Where Matflow is headed
The platform ships in research cycles, not marketing cycles. This page has three honest states — shipped, in progress and exploring — and deliberately no dates: priorities move with research feedback.
Three states, one rule: evidence
Every shipped item names the route or module behind it. Everything else is explicitly unfinished.
In the platform today
Usable now. Each card opens the feature or the documentation behind the claim.
Chain Data Synthesis → Evaluation → Prediction → Optimization with synthetic-data nodes and TEA in the loop on a visual node graph, and save a configuration as a reusable pipeline template (backend/routes/pipeline_routes.py; /pipeline).
TEA runs inside the optimization loop with scale-up costing rules feeding a net-value objective, so candidates are ranked on economics as well as performance (Studio TEA; /studio/tea).
A cross-validated ensemble returning global/local impact, permutation importance, PDP/ICE curves, conformal intervals from the same folds, and an extrapolation flag on every prediction (backend/models/prediction/generalized_predictor.py; /studio/prediction).
Structural-similarity virtual screening, six design families, and an active-learning loop with uncertainty sampling, query-by-committee and Bayesian acquisition (Studio Screening, DoE, Active Learning).
The page-aware assistant grounded in the docs, with approvals on writes and compute, plus tabular extraction from literature and instrument data (backend/routes/assistant_routes.py; /copilot).
Quality, ML-efficacy, dimensionality-reduction and anomaly reports on every generated batch — including distance-to-closest-record baseline protection and overfitting checks (backend/evaluators/main_evaluator.py).
Every registered model records task, intended use, limitations, metrics by split, calibration, baselines, artifact hash and a rollback path; a model without traceable metrics cannot be marked validated (backend/routes/model_registry_routes.py; Model Hub registers governed rows).
Datasets version with SHA-256 file hashes, benchmark runs ship downloadable reproducibility bundles, and saved datasets export as RO-Crate v1.1 ZIPs with provenance and engine cards (backend/core/provenance.py; backend/routes/rocrate_routes.py).
Organizations own projects and campaigns, members carry roles across a read/write/admin permission matrix, and verified orgs can publish a public profile that feeds the benchmark leaderboards (backend/routes/organization_routes.py; backend/core/rbac.py; /orgs/:slug).
An in-app notification center for run completion and project invites, opt-in email on run success/failure, and HMAC-signed webhooks with retries for external systems (backend/routes/notifications_routes.py; backend/core/webhooks.py).
Being built now
Actively worked areas; parts may already be live in the product.
More generative modes for Data Synthesis, additional ensemble variants and stacking candidates for prediction, and new multi-objective and Bayesian strategies for search.
DCR baseline-protection and overfitting checks already ship with evaluation. Membership-inference scores, TSTR/TRTS efficacy and optional differential privacy for shareable synthetic data are the next additions.
PDP/ICE and permutation importance are in the platform today; H-statistics for interaction strength and quantile regression for asymmetric uncertainty are being built.
Pipeline templates can be saved and reused now. Parameter sweeps across batch runs and scheduled re-optimization as new lab data arrives are in progress.
Run completion and invite events are live in-app, by opt-in email and via webhooks. Quota-level and security-event alerts, plus quiet hours, are planned.
Formula parsing, elemental properties and composition descriptors ship today (backend/core/cheminformatics.py). Pre-generation validation that warns on implausible formulations is being extended.
On the horizon
Directions we are considering — priorities shift with research feedback, so nothing here is scheduled.
Declare units per column; the platform would convert values, validate dimensional consistency and label every axis correctly. Today, unit-aware mapping exists only inside specific instrument and enrichment paths — not general columns.
English is the only shipped locale; the design system already uses logical layout properties so right-to-left layouts are supported structurally. Full locales are exploring, not scheduled.
Server-recorded 90-day uptime bars, incident timelines and maintenance notices. The /status page today probes live and keeps history in your browser only — there is no server-side historical record to publish yet.
The planning docs commit to earning new modules through customer evidence rather than shipping everything at once (planning/2026-09-replan). What gets built next is decided by observed use, not by this list.
How this page stays honest
Two rules keep the roadmap from becoming a wish list.
Help shape what comes next
Feedback from researchers working on real data decides the roadmap — not speculation. Tell us what the platform is missing.