Skills

Every skill this instance serves, as an agent reads it, sorted by name.

• aggregation-design: Aggregation + filterer slot semantics — what aggregations does, how it composes with filterers, when smart defaults fire, when to use attributes instead. Topical design; per-AGG / per-FILTER detail lives in atomic op-* skills.
• attribute-composition: Attribute composition rules — slot ordering, two-pass vs row-local attributes, the formula expression environment, attributes vs features. Topical design; per-attribute detail lives in atomic op-attr-* skills.
• cohort-null-states: How a .pulse cohort stores absence — the per-record null bitmap, no in-band sentinels, out-of-sample null promotion on inferred imports, and the null / empty / selected states of set and categorical cells across every import and export format.
• cohort-parent-groups: Parent groups in format 0x02 cohorts — storing a repeated parent block once, when a group pays for itself, declaring one at import with a key check, suggesting and gating groups, converting an existing cohort with dedup, and opt-in global-constant elision.
• cohort-schema-design: Entry point for .pulse cohort schemas — the 20-type field matrix, selection heuristics, bit-packed runs, description-length cap, the single-file vs shard-archive shapes, and which focused schema skill answers each nullability, width, parent-group, sharding and sidecar-index question. Use when picking schema types, evaluating storage layout, or interpreting a cohort returned by pulse_inspect.
• cohort-sharding: Shard archives in depth — strict structural cohesion, dictionary union-merge, set-rung auto-widen on create and add, grouped shards, the anchor syntax, union-wide memory for materializing operations, single-writer concurrency, and transport-only compression for moving a cohort.
• cohort-sidecar-index: The sidecar point-lookup index beside a cohort — the v3 file format, the O(1) seek path, three-outcome staleness, discovery through the indexes.json manifest, invalidation by a cohort rewrite, keyable types and constraints.
• cohort-width-overflow: What happens when a value outgrows its field width — inferred categorical and numeric rungs promote on the full row pass, declared widths refuse row by row, set cardinality and shard dictionary overflow, and how convert reports it.
• compose-fits: ComposedRequest.fits — named regression fits that run once before any Compose slot (sweep slots included) and bind into every slot by name, with their own rows, the response fits map, refusals and predict parity.
• compose-requests: ComposedRequest batch semantics — order-preserving slot-by-index dispatch, optional parallel execution, slot labels, post-slot Compose-overlay fold that decorates results without mutating per-slot Components.
• compose-sweeps: ComposedRequest.sweep — one request body with axis placeholders expands into many ordinary Compose slots (grid or zip), with a label pattern, per-slot overlays, limit counting and predict parity.
• crosstab-guide: Crosstab slot — rows × columns groupers, margins-from-raw-rows, normalize / normalize_level / normalize_within, matrix vs long, fused vs buffered, CrosstabComponents indexing. Topical; per-aggregator/grouper math lives in op-* atomics.
• crosstab-margin-aggregations: Auxiliary margin-only figures on a crosstab — a second base beside the cell metric in one scan, the cell-admission rule, present semantics and the display-flag gate
• derived-cohorts: Materialise a request’s derived columns (features, attributes, windows) into a new cohort so a later request can regress, group, filter or join on them — when to, the two-step recipe, what the output holds.
• expression-language: The per-row expression environment shared by the formula attribute and the expression filter — field binding by type, operators, built-in functions, set helpers, null handling, embedder functions and lookup tables. Not request templating.
• facet-design: FacetSchema endpoint — per-field summaries, histograms, additive contribution counts, streamability tradeoffs, four FACET-host overlay kinds against Population. Per-OVERLAY math in op-overlay-* atomics.
• feature-engineering: Feature slot semantics — choosing a transform, pre-filter ordering (features run before filters), per-row vs global-pass cost, target-leakage trap, train/test split. Topical design; per-feature detail lives in atomic op-feat-* skills.
• financial-cohorts: decimal128 patterns — precision/scale propagation, banker’s rounding, divide-by-zero policy, currency conventions, three consumer-facing decimal error codes. Use when a cohort has monetary fields or a response carries PULSE_DECIMAL_OVERFLOW / PULSE_PRECISION_LOSS / PULSE_DECIMAL_DIVIDE_BY_ZERO.
• glossary: Plain-language glossary of the statistical terms Pulse guidance uses. Use when a result, purpose or skill names a term you need explained.
• grouper-design: Grouper slot semantics — multi-grouper composition (key product), fused crosstab eligibility, Group.Include inclusion-list, smart defaults per field type. Topical design; per-GROUP detail lives in atomic op-group-* skills.
• intents: The closed intent taxonomy: the kinds of question an analysis answers, how each sounds, and the field shapes it fits. Use to map a user’s question to an intent ID.
• item-fan-out: for_each on a window, feature or attribute item — one item fans out in place over parameter axes (grid or zip) into ordinary items, with the sweep placeholders, a derived-column limit and predict parity.
• join-design: Pushdown hash-join surface — Request.Joins envelope, JoinSpec, OnPair key-type compatibility, As-prefix, v1 limits, and validating a join before running it.
• label-display: Translate categorical IDs to display labels at output time without rewriting the cohort. LabelBinding entries on Request / SampleRequest / FacetRequest / ExportJob. Display-only — filters / formulas / sort keys still see raw values.
• matrix-results: Vectors and matrices on a Request — naming a battery of numeric fields once, getting a covariance or correlation matrix back, the encodings, missing-data modes, warnings, per-group matrices, predict’s cost estimate and what is refused.
