Changelog
A short record of user-visible changes. Historical entries describe what shipped at the time; the newest entry and product matrix define the currently supported surface.
September 2, 2026
- Rebased Crosstabs around one permanent product boundary: always free, no account, and browser-local files and projects.
- Removed the hosted AI, cloud-project, collaboration, enterprise, billing, survey-connector, customer-Postgres, and durable-job surfaces.
- Prepared the unpublished Python 1.2.0 candidate without the legacy HTTP/SSE hosting mode or cloud deployment descriptors; both candidate MCP commands are local stdio only.
- Narrowed the local project MCP to 21 deterministic tools and the anonymous public MCP to five read-only aggregate-and-evidence tools.
- Added fixed-shape error intake, coarse aggregate product signals, an explicit product matrix, and a release boundary that rejects removed hosted surfaces.
August 29, 2026
- Outcome Studio now fails closed when a weight is selected; weighted logistic fitting and holdout inference are unavailable pending a typed, design-aware policy.
August 21, 2026
- Removed the expiring image-size advisory exception: the unused generic parser is now absent from the installed dependency tree, replaced by an integrity-pinned fail-closed package, and the release audit requires zero vulnerabilities.
- Added Research Execution Contract v1 so saved row sets, banners, filters, weights, category order, missing-data rules, and test profiles execute identically or fail explicitly across workspace, worker, wave, export, AI, and MCP paths.
- Added canonical SHA-256 input and result evidence, per-setting diagnostics, cross-surface conformance goldens, and schema-10 migration for the strict unsupported-setting policy.
- Made table and export methodology notes describe the confidence tiers, correction, minimum base, and effective-base rule that actually ran.
- Added daily, privacy-safe activation cohort evidence that waits for a full 24-hour maturity window and fails closed instead of treating missing traffic as zero conversion.
- Added clean-revision LibreOffice Writer and Impress roundtrip verification on every main release, with complete editable PPTX/DOCX proof artifacts retained for 90 days; Microsoft Office compatibility remains explicitly unclaimed.
- Added Custom Tables-style row/banner composition, nets, subtotals, suppression, base policies, category order, saved tests, governed transformations, recipe syntax, and reversible tracker schema repair.
- Added reference-tested Taylor and replicate-weight survey inference, design diagnostics, effective bases/design effects, Rao–Scott and adjusted-Wald association, Holm/BH correction, and seeded fixed-margin Monte Carlo inference.
- Added a versioned report story with page/slide outline, approvals, commentary, branding, refresh conflicts, stable manual regions, and story-ordered native PPTX/DOCX generation receipts.
- Added a fail-closed Clerk/Postgres review service with optimistic concurrency, role-governed comments/tasks/approvals, and a row-free authenticated reviewer interface.
- Repaired independent left-rail scrolling and added roving keyboard navigation, deterministic focus, 320-pixel reflow, and 200%-zoom-equivalent browser verification.
August 12, 2026
- Published a reciprocal Simplified Chinese survey-weighting and Raking guide with primary-source references, LLM evidence bindings, and an explicit non-calculator, non-design-inference boundary.
- Published a dedicated survey-weighting and raking guide with weighted-table formulas, a worked example, primary references, LLM evidence bindings, and explicit complex-survey inference limits.
- Qualified Outcome Studio with nine deterministic real-use-case datasets instead of relying on the employee-engagement sample.
- Added true CSV-upload browser journeys for churn, weighted NPS, wave tracking, concept tests, care outcomes, no-signal and rare-event controls, identifier leakage, and a 20,000-row workload.
- Made successful model exclusions visible through a Data qualification audit with exact used/excluded counts and durable reasons.
- Allowed repeated numeric survey weights to remain selectable even when their low cardinality makes them categorical table dimensions.
- Published the Outcome Studio method, measured domain matrix, workload boundary, and unsupported claims.
- Added an exact, expiring CI gate and public disclosure for two linked unpatched image-size advisories in the browser-only Office export dependency path.
August 11, 2026
- Added Outcome Studio as a first-class guided and specialist workspace stage.
