Evaluate bounded records, metrics and matching problems
Analyze supplied datasets, exact-decimal profiles, evaluation metrics and small explicitly constrained optimization problems.
How to choose
Check admitted dimensions, tie-breaking, weighting and precision. Candidate linking is not identity resolution; reported metrics describe the supplied sample and chosen definition.
No population-quality guarantee, automatic model promotion, inferred causal truth or unbounded optimization.
Check explicitly supplied decimal claims against five declared decimal-place rounding rules. Returns exact expected text, tie evidence and rational errors; no number extraction, unit inference or fact verification.
Choose for: Check rounding consistency in an agent-generated numeric report Audit exact decimal tie handling without binary floating-point artifacts
Outside this profile: Extracting facts or judging truth, semantic correctness, evidence authenticity, causal effects, or production readiness Fetching data, executing models or user code, changing datasets, authorizing actions, or contacting live services Inferring population quality, statistical significance, uncertainty intervals, or generalization from a supplied sample
Aggregate supplied weighted rubric scores with exact normalized fractions and conservative lower/upper bounds for missing ratings. Optional threshold reports certain pass, certain fail or unknown; does not assign ratings or renormalize away missing criteria.
Choose for: Aggregate an explicit agent evaluation rubric without hiding missing scores Determine whether omitted rubric ratings can change a supplied pass threshold
Outside this profile: Extracting facts or judging truth, semantic correctness, evidence authenticity, causal effects, or production readiness Fetching data, executing models or user code, changing datasets, authorizing actions, or contacting live services Inferring population quality, statistical significance, uncertainty intervals, or generalization from a supplied sample
Find exact item ID, group ID and supplied SHA-256 fingerprint overlaps across caller-declared dataset partitions. Reports missing identity coverage and capped locations; no data fetching, hash verification, near-duplicate detection or proof of leakage-free evaluation.
Choose for: Find exact train validation test overlap from supplied IDs and group keys Audit group leakage candidates before an agent evaluation run
Outside this profile: Extracting facts or judging truth, semantic correctness, evidence authenticity, causal effects, or production readiness Fetching data, executing models or user code, changing datasets, authorizing actions, or contacting live services Inferring population quality, statistical significance, uncertainty intervals, or generalization from a supplied sample
Compare supplied baseline/candidate decimal measurements against explicit directional absolute and relative regression budgets. Missing values stay unknown; every configured budget must pass. No benchmarking, unit conversion, noise model or significance test.
Choose for: Gate supplied agent benchmark changes against explicit regression budgets Compare baseline and candidate measurements while preserving missing evidence
Outside this profile: Extracting facts or judging truth, semantic correctness, evidence authenticity, causal effects, or production readiness Fetching data, executing models or user code, changing datasets, authorizing actions, or contacting live services Inferring population quality, statistical significance, uncertainty intervals, or generalization from a supplied sample
Compute exact binary Brier score and uniform-bin calibration diagnostics from supplied probabilities and nullable Boolean outcomes. Reports label coverage and per-row errors; no learned calibration, confidence intervals, forecast validity or population-quality claim.
Choose for: Evaluate supplied agent confidence probabilities against binary outcomes Inspect Brier error and reliability bins without dropping unknown labels silently
Outside this profile: Extracting facts or judging truth, semantic correctness, evidence authenticity, causal effects, or production readiness Fetching data, executing models or user code, changing datasets, authorizing actions, or contacting live services Inferring population quality, statistical significance, uncertainty intervals, or generalization from a supplied sample
Profile caller-selected JSON Pointer fields across up to 500 records: missing/null/type counts, typed distinct values, numeric bounds/median, UTF-16 string lengths and bounded duplicate-row evidence. Observations only; no inferred schema guarantee or population extrapolation.
Choose for: profile missing null fields in JSON records inspect dataset types cardinality duplicates before validation measure field completeness and mixed types
Outside this profile: Infer a guaranteed schema or population distribution Fuzzy entity matching or CSV file comparison
Check typed scalar composite foreign keys between supplied parent and child records. Reports orphan, invalid and ambiguous child links, duplicate parent keys, optional null-child skipping and unreferenced parent rows. No database connection, coercion or guessed matches.
