Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction

1 Introduction

• DCI formalizes raw-corpus search through composable terminal tools, replacing embedding models, vector indexes, and top-k APIs with agent-directed exploration, exact matching, local inspection, and refinement.The entire corpus remains available, and the agent performs semantic interpretation. Strategically capable agents can compose tools such as grep, file reads, and scripts to combine clues and enforce lexical constraints.Source: §1 · pp. 2–3
• 80.0% versus 69.0% accuracy, with costs falling from $1,440 to $1,016, follows replacing Qwen3-Embedding-8B with DCI on BrowseComp-Plus using the same Claude Sonnet 4.6 backbone.On BrowseComp-Plus (Chen et al., 2025b), this represents an accuracy improvement of 11.0 points and a cost reduction of 29.4%.Source: §1 · p. 2
• 83.0 average accuracy on multi-hop QA and 68.5 average NDCG@10 on IR ranking exceed their respective retrieval baselines by 30.7 and 21.5 points.Both results use DCI with Claude Code. Comparisons are against the strongest retrieval-agent baseline (Gao et al., 2025) and best retrieval baseline (Liu et al., 2025), respectively.Source: §1 · p. 2
• A minimal Pi/GPT-5.4 nano harness using only bash and read also outperforms QA and IR baselines while improving BrowseComp-Plus performance and reducing cost.The harness uses runtime truncation and compaction for context management. DCI requires no offline embedding or indexing and adapts to changing files in local, heterogeneous corpora.Source: §1 · pp. 2–3
• DCI’s gains primarily reflect finer-grained local search and verification of surfaced evidence, rather than simply finding more gold documents.Retrieval interface resolution denotes access to units smaller and more precise than documents or passages, supported by trajectory-level coverage and localization analyses. Benefits persist even with highly restricted tools.Source: §1 · pp. 2–3
• Fixed top-k interfaces constrain agentic exploration, motivating retrieval as an interface-design problem rather than solely a retriever-design problem (Zhang et al., 2024; Singh et al., 2025).Restricted evidence exposure hinders intermediate-entity discovery, sparse-clue aggregation, exact lexical constraints, and plan revision after local inspection. Dense and sparse retrieval nevertheless remain scalable and effective for large, static corpora.Source: §1 · pp. 2–3

Retrieval-Augmented Generation (RAG)

• Classical RAG retrieves evidence through prebuilt sparse or dense indexes before generation; subsequent improvements retain this underlying retrieval structure.Examples include LLM-based reranking (Sun et al., 2023; Zhuang et al., 2025; Weller et al., 2025) and adaptive RAG (Jeong et al., 2024).Source: §2 · p. 3

The Rise of Agentic Search

• Direct corpus interaction shifts semantic understanding to the LLM, addressing the fixed-retriever interface bottleneck in multi-step agentic search.Existing agents iteratively search, accumulate evidence, and refine plans, but still access the corpus through a fixed retriever (Team et al., 2025b,a; Li et al., 2026b).Source: §2 · p. 3

Coding Agents

• SWE-agent (Yang et al., 2024) outperforms direct prompting using terminal tools, while Agentless can match full agent frameworks with grep-and-read.CLI primitives also support precise code localization (Sutawika et al., 2026), demonstrating their usefulness beyond completing coding tasks.Source: §2 · pp. 3–4
• DCI positions terminal tools as a broader retrieval interface for agentic search, extending beyond keyword-search document QA over raw PDFs (Subramanian et al., 2025).That document-QA agent can approach vector-database RAG. Terminal tools provide precision, flexibility, and composability through pipelines (GNU Project, 2025).Source: §2 · p. 4

3 Direct Corpus Interaction

• DCI replaces retriever-ranked top-k results with direct command-line access to raw corpora, bypassing embedding models, vector indexes, and retrieval APIs.Retriever-mediated access constrains observations to evidence exposed through scoring and ranking. DCI instead uses grep, rg, find, glob, file reads, and lightweight scripts to expose matches, surrounding context, paths, counts, and metadata.Source: §3 · p. 4

