Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses

Motivation

• Harness-1 addresses inefficient retrieval-agent RL by externalizing routine state management instead of requiring the policy to reconstruct useful search state from transcripts.The policy focuses on semantic search behavior, while the harness maintains recoverable state such as candidates, evidence links, verification, and curation.Source: §1 · p. 2

Harness design

• The harness maintains a persistent working memory containing candidate and curated documents, evidence-graph summaries, verification state, compressed observations, and context-budget markers.The policy edits this state through search, curation, verification, and document-review actions; the same renderer supports teacher data, SFT, RL, and evaluation.Source: §1 · p. 2, Figure 2 · p. 3

Trainability requirements

• Stateful search harnesses require warm-started curation, compact derived-state rendering, and diversity-preserving incentives to make policy training informative and avoid repeated search.These requirements address empty early rewards, context competition from verbose state, and rewards that neglect curation or verification.Source: §1 · p. 2

Results

• 0.730 average curated recall is achieved by Harness-1 across eight retrieval benchmarks spanning web, finance, patents, and long-context multi-hop search.The system also improves downstream answer accuracy when its curated set is passed to frozen frontier generators.Source: §1 · pp. 2–3
• Harness-1 improves retrieval across specialized corpora, open-web tasks, and held-out multi-hop QA benchmarks, extending beyond its training domains.The paper presents this as evidence that explicit search-state operations support retrieval quality and modular RAG answer accuracy across eight benchmarks.Source: §1 · p. 3

2 Harness-1

• Harness-1 maintains persistent WORKINGMEMORY while the policy decides what to search, inspect, keep, verify, and when to stop.The harness executes structured actions, updates state, and renders the next observation instead of merely appending tool outputs to the prompt.Source: §2 · p. 3, Table 1 · p. 4

2.1 Policy actions as edits over working memory

• The action interface supports retrieval, memory inspection, curation, verification, and termination, with curation maintaining an importance-tagged set capped at M=30 documents.Auto-seeding initializes the curated set with the top k=8 reranked candidates at fair importance, after which the policy refines them.Source: §2.1 · p. 4

2.2 Derived-state rendering: evidence graph, compression, and deduplication

• Derived-state rendering exposes evidence structure and search progress through entity graphs, compressed results, deduplication, summaries, and explicit context-budget signals.The harness extracts entities to identify bridge documents, compresses search outputs to top K=4 sentences, and retains full text for explicit document reads.Source: §2.2 · pp. 4–5

2.3 Training: SFT and RL

• Training uses narrow SFT for harness operation followed by on-policy CISPO [3] RL over full search episodes, with terminal rewards separating discovery from final selection.RL uses SEC training queries, terminal-only rewards, a 40-turn cap, no KL anchor, and reward components for search quality, answer evidence, tool diversity, and turn cost.Source: §2.3 · pp. 5–6

3.1 Experimental Setup

• Experiments measure Recall, Trajectory Recall, and Final-Answer Recall across eight benchmarks, prioritizing evidence coverage because curated-set size is part of the agent interface.All methods share retrieval primitives, web backend, reranker, and a final budget of at most 30 returned documents.Source: §3.1 · p. 7

3.2 Results

• 0.730 average curated recall makes Harness-1 the strongest open retrieval subagent, beating Tongyi DeepResearch 30B by +11.4 points and trailing only Opus-4.6 among listed frontier retrievers.It also exceeds GPT-5.4, Sonnet-4.6, Kimi-K2.5, and GPT-OSS-120B under the evaluation protocol.Source: §3.2 · p. 7
• +17.0 recall points on held-out transfer benchmarks exceeds the +7.9-point source-family gain by 2.2×, supporting transfer of domain-general search-state operations.Held-out gains are +32.8, +18.4, +7.2, and +9.5 points across LongSealQA, Seal0QA, FRAMES, and HotpotQA.Source: Figure 3 · p. 7, §3.2 · pp. 7–8
• Disabling all harness mechanisms reduces Recall to 0.513 (−12.2%) and Final-Answer Recall to 0.624 (−6.4%), exceeding any individual ablation’s Recall loss.Six of seven individual mechanisms reduce Final-Answer Recall by −3.9% to −7.9%; content dedup instead changes Recall by +4.6% and Final-Answer Recall by +1.6%.Source: training) · p. 8, Table 3 · p. 8

