TL;DR for operators

An AI research agent can produce a correct answer without retaining enough evidence to reconstruct it. That distinction matters when its findings enter investment analysis, management reporting, or other consequential workflows.

PANINI-DR addresses this problem by organizing research into dependent subquestions, preserving intermediate findings, and checking evidence before those findings influence later work. It achieves 73.3% accuracy on BrowseComp-Plus versus SLEUTH’s 57.3% and leads a separate evidence-recoverability measure across three benchmarks.

The qualification is substantial. Its original-answer performance trails several competitors on collection-valued questions, and its estimated model-core computation is approximately 7.3 times SLEUTH’s. More verification also does not consistently improve accuracy.

The deployment lesson is to evaluate answer correctness, evidence recoverability, semantic support, calculation reproducibility, collection completeness, and inference cost separately. A more elaborate research architecture is justified only when improvements in the required dimensions outweigh its additional cost.

A correct conclusion can conceal an incomplete research record

Consider an AI-generated investment memo that identifies a company’s competitors, estimates a financial ratio, and cites several documents. The final conclusions may be correct, yet the saved investigation might not establish which source supplied each input, whether a contested entity was resolved, or whether relevant competitors were omitted.

A conventional answer-quality score does not necessarily reveal these weaknesses. An organization selecting a research assistant therefore faces two different questions: how often does the system return the right answer, and how much of that answer can be reconstructed from its retained research evidence?

The distinction becomes visible in PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces, by Turali and colleagues.1

On the MoNaCo benchmark, PANINI-DR achieves an original-answer F1 of 0.549, below HyMem’s 0.612. Yet when credited answers are evaluated for recoverability from saved evidence, PANINI-DR scores 0.458, ahead of HyMem’s 0.388 (Sections 4.1–4.2, Tables 1–2).

A system that performs worse on answer F1 can therefore perform better on evidence recoverability. That reversal motivates a closer examination of how the research itself is organized.

The research record is maintained before the final answer

Long investigations create dependencies. Identifying a company’s former subsidiary may be necessary before searching for a later acquisition. If that intermediate identity is forgotten, incorrectly inferred, or admitted without evidence, subsequent searches can become less targeted or propagate the error.

PANINI-DR uses a Generative Semantic Workspace (GSW) to represent this evolving investigation. GSW is a representation inherited from earlier research; PANINI-DR’s contribution is the execution and verification protocol built around it.

The workspace contains typed unknown values, retrieval hints, verification claims, supporting evidence, and directed dependencies. Its state combines the current research plan with records of what has been found and checked.

Three agent roles operate over this shared state. An Orchestrator assigns dependency-ready subquestions; Researchers retrieve sources and propose findings, assisted by a Reader that extracts relevant passages; and a Verifier checks whether proposed values satisfy local claims and are distinguished from alternatives. Deterministic type and citation checks complement the model-based review.

Rejected proposals can trigger further retrieval and revised evidence. Accepted findings persist for downstream assignments, while executable calculations and plan revision accommodate questions whose requirements evolve during investigation.

This matters because checking only the final response cannot reliably prevent an unsupported intermediate value from directing several subsequent searches.

The admission process is not absolute: exhausted verification attempts may produce warned findings, and final-answer recovery can synthesize an answer when the target remains unresolved. Indeed, recovery accounts for 40 of PANINI-DR’s 110 correct BrowseComp-Plus answers. A correct final response therefore does not always correspond to a verified answer-target binding.

Structured planning and verification both change outcomes

The principal comparison covers 2,289 questions across BrowseComp-Plus, FRAMES, and MoNaCo. Eight harnesses share the gpt-oss-120b backbone and benchmark-specific frozen retrieval infrastructure, although their research allowances differ.

Benchmark PANINI-DR vs. SLEUTH: official score PANINI-DR vs. SLEUTH: observed support
BrowseComp-Plus 73.3% vs. 57.3% accuracy 70.7% vs. 54.0%
FRAMES 79.1% vs. 77.2% accuracy 66.1% vs. 55.2%
MoNaCo 0.549 vs. 0.556 F1 0.458 vs. 0.396 F1

Source: Section 4, Tables 1–2. MoNaCo uses item-level F1; the other benchmarks use percentage-based accuracy and support.

The strongest official-answer improvement occurs on BrowseComp-Plus. PANINI-DR answers 26 of 50 hard questions correctly, versus 14 for SLEUTH. Its overall 16.0-percentage-point accuracy advantage has a paired 95% confidence interval of +8.0 to +24.0 points. On FRAMES, the 1.9-point advantage has an interval of −1.0 to +4.9 points, which does not exclude zero.

Component ablations provide additional evidence about the architecture. In a reduced harness, enabling structured planning and persistent research state raises BrowseComp-Plus accuracy from 41.3% to 58.0% without answer recovery. Applying recovery to both configurations preserves a 16.7-point difference, from 56.7% to 73.3%. Recorded reuse of discovered entities in later searches also rises from 32.1% to 77.4%.