3.1 DCI Agent Implementations

• DCI-Agent-Lite isolates terminal-only corpus access, while DCI-Agent-CC adds stronger prompting, tool orchestration, and context handling without changing the retrieval paradigm.Lite contains no offline indexing, dense retriever, or reranker. CC uses Claude Code (Anthropic, 2025), whose stronger runtime support improves stability on long-horizon search and heterogeneous corpora.Source: §3.1 · pp. 4–5

3.2 Runtime Context Management for DCI

• Truncation limits individual tool outputs, zero-LLM compaction replaces older outputs with placeholders, and summarization replaces compacted history while preserving recent context.Compaction operates in memory after accumulated tool output exceeds a configured threshold, preserving tool-call structure. Summarization uses a model-generated summary under additional context pressure.Source: §3.2 · p. 5, Figure 3 · p. 5
• L0–L4 vary how much tool-mediated evidence survives in working context without changing the retrieval interface.L0 disables management; L1 caps tool results at 50,000 characters; L2–L4 cap them at 20,000, with L3 additionally compacting above 240,000 accumulated characters while retaining the latest 12 turns. L4 additionally summarizes if estimated context tokens remain above a threshold after compaction, retaining the latest 20,000 tokens and suppressing further attempts after three consecutive summarization failures within a session.Source: §3.2 · p. 6

3.3 Evaluating DCI: Coverage and Localization

• coverageany, coveragemean, and coverageall measure whether trajectories surface any gold document, the average fraction of gold documents, or all gold documents, respectively.These complement answer accuracy by measuring broad evidence access rather than within-document evidence use. A document counts as surfaced when explicitly recorded as a retrieved snippet or a file returned by a tool call.Source: §3.3 · p. 6
• Localization averages each surfaced gold document’s best segment-level score, rewarding exposed snippets that are small relative to the full document.Fixed-width character segments replace lines because web and PDF exports can collapse layout into long lines. Conditional on reaching a gold document, localization measures whether the trajectory isolates a small, usable evidence span.Source: §3.3 · pp. 6–7

4.1 Experimental Setup

• Two DCI scaffolds evaluate agentic search, knowledge-intensive QA, and IR ranking under lightweight and high-capacity configurations.Lite exposes bash and read, uses GPT-5.4 nano (OpenAI, 2026), and defaults to L3; ClaudeCode uses Claude Sonnet 4.6 (Anthropic, 2026), disables web tools and subagents, and blocks the data directory to prevent leakage. Evaluation covers BrowseComp Plus (Chen et al., 2025b), six QA datasets using 2018 Wikipedia (Karpukhin et al., 2020), four BRIGHT datasets (Su et al., 2025), and two BEIR datasets.Source: §4.1 · p. 7

4.2 Main Results

• 83.0% average QA accuracy places DCI-Agent-CC 30.7 points above ASearcher-Local-14B (52.3%); DCI-Agent-Lite achieves 68.0%.The two DCI agents occupy the top two positions across all six datasets. Relative to ASearcher-Local-14B, DCI-Agent-CC gains 30 points on HotpotQA, 26 on 2Wiki, and 50 on MuSiQue.Source: Table 2 · p. 8, §4.2 · p. 9
• 68.5% average NDCG@10 puts DCI-Agent-CC 21.5 points above ReasonRank-32B (47.0%), with the best scores on all six IR datasets.DCI-Agent-Lite ranks second overall with average NDCG@10 of 56.7, exceeding ReasonRank-32B by 9.7 points.Source: Table 3 · p. 8, §4.2 · p. 9
• 80.0% versus 69.0% BrowseComp-Plus accuracy favors DCI-Agent-CC over Qwen3-Embedding-8B retrieval with the same Sonnet 4.6 backbone, while cost falls 29.4%.Cost falls from $1,440 to $1,016, and DCI-Agent-CC exceeds the strongest retrieval baseline, GPT-5 + Qwen3-Embedding-8B (71.7%), by 8.3 points. DCI-Agent-Lite achieves 62.9% accuracy for $93, versus 66.0% for o3 + Qwen3-Embedding-8B, saving $647.Source: Figure 4 · p. 9, §4.2 · p. 9

