I — Datum
Capability Manifest
Complete instrument specification for the SVEND analysis platform. 186 statistical analyses, process control, measurement system analysis, lean operations, and strategy deployment. The statistics are calibrated against NIST Statistical Reference Datasets and cross-checked against R. Anything not yet shipped is marked on the row that claims it, with the month it is due.
Contents
II — Statistical Analysis Engine
186 statistical analyses spanning hypothesis testing, regression, ANOVA, reliability, survival analysis, time series, machine learning, and Bayesian methods. Hypothesis testing, regression and capability are calibrated against published reference datasets with defined tolerance bands; families outside that scope are computed but not externally calibrated. The Workbench is a stateless calculator: it holds no project state, so an analysis is reproducible from its dataset and its configuration alone.
| Capability | Specification | Status |
|---|---|---|
| Hypothesis testing | t-tests (1-sample, 2-sample, paired), z-tests, chi-square, Fisher exact, Mann-Whitney U, Wilcoxon, Kruskal-Wallis, Friedman, McNemar, Mood's median | ● |
| ANOVA | One-way, two-way, general linear model, Welch ANOVA, repeated measures, nested, Tukey/Bonferroni/Dunnett post-hoc comparisons | ● |
| Regression | Simple linear, multiple linear, polynomial, logistic, Poisson, stepwise, ridge, LASSO, partial least squares, residual diagnostics | ● |
| Modeling | Response-type model fitting, logistic, Poisson, ordinal logistic, least squares, robust, orthogonal, nonlinear, stepwise selection, best-subsets selection | ● |
| Characterise | Descriptive summaries, outlier flagging, distribution-family fitting, coverage and confidence tolerance intervals, normal-fit assessment, Box-Cox transforms, Johnson transforms | ● |
| Reliability & survival | Weibull, lognormal, exponential fitting, Kaplan-Meier, Cox proportional hazards, accelerated life testing, MTBF/MTTF, B-life estimates | ● |
| Time series | Autocorrelation, partial autocorrelation, decomposition, trend analysis, changepoint detection (BOCPD), exponential smoothing | ● |
| Bayesian methods | Bayesian estimation, credible intervals, prior specification, posterior predictive checks, Bayesian ANOVA, Bayesian regression, model comparison (BIC, WAIC) | ● |
| Distribution analysis | Normality tests (Anderson-Darling, Shapiro-Wilk, D'Agostino), distribution fitting, probability plots, goodness-of-fit, Box-Cox transformation | ● |
| Correlation | Pearson, Spearman, Kendall, partial correlation, correlation matrix, multicollinearity diagnostics (VIF) | ● |
| Sample size & power | Power analysis for means, proportions, ANOVA, regression, equivalence/non-inferiority, acceptance sampling (AQL/LTPD) | ● |
| Multivariate | PCA, factor analysis, discriminant analysis, cluster analysis, Hotelling T², multivariate capability | ● |
| Machine learning | Gradient boosting (XGBoost, LightGBM), random forest, regularised regression, model comparison, feature importance, anomaly detection | ● |
| Data preparation | Merge two datasets, melt wide columns into long form, filter rows, and derive a new column from an expression. Each step previews before it is applied, and the result is saved as its own dataset. | ● |
| Causal discovery | PC algorithm and LiNGAM. Proposes which factors plausibly drive which. | ● |
Tolerance bands: p-values ±0.005, correlation coefficients ±0.02, test statistics ±0.01. Checked on every build.