- Added weighted, regularized logistic key-driver models with deterministic 80/20 holdout validation, AUC, accuracy, Brier score, and calibration evidence.
- Added observed driver segments, bounded exploratory cluster diagnostics, modeled scenario scoring, and a synthetic cohort sandbox that cannot be mistaken for respondent evidence.
- Added evidence-linked recommended investigations with explicit human approve, reject, and undo review history.
- Added portable Outcome Studio project state, schema migration, local worker execution, and a published 40-term/1,500-row cluster-diagnostic workload envelope.
August 10, 2026
- Added checksummed, versioned .ctproj export/import with migration preview, non-overwriting imports, and a definition-only manifest option.
- Removed the arbitrary eight-project eviction rule; browser quota failures are reported instead of silently deleting older work.
- Published the complete local/cloud/AI boundary, supported formats, storage model, offline behavior, and unsupported cases.
- Added guided and specialist workspace modes, persistent specialist controls, research-default visibility, and an independently scrolling searchable left rail.
- Added hash-scoped, evidence-bound AI change plans with dependency-safe partial approval, reject-all, atomic apply, and conflict-safe undo.
- Added theme-level open-end review with approve/reject decisions, examples, approved-only variable materialization, and exact safe undo.
- Added selective editable PPTX/DOCX refresh with stable anchors and preservation evidence for supported manual edits.
- Published the multidimensional capacity envelope and adversarial cardinality stress evidence instead of an unlimited-row claim.
- Consolidated AI analysis on the production gateway with provider failover and removed the browser-local model path.
- Added exact aggregate-context disclosure before every AI request.
- Added evidence links from AI responses to supporting crosstab cells.
- Added read-only AI activity history and undo for the latest exchange.
- Tightened the browser content-security policy after removing the local model runtime.
August 1, 2026
- Published support@crosstabs.com as the canonical support and correction contact address, without adding a response-time, availability, or SLA claim.
- Replaced inaccessible public-GitHub claims with truthful PyPI source-distribution and private-development-repository wording.
- Replaced the anonymous methods-team review claim with an explicit named-review-pending state, a correction policy, formulas, assumptions, and primary references.
- Added a live no-cache application and known-answer statistics health check, a status surface, structured health logs, and production Speed Insights.
- Corrected directional Somers' d labels and formulas against the published SciPy reference example.
- Replaced unbounded Fisher factorial/enumeration work with a bounded log-recurrence implementation, safe-integer input rules, explicit API unavailability states, and an O(1) odds-ratio path.
- Corrected equal-probability central-mode handling in the two-sided Fisher–Irwin p-value and added zero-cell, skewed-margin, and large-count SciPy parity cases.
- Published a machine-readable SciPy parity manifest and an explicitly scoped statistical validation page.
- Added a machine-readable trust-release contract that reports missing external evidence in previews and fails closed for an explicit non-main or unready production build.
July 31, 2026
- Added a reload-safe local project library with homepage re-entry, explicit naming, and IndexedDB migration.
- Turned the workspace URL into a real start-or-continue entry instead of redirecting empty visits.
- Added view-only aggregate snapshot links and local snapshot-file fallback.
- Kept source rows and filenames out of snapshot payloads, with explicit bearer-link disclosure and opt-in context.
- Added citation and export actions, recipient handoffs into the workspace or calculator, and a dedicated Share Crosstabs guide.
- Documented the sharing boundary in the sitemap, LLM discovery routes, README, and analytics architecture.
- Added the Crosstab Atlas with worked examples, provenance, and deterministic CSV/JSON downloads.
- Improved mobile workspace controls, table navigation, result summaries, and export accuracy.
- Hardened statistical p-values, worker freshness, request-body limits, CI coverage, and noindex boundaries.
May 2, 2026
- Made optional AI summaries available from aggregate crosstab context.
- Made AI analysis explicitly opt-in.
- Removed unsupported marketing claims and fake preview analysis.
- Added weighted-table caveats and invalid-weight exclusion.
- Replaced vulnerable spreadsheet and Mermaid dependencies.
- Added privacy, security, methods, sample, and changelog pages.