4.3 Controlled Ablations and Mechanism Analysis

• 48.4 versus 21.7 localization scores favor DCI-Agent-Lite over Qwen3-Embedding-8B despite lower mean gold-document coverage (28.0 versus 56.7).DCI-Agent-CC wins 176 matched questions versus 76 losses; in 142 wins, retrieval already surfaced gold evidence, indicating partial-chain or post-retrieval failures. Bash interactions compose lexical constraints, inspect local context, and verify exact spans; Lite gains 28 accuracy points with comparable any-document coverage (70.0 versus 74.0) but roughly twice the tool calls.Source: §4.3 · pp. 9–10, Figure 4 · p. 9, Table 4 · p. 10
• 61% accuracy with only read + grep exceeds the Qwen3-Embedding-8B retrieval-agent baseline (45%) by 16 points at nearly identical tool-call intensity.Enabling bash adds another 12 accuracy points but substantially increases tool usage, latency, and compute. The ablation indicates that constrained direct corpus access already helps, while greater tool expressivity trades efficiency for additional gains.Source: Table 5 · p. 10, §4.3 · p. 11
• 13.6 points of accuracy are lost when DCI-Agent-CC’s corpus grows from 100K to 200K documents, while latency and cost more than double.Tool calls rise from 38.5 to 86.9 per question as additional distractors expand the search space. At 400K documents, accuracy falls to 37.5%, average tool usage reaches 122.4 calls, and 20 examples exhaust the maximum tool budget.Source: Figure 5 · p. 10, §4.3 · p. 10
• 77 answer accuracy makes L3 the best context-management policy, despite lower retained gold-evidence coverage (27.0) than the fastest policy, L1 (31.3).L2 has the lowest cost but the worst accuracy (69), and L4 declines after the L3 peak. Selective forgetting supports continued hypothesis revision, whereas insufficient compression permits drift and overly blunt compression loses useful intermediate structure.Source: Table 6 · p. 11, §4.3 · p. 11

5 Conclusion

• DCI is competitive across ranking-oriented IR, multi-hop QA, and end-to-end agentic search without an offline index.Agents search and verify the raw corpus using general-purpose terminal tools rather than conventional retrievers. Controlled analyses attribute the advantage to retrieval-interface resolution, enabling higher-value inspection, verification, and compositional search rather than necessarily recovering more of the gold chain.Source: §5 · p. 11

A.1 Benchmarks

• 50 examples per dataset are randomly sampled, except BRIGHT Biology, Earth Science, Economics, Robotics, and Bamboogle, which use full test sets.Sampling is used for evaluation efficiency.Source: §A.1 · p. 18

Agentic Search

• BrowseComp-Plus (Chen et al., 2025b) compares indexed retrieval and index-free DCI on the same fixed, curated corpus for multi-document research questions.The corpus includes human-verified supporting documents and mined hard negatives. Embedding-based baselines use an offline search engine with a BM25/Qwen3-Embedding-8B FAISS index, while DCI accesses the same corpus through terminal tools.Source: Agentic Search · p. 18

Knowledge-Intensive QA

• Knowledge-intensive QA spans single-hop retrieval and multi-hop reasoning, excluding ambiguous questions and cases whose answers may have changed since corpus collection.NQ and TriviaQA (Joshi et al., 2017) assess single-hop retrieval. Bamboogle (Press et al., 2023) requires exactly two inferential steps across documents; HotpotQA (Yang et al., 2018) includes bridge and comparison questions; 2WikiMultiHopQA (Ho et al., 2020) supplies annotated reasoning chains; MuSiQue (Trivedi et al., 2022) tests dependent reasoning chains designed to minimize shortcuts.Source: Knowledge-Intensive QA · p. 18