III — Statistical Process Control
Classical and Bayesian control charts with automatic rule-based detection. Process capability indices with frequentist and Bayesian estimation. Measurement system analysis for validation before capability studies.
| Capability | Specification | Status |
|---|---|---|
| Variables control charts | I-MR (individuals & moving range), X-bar/R, X-bar/S, zone charts, with automatic subgroup sizing | ● |
| Attributes control charts | p (proportion defective), np (count defective), c (defects per unit), u (defects per unit variable size) | ● |
| Advanced control charts | CUSUM (cumulative sum), EWMA (exponentially weighted moving average), Bayesian control charts with posterior probability of shift | ● |
| Nelson rules | All 8 Nelson rules with automatic detection, configurable rule selection, out-of-control point annotation | ● |
| Process capability (frequentist) | Cp, Cpk, Pp, Ppk, Cpm, sigma level, DPMO, yield (FPY, RTY), confidence intervals. Every answer states the methods it was computed by — the location estimator, the dispersion estimator and the number of values — and states the restriction alongside them: the indices carry the normal substitution, which ISO 22514-2 admits for a process whose location and dispersion are constant. Tool wear and lot-to-lot variation are not that process, and the answer says so | ● |
| Process capability (Bayesian) | Bayesian Cpk with posterior distribution, P(Cpk > 1.33), credible intervals, prior specification, small-sample robust | ● |
| Gage R&R | Crossed design by ANOVA: %GRR, %Contribution, ndc (number of distinct categories), operator-by-part interaction. Gage linearity and bias runs here too. Nested (destructive) Gage R&R is not available — the study bench is section IV | ● |
| Changepoint detection | Bayesian online changepoint detection (BOCPD), evidence accumulation, shift probability estimation | ● |
Capability indices validated within ±0.05 of reference values. Ref: AIAG SPC Manual, 2nd Ed.
IV — Measurement Systems
A gage study is planned before it is measured, so it does not fit the stateless workbench in section II and has a bench of its own. Ten parts, three appraisers, three trials, randomized — issued as a run sheet, filled in on the floor hours or weeks later, then judged whole. How the study bench works →
| Capability | Specification | Status |
|---|---|---|
| Study plan | Parts as labels, named appraisers, trial count, characteristic and units, optional specification limits. Up to 40 parts, 10 appraisers, 6 trials and 600 readings | ● |
| Randomized run sheet | A fresh permutation for every appraiser and every trial, generated from a seed recorded with the study so a reprinted sheet matches a half-worked one. Re-planning over recorded readings is refused | ● |
| Keyboard-wedge capture | Any digital gage or receiver presenting as a keyboard: focus opens on the first unmeasured row, Enter or Tab advances by run order, and text that is not a reading holds the row. Input contract → | ● |
| Tamper-evident readings | Each submission is recorded as an act — who submitted what, against which rows, when — and those acts are hash-chained. A correction is a later event; the study declaration is deliberately outside the chain | ● |
| Instrument register | A typed gage name resolves to a device record, so a study names an instrument | ● |
| Judgment | The completed study is judged through the same engine door as every other analysis. A part-worked sheet is refused, naming what is missing — a crossed study is judged whole | ● |
| Nested Gage R&R | Destructive testing, where no part can be measured twice — each appraiser gets its own parts. Available on the study bench | ● |
| Gage stability | Repeated measurement of a reference part over time as a dedicated study. Available on its own bench, Team tier — it moved into the Device Trust System on 2026-08-10, because checking a master against its known value over months is calibration work | ● |
| Type 1 gage study | Cg and Cgk against a reference part and a declared tolerance. Runs in the Analysis Workbench on readings you already have — a type 1 study measures one part repeatedly and needs no run sheet issued across appraisers | ● |
| Identified device capture | Readings carrying the instrument that produced them. Specified against the MobileCollect EVO serial interface; requires hardware validation and is published as a specification | ○ |
| SVEND Device Trust System | The instrument’s own record: a device register keyed to the individual instrument, calibration records carrying as-found and as-left values per check point with uncertainty, the traceable standard and the decision rule, and the issuing authority. Team tier | ● |
| Impact assessment | When a certificate comes back out of tolerance, which work that instrument measured is put in question — answered as a query. Distinct from calibration recall, which is calling an instrument in before its interval lapses; both are here and metrologists read the two words differently. Team tier | ● |
| Certificate, checkable | A calibration certificate as a document: hash-chained to its record, carrying a code that resolves to a verification page, and signed with the issuing organization’s own digital ID — SVEND is not the signatory and cannot be. Scanning tells you whether the document matches the issuer’s record; opening it in a PDF reader tells you whether the bytes changed since it was signed. Detection only: no certificate is impossible to imitate. What a signature proves → Team tier | ● |
| Digital Calibration Certificate | The machine-readable XML calibration certificate led by Germany’s PTB, read and issued. A laboratory’s DCC can be read in directly, and a SVEND calibration can be served as one at its own address, validated against PTB’s schema and carrying the decision rule with it. Units travel as D-SI, so a value keeps its unit through the exchange. Adoption today is largely European and institutional; this is here so that a shop receiving one is not retyping it. Team tier | ● |
The study bench is available from Pro. Gage stability and the Device Trust System are Team, because calibrating an instrument is a different job from analyzing the readings it produced. The free Gage R&R calculator needs no account and runs the same crossed ANOVA on readings you bring. Ref: AIAG MSA Manual, 4th Ed.
