An Interactive Reading of

Benchmark Radar:
A Living Database and Search Engine
for AI Benchmarks and Evaluation

The paper, in plain English

There are now more AI benchmarks than any one person can hold in their head — and nobody knows where they all are. A single evaluation might have a paper on arXiv, a dataset on Hugging Face, a harness on GitHub, a row in one leaderboard, and a mention buried in a model card. Finding the right benchmark for a task, or checking whether the one you're designing already exists, means rummaging through half a dozen disconnected places. This paper's answer is Benchmark Radar: a search engine for benchmarks themselves, refreshed daily from 37 public sources.

The system runs on two streams that deliberately never merge into one number. A catalog keeps 1,283 benchmark records from four sources (LLM Stats, OpenCompass Hub, Artificial Analysis, and model reports), preserving every record's source identity, citations, and score history — even when a record has no score at all. A discovery feed collects dated snapshots of new papers, repos, datasets, and releases (11,068 observations across 46 daily snapshots). The two are joined by shared benchmark IDs and a transparent BM25F search ranking that shows you exactly which words matched, in which fields, and how the score was composed — no semantic black box.

The most quietly alarming result is the census itself. Of 790 records with numeric scores, only 82 declare a percentage unit with a known direction and values in range — the minimum bar for even defining "headroom to 100." And 668 of 1,283 records have no recorded release date, while not a single one of the 12,916 score observations carries an actual evaluation date. The paper's lesson: before you compare two benchmark scores, you have to be able to find the settings behind them. Now, for the first time, you can.

I
Search That Shows Its Work
Retrieval runs on BM25F — field-weighted word matching with bounded name and phrase boosts. Every result exposes its matched words, fields, and score components. You can audit why a benchmark ranked where it did.
II
Records Before Rankings
The catalog refuses to collapse into a single leaderboard: 493 unscored records stay (464 still link to paper, repo, or dataset), and source identities are preserved so measurements are never silently averaged away.
III
The Comparability Census
Only 82 of 790 scored records meet the percentage-scale rule for computing headroom; 668 records have no release date; zero scores carry an evaluation date. The frontier view separates verified scales from the rest.
Chapter 1

A Thousand Benchmarks, Nowhere to Look

GPT-6 Astra and Claude Fable 5.1 are built for coding, research, and multi-tool work. Comparing them requires benchmarks — but first you have to find the benchmarks.

The paper opens with a problem every evaluation researcher knows: finding a benchmark, its dataset or repository, and reports of its use means searching paper servers, code hosts, dataset hubs, vendor releases, blogs, and catalogs. Existing resources — LLM Stats, OpenCompass, Artificial Analysis — each hold a piece of the picture. Connecting a newly released benchmark to its task materials, and then to its later use in model reports, "can still require consulting several separate resources."

Benchmark Radar's answer is to ingest all four and refuse to flatten them. The v0.11.0 catalog holds 1,283 source records — one record per benchmark per source — carrying 12,916 numeric score observations on 790 records. Crucially, the 493 records without numeric scores are not deleted: 464 of them still link to at least one paper, repository, or dataset. In the full catalog, 475 records link to papers, 506 to repositories, and 293 to datasets.

The preservation rule (Section 3.3)
$$\text{keep}(r) = \texttt{true} \quad \text{for all } r \in \text{Catalog}, \qquad \text{Catalog} = \bigsqcup_{s \in \{\text{LLM Stats, OC Hub, AA, MR}\}} R_s$$
Rssource records from source s (one benchmark entry from one contributing source)
|r|687 + 461 + 25 + 110 = 1,283 — a record remains if a score, date, or citation is absent
Interactive — records vs. scores, source by source
Live updates as you drag. Toggle the log scale to see all four sources at once — Artificial Analysis's 7,050 scores would otherwise flatten everything else.
Catalog records
1,283
4 sources · v0.11.0
Numeric scores
12,916
on 790 records
Unscored but linked
464/493
paper · repo · dataset
Most lopsided source
7,050 / 25
scores / records (AA)
Data: Tables 3 and 8 of the paper. Records count source-specific benchmark entries; scores count numeric observations, not distinct models. Medians/max model counts per source: LLM Stats 3/239 · Artificial Analysis 245/586 · Model reports 2/19 · OpenCompass Hub unknown (no scores).
Why this matters

The four sources contribute in opposite directions. OpenCompass Hub brings 461 records and zero scores; Artificial Analysis brings 25 records and 7,050 scores — a median of 245 scored models per benchmark, peaking at 586 for GPQA Diamond. Any "average benchmark" statistic that mixes these populations without care is meaningless.

