Platform release notes

What shipped, month by month

Eleven monthly releases from October 2025 to August 2026. Releases are identified by the month they shipped — no version numbers have been assigned retrospectively. Notes are grouped by what the capability does, not by which component delivered it.

Proven

Running on client data at production scale.

Built

Running in the platform today.

Scoped

Designed and sequenced — not represented as available.

2026.08 · August 2026

Measures derived, not authored, and the governance rails land

The semantic layer became the thing every surface and agent reads from — and it stopped being hand-authored. Measures are now derived from the ERP dictionary and validated against imported statement arithmetic.

Mantrix SemanticBuilt

Measures derived from the ERP dictionary instead of authored. What counts as revenue, direct cost or overhead comes from the customer's own configuration and data dictionary, not from a definition someone typed.

Statement arithmetic imported from XBRL, with SAP® account classification mapped onto the statement model and the mapping measured rather than asserted.

Typed intermediate representation with a compiled lane and declared joins. A model may propose a typed plan — the intermediate representation still decides what executes. A vocabulary fix in this work removed a certified wrong answer, which is the point of having the layer at all.

The feedback loop closes. 1,173 generated questions became answerable, and subject-matter-expert answers now apply back to the model rather than sitting in a document.

AXIS.AIBuilt

Natural language to SQL grounded in the semantic layer, widened across the platform, with a governed query lane.

Cross-source answers spanning the analytics store and the relational store in one question, with cross-source lineage shown in the interface — which exposed four real defects in the process, all closed.

Governance railsBuilt

Database-backed hash-chained agent ledger — the tamper-evident record every consequential agent action writes to.

Agent runs report what they actually spent, attributed per run rather than estimated in aggregate.

Maturity-ladder registry, so autonomy level is recorded state per unit rather than a configuration flag. Guardian agents run on cadence against a platform events and attribution spine.

Honesty pass across the product surface

A deliberate sweep to make the surface tell the truth about its own data:

  • The operations console reworked to report actual state
  • Per-drilldown enablement, not just per-module — an individual drilldown with no data behind it is disabled on its own
  • The agent fleet reflects the gated module set rather than advertising agents for modules that are off
  • Superseded surfaces removed rather than left in place

Design system

One token layer across the platform — primitives, tokens and lint enforcement, a shared table system, standardised page headers, fonts and sidebar across every screen, and a centralised tile grid instead of per-file layout.

2026.07 · July 2026

Margin at document and product grain, and determinism becomes a gate

Margin reached the two grains project accounting normally destroys: the individual settlement document, and the finished product. Reproducibility stopped being a property and became a build gate.

Document MarginNewProven

Every settlement document explained to source. The settlement equation tied to the posted amount, with the recognition method resolved from the customer's own configuration rather than inferred from the account.

Trace v2 — settlement equation, method dispatch, the work-in-process bridge, and drills to the purchase order, activity and customer behind each figure.

Product MarginNewProven

True margin per finished product across programs, inception to date. Recognised lot results attributed by billed-revenue share — with the attribution basis labelled permanently at the point of display rather than buried in a footnote.

Golden-verified, 39 of 39. Every measure reproduces the verified system value to the dollar.

Program Margin CockpitProven

Recognised-versus-actual on one screen, with a reconciliation badge and actual billed and actual cost shown alongside the recognised figures.

Three-way reconciliation across recognition math, a general-ledger lens and the configuration surface, with full ledger control by document type and transaction.

Genuine-or-nothing. A measure with no posted data renders as an explicit no-activity state, never as zero.

Server-side ranking, top-N and search, so a large portfolio stays responsive.

Reproducibility

Double-run determinism gate. The full pipeline runs twice on the same input under isolated output paths, and the build fails if the two runs disagree. Determinism is now asserted by test rather than claimed.

Mantrix Agents — evaluation frameworkBuilt

A complete evaluation system landed in three phases:

  • Scorers — deterministic and model-as-judge scorers, correctness, helpfulness and faithfulness judges, and a trajectory scorer covering sequence, efficiency and tool errors
  • Governed scorers — no-PII-leak, routing-held and policy-respected, with traces enriched by events, audit and routing data
  • Regression control — baselines, regression diff, multi-run pass@k, a flakiness metric, cost aggregates and a CLI with CI exit codes
  • Online evaluation — scoring sampled production traffic, bridged into a reinforcement-learning replay buffer

Mantrix SemanticBuilt

SAP conventions expanded from 16 rules to 34, with eight new executable helpers.