Training Dynamics

• With tool-diversity reward wdiv, diversity stabilizes around ∼4.30 and curated recall reaches ∼0.60, versus ∼3.5 diversity and ∼0.53 recall without it.Without wdiv, the policy favors fan-out search and bypasses curation and verification; with wdiv, it combines search, curation, and verification.Source: Training Dynamics · p. 9, Figure 5 · p. 9
• Harness-1’s curated sets improve downstream modular RAG answer accuracy and outperform open subagents when passed to frozen frontier generators.The evaluation spans web, finance, patents, and multi-hop QA, with four benchmarks held out from Harness-1’s SFT and RL.Source: §3.1 · p. 6, Training Dynamics · p. 9

4 Conclusion

• Harness-1 separates semantic decisions—search, curation, verification, and submission—from recoverable bookkeeping maintained in environment-side state.The harness maintains candidate pools, importance-tagged evidence, verification records, evidence links, and budget-aware context rendering.Source: §4 · p. 9
• Across eight benchmarks, Harness-1 achieves the strongest average recall among evaluated open search agents and remains competitive with much larger frontier-model searchers.Its gains on held-out transfer benchmarks and component ablations suggest the harness is central to what the policy learns to use, rather than merely an implementation detail.Source: §4 · p. 9
• Held-out transfer gains and component ablations suggest the stateful harness is central to what the policy learns to use, rather than merely an implementation detail.The conclusion identifies stateful harness design as an important direction for retrieval-agent reinforcement learning.Source: §4 · p. 9

Contents of Appendix

• Appendix A discusses connections to agentic search, tool orchestration, and harness engineering, while Appendix B covers limitations, ethics, and broader impact.Future work could replace regex-based extraction with learned entity linking, relation extraction, and uncertainty-aware evidence organization.Source: §4 · p. 9

A Related Work

• Harness engineering treats the environment layer and policy interface as major determinants of system performance, beyond simply improving tool use within a fixed interface.The same model can vary by tens of points across harnesses [49] [47], while tool orchestration usually fixes the interface [18].Source: A Related Work · p. 16
• Unlike hand-designed or automatically searched harnesses [49] [54] [22] [35], Harness-1 shapes a retrieval interface that offloads mechanical bookkeeping while training semantic decisions.This complementary design question concerns how a trainable policy can use the retrieval harness effectively.Source: A Related Work · p. 16
• Agentic-search RL methods improve learned query, search, and reasoning behavior, but longer tasks overload policies with planning, memory, constraint tracking, and context management.Harness-1 instead adds editable retrieval state to the interface, so transitions depend on maintained state rather than only actions and observations.Source: A Related Work · p. 16

B Limitations, Ethics, and Broader Impact

• Harness-1 does not cover breadth-oriented research, open-ended report generation, abstention under missing evidence, or adversarial web environments.Its benchmarks primarily involve needle-in-a-haystack or multi-hop tasks with annotated relevant evidence.Source: B Limitations, Ethics, and Broader… · p. 17
• The evidence graph, LLM-based verifier, and sentence-level BM25 compressor can fail on entity linking, ambiguous claims, and discourse-dependent context.These engineered components may vary in quality across domains and can respectively make extraction errors, verification mistakes, or remove useful context.Source: B Limitations, Ethics, and Broader… · p. 17
• Benchmark size, annotation coverage, confidence intervals, duplicate or incomplete qrels, and harness conditions limit how broadly retrieval metrics indicate real-world search reliability.The paper therefore reports multiple metrics, trajectory diagnostics, ablations, and modular RAG answer accuracy, while treating them as evidence of behavior under specified conditions.Source: B Limitations, Ethics, and Broader… · p. 17
• Explicit curated documents make retrieval-augmented systems more transparent and auditable than relying only on parametric memory.The documents can be inspected, verified, and used by downstream models or human users.Source: B Limitations, Ethics, and Broader… · p. 17
• Unrestricted deployment could expose sensitive or inappropriate information, amplify biased or adversarial evidence, and enable privacy-invasive search or evidence cherry-picking.Responsible use requires corpus access controls, logging or monitoring, rate limits, privacy filters, verification, and human oversight, especially for sensitive domains.Source: B Limitations, Ethics, and Broader… · p. 17
• Harness-1 is not a source of truth or a replacement for expert judgment, particularly in medicine, law, finance, or public policy.Downstream answers should be reviewed by qualified experts and validated against authoritative sources; the planned release is limited to the retrieval subagent, harness code, data-generation pipeline, and evaluation recipe.Source: B Limitations, Ethics, and Broader… · p. 17

C Tool Signatures

• The policy selects tools for retrieving candidates, revisiting evidence, editing the curated set, verifying claims, and ending episodes, while the harness updates working memory.Table 4 lists the functional signatures used across teacher, SFT, RL, and evaluation runs.Source: C Tool Signatures · pp. 17–18, Table 4 · p. 18