IR Ranking

• Four BRIGHT (Su et al., 2025) datasets assess reasoning-intensive retrieval; BEIR contributes counter-argument retrieval and scientific fact-checking tasks.BRIGHT includes Biology (103 queries), Earth Science (116), Economics (103), and Robotics (101). BEIR includes ArguAna (Wachsmuth et al., 2018; 1,406 queries, 50 sampled) and SciFact (Wadden et al., 2020; 300 queries, 50 sampled).Source: IR Ranking · p. 18

Proprietary Models

• GPT-5.4 nano (OpenAI, 2026) powers DCI-Agent-Lite, while Claude Sonnet 4.6 (Anthropic, 2026) powers DCI-Agent-CC.Among BrowseComp-Plus baseline agents, o3 delivers the strongest retrieval-augmented performance, while GPT-5.2 serves as a standard retrieval backbone. Other evaluated backbones include Claude Sonnet 4.5 (Anthropic, 2025), Claude Haiku 4.5, GLM-4.7 (Z.AI, 2025), and Kimi K2.Source: Proprietary Models · p. 19

Retrieval Agents

• Open-weight retrieval-agent baselines span search-invoking reinforcement learning, simulated-search training, stateful tool use, and asynchronous deep research.R1-Searcher-7B (Song et al., 2025) uses two-stage outcome-supervised RL without process rewards or distillation; Search-R1-32B (Jin et al., 2025) interleaves reasoning and retrieval through RL; ZeroSearch-7B (Sun et al., 2025) simulates retrieval using SFT and progressively degraded documents. Verl-Tool-Search-7B-DAPO (Jiang et al., 2025) supports stateful interactions through asynchronous rollouts and standardized APIs, using DAPO (Yu et al., 2025); ASearcher-Local-14B (Gao et al., 2025) decouples trajectory collection from training through asynchronous RL and supports up to 128 turns.Source: Retrieval Agents · p. 19

Sparse & Dense Retrieval

• Retrieval baselines cover lexical matching, indexed dense embeddings, and explicit reasoning over candidate documents.BM25 provides sparse retrieval for IR ranking and BrowseComp-Plus; text-embedding-3-large provides indexed dense retrieval on BRIGHT and BEIR. E5 (Wang et al., 2022) is the primary QA retriever, Qwen3-Embedding-8B (Zhang et al., 2025) is the primary BrowseComp-Plus dense retriever, and Rank-R1-14B (Zhuang et al., 2025) reranks through explicit reasoning learned with RL from limited relevance labels without reasoning supervision.Source: Sparse & Dense Retrieval · pp. 19–20

A.3 Metrics

• GPT-4.1 judges BrowseComp-Plus and knowledge-intensive QA answers by comparing predictions with reference answers only.Short, well-defined references make the comparison largely unambiguous while accommodating semantic and numerical equivalence.Source: §A.3 · p. 20
• NDCG@10 is the main evaluation metric for IR ranking benchmarks.This differs from the final-answer correctness assessment used for BrowseComp-Plus and knowledge-intensive QA.Source: §A.3 · p. 20
• Retriever-mediated resolution measurements treat each returned document and its preview as one observation-level candidate, with top-k retrieval producing k candidates.Snippet lengths are the lengths of the returned previews.Source: §A.3 · p. 20
• DCI candidate sets are reconstructed by aligning raw tool observations to gold documents through explicit file paths or matched local text.Path-only surfacing or unmatched text still counts toward document coverage but receives a full-document snippet length. This yields a low localization score rather than credit for precise evidence exposure.Source: §A.3 · p. 20
• Grep-style DCI searches treat every aligned matched line as a separate candidate whose snippet is that line.The metric retains the best seg-score across multiple aligned lines from one gold document. Across 100 cases and 3,168 commands, aggregating correct and incorrect runs, rg | head and rg | rg predominate while full-document reads are rare.Source: §A.3 · p. 20, Figure 6 · p. 21
• Read-style DCI inspection uses the exposed read span as its snippet only when a gold-document read sufficiently overlaps annotated gold evidence.Snippet assignment is therefore conditional on evidence overlap, not merely on opening the correct file.Source: §A.3 · p. 20