Choose for: validate foreign keys between supplied JSON tables find orphan child records and ambiguous duplicate parent keys check composite referential integrity without a database
Outside this profile: Query or mutate a live database Fuzzy entity resolution or automatic record joins
Evaluate up to 1,000 caller-supplied single-label expected/predicted pairs: confusion matrix, per-label precision/recall/F1/support, macro/weighted/micro summaries, balanced accuracy and mismatch evidence. Undefined per-label values use zero with explicit flags; labels are not verified ground truth.
Choose for: calculate confusion matrix precision recall f1 from evaluation labels score agent routing classification predictions against supplied labels inspect macro micro weighted metrics and error samples
Outside this profile: Train or execute a model or infer factual ground truth Multilabel classification probability calibration or deployment certification
Score ranked document IDs against caller-supplied graded judgments for up to 50 queries: precision/recall, reciprocal rank, truncated average precision and NDCG at k. Explicit gain, unjudged and zero-denominator conventions; evidence lists ranked judgments and missed relevant IDs. No search or relevance inference.
Choose for: evaluate retrieval ranking ndcg map mrr precision recall at k score search results against supplied relevance judgments compare RAG retrieval quality using judged document ids
Outside this profile: Search the web or determine semantic document relevance Claim benchmark completeness from partial judgments
Evaluate supplied labeled text spans with exact boundaries or thresholded intersection-over-union. Bounded deterministic maximum-cardinality one-to-one matching returns precision/recall/F1, matches and unmatched IDs. UTF-16 half-open offsets are validated; labels and text meaning are not judged.
Choose for: evaluate named entity extraction annotation spans match predicted and reference text spans using exact offsets or iou calculate one-to-one span precision recall and unmatched evidence
Outside this profile: Semantic factuality or entailment checking Tokenization byte offsets or globally maximum total IoU matching
Inspect up to 1,000 supplied timestamp/value points for ordering, duplicate timestamps, null values and optional cadence deviations. Returns bounded interval evidence, aligned missing-point estimates and descriptive numeric summaries. No clock, interpolation, anomaly prediction or forecasting.
Choose for: inspect time series cadence gaps duplicate timestamps and null readings validate chronological telemetry points before analysis find irregular intervals and missing observations in supplied data
Outside this profile: Forecast prices predict anomalies or make trading decisions Fetch live telemetry or infer wall-clock freshness
Analyze up to 300 integer half-open intervals inside an explicit window: clipped union, gaps, covered/multiply-covered duration, concurrency segments and overlap-pair evidence. Touching endpoints do not overlap. No calendar access, reservations or scheduling actions.
Choose for: calculate interval coverage union gaps overlaps and concurrency inspect execution windows for overlaps and uncovered time find coverage gaps in supplied numeric ranges
Outside this profile: Book meetings mutate calendars or schedule execution Interpret timezone recurrences or continuous noninteger geometry
Allocate a declared integer total among exact nonnegative decimal weights using floors and largest remainders. Conserves units with exact quota evidence and ID tie breaks; no minima, caps, fairness certification, payments or monotonicity guarantee.
Choose for: Divide a fixed total of indivisible units among weighted categories Produce integer weighted allocations whose sum equals the declared total
Outside this profile: Fairness certification, eligibility decisions about people, payments, or resource reservations Monotonic allocation as total changes, recipient minimums or maximums Live execution, network lookups, payments, reservation, or inferring factual correctness of supplied data
Compute exact reduced-rational descriptive statistics and R type-7 quantiles from supplied decimal strings. Explicit null/error-or-omit policy, empty and singleton results, exact population/sample variance. No square roots, weighting, units, uncertainty or population inference.
Choose for: Compute exact descriptive mean, variance, median and R7 percentiles from supplied decimal measurements Summarize decimal observations with explicit missing-value omission evidence
Outside this profile: Significance testing, confidence intervals, population estimates, verified measurements, standard deviation or automatic missing-value imputation Network calls, external actions, supplied-evidence authenticity or factual correctness certification
Compose union, intersection, A-minus-B difference or XOR of declared axis-aligned half-open integer rectangles into deterministic x-first disjoint rectangles, exact integer area and bounds.
Choose for: Exact set algebra over caller-supplied Cartesian integer regions Generating disjoint drawing or selection data from axis-aligned half-open rectangles Computing exact coordinate-squared area beyond JavaScript safe integers
Outside this profile: Rotated shapes, arbitrary polygons, geographic or geodesic area, CRS or units conversion Image modification, masking, redaction or identifying actual objects Click safety, collision physics, optimal packing or minimum-rectangle decomposition Fetching sources, writing files, model calls, target execution or changing external state
Compose first-occurrence typed row/column axes, dense counts, exact decimal-string sums and per-measure missing/null/valid coverage with aligned margins.