Reward configuration

• Table 5 defines the reward weights and penalty coefficients used in the deployed Harness-1 RL run.The table presents these values alongside their configuration names.Source: D Reward & Training Hyperparameters · p. 18
• At termination with an empty curated set, the reward is π∅ = −0.2 and bypasses the remaining formula; otherwise, rewards follow §2.3 and satisfy R ≥ 10−3.These conditions specify the empty-set short circuit and lower clipping behavior.Source: Table 5 · p. 18

Training configuration

• The released checkpoint’s configuration separates SFT, RL, search-environment, retrieval-infrastructure, and evaluation hyperparameters for reproducibility.This separation clarifies which settings belong to model optimization versus the harness and retrieval backend.Source: D Reward & Training Hyperparameters · p. 18
• SFT teaches operation of the stateful interface, after which RL runs over full search episodes using the same harness state renderer and tool set.The deployed run’s exact settings are listed in Table 6, although those values are not provided here.Source: D Reward & Training Hyperparameters · p. 18

E Benchmark Statistics

• The benchmark suite spans static Chroma-backed and live web-backed retrieval across encyclopedic search, finance filings, patents, and multi-hop question answering.The split distinguishes source-family benchmarks from held-out transfer benchmarks.Source: E Benchmark Statistics · pp. 18–19

F Two-Tier Memory and Working-Memory Rendering

• Harness-1 separates an inner WORKINGMEMORY summary that fits the prompt from an outer document store containing every retrieved chunk for review.This two-tier design makes the rendering protocol reproducible while retaining full retrieved text outside the prompt.Source: F Two-Tier Memory and Working-Memory… · p. 19
• WORKINGMEMORY renders the query, importance-ordered curated entries with doc IDs and 120-character snippets, and recent uncurated documents while compressing older ones into truncated ID lists.The document pool shows the 50 most recent uncurated documents; older IDs are capped at 30 with a (+N more) marker.Source: F Two-Tier Memory and Working-Memory… · pp. 19–20

E Benchmark Statistics

• Table 7 specifies each benchmark’s backend, test-query count, domain, and task style.These fields define the evaluation coverage used in the main experiments.Source: E Benchmark Statistics · p. 19, Table 7 · p. 20

F Two-Tier Memory and Working-Memory Rendering

• The memory also retains the last 12 searches and conditionally adds multi-document entity bridges and near-duplicate suppression notices.The evidence graph lists up to 8 frequent entities with document IDs when a bridge exists, while deduplication reports suppressed near-duplicate chunks.Source: F Two-Tier Memory and Working-Memory… · p. 20
• Each observation combines WORKINGMEMORY, up to five prior action-result pairs, and the latest tool result compressed by sentence-BM25 and content-fingerprint deduplication.Search displays are limited to 10 results per call, with 120-character snippets and four retained sentences for search-style observations.Source: F Two-Tier Memory and Working-Memory… · p. 20
• When context exceeds 30,720 prompt tokens, five degradation passes preserve curated material and recent memory as long as possible and guarantee rollout completion.The process progressively truncates document pools and older analysis, drops oldest recent pairs, and finally retains only system instructions, tool descriptions, and the query.Source: F Two-Tier Memory and Working-Memory… · p. 20

G Per-Turn Programmatic Nudges (Result Summaries)

• Programmatic result summaries provide identical factual state and conditional prescriptive nudges to the teacher, SFT model, and policy during RL and inference.The harness—not the trained policy—provides these nudges, and disabling all Harness-1 mechanisms removes them.Source: G Per-Turn Programmatic Nudges (Result… · pp. 20–21

H Harness Algorithms

• The stage driver renders shared state, executes or replays structured actions, dispatches observation, curation, verification, and review handlers, and records summaries and memory snapshots.The same harness interface supports teacher trajectories, SFT replay, RL rollouts, and inference.Source: G Per-Turn Programmatic Nudges (Result… · p. 21, H Harness Algorithms · p. 21
• The harness maintains a capped curated set, importance map, document and evidence stores, verification cache, history, deduplication index, and auto-seeding state while the policy controls semantic choices.The curated set cap is M=30, with importance levels very high, high, fair, and low.Source: H Harness Algorithms · p. 21
• Observation processing keeps up to four BM25-ranked sentences, suppresses duplicates, stores full text, extracts entities and dates into the evidence graph, and auto-seeds eight fair candidates.Existing documents remain accounted for without being re-added, while near-duplicates are suppressed from rendered memory.Source: H Harness Algorithms · p. 22

I.1 Agent System Prompt (Harness-1)