B.1 In-Depth Analysis of DCI-Agent-Lite Search Patterns

• 56.2% of Bash interactions use search-plus-limit operations, while 20.6% use chained searches, making these the dominant DCI-Agent-Lite command patterns.Full-document reads account for only 0.1%. The authors interpret this distribution as bounded snippet inspection and iterative exact-match verification rather than broad document reading.Source: §B.1 · p. 21
• Six operation patterns span corpus exploration, broad keyword search, iterative narrowing, targeted document reading, in-document deep search, and cross-document comparison.These patterns roughly follow chronological order, progressing from locating candidates to localized inspection and cross-document verification.Source: §B.1 · p. 21

C.1 QA Instruction

• QA instructions restrict agents to corpus-only ripgrep/Bash searches, require parallel targeted searches, and demand ruling out competing answers before committing.They explicitly prohibit web browsing, Agent-tool use, and spawning subagents.Source: §C.1 · p. 22
• QA responses must cite supporting document paths inline and provide an explanation, concise exact answer, and confidence score.Confidence must fall below 50% when evidence is weak, ambiguous, or missing.Source: §C.1 · p. 22

C.2 IR Instruction

• IR instructions require parallel corpus-only Grep/Bash searches, diverse keywords, gap-driven follow-ups, exhaustive search angles, and careful review of candidate documents.Agents must continue beyond finding a few documents and inspect each candidate before including it.Source: §C.2 · p. 22
• IR output balances recall and precision under NDCG, excludes tangential documents, and ranks at most 20 directly relevant documents by usefulness.The instructions seek every genuinely relevant document, defining relevance as directly answering the question or providing essential supporting evidence.Source: §C.2 · pp. 22–23

C.3 LLM as Judge

• The LLM judge extracts the final answer and checks equivalence against the supplied correct answer, allowing a small numerical margin of error.It evaluates answer matching rather than solving the problem; absent final answers become ‘None’. It extracts reported confidence, defaulting to 100% when none is available.Source: §C.3 · p. 23

D Case Study

• The case-study introduction presents DCI-Agent-Lite and DCI-Agent-CC as demonstrations of more elegant, efficient, and robust deep research.The supplied introduction provides no concrete examples or measurements supporting these characterizations.Source: D Case Study · p. 24

D.1 Case 1: Demonstration Example

• The agent searches the closed corpus directly using precise lexical matching, beginning with a Grep query for “Rocinante.”The initial query requests matching document content with a ten-result limit.Source: §D.1 · p. 24
• After a broader Grep query encounters corpus-size limitations, the agent switches to a bounded Bash search to confirm the answer.The Bash command stops after three “Rocinante” matches and limits output to 1,500 characters.Source: §D.1 · pp. 24–25
• The agent answers Rocinante, supported by multiple wiki-corpus documents explicitly identifying Don Quixote’s horse.The final answer cites documents 135082 and 339679 as corroborating evidence.Source: §D.1 · pp. 24–25