V — Design of Experiments
Both halves work today. Declare your factors and their real settings, choose a design, and get back a coded matrix and a randomized run sheet to carry to the line. Type the results in when they come back and the fit runs through the same engine as everything else in section II.
| Capability | Specification | Status |
|---|---|---|
| Factorial analysis | Main effects and interactions via ANOVA and general linear model, two-way and nested designs, post-hoc comparison, residual diagnostics | ● |
| Response surface fitting | Polynomial regression, curvature and lack-of-fit testing, multiple regression with interaction terms, model comparison | ● |
| Power & sample size | Power analysis for factorial designs, runs required per group at a target effect size, computed power curve across sample sizes | ● |
| Design generation | Full factorial, fractional factorial, Plackett-Burman, definitive screening, central composite, Box-Behnken and Taguchi designs, with a coded matrix and a randomized run sheet in the operator's own units. Mixture and split-plot generation are not among them; split-plot analysis is | ● |
Available from Pro. The full design reference covers the seven families, the run-sheet contract and the boundaries. Ref: Montgomery, Design and Analysis of Experiments, 8th Ed.
VI — Lean Operations
Value stream mapping, line balancing, OEE, takt time, kanban sizing, and a set of free operations calculators. Map current state, identify waste (DOWNTIME), simulate the line as a discrete-event model, and carry kaizen bursts forward into strategy.
The method these tools support is written up in Force Preservation: A Campaign Model for Operational Excellence — shaping, a narrow breach, consolidation, and expansion paid for by the capacity consolidated areas return.
| Capability | Specification | Status |
|---|---|---|
| Value stream mapping | Current state / future state pairing, process box editor, inventory triangles, two-level lead time ladder, discrete-event simulation of the mapped line, overhead floor layout | ● |
| Line balance & simulation | Cycle time analysis, operator balance charts, line efficiency, bottleneck identification, what-if simulation | ● |
| OEE | Availability, performance, quality breakdown, six big losses classification, OEE trending, world-class benchmarking | ● |
| Takt time | Available time, customer demand, takt calculation, pitch, planned cycle time, efficiency ratio | ● |
| Kanban sizing | Kanban quantity calculation, safety factor, replenishment lead time, container sizing, signal kanban | ● |
| Free calculators | 18 calculators, no account required: Cpk, Bayesian Cpk, control chart, Gage R&R, sigma level, FPY/RTY, sample size, OEE, takt time, Little’s law, EOQ, Pareto, kanban cards, FMEA RPN, cost of poor quality, A3 report, ISO 9001 audit checklist, ISO document creator | ● |
| Kaizen bursts | Bursts placed on the map, waste classification, lot sizing and SMED analysis per step, savings estimated from the step's own data | ● |
| Kaizen charters into a vandrebok | An approved burst is promoted to a chartered entry in the practitioner's vandrebok, carrying its charter and the map and burst it came from. Promoting the same burst twice returns the same entry | ● |
| Process analysis | Takt feasibility, line balance, queueing and bottleneck analysis against a value stream map, ordered so the cheapest question is asked first. Fits input distributions, so a simulated number can be traced to the data it came from. | ● |
Free tools available at svend.ai/tools — no account required.