• The Harness-1 system prompt defines retrieval-only behavior, diverse search and inspection tools, immediate curation, verification, importance rules, evidence-graph follow-up, and budget-aware stopping.Episodes allow at most 40 turns; context markers encourage wrapping up above 75% and ending search above 90%.Source: §I.1 · pp. 22–23

I.2 verify Sub-Prompt

• Verification checks each requested document against every claim constraint and stores conservative yes/no judgments with short rationales in the verification cache.The verifier answers yes only when the document directly supports every part of the claim.Source: §I.1 · p. 23, §I.2 · p. 23

I.3 Answer-Generator Prompt (modular RAG)

• The modular answer-generator prompt uses retrieved documents for reasoning, while Closed-Book removes them and Naive RAG replaces the curated set with top-10 BM25+dense hybrid chunks.All variants require the final answer on the last line, with the modular prompt using the retrieved documents as its evidence input.Source: §I.3 · p. 24

I.4 LLM-as-Judge Prompt

• The LLM judge extracts the final answer, explains meaningful differences, and returns a strict correctness decision with confidence against the gold answer.GPT-5.4 at temperature 0 supplies the strict CORRECT rate reported as answer accuracy.Source: §I.4 · p. 24

J SFT Data Generation Recipe

• GPT-5.4 generates SFT trajectories natively in the harness using the policy’s system prompt, observation layout, tool schemas, and structured tool calls.Each turn includes working memory, the latest K=3 turn transcripts, older-turn summaries, and all 7+1 tool schemas.Source: J SFT Data Generation Recipe · pp. 24–25
• Turn-level guidance enforces search→curate rhythms while triggering verification, backtracking, tool-diversity, urgency, and curate-aggressiveness prompts under specified conditions.Verification nudges apply after at least 6 turns and 3 curated documents; backtracking follows three searches yielding at most one new document.Source: J SFT Data Generation Recipe · p. 25
• ~26K training examples result from 899 filtered trajectories generated after a 0.10 recall gate from approximately 1K raw trajectories.Per-dataset raw quotas cover BC+, SEC, Patents, Web, Web-simple, and SEC-simple.Source: J SFT Data Generation Recipe · p. 25

K Importance-Aware Curation: Algorithmic Detail

• At capacity M=30, curation evicts the worst-ranked document only when an incoming document has higher importance; otherwise the addition is rejected with a [CAPACITY] marker.Importance ranks are very high=0, high=1, fair=2, and low=3; auto-populated documents start at fair.Source: K Importance-Aware Curation:… · p. 25
• Existing curated documents can be re-tagged in place without eviction, enabling promotion after successful verification, while rendering groups documents from very high to low importance.The visible importance tags keep the ordering explicit in working-memory output.Source: K Importance-Aware Curation:… · p. 26

L Content Deduplication: Algorithmic Detail

• MinHash-LSH deduplicates chunks using 64 permutations, 5-gram shingles, and a 0.85 Jaccard threshold, dropping near-duplicates while retaining them in trajectory recall accounting.The fallback computes SHA-1 over the first 4,000 normalized characters and catches exact duplicates only when datasketch is unavailable.Source: L Content Deduplication: Algorithmic… · p. 26
• The harness exposes suppressed-duplicate counts through a [Dedup] working-memory line, allowing the policy to detect heavy corpus duplication.This diagnostic reports the number of near-duplicate chunks auto-suppressed.Source: L Content Deduplication: Algorithmic… · p. 26

M Component Ablation: Error Analysis

• On 100 BrowseComp+ test queries, each analysis compares queries solved by full Harness-1 but failed by an ablated condition, without retraining.The paired criterion is FA Recall ≥0.5 for full Harness-1 versus FA Recall =0 for the ablation.Source: M Component Ablation: Error Analysis · p. 26
• Removing a single state mechanism changes policy behavior rather than simply eliminating information.The trajectory-level comparisons examine action mixes and representative paired failures.Source: M Component Ablation: Error Analysis · p. 26
• Search corpus actions increase by 3–7 points on failing paired queries after a state mechanism is removed.This shift is accompanied by reduced document reading and verification.Source: M Component Ablation: Error Analysis · p. 26

M.1 Per-mechanism failure modes

• Read-document and verify actions drop by 2–6× in the ablated trajectories.The resulting pattern is characterized as a retrieval failure rather than a curation failure.Source: M Component Ablation: Error Analysis · p. 26
• The trained policy reverts to a wide, shallow, search-dominated mode when a state mechanism is removed.The per-mechanism analyses illustrate this pattern using representative failing queries.Source: M Component Ablation: Error Analysis · p. 26
• In the failing trajectories, the policy never reaches the relevant documents in the first place.This explains why the observed degradation is attributed to retrieval rather than subsequent curation.Source: M Component Ablation: Error Analysis · p. 26