• multiplicity-correction: Multiple-comparison correction on a Request — the multiplicity block (method, family, alpha), which p-values are corrected together, the per-surface family table and its refusals, precedence, the additive adjusted outputs, row and column geometry on pairwise overlays, and what is never corrected.
• multivariate-design: Choosing a multivariate operator on a battery of numeric fields — which matrix answers which question, what to do with missing data and weights, when a decomposition needs a repaired matrix, and where p-values stand. Topical design; per-operator params in the atomic op-mat-* skills.
• multivariate-outputs: Long-form output and parameter detail for the multivariate operators — how rotated PCA and factor loadings, the maximum-likelihood fit indices, centroid distances, row scores and the vector tests lay out their figures. Topical design; the atomic op-mat-, op-attr- and op-test-* skills keep the short form.
• op-agg-average: Arithmetic mean of a numeric field across the input set.
• op-agg-ci-lower: Lower bound of the confidence interval for the mean.
• op-agg-ci-upper: Upper bound of the confidence interval for the mean.
• op-agg-count: Count records that pass the active filter, optionally per group.
• op-agg-distinct-count: Count distinct non-null values across the input set.
• op-agg-distinct-sum: Sum a value field once per distinct key.
• op-agg-frequency: Value count — how many non-null rows hold one chosen value (params.value); one float64 per output row.
• op-agg-kurtosis: Population excess kurtosis g2 (m4 / m2^2 - 3, dividing by n) via online moments.
• op-agg-mae: Mean absolute error of a prediction against an actual — mean(|y - prediction|).
• op-agg-mape: Mean absolute percentage error of a prediction against an actual — mean(|y - prediction| / |y|), as a fraction.
• op-agg-max: Largest non-null value of the field.
• op-agg-median: 50th percentile of the field; requires sorting the full value set.
• op-agg-min: Smallest non-null value of the field.
• op-agg-mode: Most-frequent value of the field (ties broken by the smallest value).
• op-agg-mode-count: Modal count — how many rows hold the field’s most common value; one float64 per output row.
• op-agg-null-count: Count records where the field is null. The inverse of a non-null count.
• op-agg-percentile: Configurable percentile of the field; requires sorting the full value set.
• op-agg-range: Spread (max minus min) of the field across the input set.
• op-agg-ratio: Emits sum(numerator_field) / sum(denominator_field). The Aggregation’s own Field is ignored.
• op-agg-rmse: Root mean squared error of a prediction against an actual — sqrt(mean((y - prediction)^2)).
• op-agg-set-cardinality-avg: Average popcount per contributing row — typical number of selections.
• op-agg-set-cardinality-sum: Sum of popcounts across contributing rows — total label selections seen.
• op-agg-set-distinct-values: Count of distinct exact mask values seen; each combination is atomic.
• op-agg-set-frequency: Per-bit row count — how many rows had each set label selected.
• op-agg-set-intersection: Bitwise-AND a set field across rows; returns labels for every bit set in every contributing row.
• op-agg-set-union: Bitwise-OR a set field across rows; returns labels for every bit set in any contributing row.
• op-agg-skewness: Population skewness g1 (m3 / m2^1.5, dividing by n) via online moments.
• op-agg-stddev: Population standard deviation via Welford’s online algorithm.
• op-agg-sum: Sum the numeric values of a field across the input set.
• op-agg-variance: Population variance via Welford’s online algorithm.
• op-agg-weighted-mean: Weighted arithmetic mean — sum(field * weight) / sum(weight).
• op-agg-welford: Streaming Welford-Pébaÿ moment triple — running mean, sample variance (n-1), and observed count.
• op-agg-zscore: Standardized z-score aggregate — mean-centered, stddev-scaled summary.
• op-attr-code-in: Per-row 0/1 — whether a field’s value is one of a list of codes; its mean is the share of the whole base.
• op-attr-date-part: Extract a calendar component (year, month, day, year_month, …, hour) from a date or datetime field.
• op-attr-factor-score: Per-row score on one common factor of a vector of numeric members, from an inline fit with the factor-analysis matrix operator’s knobs, loadings, order and sign; regression or Bartlett scores; fit_on fits on a subset and still scores every row.
• op-attr-formula: Per-row expression evaluation against the record’s fields via expr-lang.
• op-attr-mahalanobis: Per-row Mahalanobis distance D² (or its chi-square p) across a vector of numeric members, from a cohort-wide fit; fit_on fits on a subset and still scores every row.
• op-attr-normalized: Per-row min-max normalized column — (value − min) / (max − min) ∈ [0, 1].
• op-attr-pc-score: Per-row standardized score on one principal component of a vector of numeric members, with the PCA matrix operator’s sign and order and optional varimax; fit_on fits on a subset and still scores every row.
• op-attr-percentile: Per-row percentile rank column against the post-filter value set; requires sorting.
• op-attr-reg-fitted: Per-row fitted value ŷᵢ = Xᵢ · β + β₀ from an OLS refit during the attribute prepass.
• op-attr-reg-leverage: Per-row hat-matrix diagonal hᵢᵢ from an unpenalized OLS refit — leverage diagnostic.
• op-attr-reg-predict: Per-row score from a bound model (a compose fits entry or a loaded models file) — no refit; whole model or a terms subset.
• op-attr-reg-residual: Per-row residual yᵢ − ŷᵢ from an OLS refit during the attribute prepass.
• op-attr-set-has: Per-row 0/1 — whether the configured label’s bit is set on a set field.
• op-attr-set-popcount: Per-row popcount of a set field — number of selected labels.
• op-attr-tscore: Per-row T-score column — z-score rescaled to mean 50, stddev 10.
