Changelog

Six English documents, no edition labels. The Resources page now offers the full October 2026 set as direct PDF downloads: Executive Brief, Technical Overview, Benchmark Report, Use Case Guide, Security & Deployment Datasheet and the Whitepaper. Earlier editions are withdrawn. The Turkish page offers the same English set.

Benchmark Report. Results on 3,396 real business documents held out from model training: 98.3% of labelled personal values fully protected, 97.81% of all values, 14.5 minutes of CPU batch time and 0.36 s median per document, with a nine-system comparison on the same workload and a two-page method annex covering corpus construction, labelling under human supervision with legal expertise, scoring, comparator protocol, training regime and verification.

Technical Overview. Five product surfaces on one Engine (Salus Engine, Salus Desktop, Salus Chat, a gateway relay and a meeting note-taker, the last three delivered per deployment), a page on how Salus sits alongside personal agents, a central AI platform and developer assistants, and the vision-language OCR reader.

Use Case Guide and Whitepaper. The guide adds a sales and customer-team workflow and the new surfaces in its integration routes. The whitepaper is published as a customer document: the control boundary across AI lanes, the market review, focused models, measured results, connection points and three ways to run a proof.

Benchmark online. The same results are now a web page at /benchmark/, built from the report’s source data: the headline block, the nine-system comparison, all category results and a method summary.

Publication update. The Resources page now offers direct English PDF downloads: the ten-page Technical Whitepaper, seven-page Benchmark Report and six-page Use Case Guide. The guide uses illustrative workflows and synthetic example values, not customer case studies or new performance results.

Compute and models. Core text detection runs on CPU with compact specialised models and validated rules. Determinism is scoped to the same input, pinned release, runtime, policy and settings. Scanned-page OCR and the optional Semantic Layer use GPU acceleration inside the customer’s environment.

Optional policies and reading. Multilingual vision-language OCR and finance or telecom semantic policies are available on request. Further policies and identifier formats are scoped and validated per deployment. Earlier references to seven models, nine shipped semantic categories and CPU-only processing for every stage are superseded by this scope.

Delivery and evidence. The public copy distinguishes core, configured, delivered and on-request capabilities. Release manifests and CycloneDX SBOMs are core artefacts; signing is configured. Review records and customer-controlled retention are described separately from value-free security events. No new product-performance measurement is claimed in this publication update.

What ships, by control code.

Historical copy: the deployment and capability wording below was clarified on 24 October 2026. See the newer Booklets and deployment scope entry on this page and the current detection and control catalogues.

Detection. 31 classes and a nine-category semantic layer (SC-01) found by format rules with checksums, seven in-house language models and a verifier that re-checks every candidate (SC-04). Validators for TCKN, VKN, IBAN and card numbers (SC-02); telecom identifiers IMSI, ICCID, IMEI and MSISDN (SC-03); tenant-defined classes activated at runtime (SC-05); PDF, DOCX, XLSX, PPTX, PNG, JPG and scanned pages protected in place; signatures, identity documents and faces blocked on the page image, never sent; unreadable input refused (SC-06).

Protection and restoration. Opaque per-case tokens, one per entity (SC-07); optional format-preserving mode with the key held by the customer (SC-08); restoration inside the perimeter, scoped to the case (SC-09); fail-closed on any protection failure (SC-10); streaming protect and restore for gateway-fronted traffic (SC-11); endpoint deployment with human review before send (SC-12).

Operations and evidence. Tenant-scoped API keys and organisation isolation (SC-13), rate limiting (SC-14), administrator console with local accounts (SC-15), reviewer feedback loop (SC-16), ECS security events with stable codes and no values reprinted (SC-17), PostgreSQL audit store (SC-18), OpenMetrics (SC-19), restore evidence recording who restored which token, when and in which case (SC-20), and a decision record for every finding with class, confidence, verifier outcome, action and reason code (SC-21).

Delivery. Private OCI Helm chart, six images, OpenShift 4.x, mirrored-registry install (SC-22); signed release manifest and CycloneDX SBOM (SC-23); no Salus-side data, keys or telemetry (SC-24); CPU-only inference (SC-25).

Extended identifiers. Foreign identity, tax and social-security numbers, driving licence numbers, MAC addresses, credentials, domestic banking formats, healthcare identifiers and crypto-asset addresses, each with its own validator and reference tests (SC-01). Country-aware IBAN extraction against the full registry, one casing rule for every labelled identifier, phone paths reconciled with the numbering plan (SC-02).

Semantic layer. Two purpose-trained models screen a passage for a sensitive category and locate the stated disclosure; the spans are merged with the core result and rendered as one protected text. Additive only: it never removes, re-types or re-bounds what the core found (SC-01, SC-04).