The paper's discipline is to keep the units separate: records ≠ benchmarks ≠ scores ≠ models. Twenty-two records carry scores for at least 100 distinct models each.

Next: how the system is wired
Chapter 2

Two Streams, One Index

A daily radar and a stable catalog pull in different kinds of evidence. The whole design lives or dies on keeping them separate — then joining them by ID.

Figure 2 of the paper separates two input paths. The benchmark catalog is populated by registries and model reports: 1,283 source records with normalized names, identifiers, artifact links, score observations, model identities, and cited documents. The discovery history is a stack of dated snapshots — mentions, releases, and updates found across public sources, linked to artifacts by exact identifiers such as DOIs, arXiv IDs, and repository URLs.

The join happens in a shared evidence layer: stable source-record IDs, scores, models, citations, and links back to discovery evidence. Clients — the web dashboard and the offline CLI — consume the same benchmark IDs and the same JSON response format. A shared query service supplies both interfaces, so "what you see on the web" and "what you query locally" cannot drift apart.

Interactive — the architecture, node by node
Click a node to inspect it. The dashed lines animate in the direction evidence flows — discovery snapshots stay separate from the catalog total.
Select a node
Click any box in the diagram to see the records, counts, and interfaces behind it.
Discovery sources
37
13 connectors + 24 feeds
Discovery observations
11,068
46 snapshots · 6,546 artifacts
Catalog records
1,283
discovery adds 0 to this
Shared model registry
868
model identities
Identity linking by exact identifiers (Section 3.3)
$$\text{artifact}(o) = \{\,\text{DOI},\ \text{arXiv ID},\ \text{repo URL}\,\} \cap \text{identifiers}(a) \neq \emptyset \;\Rightarrow\; o \sim a$$
oa discovery observation (one collected record: a mention, release, or update)
aan artifact: a paper, repository, dataset, release, or page
o ∼ alinked by exact identifier — never by fuzzy match; 59 of 6,546 artifacts have multi-source observations
Why this matters

The same six reader questions route through different evidence scopes: Today (what appeared or changed), Search (which benchmarks match this task — the complete catalog, unscored included), Leaderboard (which evaluations combine lower scores with broader use), Saturation (score histories and settings), Blog (a daily brief), and CLI (the same evidence, offline).

Display filters change only what a view shows; search and exports always retain the full catalog. The census in the paper applies none of the display filters.

Next: the daily harvest
Chapter 3

The Daily Harvest

Every day, 37 sources get swept in a 48-hour window. Most of what comes back is thrown away — and the run only publishes if three core sources are healthy.

Daily discovery searches a rolling 48-hour window across 13 direct connectors and 24 first-party research and engineering feeds, recording counts and errors by source, removing future-dated rows, and requiring healthy core sources before publication. At the cutoff the core three were arXiv, the Hugging Face Hub, and GitHub Search — all must be healthy before a snapshot can publish. Optional routes may fail without blocking: at the cutoff, Semantic Scholar returned a malformed payload and Brave Search lacked an API key, both recorded in the snapshot.

The 2026-09-07 snapshot shows how brutal the filter is: 1,003 fetched rows became 954 candidates after duplicate removal; only 366 qualified for publication, and 138 met the recommendation threshold. Cumulatively the feed holds 11,068 observations and 6,546 artifacts across 46 snapshots (4 of them simulated historical backfills). Five source labels account for 88.0% of all observations: Hugging Face 3,487 · GitHub 2,470 · arXiv 1,945 · Semantic Scholar 1,039 · OpenAlex 802.