• op-attr-zscore: Per-row standardized z-score column — (value − mean) / stddev via two-pass Welford.
• op-feat-bucketize: Bin a numeric column into ordered buckets — explicit boundaries or N equal-population quantiles.
• op-feat-date-features: Expand a date or datetime field into year / month / day / day-of-week / quarter (+ hour) columns.
• op-feat-frequency-encode: Replace each categorical value with its observed relative frequency in the cohort.
• op-feat-log: Per-row log1p(x) of a numeric field; emits one f64 column. Standard skew-tamer.
• op-feat-one-hot: One-hot encode a categorical field as one f64 column per dictionary entry.
• op-feat-poly: Per-row polynomial expansion x^2..x^Degree of a numeric field — the basis for a polynomial regression.
• op-feat-sqrt: Per-row sqrt(x) of a numeric field; emits one f64 column.
• op-feat-target-encode: Replace each categorical value with the smoothed mean of a numeric Target field for that category. Leakage trap — see Gotchas.
• op-feat-train-test-split: Deterministic split-assignment column (0=train, 1=val, 2=test) for ML workflows.
• op-filter-date-ranges: Keep records whose date/datetime field falls inside any of a validated set of labeled date ranges.
• op-filter-exclude: Drop records whose field value appears in the supplied Values list.
• op-filter-expression: Keep records for which an expr-lang expression evaluates truthy against record fields.
• op-filter-false: Keep records where Field is logically false. Strict packed_bool by default; opt into JS falsiness.
• op-filter-include: Keep records whose field value appears in the supplied Values list.
• op-filter-null: Keep or drop records based on null state of a field.
• op-filter-range: Keep records whose numeric field value falls within [low, high] inclusive.
• op-filter-set-contains-all: Keep records whose set field has every bit in the supplied label mask set.
• op-filter-set-contains-any: Keep records whose set field shares at least one bit with the supplied label mask.
• op-filter-set-contains-none: Keep records whose set field shares no bits with the supplied label mask.
• op-filter-set-equals: Keep records whose set field mask exactly equals the supplied label mask.
• op-filter-true: Keep records where Field is logically true. Strict packed_bool by default; opt into JS truthiness.
• op-group-category: Partition records by exact field value; ideal for categorical fields.
• op-group-date: Partition date/datetime records by calendar component (hour..year); week_start picks the week day; fiscal_offset shifts year/quarter.
• op-group-date-ranges: Bucket date records by inline labeled date ranges; the range label becomes the bucket key.
• op-group-kmeans: Segment records into k clusters by seeded, deterministic k-means over a vector of numeric members; keys cluster_1..cluster_k by descending size.
• op-group-quantile: Partition records into N equal-population quantile buckets (Q1..Q4 / D1..D10 / P1..P100).
• op-group-range: Partition numeric records into half-open ranges [a, b); Interval controls bucket width.
• op-group-rounded: Round each numeric value down to a multiple of Interval (floor, not nearest) and group by that scalar.
• op-group-set-per-element: Fan each row into one bucket per selected label (multi-key). Cardinality multiplies with set popcount.
• op-group-set-value: Partition rows by the exact set mask — one bucket per unique combination; key = sorted labels joined with “|”.
• op-mat-collinearity: Collinearity check of candidate regression predictors (no response) — VIF and tolerance per predictor, Belsley’s condition indices and variance-decomposition proportions (uncentered with the intercept, or centered), listwise or pairwise, weighted.
• op-mat-correlation: Correlation matrix of a vector’s numeric members — Pearson (default), Spearman or Kendall tau-b via params.method — listwise or pairwise, weighted; one MatrixResult per spec (per group bucket when grouped), no p-values.
• op-mat-covariance: Covariance matrix of a vector’s numeric members (listwise or pairwise, sample by default), weighted under frequency and probability weights; one MatrixResult per spec (per group bucket when grouped).
• op-mat-distance: Distances between segment centroids — a k × k matrix over the buckets a grouper (params.between: any built-in, date family included, or a registered extension grouper) cuts the rows into; euclidean, sqeuclidean, manhattan, chebyshev, cosine or mahalanobis (pooled within-bucket covariance); optional standardize; weighted centroids.
• op-mat-factor: Exploratory factor analysis of a battery — minres, principal-axis or maximum-likelihood extraction from the SMC, varimax / promax / oblimin rotation, loadings (p × k), communalities, uniquenesses, variance accounted, rmsr, the empirical chi-square, KMO and Bartlett, ML’s likelihood chi-square, TLI, RMSEA with its interval and BIC; listwise or pairwise, weighted.
• op-mat-partial-correlation: Partial correlation matrix of a vector’s numeric members — each pair with every other member, or the params.control fields, held fixed — listwise or pairwise, weighted; non-PSD input refused unless repair “nearest”, singular input refused.
• op-mat-pca: Principal component analysis of a battery — loadings and eigenvectors (p × k, rectangular), every eigenvalue with explained / cumulative shares, communalities, KMO (overall and per member) and Bartlett’s sphericity test; correlation or covariance input, listwise or pairwise, weighted.
• op-mat-reliability: Scale reliability of a battery — Cronbach’s alpha, standardized alpha, McDonald’s omega (one-factor minres), mean inter-item r, per-item item-total r / alpha if deleted / mean / sd — with reverse-keyed items flipped before the fold; listwise or pairwise, weighted.
• op-overlay-chisq-col: Per-column χ² goodness-of-fit test across the host crosstab’s contingency table.
• op-overlay-chisq-matrix: Whole-matrix χ² independence test across the host crosstab’s row × column contingency table.
