bundle and active access adapter were permitted.
- Scoring: Promisingness was a number in [0, 1] representing the pair’s predicted position on the clinical-approval spectrum and ranking candidates.The estimate had to reflect true relative promise rather than enthusiasm.
- Scoring: Confidence was a number in [0, 1] estimating prediction quality, scored as 1 - (confidence - quality)^2 with quality = 1 - prediction_error^2.The brief recommended setting confidence to expected prediction quality and lowering it when evidence was thin or conflicting.
- Scoring scale: The phase scale mapped 0.00 to preclinical only, 0.15 to passed Phase 1, 0.45 to passed Phase 2, 0.65 to passed Phase 3 or established off-label use, and 1.00 to FDA-approved on-label treatment.Successful but not-yet-approved phases were lower bounds, while terminal later-phase failures were capped at the last passed phase.
- Task definition: The task assessed Dostarlimab for Stage II Breast Cancer, requiring promisingness, confidence, and a concise evidence-based rationale.The system was instructed to investigate the drug-disease pair with available tools and follow the output contract.
D.2. Full reasoning trace
The full reasoning trace records every action, tool call, returned output, failure, and submission while withholding private chain-of-thought from judges. It also documents the isolated, metered execution environment, constrained biomedical-tool workflow, and evidence surfaced for the Dostarlimab–Stage II Breast Cancer task.
- Trace presentation: Every recorded step is shown, but model reasoning content is withheld; judges see only actions and outputs, with calls and submissions visually distinguished.Reasoning turns are grey, tool calls blue, failed calls red, and submissions gold.
- Execution environment: The environment meters and records task inspection, sandbox commands, allowlisted domain-tool calls, and final submission.The sandbox is isolated, networkless, and ephemeral; evaluated answers must be persisted with submit rather than a local file.
- Domain-tool workflow: The biomedical workflow uses one bounded Stage-1 tool_A search per drug–disease pair, treating its returned summary as the canonical decision surface.The workflow instructs agents to inspect resolved entities, providers, candidate paths, connections, evidence, relationship pages, expansion candidates, stage2, and report without dumping raw provider results or fused graphs.
- Retrieved evidence: 150 expansions were exhausted by provider_A while 3,197 frontier items remained, producing a partial weighted-best-first search without timeout.The recorded provider status was partial, with a 240-second effective timeout, 150 maximum expansions, and budget exhaustion.
- Retrieved evidence: Dostarlimab’s strongest surfaced evidence concerned colorectal and rectal malignancies, including 100% clinical complete responses in deficient mismatch repair patients.The trace also reported moderate-to-high clinical-trial support in NSCLC, directionally positive but biomarker-constrained prostate-cancer evidence, and approval-related caveats.
D.3. Process score
The process score aggregates sub-rubric performance into a rescaled capability score while accounting for critical failures and unexercised sub-rubrics. It uses weighted means over available 0/1/2 bands.
- Process score: Each capability is the weighted mean of its sub-rubric bands, scored 0/1/2 and rescaled to [0, 4].The score therefore combines sub-rubric results into a common four-point scale.
- Process score: A critical sub-rubric caps its capability when the sub-rubric falls below the top band.Critical sub-rubrics can limit the resulting capability score despite other sub-rubric performance.
- Process score: Sub-rubrics marked nr are excluded from the mean when the trajectory provided no opportunity to exercise them, rather than scored zero.This distinguishes unavailable evidence from a failed score.
C Long-horizon State Coherence
The strongest dimension scored 3.50 / 4 because key claims were traceably supported while retrieved facts, inferences, and unknowns remained distinct. One band was lost because load-bearing evidence lacked an explicit recency statement.
- C Long-horizon State Coherence: 3.50 / 4 marked the strongest dimension.The evaluation identified this as the strongest dimension.
- C Long-horizon State Coherence: Every key claim named a resolvable retrieved record with support traceable to that same record.This made the evidentiary basis of the claims directly checkable.
- C Long-horizon State Coherence: Retrieved facts, own inferences, and acknowledged unknowns were kept separate in the prose.The mechanism was described as plausible rather than established, and the class-level hint was explicitly labelled class-adjacent.
- C Long-horizon State Coherence: One band was lost because load-bearing evidence lacked an explicit recency statement.The evidence was otherwise described as traceable to the relevant retrieved records.
E Evidence Fidelity
The submission demonstrates strong evidence fidelity by tracing claims to inspected records and distinguishing established findings from inference and unknowns. Its main weakness is that the repurposing counter-case lacks concrete failure, safety, feasibility, and confidence justification.
- Evidence Fidelity: Claims consistently trace to resolvable tool commands or inspected pages, including the unresolved breast-cancer anchor, no-connection statement, and dostarlimab mechanisms.Supports are traceable to the same records used as sources.
- Evidence Fidelity: The final answer labels biological extensions as plausible, identifies the unresolved disease anchor, and distinguishes class-adjacent evidence from direct clinical support.It avoids presenting load-bearing conjectures as settled tool_A facts.
- Evidence Fidelity: 2.50 / 4 marks the weakest dimension because the submission states the evidence gap but offers little concrete analysis of repurposing failure or confidence.It omits indication-specific safety or feasibility constraints and does not justify the stated confidence value.
S Boundary & Failure Reasoning
The submission remains within research-use boundaries by qualifying its mechanistic claims and avoiding clinical treatment recommendations, while scoring 3.50/4 on documented retrieval procedures despite one invalid-parameter call.
- Boundary & scope: The submission qualifies evidence for Stage II Breast Cancer as lacking direct clinical support and describes PD-L1/TNBC evidence as only class-adjacent.It states that no complete connection was reconstructed and provides no dosing, administration, monitoring, or treatment recommendation.
- Boundary & scope: 3.50 / 4 was earned for reading manuals, issuing bounded searches, opening returned artefacts, reusing their content, and escalating search depth once before stopping.One band was lost for an avoidable invalid-parameter attempt at step 12, which also made Repair gradeable.
T Tool Use & Execution State Management · R Self-correction under Verification
The agent combined retrieval-based tool use with iterative correction under verification, but an invalid tier argument and incorrect identifiers reduced call correctness. It escalated queries, inspected limited returns, and judged the remaining evidence insufficient for a strong conclusion.
- T Tool Use & Execution State Management: The agent read the workflow/help first and correctly formatted the required initial Stage-1 call.This established compliant execution at the start of the investigation.
- T Tool Use & Execution State Management: It later used an unsupported “deep” tier, received a tool rejection, and repaired the call by retrying with the valid “max” tier.The pair was also passed by surface names rather than supplied task IDs, and the disease resolved as local/unknown.
- R Self-correction under Verification: The agent self-corrected by escalating from an extra-tier pair query to max tier, inspecting the truncated drug page, and issuing one relevant PD-L1 hint query.These actions went beyond stopping after a single weak return.
- R Self-correction under Verification: It submitted a low/moderate-confidence conclusion after explicitly judging that the exact pair lacked a reconstructed path or load-bearing evidence even at max tier.The PD-L1 breast-cancer evidence was only class-adjacent, supporting the stated below-Phase-2 score.
- T Tool Use & Execution State Management: The agent used retrieval rather than recall and selected commands aligned with the open questions.It began with the required pairwise connect for Dostarlimab and Stage II Breast Cancer.
- T Tool Use & Execution State Management: Because the pairwise result suggested no expansions, skipping expand was acceptable; the agent later followed a PD-L1 related-page hint with the correct target ID and stage2 suggest form.The invalid deep-tier attempt was a call-correctness issue rather than evidence that the subcomponent was unused.