Daily recommendation priority (Appendix C)
$$P = 0.35\,R + 0.20\,E + 0.20\,C + 0.25\,A \in [0, 100]$$
Rrelevance — match to benchmark-tracking intent
Eevidence — does the observation link artifacts (paper, repo, dataset)?
Crecency — freshness within the 48-hour window
Aadoption — observed traction of the underlying artifact
P ≥ τmeets the recommendation threshold τ; at the 2026-09-07 cutoff, 138 of 366 qualified items cleared it
Interactive — the funnel, the gate, and the priority score
Everything updates live. The three core-source switches are the publication gate; the weight sliders renormalize to sum to 100%.
CORE SOURCES:
Fetched rows
1,003
2026-09-07 snapshot
After dedup
954
−49 duplicates
Qualified
366
36.4% of candidates
Recommended
138
37.7% of qualified
The three scored items are illustrative discovery observations (paper prints the formula and cutoff counts, not per-item components). Weight sliders renormalize to 100%; the reset button restores the paper's fixed weights.
Why this matters

Source caps and collection failures shape the mix — GitHub Search was at cap (300 rows) at the cutoff, and five labels carry 88% of everything. The system's honesty about this is the feature: failed optional sources are reported in the visible coverage notice rather than silently dropped.

The discovery feed's job is evidence, not totals — those 11,068 observations add exactly nothing to the 1,283-record benchmark count.

Next: the ranking formula you can audit
Chapter 4

Ranking Without a Black Box

Most search engines ask you to trust the model. Benchmark Radar uses BM25F — a 2004 formula where every point of every score can be traced to a word, a field, and a weight you can inspect.

Catalog retrieval is lexical: BM25F, the field-weighted extension of BM25 from Robertson, Zaragoza & Taylor (2004), "with bounded boosts for name and phrase matches." Each result exposes matched and missing query words, the fields they occur in, and the score components. Source membership does not change the ranking — a record from a small source competes on equal terms with one from a big one.

The design stance is that candidate retrieval precedes suitability judgment. An analyst or agent tries focused query variants, inspects a record's tasks and score settings, and follows its citations. The engine's job is to surface candidates with their evidence attached — not to decide relevance for you.

BM25F scoring (Section 3.4, after Robertson et al. 2004)
$$\mathrm{score}(D, Q) = \sum_{t \in Q} \mathrm{idf}(t)\;\cdot\;\frac{\mathrm{tf}_w(t, D)\,(k_1 + 1)}{\mathrm{tf}_w(t, D) + k_1\!\left(1 - b + b\,\dfrac{|D|}{\mathrm{avgdl}}\right)} \;+\; \beta_{\text{name}}\,\mathbb{1}[\text{phrase} \subset \text{name}] \;+\; \beta_{\text{desc}}\,\mathbb{1}[\text{phrase} \subset \text{desc}]$$ $$\mathrm{tf}_w(t, D) = \sum_{f \in \{\text{name, aliases, desc}\}} w_f \cdot \mathrm{tf}(t, f), \qquad \mathrm{idf}(t) = \ln\!\left(1 + \frac{N - \mathrm{df}(t) + 0.5}{\mathrm{df}(t) + 0.5}\right)$$
D, Qa benchmark record (its concatenated weighted fields) and a query
wfper-field weights — the "F" in BM25F; name matches should dominate
k1, bterm-frequency saturation and length-normalization constants
βname, βdescbounded boosts — capped additive bonuses for full-phrase matches (the paper's "bounded boosts for name and phrase matches")
N, df(t)corpus size and document frequency of term t
Interactive — a live BM25F engine over the paper's benchmark corpus
Type and the results re-rank instantly. Click any result to expand its per-term score components, matched fields, and boosts. Sliders re-weight the formula itself.
Corpus
31
benchmarks named in the paper
Results with ≥1 match
score 0 records are listed last
Top score
BM25F units
Phrase boosted
β added (bounded)
Corpus: the 31 benchmarks the paper names in its introduction, taxonomy, and worked examples, with short descriptions drawn from the paper's own characterizations. Real BM25F implementation — idf, field weights, saturation, and capped boosts all computed live.
Why this matters

Lexical matching has a known cost, and the paper says so plainly: it can miss paraphrases and renamed tasks, and evaluating semantic retrieval "will require reviewed relevance judgments." In the worked example (Chapter 9), the agent supplemented local queries with web searches for exactly this reason.

But auditable retrieval is the right default for prior-art search, where a silently missed candidate is a false "we're first" claim. When the engine can't find your words, it shows you that too.

Next: the taxonomy puzzle
Chapter 5

One Axis Is Never Enough

Is SWE-bench a coding benchmark or an agent benchmark? Force one label and you lose 117 agentic benchmarks from every count of agentic evaluation.

The census classifies 1,279 of 1,283 records into 11 top-level capability domains and 63 sub-domains, using the publishers' own fields — OpenCompass Hub dimensions, LLM Stats categories, Artificial Analysis categories, and model-report registry domains. The 4 unclassified records are LLM Stats community rows whose crawl supplied no description, category, or modality; the paper records that reason rather than guessing from titles.