• op-overlay-chisq-row: Per-row χ² goodness-of-fit test across the host crosstab’s contingency table.
• op-overlay-chisq-vs-pop: χ² goodness-of-fit comparing host Facet subset distribution against the resolved population distribution.
• op-overlay-chisq-vs-ref: Compose-host whole-matrix χ² comparing target slot’s matrix against the reference slot’s matrix (rescaled to target N).
• op-overlay-delta-vs-baseline: Per-point additive delta against a fixed positional baseline of an ordered SERIES host.
• op-overlay-delta-vs-margin: Per-cell additive delta against the matching axis margin (cell − margin).
• op-overlay-delta-vs-prior: Per-point streamable windowed additive delta against the immediately preceding point of an ordered SERIES host.
• op-overlay-delta-vs-ref: Compose-host additive delta of target slot value against the matching reference slot value (matrix or series).
• op-overlay-delta-vs-sibling: Per-group additive delta against a sibling group named by (Field, Value) on the SERIES host.
• op-overlay-delta-vs-stage: Whole-chain additive delta of target stage’s result against the reference stage’s result.
• op-overlay-fisher-exact-cell: Per-cell Fisher’s exact two-sided p-value over a 2×2 contingency formed from each host crosstab cell + its margins.
• op-overlay-formula: Expression-driven projection computed via expr-lang against a per-host-shape variable namespace.
• op-overlay-index-vs-baseline: Per-point ratio index against a fixed positional baseline of an ordered SERIES host (×100).
• op-overlay-index-vs-margin: Per-cell index against the matching axis margin (100 × cell / margin).
• op-overlay-index-vs-pop: Per-value population-comparison index for a Facet host (subset_freq / pop_freq × 100).
• op-overlay-index-vs-prior: Per-point streamable windowed index against the immediately preceding point of an ordered SERIES host (×100).
• op-overlay-index-vs-ref: Compose-host ratio index of target slot value against the matching reference slot value (matrix or series).
• op-overlay-index-vs-rolling-mean: Per-point windowed index against the arithmetic mean of the W preceding points of an ordered SERIES host.
• op-overlay-index-vs-sibling: Per-group ratio index against a sibling group named by (Field, Value) on the SERIES host (×100).
• op-overlay-index-vs-stage: Whole-chain ratio index of target stage’s result against the reference stage’s result (×100).
• op-overlay-index-vs-total: Per-group streamable ratio index against the SERIES host’s grand total (×100).
• op-overlay-ks-vs-pop: Kolmogorov-Smirnov distance + p-value comparing host Facet NUMERIC distribution against the resolved population.
• op-overlay-pairwise-probit-t: Intra-matrix axis-pairwise probit t-test (Φ⁻¹ transform of each leg’s proportion, Student-t tail).
• op-overlay-pairwise-prop-z: Intra-matrix axis-pairwise pooled-SE two-proportion z-test (row-vs-row or col-vs-col within one crosstab).
• op-overlay-pairwise-two-means-z: Intra-matrix axis-pairwise two-means z-test on AGG_WELFORD cells (normal-CDF tail, no df).
• op-overlay-pairwise-weighted-two-means-z: Intra-matrix axis-pairwise two-means z-test on weighted-mean cells under a required n_basis (weights | kish).
• op-overlay-pairwise-welch-t: Intra-matrix axis-pairwise Welch–Satterthwaite t-test on AGG_WELFORD cells.
• op-overlay-panel-index-vs-ref: Compose-host multi-reference index — indexes every target slot against a shared reference; emits one layer per target.
• op-overlay-prop-z-cell: Compose-host per-cell two-proportion z-test against the reference slot’s matching cell.
• op-overlay-prop-z-panel: Compose-host multi-reference per-cell pairwise two-proportion z-test across N+1 slots.
• op-overlay-rank: Compose-host per-cell rank of each target cell within a configurable population (row / column / matrix).
• op-overlay-share-of-col: Per-cell share-of-column ratio (cell / col_margin) — raw share, sums to 1.0 per column.
• op-overlay-share-of-row: Per-cell share-of-row ratio (cell / row_margin) — raw share, sums to 1.0 per row.
• op-overlay-share-of-total: Share-of-grand-total ratio — dual-shape (matrix per-cell or series per-group); raw share sums to 1.0.
• op-overlay-t-cell: Compose-host per-cell Welch t-test against the reference slot’s matching cell; parity overlay reading from CellComponents.
• op-overlay-t-vs-ref: Compose-host per-group Welch t-test against the reference slot’s matching group (SERIES); parity overlay reading from CellComponents.
• op-overlay-yoy: Per-point year-over-year ratio against the same period one year prior; requires GROUP_DATE host.
• op-overlay-z-cell: Compose-host per-cell two-sample z-test on the means against the reference slot’s matching cell; parity overlay reading from CellComponents.
• op-overlay-z-vs-ref: Compose-host per-group two-sample z-test on the means against the reference slot’s matching group (SERIES); parity overlay reading from CellComponents.
• op-overlay-zscore-vs-margin: Per-cell standardized-margin z-score — (cell − margin) / sd where sd is the population SD within the same margin slice.
• op-overlay-zscore-vs-pop: Per-value population-comparison z-score for a Facet host — (subset_freq − pop_freq) / sd_pop.
• op-overlay-zscore-vs-rolling: Per-point windowed z-score against the rolling-window mean + SAMPLE SD of the W preceding points.
• op-overlay-zscore-vs-total: Per-group streamable z-score against the SERIES host’s grand-total distribution (population SD).