Cross-vendor proof. The same finance dataset represented against a second ERP and a third vendor's model — evidence that the canonical model is genuinely vendor-neutral rather than SAP-shaped.

2026.06 · June 2026

Program margin, computed end to end

The month the margin story became computable end to end. The Program Margin Cockpit went live against real extracts, and two new modules opened the underlying catalog to inspection.

Program Margin CockpitNewProven

Recognition recomputed in the deterministic engine. Recognised revenue and cost of sales resolved per lot from the method the customer's own configuration assigns, set against actual billing and work-breakdown spend read from the customer's Universal Journal (ACDOCA) extract.

A reduction layer that reduces the journal in SQL before it reaches the application, which is what makes the cockpit responsive at real extract volume.

Project LensNewProven

Single-work-breakdown forensic deep dive, built in three semantic layers: portfolio semantics, a reconciliation cube on real data, and a truth score combining reconciliation, recognition, period and status.

Multi-level drill from portfolio to project to work-breakdown element, with chart-click drill-through on overview, margin, plan-versus-actual, cost flow and process mining.

Honest data-source badges per tile. Each tile states where its figures come from, and tiles are grouped by what data is actually available — a tile with no underlying data says so rather than rendering zero.

GL ExplorerNewProven

Selection builder over the full journal joined to master data, with a cube layer, KPI cards and a query summary. Line-item grid virtualised for large result sets.

SQL WorkbenchNewProven

Read-only SQL editor over the full catalog, with autocomplete and query history — so a figure in any module can be checked independently by whoever needs to check it.

Engine and ingestion

Numbers tie-out harness — an independent raw-versus-platform reconciliation that verifies published figures against the source extract rather than against the platform's own intermediate state.

Join-coverage reporting for canonical enrichment, so an unmatched key is reported rather than silently producing a blank field.

Module gating with failure isolation. A module whose upstream data has not been produced returns an explicit not-ready state instead of failing the page.

2026.05 · May 2026

Forensics gets a face, and ExcelHub gets an audit trail

T-Account Forensics arrived as a working investigation surface backed by real endpoints rather than mock data, the engine gained enterprise authentication, and ExcelHub became defensible.

T-Account ForensicsNewBuilt

The CFO financial X-ray. Measure-to-T-account bridge with chained offsets, offset DNA and an audit workbench across the full ledger.

19 endpoints backed by the analytical engine, and every view moved from mock data to live fetches — the point at which the module became an investigation tool rather than a demonstration.

Evidence packs. Open any case to its document header and lines, with the five-part who / when / where narrative populated from real values per case rather than templated text.

PROJECT-LEDGER.AINewBuilt

Work-breakdown resolution, trial balance and project profit and loss on the CORE.AI surface.

Deployment and access

Enterprise single sign-on with per-user role-based access control, plus pre-install documentation for federated identity integration.

HTTPS termination with automatic certificate handling, a host runbook and a cloud-init script — a documented path from bare host to running engine.

Preflight checks, so a misconfigured environment fails at setup rather than mid-pipeline.

ExcelHub — governance and evidenceBuilt

Hash-chained audit trail with per-cell optimistic locking. Concurrent edits cannot silently overwrite each other, and every change is chained.

Audit chain export to CSV and JSON as compliance evidence, extended to cover rule runs and error resolution.

Customer-specific field registry — detection with a tenant-scoped index and a searchable drawer showing every occurrence.

Dashboards as ordered groups of saved pivot specifications, with per-sheet validation history and workspace export.

2026.04 · April 2026

The deterministic engine arrives

The most significant month in the platform record. The deterministic computation engine for SAP S/4HANA® data entered the platform, ExcelHub grew from a file processor into a full workspace, and the agent runtime reached SDK parity across five languages.

FORENSX.AI — the deterministic coreNew

The deterministic computation engine for SAP S/4HANA® data enters the platform. It requires no connection to the SAP system at all: it reads standard SAP table extracts that the customer produces themselves, delivered to a file drop — no RFC, no agent in the SAP landscape, no transport, no custom ABAP, and no login to the SAP system. The customer produces the extract on their own schedule. This is the forensics path; platform connectors reach SAP separately, through SAP-permitted interfaces.