The design tension: interaction paradigm and input modality are recorded as facets — properties held beside a record's domain, not inside it. Every record carries exactly one Level-1 class and, independently, any number of facet values, so facet counts don't sum to the population. The paper's example is exact: "a benchmark that resolves repository issues is a coding benchmark run as an agent, not an agent benchmark." Of 345 records with the agentic facet, only 128 take Agentic & Tool Use as their Level-1 class; 117 sit under Coding & Software Engineering, with the rest spread across 6 further classes.

Single axis vs. facets (Section 4.2)
$$\underbrace{\sum_{d \in \mathcal{D}} n_d = 1{,}283}_{\text{Level-1 domains partition the catalog}} \qquad\qquad \underbrace{\sum_{f \in \mathcal{F}} n_f \;\neq\; 1{,}283}_{\text{facets overlap each other and the domains}}$$
nagentic345 records carry the agentic facet (any Level-1 class)
one-axis count128 — records whose Level-1 class is Agentic & Tool Use
hidden agents117 under Coding & Software Engineering + 100 spread across 6 further classes
Interactive — the 11 domains, 63 sub-domains, and the hidden agents
Click a domain bar to drill into its sub-domains. Then flip the counting mode and watch 345 agentic records collapse to 128.
Classified records
1,279/1,283
11 domains · 63 sub-domains
Largest domain
267
Multimodal (20.8%)
Agentic: facet view
345
26.9% of catalog
Agentic: one-axis view
128
117 hidden in Coding
Counts from Figure 5 and Section 4.2 of the paper. The "6 further classes" bucket aggregates the 100 remaining agentic-facet records the paper reports only in aggregate.
Why this matters

Every "how many agent benchmarks shipped this year?" statistic you've read depended on this labeling choice, and most datasets make it silently. With a facet scheme, the claim becomes inspectable: you can list the 117 coding benchmarks that are run as agents and check them yourself.

The paper is equally careful about the limits: task-capability classification "remains unvalidated in this rebuild," and a deterministic null extractor assigns no capability levels to its 5,863 discovery-derived tracks.

Next: the undated half of the catalog
Chapter 6

The Undated Half

668 of 1,283 records have no release date. And of 12,916 score observations, not one carries the date the evaluation actually ran.

Benchmark release dates are known for 615 records. The dated mass clusters heavily in recent years: n=49 records in 2023, n=179 in 2024, n=279 in 2025 — with the remaining 108 dated records spread across all other years. Figure 6 keeps the 668 undated records in a separate column with its own scale rather than leaving the figure, because a year's share "is read only where the evidence supports one."

Even that isn't enough. The OpenCompass Hub crawl stops at the discovery cutoff, so the truncated final year is drawn mostly from model reports — its share would measure a change of catalog, not a change in the field. So the paper additionally requires a year's source mix to stay close to the pooled mix, and checks: reweighting each year to a common source composition moves the agentic share by at most 1.1 percentage points over the reported years. The rise in agentic benchmarks is real, not an artifact of who supplied which year.

Source-mix reweighting check (Section 4.2)
$$\widehat{\text{share}}_{\text{agentic}}(y) \;=\; \sum_{s} \bar{w}_s \cdot \text{share}_{\text{agentic}}(y, s), \qquad \max_{y \in \{2023, 2024, 2025\}} \left| \widehat{\text{share}}(y) - \text{share}(y) \right| \;\leq\; 1.1\,\text{pp}$$
yrelease year with n2023=49, n2024=179, n2025=279 dated records
̄wsthe pooled source composition (fraction of dated records from each of the 4 catalog sources)
≤ 1.1 ppthe paper's empirical bound: the agentic rise survives reweighting to a common source mix
Interactive — what's dated, what's not, and what dates scores carry
Toggle the undated column to see the hidden half. The right chart breaks down the date basis attached to all 12,916 score observations.
Dated records
615
47.9% of catalog
Undated
668
kept, separate column
Scores w/ model-announce date
12,594
proxy only
Scores w/ evaluation date
0
not recorded anywhere
Left chart: 2023–2025 counts and the aggregate "other years" bucket (108 dated records across every year except 2023–25) are exact from the paper. The undated column (668) is drawn on its own scale as in Figure 6A — flip the toggle to put it on the same scale and see the hidden half at full height. The reweight annotation states the paper's measured bound: reweighting each year to a common source composition moves the agentic share by at most 1.1 percentage points.
Why this matters

Any trend claim about benchmarks — "agentic evals are exploding," "multimodal is saturated" — is a claim about dates. This catalog makes the evidence base visible: which half is dated, which mix each year draws from, and which date basis each score carries.