Choose for: Producing typed crossed categories with exact decimal sums from supplied records Preserving missing versus null and scalar-type distinctions Preparing aligned count grids and row/column/grand totals
Outside this profile: Accounting advice, statistical inference or verification of supplied observations Floating-point measurements, automatic numeric string coercion or implicit chart conversion Locale ordering, category normalization or inventing absent categories Fetching sources, changing external state or executing supplied data
Clip each directed source segment against a closed integer rectangle and return exact reduced rational parameters and endpoints, preserving point contacts and per-segment provenance.
Choose for: Exact viewport clipping on a caller-supplied integer Cartesian grid Obtaining ordered segment fragments with exact source parameters Keeping boundary and tangent contacts without rounded coordinates
Outside this profile: Half-open rectangle set operations or polygon clipping Geodesic distances, CRS conversion, antimeridian handling or physical units Merging fragments, topology guarantees or direct floating-point renderer output Fetching sources, changing external state or executing supplied data
Simplify a supplied integer polyline with exact point-to-segment Ramer–Douglas–Peucker decisions, original retained indexes and per-final-segment squared vertex-deviation witnesses.
Choose for: Reducing original-vertex polylines for previews or vector exports Requiring exact inclusive tolerance decisions and earliest-index distance ties Checking original removed vertices against their final enclosing segments
Outside this profile: Minimum-vertex optimization, topology or area preservation Curve smoothing, map matching or global geometry equivalence General Hausdorff certificates, geodesic distances or coordinate conversion Fetching sources, changing external state or executing supplied data
Fold supplied ordered revision-checked put/delete events into a fresh complete keyed snapshot with tombstones and receipts, or return the first tentative-fold conflict without partial state. Valid conflicts are normal paid results; never reads or writes a database.
Choose for: Preparing a complete next snapshot from caller-supplied records and events Keeping exact typed tuple identity and explicit revision arithmetic Inspecting an offline ordered event fold before separately authorized persistence
Outside this profile: Live CDC ingestion, database CAS or transactions, cursor advancement or durable writes Historical event deduplication, exactly-once execution or authenticated event provenance Partial record patches, last-writer-wins, tombstone garbage collection or CRDT convergence Fetching sources, changing external state or executing supplied data
Select the lexicographically highest jointly compatible stable versions in caller-declared package priority from a complete supplied finite catalogue, with first-failure witnesses for every earlier combination. Valid infeasibility is a normal paid result; never contacts a registry or installs packages.
Choose for: Choosing one supplied version of every package for a bounded offline release fixture Resolving conditional version requirements, including self-dependencies and cycles Regenerating and checking every rejected higher-priority combination from explicit evidence
Outside this profile: Registry lookup, download, installation, native npm lockfile generation or unseen releases Prerelease/build metadata or npm/pnpm/yarn range, peer, optional, workspace, platform or engine semantics Authenticity, vulnerability or license checks; minimum-change objectives; installation ordering Fetching sources, changing external state or executing supplied data
Produce every admissible cross-table candidate using explicitly normalized Unicode-scalar Levenshtein distances and exact integer weighted costs. Includes all tied best candidates; never selects an identity, asserts duplicates or merges records.
Choose for: Return a complete bounded candidate queue with exact field-distance evidence
Outside this profile: Identity resolution, true-duplicate assertions, automatic merging, confidence probabilities or semantic similarity Medical, employment, credit or other high-impact decisions about people Large-table nearest neighbors, top-k retrieval or transposition distance Network requests, file changes, external execution, publishing or deployment
Generate a minimum unit-edit alignment of two supplied short texts with deterministic ties, scalar/UTF-16/UTF-8 spans, exact digests and a fully guarded text-edits-apply request. Preserves every supplied scalar and newline without normalization.
Choose for: Produce a complete reproducible minimum unit-edit script and consumer-ready guarded edit request
Outside this profile: Line, token, grapheme or semantic alignment; fuzzy anchoring or transposition distance Long documents or more than 200 separated changed blocks Network requests, file changes, external execution, publishing or deployment