D.2 Case 2a: DCI-Agent-CC on Agentic Search

• A World Cup hypothesis guided shell-based corpus exploration, with filename inspection surfacing candidate records before direct reads of match and interview files.The question contained twelve interlocking clues, and the full trace spanned 80 turns.Source: §D.2 · pp. 26–27
• 34,520 spectators, one yellow card, two substitutions per team, and officials from three different countries matched clues for England’s 1-0 victory over Belgium.The match record named David Platt as the goalscorer and Stadio Renato Dall’Ara as the venue.Source: §D.2 · p. 27
• Platt’s interview confirmed the “technique” clue, while Paul Parker named Scifo and Chris Waddle connected him to Auxerre.Waddle recalled having seen Scifo perform there, supplying the club connection required by the question.Source: §D.2 · pp. 27–28
• 20 days separate Enzo Scifo’s and Stéphane Demol’s birth dates; both share David Platt’s 1966 birth year, identifying Scifo as the older candidate.Scifo was born on 19 February, Demol on 11 March, and Platt on 10 June.Source: §D.2 · p. 28
• AJ Auxerre is the final answer, established by identifying Enzo Scifo as the relevant Belgian player and connecting him to Chris Waddle’s interview recollection.The identified match was England’s 1-0 victory over Belgium in the 1990 World Cup round of 16.Source: §D.2 · p. 28

D.3 Case 2b: DCI-Agent-Lite on Agentic Search

• Targeted rg searches for scoreboard, principal, and tank progressively narrow the search space toward candidate documents.The agent initially selects scoreboard because it might be distinctive enough to narrow the corpus.Source: §D.3 · pp. 29–30
• Candidate-document reading supports verification of publication timing, plot details, and the author’s background before the final answer.The agent’s reasoning aligns May 19, 2015 with the lobster tank, blue-eyed boy, and scoreboard clues. It reads an author interview and a book listing while checking the remaining constraints.Source: §D.3 · pp. 29–31
• DCI-Agent-Lite identifies Made You Up by Francesca Zappia using only Bash and Read to solve a multi-constraint BrowseComp-Plus query.The question combines publication timing, plot details, and author-biography clues.Source: §D.3 · pp. 29, 31

D.4 Case 3a: DCI-Agent-CC on Knowledge-Intensive QA

• After finding a truncated film article, the agent refines its search toward the virgin character, Jehnna, and Olivia d’Abo.The agent explicitly recognizes that the initial result lacks the relevant character information.Source: §D.4 · p. 32
• The film passage names Jehnna as a virgin played by Olivia d’Abo, while the biography passage confirms her role as Princess Jehnna.The biography also identifies this supporting role as d’Abo’s film debut.Source: §D.4 · p. 32
• DCI-Agent-CC answers Olivia d’Abo, identifying her character Jehnna as the virgin in Conan the Destroyer and citing two corpus documents.The final answer cites documents 2205279 and 4851059.Source: §D.4 · p. 32

D.5 Case 3b: DCI-Agent-Lite on Knowledge-Intensive QA

• DCI-Agent-Lite searched for St. Louis Cardinals with football context, then inspected neighboring text to extract the count.Football context distinguished the team from the identically named baseball team.Source: §D.5 · p. 33
• 4 NFL teams were identified for St. Louis: the All-Stars, Gunners, Cardinals, and Rams.The extracted Wikipedia sentence explicitly stated the count and enumerated the teams.Source: §D.5 · p. 34

D.6 Case 4a: DCI-Agent-CC on Information Retrieval

• DCI-Agent-CC searched jointly for phototaxis, light stimulus, positive phototaxis, proximate, and ultimate causation to locate relevant files.The matches included both phototaxis and causation documents, connecting behavioral evidence with the conceptual distinction needed by the query.Source: §D.6 · pp. 35–36
• The phototaxis evidence describes directed movement toward light and innate positive responses in flying insects.These observations support light itself as the proximate attractant.Source: §D.6 · pp. 36–37
• Five gold documents spanning phototaxis and proximate/ultimate causation were returned by DCI-Agent-CC for the BRIGHT-Biology insect-light query.The returned evidence comprised three phototaxis documents and two proximate/ultimate causation documents.Source: §D.6 · pp. 35, 37
• The LED argument excludes heat as the immediate cue without ruling out an evolutionary association between light and heat.The answer distinguishes light-driven orientation as a proximate mechanism from a separate ultimate evolutionary explanation.Source: §D.6 · pp. 35, 37