M.1.1 Importance tags

• ∆Recall −4.1% and ∆FA −7.9% when importance tags are removed, with read-document actions falling 7× and verification reaching 0% on failing queries.Binary FIFO curation loses confidence signals, preventing the policy from selecting candidates for follow-up.Source: §M.1.1 · pp. 26–27

M.1.2 Sentence-BM25 compression

• ∆Recall +0.2% but ∆FA −7.0% without sentence-BM25 compression, as raw noisy returns remove bridge sentences and reduce failing-query reads from 4.1% to 2.6%.Aggregate Recall is nearly unchanged, while FA Recall suffers because follow-up reads lose compressed queryable evidence.Source: §M.1.2 · p. 27

M.1.3 Auto-seed

• ∆Recall −0.3% and ∆FA −6.4% with an empty initial curated set, because candidates lack comparison context during the cold-start window.Aggregate Recall barely changes, but early wrong-document retention sharply reduces FA Recall.Source: §M.1.3 · p. 27

M.1.4 Evidence graph

• ∆Recall −2.6% and ∆FA −5.4% when the evidence graph is hidden, increasing search actions to 91.8% and reducing document reads 3× on failing queries.Without entity-bridge structure, the curated set lacks the chunks needed to close multi-document chains.Source: §M.1.4 · p. 27

M.1.5 verify tool

• When verify returns unavailable, ∆Recall falls 3.1% and ∆FA 3.9% relative (12 hard fails), while failing-query document reads drop from 5.1% to 0.8% and search rises to 94.6%.Source: §M.1.5 · p. 27

M.1.6 review docs tool

• ∆Recall +2.4% but ∆FA −3.9% without review docs, as turn-level reads replace bulk curated-set scans and failing-query reads fall 4×.The policy cannot economically combine partial information across many documents and searches rise to 93.6%.Source: §M.1.6 · p. 28

M.1.7 Content-fingerprint dedup

• ∆Recall +4.6% and ∆FA +1.6% when content-fingerprint deduplication is disabled, recovering near-duplicate gold IDs while inflating curated context with redundant chunks.Deduplication is a token-budget mechanism rather than a recall mechanism, and answer accuracy prefers the deduped variant in most paired queries.Source: §M.1.7 · p. 28

M.2 All harness mechanisms disabled

• ∆Recall −12.2% and ∆FA −6.4% with all harness mechanisms disabled; Recall falls to 0.513 and FA recall to 0.624 as searching persists without candidate discrimination.The same trained checkpoint searches comparably hard but cannot rank what it has seen without compact harness state.Source: §M.2 · p. 28

N Evaluation Recipe

• Final-Answer Recall measures answer-bearing annotations, whereas Recall measures the broader relevant set, so either metric can exceed the other.When answer-bearing documents are a subset of relevant documents, the metrics use different denominators; otherwise Aq equals Rq.Source: N Evaluation Recipe · pp. 28–29
• Trajectory Recall measures discovery from the full episode pool and can exceed final Recall because encountered evidence may later be omitted from the curated set.The final curated set is Cq, while the trajectory pool Pq contains every encountered document after chunk normalization.Source: N Evaluation Recipe · pp. 28–29
• Scores are macro-averaged over queries, with fact-level datasets counting a fact as covered when any annotated chunk is retrieved.The same recall definitions apply after substituting facts for documents.Source: N Evaluation Recipe · p. 28
• Released Search-R1 and Tongyi DeepResearch harnesses are standardized with a shared reranker that retains the full pool up to 30 documents and otherwise keeps the top 30.Recall and Final-Answer Recall use the standardized curated set, while Trajectory Recall uses the unrereanked trajectory pool.Source: N Evaluation Recipe · p. 29
• Search-R1’s Recall and Trajectory Recall are nearly identical across benchmarks because its short four-to-five-round trajectories rarely exceed the final document budget.Its training optimizes short search-and-answer trajectories rather than long-horizon evidence collection and curation.Source: N Evaluation Recipe · p. 29

O Modular RAG: Curated Sets Yield Higher Answer Accuracy

• Harness-1’s curated sets dominate open subagents across all four modular-RAG generators, and better curated sets broadly yield higher answer accuracy.The generator sees only the query and curated documents, so subagent-row accuracy differences isolate curated-set quality.Source: O Modular RAG: Curated Sets Yield… · p. 29, N Evaluation Recipe · p. 29
• Closed-Book and Naive RAG are weak on BC+, while generator variation is modest for strong searchers and larger for weaker ones.These patterns indicate that difficult tasks require agentic search and that missing evidence increases the generator’s need to compensate.Source: O Modular RAG: Curated Sets Yield… · p. 29