VII — The Vandrebok
A practitioner's book of improvement work — a running record of what you have actually done. An entry arrives by promotion from a value stream map, carries the map and burst that produced it, and is measured against the map. It is an output of value stream mapping: current state, future state, waste register, review, then the book. What a vandrebok is, and how it differs from hoshin kanri →
| Capability | Specification | Status |
|---|---|---|
| Entries | Chartered improvement work: title, charter, class and type, year, and the value stream it belongs to | ● |
| Provenance | Which map and which kaizen burst produced the entry, recorded on the entry itself and used as its identity | ● |
| Action items | The work inside an entry, owned and dated | ● |
| Commitments | Who has committed to which entry, with over-commitment detection across a facilitator's whole load | ● |
| Templates | Start an entry from a standard improvement pattern | ● |
| Sites & people | Multi-site organization model, per-site access control, employee assignment to entries | ● |
| The book screen | Your own logbook: the work you did in SVEND, indexed by you, and readable across every project you have worked on. It is a log — it carries no tamper-evidence and makes no audit claim, which is what lets ordinary retention apply to it. In build for November 2026 | ○ |
Available from Pro. There is deliberately no savings estimator and no X-matrix — value is claimed against a measure on the map.
VIII — Platform
Multi-tenant architecture with tenant isolation enforced at the database level by Postgres row-level security, statistical results calibrated against externally certified reference values, and a public record of the result.
| Capability | Specification | Status |
|---|---|---|
| Managed datasets | Upload once into a managed library, clean and transform on a dedicated screen, then pull the finished dataset into any analysis. In build for November 2026 | ○ |
| Team collaboration | Shared work across a team — analysis sessions and improvement entries visible to the people who own them together. Deliberately held for November 2026: it is the feature most likely to be built as an activity feed nobody reads, and the record-and-provenance work this platform is built on gives it a better shape than that. In build for November 2026 | ○ |
| Audit trail | Tamper-evident hash chaining ships for gage studies, where every reading submission is recorded as an act and chained (section IV). It does not yet cover workbench analyses. The claim stays narrow on purpose: a chain shows the covered acts have not been altered since they were chained, and attests nothing before its first entry. In build for November 2026 for analytical entries | ○ |
| Data export | CSV, Excel and JSON export for any analysis. Printable report per analysis. 90-day retention after cancellation | ● |
| Security | Encryption at rest and in transit. Postgres row-level security for tenant isolation. Argon2 password hashing, TOTP multi-factor authentication, session management, sign-in rate limiting | ● |
| Multi-tenant | Multi-plant / multi-site model, per-site access control, tenant isolation enforced in Postgres by row-level security | ● |
| Calibration system | NIST Statistical Reference Datasets (14 certified sets, 15 significant digits) · R cross-check: control charts and capability against qcc 28/28, hypothesis tests and regression against base R 40/40 · 149 golden cases in the statistics library · defined tolerance bands | ● |
| Asset register | A record of each production machine, its run conditions, and the settings it holds. Available from Team. | ● |
| Fleet comparison | Empirical-Bayes partial pooling across sites: each site mean becomes a weighted combination of itself and the fleet mean, weighted by that site’s precision. A site with two runs is pulled close to the fleet; a site with thirty keeps its own answer. Available from Pro. | ● |
Team collaboration is the Team tier. Everything else on this page is Free or Pro. Deep customisation — your own domain namespace, your own branding, a bespoke deployment — is a conversation: see pricing.