• op-reg-bayes-linear: Bayesian linear regression with a conjugate Normal-Inverse-Gamma prior; emits posterior means, std errors, and credible intervals. Streams the same sufficient statistics as an OLS fit.
• op-reg-glm: Generalized linear model via IRLS; supports binomial (logistic), poisson, and gamma families with family-specific link functions. Always buffered.
• op-reg-mod-resample: Spec-level resampling modifier (jackknife / bootstrap) that composes with an OLS or GLM fit; replaces analytical std errors with empirical estimates. Non-empty value forces the buffered path.
• op-reg-mod-selection: Spec-level subset-selection modifier (forward / backward / stepwise) that composes with an OLS or GLM fit; drives a greedy search over predictors against an information criterion. Non-empty value forces the buffered path.
• op-reg-ols: Ordinary least squares with optional l1/l2/elasticnet penalty; covers simple, multiple, ridge, lasso, and elastic-net regression over streaming sufficient statistics.
• op-synth-bernoulli: Bernoulli samples emitting 0 or 1 with probability p; pairs with packed_bool or unsigned int fields.
• op-synth-constant: Emit a constant value on every row; useful for unit testing, sentinel columns, and placeholder fields.
• op-synth-discrete: Exact per-level histogram for an integer column; what a profile reconstructs every u4/u8/u16/u32/u64 field from.
• op-synth-exponential: Exponential samples with rate parameter lambda; mean is 1/lambda.
• op-synth-lognormal: Log-normal samples parameterised by log-space mu and sigma; output strictly positive.
• op-synth-mixture: Mixture-of-normals samples — per-row component draw by weight, then a Gaussian from that component; reproduces bimodal/multimodal shapes.
• op-synth-monotonic-from: Strictly increasing or decreasing integer counter; deterministic per row, ignores RNG. Synthetic primary keys.
• op-synth-normal: Gaussian samples with optional [min, max] clamp; default mean=0, std=1.
• op-synth-pareto: Heavy-tailed Pareto samples; xm is the scale (minimum), alpha the shape (tail weight).
• op-synth-poisson: Poisson-distributed non-negative integer counts with mean lambda; Knuth multiplicative below 30, normal-approximation above.
• op-synth-regex: String samples generated from a Perl/RE2 regex pattern; walks regexp/syntax AST with bounded repetition.
• op-synth-set-bernoulli: Multi-select set_* bitmask generation — independent per-option Bernoulli draws, or a joint-structure resample when a paired field is declared.
• op-synth-uniform: Uniform real-valued samples in [min, max); default [0, 1).
• op-synth-uniform-date: Uniform calendar-date samples in [start, end] inclusive; days-since-epoch internally.
• op-synth-weighted-categorical: Discrete categorical draws from a values list with optional weights; uniform when weights absent.
• op-test-anova-f: One-way ANOVA F-test comparing the means of a numeric Field across k groups defined by SplitBy.
• op-test-anova-rm: Repeated-measures one-way ANOVA; rows pivoted on SubjectField × SplitBy condition.
• op-test-anova-welch: Heteroscedasticity-robust one-way ANOVA with Welch-Satterthwaite df correction; streamable per-group Welford.
• op-test-bartlett-sphericity: Bartlett’s sphericity test on a vector’s correlation matrix (is it the identity?); the factorability check, bit-equal to the PCA operator’s Bartlett scalars.
• op-test-box-m: Box’s M test of one shared covariance across a SplitBy of 2+ levels; chi-square headline, F approximation in details. The MANOVA / Hotelling assumption check.
• op-test-brown-forsythe: Homogeneity-of-variance test; one-way ANOVA on absolute deviations from per-group medians.
• op-test-chisq: Chi-square independence test on a 2D Rows × Cols contingency table.
• op-test-fisher-exact: Exact two-sided p-value for a 2×2 Rows × Cols contingency table; small-sample CHISQ alternative.
• op-test-hotelling-t2: Hotelling’s T² on a mean vector of p numeric members; one-sample or paired (difference vector) against params.mu, or two-sample over a 2-level SplitBy with pooled or unequal covariance.
• op-test-kendall-tau: Concordance-based correlation between Field and Field2; buffered O(n²) pair count under tie correction.
• op-test-kruskal-wallis: Nonparametric k-group location test; ANOVA-like rank sums under tie correction.
• op-test-ks: Kolmogorov-Smirnov two-sample distribution test on Field partitioned by SplitBy.
• op-test-mann-whitney-u: Nonparametric two-sample location test; buffered tie-corrected rank sum (Mann-Whitney U).
• op-test-manova: One-way MANOVA of a mean vector of p numeric members across a SplitBy of 2+ levels; Pillai’s trace headline, with Wilks, Hotelling–Lawley and Roy in details.
• op-test-mardia: Mardia’s multivariate normality test on a vector — skewness χ² and kurtosis z; buffered.
• op-test-paired-t: Paired-sample t-test on the per-row difference d = Field − Field2; streamable via Welford on d.
• op-test-pearson-r: Parametric Pearson correlation test between two numeric fields; streamable via online cross-product.
• op-test-prop-z: Two-proportion z-test on the success rate of Field across two SplitBy groups; streamable via success / total counts.
• op-test-shapiro-wilk: Shapiro-Francia (W′) normality test on Field; runs per-group when SplitBy is set; p advisory outside 5..5000.
• op-test-spearman-r: Rank-based correlation between Field and Field2 (monotonic association); buffered mid-ranks then Pearson.
• op-test-t: One-sample or two-sample Welch t-test on a numeric field; SplitBy switches to the two-sample variant.