The punchline for saturation research: measuring it "requires comparable test versions and settings, with dates tied to score reporting or evaluation." In this catalog, that last part doesn't exist yet.

Next: the score you can't compare
Chapter 7

The Uncomparable Score

12,916 numbers, and only 82 records' worth can even define "distance from 100." The rest are scores without a scale you can verify.

Section 3.5 sets a deliberately strict eligibility rule. To summarize a record on a percentage scale, the census requires: a declared percentage unit, a known score direction, and numeric values within 0–100. Rescaling a displayed value or reading an aggregator's declared maximum does not establish the unit. And matching scales still don't establish matching test versions, prompts, tools, attempts, or evaluators.

Under that rule, only 82 of 790 scored records qualify. The remaining 708 keep their numeric observations — they're not deleted — but they don't support a shared percentage-headroom calculation, and the 493 unscored records have unknown headroom by definition. Headroom itself is a narrow calculation: $100 -$ a record's highest value, only when the three gates pass.

Percentage-scale eligibility and headroom (Sections 3.5, B.2)
$$\mathrm{eligible}(r) \;=\; \mathbb{1}\!\left[\,\text{unit}_r = \%\,\right] \cdot \mathbb{1}\!\left[\,\text{direction}_r \;\text{known}\,\right] \cdot \mathbb{1}\!\left[\,0 \le v \le 100 \;\; \forall v \in \text{scores}(r)\,\right]$$ $$H(r) \;=\; \begin{cases} 100 - \max \text{scores}(r) & \text{if } \mathrm{eligible}(r) \\[4pt] \text{unknown} & \text{otherwise} \end{cases}$$
unitr = %declared in the source — a displayed multiplier or an aggregator's maximum does not establish it
directionhigher-is-better vs. lower-is-better must be recorded; a numeric maximum need not be the best result
0 ≤ v ≤ 100every numeric observation in range
82 / 708 / 493eligible / other-or-unverified-scale / no numeric score — all 1,283 records remain accounted for
Interactive — the eligibility cascade and a headroom calculator
Pick a record, flip its gates, and watch the verdict change live. The cascade chart shows where all 1,283 records land.
RecordMax
Headroom H(r)
100 − max, if eligible
Full population
1,283
every record accounted for
Scored, % scale OK
82
10.4% of scored
Scored, unverifiable
708
scores retained
No numeric score
493
headroom unknown
The four calculator records are constructed to exercise each rule from Sections 3.5/B.2 (record A mirrors a percentage-scale record like those among the 82; B/C/D mirror the failure modes the paper names: undeclared unit, unknown direction, out-of-range values). Humanity's Last Exam's displayed max of 55.47 is real (Figure 4); its eligibility state is not stated in the paper.
Why this matters

When the next paper claims "benchmark X is saturated — only 4 points of headroom," this census is how you check whether those 4 points mean anything. A small gap to a ceiling can describe a reported setup; catalog-wide headroom estimates require the eligible measurement coverage alongside the statistic.

The deeper point is evidentiary: scores are claims, and claims have scopes. Test versions, reasoning budgets, tools, attempts, and evaluators must still be checked before comparing results — the census tells you which records even start that conversation.

Next: the frontier you can still draw
Chapter 8

The Frontier

Scores alone mislead. The Leaderboard's answer: plot reported score against measured use — how many distinct models have actually been scored — and trace the Pareto frontier.

The Benchmark Frontier view (Figure 3) combines reported scores, scored-model counts, and benchmark release dates or first-score proxies. Gold rings mark Pareto candidates: benchmarks no other record beats on both axes. Hollow marks flag unverified scales or counts — those stay visible but are kept outside the Pareto calculation. Dotted outlines flag model-release date proxies. The view requires a numeric score, excludes known pre-2024 benchmarks, and starts from a score cutoff of 70.