P Harness as Confound: Same LLM, Different Harness

• +4.2 points in recall accrue to GPT-5.4 when switching from Context-1’s harness to Harness-1’s, with Curated Recall rising from 0.511 to 0.849 and Final-Answer Recall from 0.612 to 0.876.All four retrieval metrics improve monotonically from naive search-add to Context-1 to Harness-1, showing that the harness materially changes discovery without additional RL training.Source: P Harness as Confound: Same LLM,… · p. 30, Figure 7 · p. 30

Q Qualitative Case Studies

• The harness changes search by auto-seeding editable state, promoting evidence, recovering needles with exact search, removing false leads through backtracking, and resolving answers through full-document reading.Routine boilerplate, redundant calls, false positives, and repeated metadata are omitted from the abridged trajectories.Source: Q Qualitative Case Studies · p. 31

Q.1 Case Study A: Stateful Search over a Drug-Approval Query

• 18 turns achieve 1.0 curated recall and 1.0 final-answer recall for a drug-approval query whose gold answer is Alicia Moruf.The trajectory links label constraints, a pharmacology study, an NDA package, and an approval letter naming the Regulatory Project Manager.Source: §Q.1 · p. 31

Phase 1 - Auto-seeded working memory from a broad medical search

• The broad medical search auto-seeds a six-document editable set and surfaces an FDA Livtencity submission fitting the oral-administration, nausea, and age/weight clues.The initial evidence identifies maribavir and establishes the regulatory anchor for subsequent searches.Source: Phase 1 - Auto-seeded working memory… · p. 32

Phase 2 - The policy pivots to pharmacokinetic evidence

• The policy pivots to a 2022 maribavir study and curated pharmacokinetic evidence showing hepatic metabolism and a 4.32-hour half-life.The search connects the study constraints to the approval-letter evidence while preserving both regulatory and pharmacology candidates in curated state.Source: Phase 2 - The policy pivots to… · p. 33

Phase 3 - Exact search misses, so state prevents a blind loop

• When the exact “Maribavir.*Regulatory Project Manager” search returns no results, the policy recovers by searching for NDA 215596 and finds Alicia Moruf in the approval letter.The state prevents a blind loop after the brittle exact match fails.Source: Phase 3 - Exact search misses, so… · pp. 34–35

Phase 4 - Full-document reading confirms the final evidence

• Full-document reading confirms the decisive approval-letter sentence naming Alicia Moruf as the Regulatory Project Manager and keeps that evidence highly curated.Reading the complete document exposes the answer text after recovery by NDA number.Source: Phase 4 - Full-document reading… · p. 36

Q.2 Case Study C: Exact-Date Recovery in SEC Filings

• 9 turns achieve 1.0 curated recall and 1.0 final-answer recall while recovering the exact SEC-transition date July 3, 2025.Auto-seeding finds the relevant filing family, curation preserves 8-Ks, document reading resolves succession details, and later grep corroborates the date.Source: §Q.2 · p. 38

Phase 1 - Auto-seeding finds the right company and date

• On January 27, 2025, Shaun P. Mara notified Simply Good Foods of his retirement effective July 3, 2025, while remaining CFO until retirement.The filing also announced Christopher J. Bealer as Senior Vice President of Finance effective April 1, 2025, expected to succeed Mara as CFO.Source: Phase 1 - Auto-seeding finds the right… · p. 39

Phase 2 - Full-document reading keeps the answer grounded

• Full-document reading confirmed that Simply Good Foods appointed Christopher J. Bealer CFO effective July 3, 2025, after he joined as Senior Vice President, Finance.The document states that Bealer would continue reporting to President and CEO Geoff E. Tanner.Source: Phase 2 - Full-document reading keeps… · p. 40

Phase 3 - Exact grep adds corroborating filings

• Exact grep corroborated the transition: Simply Good Foods reported that Bealer became an NEO in fiscal 2025 and Mara retired and resigned effective July 3, 2025.The corroborating filing also states that Bealer was hired on March 31, 2025.Source: Phase 3 - Exact grep adds… · p. 41

Q.3 Case Study B: Fan-Out Search Resolves an Ambiguous Transfer Query

• For the ambiguous American Pie commencement-address query, the trajectory kept competing actor hypotheses, pivoted from Seann William Scott to Eugene Levy, and used exact search and full-document reading to expose Dalhousie address evidence.Source: §Q.3 · p. 42

