TRUST CENTER

Evidence, not seals.

We hold neither ISO 27001 nor SOC 2. What we do have is listed here — with the document behind it.

A seal is a third party's statement about us. As long as we hold none, the most honest answer is the list of what actually exists: registers, procedures, measurements, and where to read them.

Management system: ISO/IEC 27001 + 42001, built
Certification: none
Controller: TACP Operations GmbH & Co. KG
Registered office: Detmold, Germany

LIMITS FIRSTWe do not claim zero-knowledge (our operations team can reach content during an incident — our privacy notice says so), no confidential computing, and no critical-infrastructure suitability. Where a claim depends on a review that has not happened, it is not on this page.

WHAT EXISTS

Six registers we keep — and where they lead.

All are working documents, not brochures: every entry is written in the same step as the change it describes.

ISMSManagement systemPolicies, risk method and register, statements of applicability for 93 controls (ISO/IEC 27001:2022) and 38 controls (ISO/IEC 42001:2023), suppliers, incident management, business continuity, training, internal audits, management review.

PRIVACYImpact assessmentA DPIA under Art. 35 GDPR with a risk register per processing activity, measures, and residual risk. Outcome: required, residual risk medium, not an Art. 36 case.

AIAI system registerEvery AI system with its intended purpose, Art. 5 and Annex III review, models, data classes, human oversight and limits. Every new feature answers whether it changes the classification.

SECURITYSecurity registerEvery deliberate security trade-off with date, scope, test and named residual risk — in the same commit as the implementation. Currently 56 entries.

SUPPLY CHAINSoftware bill of materialsA CycloneDX bill of materials per shipped image, generated from the registry and signed as an aggregate; the check against known vulnerabilities runs on every build.

CONTRACTSub-processorsWhoever processes on our behalf is named, with purpose and location, in a public document — and whoever does not is not listed there either.

INTERNAL AUDITS

Five rounds, and what they found.

An external penetration test is scheduled for Q2 2027. Until then we audit ourselves — with a method designed not to go easy on us.

METHOD01Several reviewers, then a counter-reviewEach round runs in parallel tracks; afterwards every finding is re-checked by someone who did not find it. In round 5 that lowered several severities, raised one, and corrected two diagnoses.
RESULT02Findings are counted, not talked awayRound 4: 67 findings, 10 of them high. Round 5: 62 findings, 4 of them high — all four closed. What stays open stays open with a reason, and the reason is in the report.
PATTERN03The most uncomfortable class firstRound 5 introduced a class of its own: promises without a mechanism — a deadline without a clock, a role without a person, an evidence link without the evidence. Twelve of 62 findings fell into it.

MEASUREMENT REPORTS

Numbers with a method, not without.

Every number we quote has a report with setup, metrics, raw data and a counter-check. Where we have not measured, we say so.

LOADA thousand boxes on the real protocolThe fleet load test ran a thousand simulated boxes against the platform using the real protocol. The database was not the limit — memory per open connection was, and it sits at three times the target figure.

MEDIA272 pages in one goA maximum test ingested a 272-page scan without a text layer in full and then answered eighteen questions from it — counting hits and refusals honestly.

AVAILABILITYMeasured, not promisedTwo independent probe locations measure every minute; a missing sample counts as downtime. The figures and their measurement basis are on the status page.

MODELSBake-offs instead of gut feelingWhich model runs is decided by a comparison with measured figures and a regression set that runs again on every switch. We do not train or fine-tune the large language model; we select it. Our small decision models we develop and train ourselves, on data we generate.

DETERMINISMReproducible means: the decisionOur inference hardware is not bit-stable at temperature=0. We measured where it flips and why; decisions have since been taken in a short call of their own. With whitepaper and raw data.DECISION MODELSExplicio pair4, built in-houseOur own model for the three decisions before every search, measured against large language models on the same test sets. With benchmark, limits and whitepaper.

QUESTIONS FROM A REVIEW

Reviewing us and need a document?

Tell us which one. What can be public is here; the rest you get under confidentiality.

Request a document

Disclosure: security.txt
Sub-processors: Document
Privacy: Notice
Operations: Status page