Why "measured use"? Because a benchmark's score in isolation says nothing about whether anyone relies on it. The Artificial Analysis record for GPQA Diamond carries scores for 586 distinct models; Humanity's Last Exam and SciCode for 577 each; CritPt for 492. Median scored-model counts tell the same story of skew: 245 for Artificial Analysis records, 3 for LLM Stats, 2 for model reports. Twenty-two records have ≥100 distinct scored models.

Pareto dominance on the frontier (Figure 3, Appendix C)
$$i \succ j \;\iff\; \big(s_i \ge s_j\big) \wedge \big(m_i \ge m_j\big) \wedge \big(s_i > s_j \;\vee\; m_i > m_j\big), \qquad \mathrm{Pareto} = \{\, r : \nexists\, r'\, \text{with}\, r' \succ r \,\}$$
srreported score of record r (highest numeric value on the source's native scale)
mrdistinct scored models (measured use) — counted by source model ID within the record
filternumeric score required · known pre-2024 excluded · initial cutoff s ≥ 70 · unverified scales excluded from Pareto (shown hollow)
Interactive — drive the frontier and scrub a score history
The cutoff slider and the verified-scale toggle reshape the frontier live. On the right, scrub through 577 scored models on one record — the line tracks successive best scores by model release date, ending at 55.47.
Records in view
after cutoff + filters
Pareto candidates
gold rings
Hollow (unverified)
visible, not ranked
Best score @ 577 models
55.47
Humanity's Last Exam (AA)
Frontier scatter: anchor points (GPQA Diamond 586 models; HLE 577 @ 55.47; SciCode 577; CritPt 492; LLM Stats median 3 / max 239 models) are real from Tables 5/8; surrounding points are calibrated illustrations so the frontier shape is visible — marked as such here, unlike the paper's real data. Right chart mirrors Figure 4: 577 observations, best-line to 55.47, deterministically generated trajectory.
Why this matters

The frontier is the paper's alternative to the naive leaderboard: it answers "which evaluations combine lower scores with broader measured use" — the benchmarks where headroom and evidence actually coexist. It also encodes the honesty rules from Chapter 7 as geometry: unverified scales hover at the edge, visible but unable to define the frontier.

Note the frontier's own date caveat: positions use "benchmark release dates or first-score date proxies" — and we know from Chapter 6 how thin dated coverage is.

Next: the whole thing in action
Chapter 9

Will It Duplicate?

A contributor is about to design a new evaluation of credit assignment in agentic training. Before building it, a coding agent interrogates the catalog: does this already exist?

Section 5 walks through a complete prior-art search. The requirement: small Qwen-series models as reproducible baselines, work published in August 2026. A coding agent installed the Benchmark Radar client and its public Skill, downloaded the corpus, and searched locally — inspecting recorded paper, repository, and dataset links, then supplementing with web searches for work described in different terms (lexical matching can miss paraphrases). Five works surfaced: SRPO, ContextPilot, SkillGate, CIPO, and MoRSE — each a different attack on credit assignment, each using different agentic benchmarks.

The workflow's shape matters more than the specific hits: candidate retrieval is separated from comparison. Benchmark Radar retrieves candidates and exposes their evidence; the researcher or agent judges relevance and compares designs. The paper also shows the counterfactual: for an earlier benchmark-design effort, one contributor hand-assembled an 11-row prior-art comparison table (MLE-Bench, MLGym-Bench, EXP-Bench, ResearchCodeBench, MLR-Bench, PaperBench, AstaBench, InnovatorBench, AIRS-Bench, COMPOSITE-Stem, ScienceBoard) — "more effort than any part of that project except producing the benchmark data."

Interactive — re-run the prior-art query yourself
The query box drives the same BM25F engine from Chapter 4, over the five works the agent found. Toggle query terms to see how each one steers the ranking.
Corpus
5
August 2026 works (Table 4)
All Qwen-based
5/5
baseline requirement met
Top match
by BM25F score
Duplicates found
0
overlaps, not duplicates
Corpus text condensed from Table 4 of the paper: base models, credit-assignment focus, and benchmarks used for each work.
Why this matters

This is the whole system compressed into one use case: discovery found the papers, retrieval surfaced them with visible word matches, evidence links let the agent read the sources, and the judgment — is this the same idea? — stayed with the human. The paper is careful not to overclaim: retrieval precision, task suitability, and time saved "remain to be evaluated"; the example has no controlled baseline.

But the shape of the claim is right: the way to prevent duplicate benchmarks is not better memory — it's a searchable, source-preserving, auditable index.

Back to the top