What we shipped

The changelog.

Every entry below cites a model, a feature, a runbook or a spec — never marketing. Sorted by date, grouped by month, oldest at the bottom.

Latest

  1. Security

    Credit models now refuse synthetic training labels, end to end.

    A fail-closed provenance gate blocks any credit model trained on synthetic labels from being promoted or served to a borrower; a real forward-outcome label substrate replaces them. Nothing borrower-facing activates until the compliance ceremony completes.

  2. Security

    Account Aggregator data is now purpose-scoped by construction.

    AA-sourced data carries consent purpose codes, and a fail-closed gate keeps it out of credit decisioning unless the borrower separately consents to that purpose. The machinery ships dormant; consumer-visible behaviour is unchanged until activation.

  3. Feature

    Data Hub uploads suggest their own column mappings.

    Uploads now auto-suggest schema mappings with reranked alternates, extract more from messy statements, and pseudonymize per-account identifiers with dedicated keys. Mapping mistakes stopped looping; label integrity is enforced at ingestion.

  4. Security

    Every external response path now passes one egress gate.

    B2B responses and consumer streaming replies are buffered through a single gate that enforces grounding checks and PII redaction before any byte leaves the platform.

  5. Feature

    First live-capable market-data feeds: ECB forex + RBI rates.

    The forex reference-rate and RBI policy-rate feeds are live-capable with their license bases verified and documented. Feed errors now raise loudly instead of silently serving cached values as fresh.

  6. Model

    The model catalog now declares an honest served_mode per model.

    Each registered model states how it is actually served — trained artifact, heuristic, or mock — and product surfaces can only claim what the registry backs. The transaction categorizer is the first fully trained-and-served model under this regime.

Older

June 2026

  1. Feature

    OpenTelemetry tracing, review-SLA sweeps, and a crisis-event consumer.

    A platform-wide tracing spine landed alongside a human-review SLA breach sweep and a dedicated crisis-event consumer, tightening the loop between detection and human response.

May 2026

  1. Feature

    Partner usage and cost statements landed in portal and admin.

    Partners can now view a per-month, per-product usage and cost statement, with a period selector and CSV export. The statement is informational, not a tax invoice; GST-invoice issuance stays gated on separate legal and finance preconditions.

  2. Security

    KYC document persistence: storage, corrections, forgery, audit.

    Uploaded KYC documents now persist through a UIDAI masking gate into AES-256-GCM encrypted object storage, with an immutable corrections overlay, forgery-check records, and an audit-event trail. All four tables are covered by the DPDP erasure cascade.

  3. Model

    TimesFM quantile forecasting shipped end-to-end (dormant).

    An eleven-PR track shipped p10/p50/p90 quantile forecasts on the Apache-2.0 TimesFM foundation model: trainers, a consumer route behind a consent gate with active erasure, public market-data forecasts, and mobile cards. Production cut-over stays operator-gated.

  4. Docs

    Hreflang infrastructure shipped (English-only, 8-locale scaffold).

    Sitemap and per-route metadata now emit language alternates once two or more locales ship. Today only English is live; adding another locale fans the alternates out across every URL automatically.

  5. Security

    User-state change events get HMAC + boot-time drift guards.

    User-state change events are now HMAC-signed end-to-end, and boot-time drift guards lock the materializer's invariants so a regression fails the deploy rather than shipping silently.

  6. Security

    Outbox relay leader-election + view-lag histogram + replay CLI shipped.

    Outbox relay now elects a single leader and ships an operator replay tool; observability dashboards gained a view-lag metric.

  7. Feature

    Anthropic-shape /api/v1/messages on FinAi Qwen + vLLM (dormant).

    An Anthropic-shape Messages API over our self-hosted Qwen + vLLM cluster shipped dormant behind a feature flag. Signed DPIA, infrastructure capacity, and SDK regeneration remain independent operator gates before activation.

  8. Feature

    Anthropic-shape MessagesRequest v2 + tool_use shipped dormant.

    The Anthropic-shape tool-use surface (tools and tool-choice) is accepted by the validator with strict-schema constrained decoding behind a flag. Shipped dormant — no consumer flipped yet.

  9. Security

    Security-event normalization follow-up bundle landed.

    A follow-up bundle for the security-event normalization pipeline landed: a severity hotfix, hardening, event wire-up, and a four-expert review reconciled before merge.