D.7 Case 4b: DCI-Agent-Lite on Information Retrieval

• The agent prioritized recall with targeted local searches, then refined several thyroid-related hits using the exact disease phrase and thyroid-volume outcome.The refined search surfaced 9745001.txt with both the target population and explicit volume-reduction evidence.Source: §D.7 · p. 38
• Before final ranking, the agent listed multinodular-goitre candidates and read 9745001.txt to confirm the treatment, population, and measured outcome.Disease-phrase matching returned 9745001.txt, 6751418.txt, 43122426.txt, and 26026009.txt.Source: §D.7 · pp. 38–39
• 60% median thyroid-volume reduction (73 ml to 29 ml) at 24 months supported radioiodine treatment’s effect in non-toxic multinodular goitre (9745001.txt).The reported significance was p < 0.0001; the paper described radioiodine as an attractive alternative to surgery.Source: §D.7 · p. 39

D.8 Case 5a: Failure Case of DCI-Agent-CC in Agentic Search

• DCI-Agent-CC identifies Denis Suarez and FC Barcelona but misattributes the bench appearance’s opponent; the supplied correct answer is FC Krasnodar.The query requires linking a 2019 English football water-bottle scuffle to a player’s loan history and a European bench appearance for his parent club.Source: §D.8 · p. 40
• Extensive corpus searches span match reports, player profiles, and club archives, while the agent explores exact match statistics and an alternative Andre Gomes hypothesis.The statistical clues specify six away-team shots on goal, four yellow cards, and four corners; the Gomes hypothesis prompts further speculation about loan timing.Source: §D.8 · pp. 41–42
• The agent uses Suarez’s January 2019 Arsenal loan date to narrow the opponent candidates to PSV Eindhoven, Tottenham Hotspur, and Inter Milan.This establishes a possible timeline, but the agent still needs evidence identifying a specific bench appearance.Source: §D.8 · p. 43
• The agent acknowledges lacking concrete evidence of Suarez’s Champions League bench appearances and recognizes the need to search directly for the water-bottle incident.It explicitly contrasts direct incident verification with continued speculation about possible teams.Source: §D.8 · p. 45

• 6 shots on goal, 4 yellow cards, and 4 corner kicks became targeted away-team clues as the agent reconsidered the player and pursued incident-based searches.It proposed searching for water-bottle incidents or football altercations during 2019 rather than guessing the opponent.Source: §D.8 · p. 46
• Missing squad sheets and incident evidence left the agent unable to verify the fixture as its search budget ran low.The agent instead reconstructed Barcelona’s goalscorers against Inter Milan, distinguishing the home and away matches without establishing Suarez’s bench appearance.Source: §D.8 · pp. 46–47
• The final answer named Inter Milan as Barcelona’s opponent in a 2018-19 Champions League Group B match, despite unresolved evidence about the bench appearance.Source: §D.8 · pp. 47–48

• Repeatedly broad rg patterns returned irrelevant documents and failed to converge on the correct corpus files.Source: §D.9 · p. 49
• The clues combined a 2000s release, wealthy and poor protagonists, actors born in the 1960s, and a director’s dispute with an actor’s brother as of November 2023.Source: §D.9 · p. 49
• The agent recognized its queries were too broad and proposed actor-brother phrases and line-based dispute filters.It later intersected file lists for date, dispute, director, brother, and actor patterns.Source: §D.9 · pp. 50–51
• DCI-Agent-Lite incorrectly answered The Family Man; the correct answer was Dosti: Friends Forever.The case description characterizes the answer as hallucinated, matching only some surface-level keywords.Source: §D.9 · pp. 49, 51