• op-test-trend: Mann-Kendall trend test over an ordered numeric series; tier-2 over windowed result rows.
• op-test-tukey-hsd: Tier-2 post-hoc pairwise comparison of group means using Tukey’s Honestly Significant Difference (Tukey-Kramer).
• op-test-welch: Explicit two-sample Welch t-test on a numeric field across a categorical SplitBy partition.
• op-test-wilcoxon-sr: Wilcoxon signed-rank test on per-row difference d = Field − Field2; buffered tie-corrected sign-rank.
• op-test-z-two-sample: Two-sample z-test on Field means across two SplitBy groups; identical SE to Welch’s t-test but p-value via standard normal Φ.
• op-win-delta: Point difference against the row periods positions earlier in the ordered partition — the subtraction counterpart of a percent change.
• op-win-dense-rank: Dense rank with no gaps after ties (1, 2, 2, 3, …) within the ordered partition.
• op-win-ewma: Exponentially weighted moving average; s_i = alpha*x_i + (1-alpha)*s_{i-1}.
• op-win-lag: Per-row value of Field from N rows earlier in the ordered partition.
• op-win-lead: Per-row value of Field from N rows later in the ordered partition.
• op-win-moving-avg: Moving average of Field over a bounded Frame; both ends must be set.
• op-win-pct-change: Fractional change (0.05 = 5%) against the row periods positions earlier in the ordered partition.
• op-win-rank: Sparse rank with gaps after ties (1, 2, 2, 4, …) within the ordered partition.
• op-win-row-number: 1-based row index within the ordered partition; never ties.
• op-win-running-avg: Running average of Field over the configured Frame within the ordered partition.
• op-win-running-sum: Running total of Field over the configured Frame within the ordered partition.
• overlay-system: Overlay framework — OverlaySpec composition, six reference families, three payload shapes, host-arm wiring (SERIES / FACET / CHAIN / FORMULA), parity overlays + Welford migration. Per-kind detail lives in op-overlay-* atomics.
• pairwise-n-sources: The sample-size vocabulary shared by every OVERLAY_PAIRWISE_* kind — the nine n_source modes, the three distinct-key modes and their cell-aggregator admission, the slab partition gate, the null rules, and p_source.
• predict-advisories: Predict advisories — coded, non-blocking notes in a predict result that the chosen analysis may not fit the data (a two-group test over many groups, many uncorrected p-values, an SPSS nominal or ordinal field treated as a quantity, an unused SPSS weight); how they differ from warnings, what each code means, what it suggests, and how an embedder suppresses one.
• process-chain: ChainRequest source-rooted linear pipeline — mergeable-only v1 gate, per-stage Response.Components, dual-slot overlays (per-stage + whole-chain), StageRef resolution, shape-divergence warning.
• regression-inference: Coefficient confidence intervals (ci_lower / ci_upper at conf_level) and the coefficient covariance and correlation on a regression — the per-family critical value, the opt-in vcov flag, where each engine’s and modifier’s matrix comes from, how it relates to the standard errors, weighted fits, the refusals, and what R’s vcov does differently. Topical design; per-model params in atomic op-reg-* skills.
• regression-modeling: Choosing and composing a regression — outcome type, priors, penalties, the resampling and stepwise modifiers, polynomial terms built upstream, and how textbook regression names map to specs. Topical design; per-model detail in atomic op-reg-* skills.
• request-envelope: Request shapes, envelope contract, slot keys, smart defaults, streamability and time zones. Use when authoring or adapting any Pulse Request.
• request-templating: Stored parameterised JSON requests — the {target, variables, body} wrapper, $var / {{}} / $when substitution, the nine variable types, directory discovery, hot reload, and the PULSE_TEMPLATE_* family. Not expr-lang.
• response-components: How Response.Components carries the constituent parts of every aggregation, grouper, filterer, and crosstab cell
• response-shaping: The request return block — choose which parts of a response come back (presets, include / exclude paths) and how many significant digits floats carry on the wire; what is projected, what stays exact, and how it composes with streaming and instance defaults.
• session-bootstrap: Canonical MCP session order — manifest once, then examples → predict → process → errors_lookup. Skill-name derivation from operator names. Use first on every new MCP session.
• session-format-flags: Per-format CLI flags an agent must know — source-side read knobs (Excel sheet, SPSS missing-value mode and charset), parent-group declaration on the managed import path, the import source zone, the export render zone, and the four .sav write knobs with their refusal defaults.
• session-skill-routing: Which skill to fetch for a given trigger — operator family prefixes, manifest catalog entries, request and response slots, request roots, CLI leaves and plain-language questions — plus how to read data.components on first sight.
• session-synth-flags: The profile-capture and synth-generation CLI flags an agent must know — the five additive profile create sections, emitting the generating spec, supplying and suggesting structural rules — and why a model coefficient is a latent-scale quantity.
• spss-cohorts: Entry point for SPSS .sav / .zsav cohorts — the three surprises (extra derived columns, codes-not-labels dictionaries, the label-tables directory trap), why dictionaries hold codes, and which focused SPSS skill answers each import, missing-value, sidecar and .sav-writing question. Read this when a cohort came from SPSS or when emitting a .sav.
• spss-export: Writing .sav — the dictionary + data-section writer, bytecode compression, the four write flags, why the writer takes the whole cohort rather than rendered rows, how a .sav-to-.sav convert carries the source through, and asking export predict first.
• spss-export-fidelity: How far the .sav round-trip claim is proven — what CI gates, what was checked only locally against independent readers, what nothing corroborates — and the write-side charset rules (source charset kept, no replacement characters, no silent truncation, encode-measure-segment).