Phase 1 - Fan-out keeps competing hypotheses alive

• The fan-out trajectory kept competing actor hypotheses alive, then promoted Eugene Levy after evidence showed Noah Levenstein appeared in eight of nine released films.The policy tested Seann William Scott and Eugene Levy, expanded the candidate set, and used tagged curation to protect stronger evidence from noise.Source: §Q.3 · p. 42, Phase 1 - Fan-out keeps competing… · pp. 44–45
• The search process maintained competing hypotheses by collecting a broad candidate pool before later exact searches and curation narrowed the evidence.The visible candidate list includes multiple actors and related pages, illustrating the fan-out stage’s retained alternatives.Source: Phase 1 - Fan-out keeps competing… · p. 46

Phase 3 - Exact search tests the candidate bridge

• Exact searches tested the candidate bridge with commencement-related queries for Seann William Scott and Eugene Levy, while the pool reached 74 documents and the curated set 30/30.The exact-search stage used grep patterns for commencement and address terms before returning candidate documents and updating the curated set.Source: Phase 3 - Exact search tests the… · pp. 46–47

Phase 4 - Full-document reading exposes the answer sentence

• Full-document reading exposed that Eugene Levy delivered a commencement address at Dalhousie University on May 22, 2012.The answer sentence also states that Levy received an honorary Doctor of Laws degree at the event.Source: Phase 4 - Full-document reading… · p. 47

Final curation update

• The final curation update promoted Eugene Levy to very high importance, retained Dalhousie University as high importance, and evicted one low-importance document.The final set remained full at 30/30, with one very-high, five high, fourteen fair, and ten low-importance items.Source: Final curation update · p. 48

Q.4 Case Study D: Held-Out Multi-Hop QA via Search-State Editing

• 1.0 curated recall and 1.0 final-answer recall were achieved in 11 turns with five tool types for the Garrett Morris query.The case study asks which comedian was an original Saturday Night Live cast member and played Stan Winters in Martin; the gold answer is Garrett Morris.Source: §Q.4 · p. 49
• The trajectory applied search, curation, full-document reading, exact grep, and final evidence promotion to a held-out HotpotQA query outside the SFT and RL families.This case is presented as evidence that the same search-state operations can transfer beyond the training domains and task families.Source: §Q.4 · p. 49

Phase 1 - Joint search identifies the likely entity

• Joint search linked Stan Winters to Garrett Morris, whose page identified his role in Martin as Stan Winters from 1992 to 1995.The initial search query combined the Stan Winters, Martin, and original-cast clues, and the result promoted Garrett Morris into the curated set.Source: Phase 1 - Joint search identifies the… · p. 49

Phase 2 - Second-hop search checks original SNL cast membership

• A second-hop search for Garrett Morris and the original Saturday Night Live cast returned supporting cast evidence, including his listing among the original cast.The retrieved sources included the cast-members list, the 15th Anniversary Special, and season-one materials; the curated set grew to 12/30.Source: Phase 2 - Second-hop search checks… · pp. 50–51

Phase 3 - Full-document reading unifies both clues

• Garrett Morris’s biography identifies him as Stan Winters on Martin from 1992 to 1995, until he suffered an injury.The full-document summary also records his work on Saturday Night Live, The Jeffersons, The Jamie Foxx Show, and 2 Broke Girls.Source: Phase 3 - Full-document reading… · p. 51
• A corpus grep for “Stan Winters” retrieves both Garrett Morris’s biography and the Martin article, unifying the two clues.The retrieved Martin passage explicitly lists Garrett Morris as Stan Winters across seasons 1–2 and as a guest in season 3.Source: Phase 3 - Full-document reading… · p. 52
• The resulting pool contains 23 documents, with 14 of 30 curated slots filled and three documents marked high importance.The curated set includes Garrett Morris, Saturday Night Live cast, and Saturday Night Live anniversary sources as high-importance entries.Source: Phase 3 - Full-document reading… · p. 52

Q.5 Case Study E: Backtracking after Rejecting the Right Entity

• 40 turns, tool diversity 5, curated recall 1.0, and final-answer recall 1.0 yield the correct answer: October 17, 2006.The query concerns the PlayStation 2 release date of the game in which Drew Gehling voices Gord.Source: §Q.5 · p. 53

Phase 2 - A delayed read overturns the wrong branch

• Bully remained visible at low importance while generic Gord, Duck Tales, and unrelated searches failed to resolve the query.The mistaken hypothesis persisted across several turns rather than being corrected immediately.Source: §Q.5 · p. 53, Phase 2 - A delayed read overturns the… · p. 54
• The harness corrected state by removing stale foreign-language Bully variants, retaining the English Bully page, and promoting Drew Gehling’s page to high importance.The curated set retained Bully at low importance alongside the corrected entity and Gord-related evidence.Source: §Q.5 · p. 53, Phase 2 - A delayed read overturns the… · p. 55