A separate verification ablation raises BrowseComp-Plus accuracy from 57.3% to 73.3% and observed support from 54.7% to 70.7%. Among 101 subsequently endorsed proposals following a rejection, 77 change their candidate value or cited sources.

These are component-package experiments, not clean isolation of individual mechanisms. The planning intervention changes persistent state and revision capabilities together; the verification intervention changes model review alongside citation and evidence-attachment checks. The recorded behavior is consistent with the proposed mechanism, but does not identify each component’s independent contribution.

Evidence recoverability is measurable, but it is not entailment

The paper introduces observed support to assess whether officially credited answer content can be recovered from saved research evidence. On BrowseComp-Plus and FRAMES, a question receives support credit only if the answer is correct and every checkable answer part passes deterministic evidence matching. MoNaCo retains item-level partial credit.

This measure identifies whether the relevant content appears in the retained record. It does not prove that the cited material logically entails the conclusion. A date appearing in a retrieved document, for example, may concern the wrong event.

The authors examine this limitation through complementary audits rather than treating observed support as a complete grounding test.

A model-based claim audit retains only claims judged supported and regrades the reconstructed answers. PANINI-DR leads SLEUTH and HyMem across all three benchmarks under this grounded-only test (Appendix C.2, Table 5). A separate common-reader evaluation asks the same model to answer from each system’s saved evidence without additional retrieval; PANINI-DR again leads the compared systems under both answer-present and uncertain-evidence prompts (Appendix C.3, Table 6). Neither audit removes dependence on model judgment, and evidence volumes are not equalized.

Derived figures require another distinction. A calculation can be properly supported even when its result does not appear literally in a source. Replaying PANINI-DR’s recorded operations using source-matched inputs raises FRAMES observed support from 66.1% to 72.3% and MoNaCo support F1 from 0.458 to 0.498 (Appendix C.4, Table 7). This is a PANINI-DR-only audit because equivalent baseline calculation chains were unavailable.

Together, the audits support retaining both source evidence and computation provenance. They do not establish that deterministic matching can replace contextual validation.

More checking has a cost, and completeness remains unresolved

Verification intensity is a design variable rather than an unconditional improvement. Increasing the retry cap from five to ten reduces BrowseComp-Plus accuracy from 73.3% to 66.0%, even as the number of warned resolved bindings falls to zero in the ten-retry configuration (Appendix D.2.3, Table 13). Additional checking can consume resources without improving the eventual answer.

The computation difference is material. In separate engine measurements, PANINI-DR submits approximately 19.2 times SLEUTH’s input tokens, with 66.5% served from prefix cache. After accounting for cached inputs, estimated model-core computation remains approximately 7.3 times greater. Eight independently sampled SLEUTH runs with answer voting consume 1.12 times PANINI-DR’s estimated model-core computation but achieve only 64.7% BrowseComp-Plus accuracy, versus PANINI-DR’s 73.3%.

These estimates exclude attention and runtime overhead. They are not measurements of latency, energy use, or deployed infrastructure expense.

Completeness presents a different limitation. On MoNaCo collection questions, PANINI-DR achieves the highest collection precision, 0.690, but recall of only 0.462. HyMem reaches 0.566 collection recall. The architecture’s stronger evidence discipline does not ensure that every relevant item has been discovered.

The evidence also comes from frozen corpora, one shared backbone, and comparisons with unequal harness budgets and some reconstructed baseline implementations. Open-web performance and general deployment economics remain untested.

Deployment requires separate acceptance criteria

For an investment-research or enterprise-reporting team, the relevant decision is not simply whether to purchase a more capable research agent. It is whether a structured research workflow provides sufficient improvements on the organization’s actual questions to justify its additional inference and engineering costs.

Cognaptus infers that evaluation should distinguish six outcomes: final-answer correctness, recoverability from retained evidence, semantic support for claims, reproducibility of derived figures, completeness when collecting multiple items, and actual inference expense. Each requires its own acceptance threshold and failure analysis.

Where research depends on multiple linked facts, persistent node-level findings and pre-reuse verification deserve testing. Where outputs enumerate all eligible companies, suppliers, or regulatory requirements, explicit coverage and recall checks are necessary. Where calculations affect decisions, source-linked inputs and replayable operations should be retained.

PANINI-DR demonstrates that research-state design can improve both answer quality and evidence recoverability within a controlled evaluation. It also shows why neither metric alone determines readiness for a consequential workflow. The purchasing or architecture decision should be based on the combination of evidence quality, task completeness, and measured deployment cost—not the answer leaderboard alone.


Cognaptus: Automate the Present, Incubate the Future.


  1. M. Yigit Turali and Shreyas Rajesh and M. Efe Lorasdagi and Chenda Duan and Kartik Sharma and Vwani Roychowdhury (2026). PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces. arXiv:2610.15208. https://arxiv.org/abs/2610.15208 ↩︎