• spss-export-values: What a .sav export writes for each value — the variable-name policy and –sanitize-names, set_* masks at every width, the missing state each null takes, original codes versus dictionary positions, byte order, and labels that live only on user-missing codes.
• spss-import-schema: How a .sav / .zsav file imports — the declared dictionary replaces inference, the SPSS-to-Pulse type mapping, the three data encodings, charset and byte-order rules, and the fatal-vs-warning damage diagnostics. Read this when an SPSS import types a column unexpectedly or refuses a file.
• spss-metadata-sidecar: The cohort.pulse.spss.json sidecar an SPSS import writes — what it records that the .pulse format cannot, the code-label-dictionary-ID triple labels come from, its fingerprint, and the absent / stale / invalid / ignored read verdicts.
• spss-missing-values: How SPSS user-missing and system-missing values land in a cohort — the numeric <var>_missing sibling versus the flagged categorical code, why the two arms deliberately differ, –spss-missing, and filtering on the code rather than the label.
• spss-response-sets: SPSS multiple-response sets and the columns an import synthesises — the derived set_* mask beside a multiple-dichotomy set’s constituents, why multiple-category sets derive nothing, and the closed derived registry the export folds by. Read this when a cohort has more columns than its .sav had variables.
• statistical-testing: Tier-1 (tests) vs tier-2 (post_tests) pairing, assumption gates, ANOVA + post-hoc composition, p-value conventions. Topical design; per-TEST detail in atomic op-test-* skills.
• streaming-and-watching: Reactive Pulse primitives — Request.Hash() cache keys, StreamResult[T] incremental output with per-chunk + terminal Components, Watch / WatchDir mutation observation, FilterToFileWithRequest deterministic derived cohorts, manifest CommandAnnotations caching policy.
• synth-conflicts: Synth conditional relationship conflicts — the shared claim space, the fixed priority order resolveConflicts applies once per spec, how rule / model / shape pre-claims compose, and where the dropped-relationship warnings surface.
• synth-correlations: Synth spec correlations — the Gaussian-copula construction, which marginals and field types may participate, how rank and Pearson correlation survive, and how unmeasured matrix entries are completed and reported.
• synth-determinism: The synth determinism contract — byte-identical .pulse output per spec and seed, byte-identical profile documents per input and seed, and the two non-RNG threats to capture (map iteration order, float fusion).
• synth-fidelity-report: The pulse synth from-profile tagged top-up contract (_synthetic column, new output path, --rows counts new rows) and the --fidelity-report sections — marginals, pairs, set fields — scored only over relationships generation actually applied.
• synth-marginals: Why a profiled packed_bool reconstructs as bernoulli and a small integer as discrete rather than a clamped normal — the three writers each arm covers, probit and ordered-probit model draws, and what that does to recovery and correlation.
• synth-model-draw: How synth generates a modelled numeric — the composed latent draw, firing terms and the “other” catch-all, the single clamp, draw ordering, and why a coefficient is latent-scale and non-linear in value space.
• synth-model-recovery: The synth fidelity report’s models section — refitting each applied model on the generated rows, comparing on the latent scale (or a calibrated probit score for step targets), flagging bands, and how to read a flag on a quantized target.
• synth-model-selection: How synth --fit-models chooses predictors — the top-K collapse, the variance-explained floor, marginal scoring and refit-on-rank-deficiency, main effects only, zero-predictor models — and how thin levels are shrunk by ridge rather than dropped.
• synth-models: Entry point for synth --fit-models — per-numeric linear model capture and its wire shape, the latent-scale reading of a coefficient, what a model retires, and which focused skill covers selection, the draw, residual correlations and recovery.
• synth-profile-capture: What pulse profile create captures per field and per pair (conditional pairs, thin-pair warning, run-continuation, shape fitting), how SpecFromProfile reconstructs each field, and how categorical joint structure is generated.
• synth-residual-correlations: Synth --residual-correlations — capturing the correlation among fitted model residuals (full submatrix, measured vs unmeasured), the correlated residual draw that lets a numeric be both conditioned and correlated, and its zero-predictor boundary.
• synth-residual-recovery: The synth fidelity report’s model_residual_correlations section — captured vs recovered residual correlation per applied pair, why it is not pairwise, its bounded worst-first listing, and why step-target endpoints stay unmeasured.
• synth-rule-claims: Which synth rules pre-claim a field ahead of its linear model, conditional pairs and residual correlations, the four exclusions that are silent if got backwards, and why the fidelity report is the reason.
• synth-rule-detectors: The --suggest-rules detector rules that are silent if got backwards — rounded gates, co-missingness, near and always-null blocks, the narrow dependency search and discovered band edges — and why emitted candidate order is applied order.
• synth-rule-expressions: Synth constraints[] and the expression side of rules[] — the row environment’s types, isnull, the set_expr coercion matrix, set-option addressing, and the pre-rounding gotcha with its round (not int) remedy.
• synth-rule-nulls: Synth rule null handling — null_together as one null decision per question block (first field wins), and owns_nulls making a gate the field’s only source of absence, with the divergence warnings each emits.
• synth-rule-validation: Synth rule validation and diagnostics — the eight eager PULSE_SYNTH_RULE_* refusals, the non-nullable-target refusal, and the never-fired warning with its three cause arms.
• synth-rules-from-profile: Reaching synth rules from the profile path — --rules (replace semantics, load-time validation) and --emit-spec (the spec that generated) — and profile create --suggest-rules: proposed candidates, inert evidence, fixed thresholds and the three detectors.
