Every API change, documented.

2026-05-23

v1.0.8.2

Simulation detail page render fix

Two surgical bugs caught during the prod smoke. (1) The detail page crashed on every render because it passed a plain JavaScript callback across the server-to-client component boundary; the boundary only accepts serializable values. (2) The timestamp on the page used a locale-aware formatter that produced different output server-side and client-side, triggering a hydration warning on every load. Both fixed. The page now renders cleanly for every simulation.

2026-05-23

v1.0.8.1

Geography and demographics endpoints serve the dashboard

Cookie-authenticated dashboard users on the sandbox tier hit a 422 on every geography search and demographics call because the feature flag default-on path only matched API-key sessions. Extended the check to also consult the session plan tier. Live-tier callers are still gated by the AUDIENCE_PICKER_ENABLED env var for the post-launch observation window.

2026-05-23

v1.0.8

Geography picker in the Composer

A new line above the methodology preview lets you refine the audience to specific ZIPs, cities, counties, states, or regions. Multi-select within and across entity types. Live reachable-agent count as you select. A calm low-coverage warning chip when your selection is too narrow for the requested sample. Defaults to U.S. adults; refining is optional.

2026-05-23

v1.0.8

Post-run demographic exploration on the simulation detail page

After a run completes, the detail page surfaces twelve filter chips for income, age, parental status, education, household composition, and occupation. A verdict-by-demographic cross-tab with a switchable dimension. Three population-shape histograms. A side-by-side cohort compare. A per-agent drawer that opens on click with the agent's full profile and response, gated by an inline privacy footer. Sensitive dimensions (race, religion, political lean) are excluded from the default filter chips per ADR-034.

2026-05-23

v1.0.8

Detail page hierarchy puts your question first

The simulation detail page now opens with your literal question in serif italic as the largest text on the page. The matched methodology and the calibration match score are subordinated to a smaller row below it. Easier to scan when you have a history of runs against the same methodology.

2026-05-23

v1.0.8

Data-source language scrubbed across customer surfaces

Every customer-facing page now describes the audience as 'demographically grounded' rather than naming the upstream data sources. The dimensions we audit are listed openly on the methodology page; the full underlying data provenance is available under NDA via security@simulix.com. A CI guard blocks the older language from re-appearing.

2026-05-23

v1.0.8

Transcript no longer chops the start of agent responses

A regex sentence-splitter was eating leading prose whenever an agent response mentioned a URL (e.g. 'I haven't heard of madcha.com before, so...' came out as 'com before, so...'). New excerpt function handles URLs, decimals, abbreviations, ellipses, and em-dashes correctly. Thirteen unit tests cover the regression.

2026-05-23

v1.0.8

URL-context interest capture

When your question contains a URL the synthesized population has no prior knowledge of (e.g. asking about a new brand), the agents now honestly answer 'I haven't heard of that'. The real fix — fetching the URL, summarizing, and giving the agents a brand brief — is on the v1.0.9 roadmap (ADR-035). In the meantime, a /v1/feedback/url-context-interest endpoint captures which brands customers are trying to ask about.

2026-05-22

v1.0.7c

Your question is now what the agents see

Through v1.0.6, the auto-router treated your natural-language question as a signal for picking the right calibrated workflow, then asked the agents the workflow's own published question instead of yours. v1.0.7c flips that: the router still picks the closest methodology (its population, segments, sample profile, math), but the agents now see your literal question. A new calibration-boundary banner discloses which of three tiers the verdict belongs to (calibrated_match, methodology_match, no_match) and explains what the published accuracy figure applies to. ADR-029.

2026-05-21

v1.0.6

v1.0.6 audit + Path A routing-explainer modal

Full pre-launch audit cycle. New pre-mint routing-confirm modal lets you review which calibrated workflow the auto-router picked and the match score before the simulation queues. New post-run banner discloses calibration boundary. Operator audit feed at /v1/admin/auto-publish-feed.

2026-05-19

v1.0.0-rc2

/v1/status hardens against Redis latency

Status endpoint now wraps Redis cache reads and writes in a 2-second hard timeout matching the per-probe pattern. Eliminates the failure mode where a slow Redis cold-start hung the whole endpoint before the parallel probes started.

2026-05-19

v1.0.0-rc1

Customer training surface and four /v1/training/* endpoints

Upload a calibration CSV, run training synchronously, and read back preview results. Console at /dashboard/training; programmatic at /v1/training/uploads, /v1/training/:id/start, /v1/training/:id, /v1/training/:id/results. Every result carries the preview_only flag per the ADR-028 mock-label discipline.

2026-05-19

v1.0.0-rc1

Dashboard: single-sim composer, workflow catalog, side-by-side compare

Three new console surfaces. A single-simulation composer with hybrid Ask/Pick interaction, a filterable workflow catalog, and a side-by-side compare view for two runs.

2026-05-19

v1.0.0

Initial public release

Census-grounded simulation platform. Free tier, paid, enterprise. The public benchmark page is canonical and live. The REST API is the integration surface today; SDKs (Python, TypeScript, Go) publish during the launch window.