Q.5 Case Study E: Backtracking after Rejecting the Right Entity

• After backtracking, search targeted the corrected game and promoted release-date evidence from Bully and Canis Canem Edit before verification.The agent then attempted to verify the release-date claim over the curated candidate documents.Source: §Q.5 · p. 53
• Date-only grep was rejected because matching October 17 pages were unrelated, preventing the date string alone from being treated as evidence.This demonstrates verification restraint after the search had returned the corrected entity.Source: §Q.5 · p. 53
• The first search surfaced Bully, but the policy wrongly rejected it and pursued a separate Gord interpretation.Its initial decomposition searched Drew Gehling, Gord, and PS2 together, yet the policy judged Bully likely wrong because of a mistaken voice-actor inference.Source: §Q.5 · pp. 53–54

Phase 2 - A delayed read overturns the wrong branch

• A delayed read of Drew Gehling’s page overturned the rejection by confirming that he voiced Gord in the 2006 game Bully.This read provided the stronger evidence needed to identify the discarded game as correct.Source: §Q.5 · p. 53, Phase 2 - A delayed read overturns the… · p. 55

Phase 3 - The corrected entity drives the answer search

• The identity correction redirects search from generic Gord pages to Bully and the PlayStation 2 release-date facet.The corrected entity is used to guide the answer search rather than continuing with the initial hypothesis.Source: Phase 3 - The corrected entity drives… · p. 55
• 17 October 2006 is supported by Bully title-and-platform records identifying the PlayStation 2 release.The evidence appears in multiple retrieved records, including Simple Wikipedia and Rockstar Games’ published-games list.Source: Phase 3 - The corrected entity drives… · p. 55
• The harness promotes corrected answer evidence by assigning high importance to Bully and Canis Canem Edit sources.The curated set retains five sources, with the corrected and supporting entities marked high importance while stale Gord and Bully variants remain low.Source: Phase 3 - The corrected entity drives… · p. 56
• A date-only grep is insufficient because identical dates produce false positives from unrelated current-events pages, films, and titles.The agent therefore keeps title-and-platform evidence as the answer path instead of treating the date string alone as proof.Source: Phase 3 - The corrected entity drives… · p. 56
• The delayed correction makes backtracking visible in external state: the policy preserves the early hypothesis at low importance, then edits, replaces, and redirects it after contradiction.This state edit lets the harness retain the corrected entity, discard stale variants, and redirect search toward the release-date facet.Source: Phase 3 - The corrected entity drives… · p. 56

NeurIPS Paper Checklist

• The checklist states that the abstract and introduction present the stateful retrieval harness and SFT/RL recipe as the paper’s contributions.It also says the results cover eight benchmarks and that the limitations paragraph qualifies the claims’ scope.Source: NeurIPS Paper Checklist · p. 57
• The checklist requires claims to match theoretical and experimental results while accurately reflecting assumptions, limitations, and expected generalization.It distinguishes clearly stated contributions from aspirational goals that the paper does not actually attain.Source: NeurIPS Paper Checklist · p. 57

• The limitations include task scope, benchmark style and size, annotation coverage, regex-based evidence graphs, LLM-based verification, and sentence-BM25 compression.These constraints bound the evaluation and introduce dependencies on the harness’s evidence-processing components.Source: §2 · p. 57
• The paper is empirical rather than theoretical, and it declares LLM usage because language models are central to training, retrieval, baselines, answer generation, and evaluation.The paper does not present theorem-level claims, proofs, or human-subject research.Source: §2 · pp. 58, 63
• The method, harness algorithms, rewards, hyperparameters, prompts, benchmark statistics, and evaluation recipe are specified, but public access to all planned assets is not yet available.The authors plan to release weights, harness code, the data-generation pipeline, and the reinforcement-learning recipe.Source: §2 · pp. 58–59
• The empirical results report aggregate point estimates without formal confidence intervals or significance tests for all main comparisons.Repeated-run evaluation is computationally and financially expensive, especially for smaller benchmarks where uncertainty may be greater.Source: §2 · p. 60
• The work uses benchmark and public-domain or cited corpus sources, avoids human-subject experiments, and discusses risks from unsupported or privacy-sensitive retrieval.The planned retrieval-subagent release will document intended use, dataset provenance, and evaluation protocols.Source: §2 · p. 61