• synth-set-fields: How set_* (multi-select) fields are profiled per option, captured jointly per option pair, read safely on the wide rungs, and generated deterministically with a pre-registered dictionary.
• synth-shape-fit: Synth --fit-shape — the two-component mixture capture and its keep rule, and how a shape-fitted numeric still accepts model conditioning through its mixture quantile, with a fixed-count bisection for determinism.
• synth-structural-rules: Entry point for synth spec rules[] and constraints[] — the five rule slots, declaration order and last-write-wins, why the rule pass runs last, and which focused skill covers expressions, null handling, validation, claims and profile-derived rules.
• synthetic-data: Entry point for synthetic data — pulse synth from-schema vs pulse synth from-profile (CLI + library; no MCP tool), the spec shape, the distribution registry, the determinism rule, and which focused synth skill answers each question.
• time-zones: Time zones end to end — which zone a request uses, how a datetime buckets by local calendar day, import in a source zone with DST policy, native timestamps, local-offset export and the tz database version. Topical design.
• tool-compose: Execute a batch of processing requests in one round-trip.
• tool-dedup: Deduplicate an existing cohort’s repeated parent blocks into parent groups (format 0x02), or suggest groups read-only.
• tool-drop: Remove a managed-import handle — delete its .pulse file and sidecar.
• tool-errors-lookup: Look up Pulse error code metadata — message and fixup hints.
• tool-examples-get: Fetch one runnable request example from the embedded library by name.
• tool-examples-search: Search the runnable request-example library for templates matching a question.
• tool-explain: Say in plain words what a request will do, or what a result found — closed verdicts, conventional strength bands and caveats.
• tool-facet: Return distinct values for one field in a cohort.
• tool-facet-schema: Multi-field rich facet — counts, nulls, percentiles, histograms, additive contributions.
• tool-import: Import a tabular source file (or pass through .pulse) into a managed handle.
• tool-imports-list: Enumerate every managed-import handle with sidecar metadata.
• tool-inspect: Read header + schema of a .pulse file without touching record data.
• tool-label-resolve: Reverse-resolve a human-readable name (typo-tolerant) to raw categorical key(s).
• tool-label-tables: List registered label tables — ID→display-name dictionaries for categorical fields.
• tool-lookup: Resolve a point lookup against a cohort’s prebuilt sidecar index.
• tool-manifest: Bootstrap blob — call once per session for the operator + capability catalog.
• tool-materialize: Write a request’s derived columns (features, attributes, windows) as ordinary fields of a new cohort that later requests read.
• tool-predict: Validate a request against a cohort schema without executing.
• tool-process: Execute one pre-built processing request against a cohort.
• tool-process-chain: Linear chain of stages; stage N+1 consumes stage N’s output rows.
• tool-range-tables: List registered range tables — named, reusable sets of labeled date ranges.
• tool-recommend: Turn a question kind (an intent ID) into ranked draft requests — placeholder skeletons, or predict-checked drafts bound to a cohort’s fields.
• tool-sample: Return up to N rows from a cohort for eyeball / preview.
• tool-skills-get: Fetch the markdown body of one named skill.
• tool-skills-list: List the skill pack — domain guides and atomic operator/type/tool refs, deployment-added skills included.
• type-categorical-u16: 2-byte categorical with inline dictionary up to 65,536 labels.
• type-categorical-u32: 4-byte categorical with inline dictionary up to ~4.29B labels.
• type-categorical-u8: 1-byte categorical with inline dictionary up to 256 labels.
• type-date: Day-granularity date stored as epoch days (signed since 1970-01-01).
• type-datetime: Second-granularity instant stored as signed epoch seconds (int64, negative = pre-1970), naive UTC.
• type-decimal128: 128-bit fixed-point decimal with per-field (precision, scale); exact base-10 arithmetic.
• type-f32: IEEE-754 single-precision float; ~7 significant digits.
• type-f64: IEEE-754 double-precision float; ~15 significant digits.
• type-packed-bool: 1-bit boolean; bit-packed with adjacent packed_bool / u4 fields.
• type-set-u128: 16-byte multi-select bitmask over an inline dictionary of up to 128 labels.
• type-set-u16: 2-byte multi-select bitmask over an inline dictionary of up to 16 labels.
• type-set-u256: 32-byte multi-select bitmask over an inline dictionary of up to 256 labels — the widest set rung.
• type-set-u32: 4-byte multi-select bitmask over an inline dictionary of up to 32 labels.
• type-set-u64: 8-byte multi-select bitmask over an inline dictionary of up to 64 labels.
• type-set-u8: 1-byte multi-select bitmask over an inline dictionary of up to 8 labels.
• type-u16: 2-byte unsigned integer; fits cardinals through 65,535.
• type-u32: 4-byte unsigned integer; the common wide-cardinal default.
• type-u4: 4-bit unsigned integer; bit-packed with adjacent u4 / packed_bool fields.
• type-u64: 8-byte unsigned integer; for sub-day timestamps and very wide cardinals.
• type-u8: 1-byte unsigned integer; fits cardinals through 255.
• weighting: Row weights on a Request — the request / per-slot weight and the instance default, probability vs frequency kinds, resolution order, invalid-weight exclusion, which operators honour, skip or refuse a weight, weighted inference on the effective sample size, the unweighted-base recipe and the weighted floor keys.
• window-design: Window slot semantics — partition / order / frame, what window operators share conceptually, streamability per window. Topical design; per-WIN detail lives in atomic op-win-* skills.