v0.1.0-alpha.492

Released 2026-10-19.

Features

Column-Level Masking

• A meta-domain view, sensitive_coverage, lists every sensitive column of the org (any tag with the Sensitive data option, pii included), one row each, read through the same statement that decides a Test (fake) environment's uncovered columns (provisa.security.sensitive.sensitive_assignments): its tags; has_fake and has_synthetic_rule; the fake as declared, its kind, that it is deterministic, whether it is stable across engines and whether it is computed from the real value; the synthetic rule as declared; and a status of covered, no_fake or inadequate with the reason.
  • sensitive_coverage meta view and fake adequacy checks
  • fake_adequacy_checks is classified, inventoried and in schema.sql

Demo Provisioning

• A configuration may declare a replica's rows instead of landing them from its source: replica_rows//..tsv beside its config.yaml, its first line the table's columns in their order, tab separated, each line after one row (\N is NULL; \t, \n, \r and \ a tab, newline, carriage return and backslash; an integer, floating or boolean column's cells read as that type, every other as its text).
  • A config's declared replica rows land as its replica; its source is never read
  • A config's declared rows land in its model build, beside the config a test server serves
  • Declared rows land before the schema build publishes its replica routes
  • Hasuramigration and provisioning bundles build

  MCP & AI Integration

  • The MCP server offers five statistics tools over a dataset the caller defines in SQL: describe (per column count, missing, distinct, mean, sd, quantiles and top values), correlate (pearson or spearman matrix and n), regress (linear: coefficients, standard errors, p-values, R²; logistic: the same with pseudo R² and AUC), compare (Welch's t-test for two groups, one-way ANOVA for more, with Cohen's d or eta squared) and forecast (exponential smoothing, ETS, with prediction intervals).
    • describe, correlate, regress, compare and forecast over governed SQL
    • stats.forecast casts its ETS fit to its results, which pyright refused as a union

  Security

  • Central redaction: one list of rules is applied to text before it leaves Provisa.
    • Central redaction of spans, logs, audit text and client errors
    • Redaction rules -- one name column width, monospace values with plain labels
    • Redaction -- a bearer token is 16+ token characters; names resolve through the backend
  • Judgments under review (docs/arch/judgment-review-design.md, phases 1-3).
    • Judgments under review -- why, bases, findings, the review and posture
    • Judgment review -- nullable JSON stores SQL NULL; exports compare without the server's stamps
    • A profiler member never run gives no evidence instead of failing the seed
  • A named demo declares its secrets in secrets.yaml beside its config.yaml (secrets: a mapping of each name to its value), carried in its bundle with the rest of its directory.
    • A demo seeds its secrets; testdata reads prod as Provisa's own secret

  Source Connectors

  • Package registries reach a proxy such as Artifactory and the Git servers GitLab, Bitbucket Cloud and Bitbucket Data Center as well as GitHub.
    • Package registries reach Artifactory, GitLab and Bitbucket, behind a private CA

  Changes

  bench

  • Governance knobs in the benchmark contract, set up in the deployment for each run

  bundles

  • Concurrent fetches of one connector bundle download it once and each finds it whole

  codesource

  • A host's hello names its protocol and code hash, refused by name otherwise
  • A remote host over mutual TLS, paired once, dialled by each Provisa process
  • A Python source's rows travel as Arrow IPC batches, protocol 2
  • A Node source's rows travel as Arrow IPC batches, protocol 2

  compiler

  • A lowered reference names its own source's table, not another of the same name
  • A draft is refused under the catalog-physical name its served table answers to

  demo

  • Cards for provisioning and retail, from their real answers
  • Agriculture -- the notebook measures how much of the portfolio's losses the weather explains
  • Agriculture -- the 2012-like scenario prices the portfolio's latest book, and each empty step says which
  • Agriculture -- the scenario is the data's driest season, priced at each county's latest premium; the answer is captured before the notebook runs
  • Agriculture -- a pull of every county season (320 rows) is not cut at 100
  • Agriculture -- the card is the notebook's plot, and the tour says what the notebook measured

  demo bundle

  • A build's servers are given the stores' addresses, not the build's whole environment

  demo-stores

  • A demo's stores start at its first open, its data in one shared instance of each store type

  demos

  • opioids' public dashboard is a view that reads no allocation
  • Publish provisioning and hasuramigration
  • Publish insurance
  • Start here, the standard demo's objects with every Deep Dive
  • Start here's links to the standard demo's files are tracked
  • Publish Start here, listed first
  • Live tour steps -- profile runs reach every sandbox, the REST step waits for its docs, a table step opens the panel it describes
  • Starthere ships its example query; the card images site/demos.html names are committed
  • The data room, a retailer's own systems shown to each bidder with no exports
  • Testdata over three environments; a demo's Test (synthetic) environment is generated once and carried to every sandbox

  docs

  • Each locale's search reads only that locale's pages

  events

  • A view over a view not built yet is not probed; it awaits that input's first build

  install

  • A rerun of first launch keeps the node's provisa.env settings it did not generate

  mcp

  • The forecast's ETS fit is typed as its results

  mv

  • A view is materialized in its own org's store schema, not the deployment org's

  pipeline

  • A statement waits for no server of a source read from its landed replicas alone

  push

  • A push listener stops with its runtime and its org's next build wires it again
  • A retired runtime's listeners are stopped through a hook the app registers

  regions tier

  • Its auth block is the sandbox tier's

  runtime

  • A background job walks only the org runtimes that are built
  • A runtime being built is resolved only by its own build

  sandbox

  • A bundle ready after the tier started has its template built
  • A demo sandbox keeps its template's code uploads

  sandboxes

  • Opening a sandbox counts places under the control plane's advisory lock, which holds across request threads
  • Two visitors opening one demo while its template is replaced both wait their turn
  • The tier builds its demos' templates a few at a time, an opened demo first
  • A demo's readiness is asked without opening it
  • A restarted sandbox tier serves its unchanged templates as they were built

  security

  • A demo sandbox visitor's token naming a role it does not hold is refused and recorded, not suspended

  sources

  • An engine whose driver library is absent is unavailable, and the Sources page still answers

  subscriptions

  • A sandbox's live subscription reaches the browser frame by frame

  tour

  • An image step's popover is placed again once its picture loads, and the picture fits the window

  ui

  • The splash halo centres on the P's loop, not the logo's box