Nine-phase pipeline, from column mapping through canonical journal construction, feature engineering, T-account reconstruction, recognition and margin, the rule catalog, statistical forensics, seeded machine learning, and case and evidence assembly.

Detection rules across five families — journal entries, project systems, procure-to-pay, user activity and machine-learning models.

Specifications published alongside the engine, with an extract questionnaire carrying sensible defaults — so the data request is a contract rather than an email thread.

ExcelHub — spreadsheet workspaceBuilt

A six-part build turned spreadsheet handling into a governed workspace:

  • Multi-file workspace with a spreadsheet grid and data grid
  • Visual join canvas compiling to an analytical query engine
  • Natural-language bridge into AXIS, so a workspace question is answered against the workspace
  • Transform intermediate representation, so transformations are declared and re-runnable rather than click-recorded
  • Multi-format ingest with encoding and locale detection — the failure mode where a European CSV silently mis-parses

Mantrix AgentsBuilt

Five-language SDK parity. Go, Java and C#/.NET joined Python and TypeScript, with full feature parity and integration tests against real APIs.

Enterprise primitives — transactions, triggers, approvals and an explicit reasoning boundary — plus a compliance-grade observability pipeline.

Production memory system across vector stores, intelligent memory, graph memory and tiered retrieval, and a tool library of 41 tools across 10 categories.

DOC.AIBuilt

Purchase-order extraction pipeline, with a reference-template system for few-shot extraction and structured table rendering.

Note. FORENSX.AI entered the platform this month and reached Proven status on client extracts in a later release.

2026.03 · March 2026

AXIS gains an analytics spine, and test coverage goes deep

AXIS.AI moved from a query interface to a governed analytics path with its own warehouse layer and a feedback loop. Alongside it, a sustained coverage push added roughly 500 tests.

AXIS.AIBuilt

Analytics layer with models, tests, macros and documentation, and a feedback-weighted learning loop so corrections improve subsequent answers rather than being discarded.

Conversation context. Follow-up questions resolve against the prior turn, so “and the quarter before that?” works without restating the question.

Answer quality — result summaries, refined explanations, and chart intelligence that selects a chart form suited to the result rather than defaulting to one.

PROJECTS.AINewProven

Portfolio, program and project control on the CORE.AI surface.

Mantrix AgentsBuilt

705 real-API integration tests. Not mocks — the runtime is exercised against live provider APIs.

Production robustness hardening across six workstreams, plus stateful agents and a durable-execution module for long-running work.

Platform quality

Roughly 500 tests added across authentication, request logging, database permissions and access control, conversation and chat routes, cross-database query validation, user profiles, API models, drill-path detection and column formatting.

2026.02 · February 2026

Payables, deployment infrastructure, and compliance groundwork

A month of consolidation. AP.AI arrived as a working surface, the platform got its production deployment path, and the agent runtime gained the memory and planning machinery it needed to run unattended work.

AP.AINewBuilt

Five operational tiles, with an embedded document viewer and a workflow stepper — the invoice, the exception and the resolution path in one place rather than three systems.

Line-item matching engine, with inline drilldowns from any tile to the lines behind it.

Mantrix AgentsBuilt

Provider-agnostic programmatic tool calling. Model-generated code calls tools inside a sandboxed subprocess, and the same mechanism works across providers rather than depending on one vendor's implementation.

Memory consolidation and persistent graph stores, so agent memory survives restarts and can be inspected.

Strategy selection surfaced in the interface. Choosing how agents collaborate became an operator decision rather than a code change — with strategy pickers, per-strategy visualisations, side-by-side comparison and inline documentation.

ConnectivityBuilt

Supply-chain planning connector and universal webhook support, bulk API support for large batch operations, and the SAP connector path extended beyond service interfaces — read and write through SAP-permitted interfaces, with no core modifications.

Field-mapping transparency completed across every connector.

Platform infrastructure

Repeatable cloud deployment with scale-up, scale-down and teardown workflows — the platform gained a documented path to a running environment.

2026.01 · January 2026

SAP goes deep, and the connector estate more than doubles

The month the connector fabric stopped being a finance-systems integration and became an enterprise one. Two new fabrics — communications and industrial IoT — came online alongside it.

Mantrix WeaveBuilt

ERP support expanded from 18 systems to 41. Connection and setup pages shipped alongside, so a newly supported system is configurable on the day it lands rather than by ticket.

SAP read and write through permitted interfaces. The SAP connector path uses SAP-permitted integration interfaces and makes no core modifications. Write-back stays approval-gated — nothing posts without a named human approver.

Field-mapping transparency. Every connector exposes how a source field reaches the canonical model. Where a mapping is applied, it is inspectable.

Authentication reliability — a shared retry path across all OAuth connectors with automatic token refresh, closing the failure mode where a long-running sync dies on an expired token.

Reliable event delivery through a queue-backed dispatcher, so downstream systems receive connector events rather than best-effort callbacks.

CommBridgeNewBuilt

One API across communication channels, with service implementations, SDKs and a voice connector in the first release.

IOTBridgeNewBuilt

Operational-technology connector service for industrial protocol support, wired through to the device management dashboard, with device decoders for tracking hardware.

HeadroomBuilt

Adaptive compression sizing and syntax-aware code compression — compression that understands what it is compressing rather than treating every payload as text.

Bounded memory throughout. Vector and graph storage moved to bounded, persistent backends, so long-running sessions cannot grow without limit.

2025.12 · December 2025

The agent runtime, and connectors go multi-ERP

Two foundations landed that everything later depends on: the runtime every agent in the platform now runs on, and the connector fabric that reaches business systems beyond SAP.

Mantrix AgentsNewBuilt

The single runtime for every agent in the platform. All agent creation flows through it; there is no second path.

17 context-sharing strategies, with native-code acceleration on the hot paths, and 18 handoff strategies for routing and workflow patterns — including tiered-access handoff, so an agent handing work to another agent cannot widen the data scope in the process.

Multi-language SDKs with native providers, cross-language bindings, audit streaming and a published benchmark suite.

Mantrix WeaveBuilt

Accounting and finance connectors, plus a generic OData connector for custom and in-house systems — coverage without waiting for a bespoke connector.

Visual field mapping. Map source fields to the canonical model in an interface rather than a configuration file, with the mapping stored as configuration.

Per-organisation encryption keys and durable persistence replacing in-memory stores — the point at which the fabric became operable rather than demonstrable.

2025.11 · November 2025

Document intelligence and the analyst's desk

The month the platform stopped being only dashboards. Document extraction, spreadsheet processing and correspondence handling all landed, and the monitoring surface got its first templates.

DOC.AINewBuilt

Template-aware extraction. Point it at a document type once; subsequent documents of that type extract against the same template, with columns derived dynamically rather than fixed in advance.

Extraction results are reviewable and correctable by the person who owns the document, not only by an engineer.

ExcelHubNewBuilt

Template persistence. The spreadsheet a controller already uses becomes a repeatable pipeline instead of a monthly re-keying exercise.

PULSEBuilt

Monitoring templates for inventory and margin, so a new tenant starts with working detection rather than an empty rule set.

STOX.AIBuilt

Supply chain map with marker clustering — a geographic view that stays legible when many sites sit close together — plus an inventory health dashboard and additional optimisation tiles.

2025.10 · October 2025

Supply chain and equipment surfaces take shape

The first month of the platform record. Work concentrated on the operational surfaces — inventory, distribution and equipment — with the tile-based navigation model the rest of the platform later adopted.

STOX.AIBuilt

Supply chain modules for multi-site distribution, with search wired across modules so a part number resolves from anywhere in the workspace.

What-if scenario planner. Model a change to demand, supply or lead time and see the inventory consequence before committing to it.

Equipment and distributionBuilt

Equipment optimisation — utilisation and availability for revenue equipment — and distribution centre components, with module naming brought into line across the platform so navigation is consistent rather than per-module.

Every entry traces to shipped code in the platform repositories. Where a figure is cited — a connector count, a measured timing, a verified measure — it comes from the component that produces it. Nothing here is a projection.

SAP, SAP S/4HANA and other SAP product names referenced here are trademarks or registered trademarks of SAP SE (or an SAP affiliate company) in Germany and other countries. Cloud Mantra is not affiliated with, endorsed by, or sponsored by SAP SE. SAP product names are used solely to identify the systems this platform reads from and writes to.