TALLWRIGHT - TECHNICAL AND ORGANISATIONAL MEASURES Version toms-1.3. In force from 2026-11-12. Replaces: fcc23b5d489352b41828ccb27ce6f2ba0a3630300fc9baef9b0c788fec29f21e The SHA-256 of this file, with CRLF normalised to LF, is listed in https://app.tallwright.com/legal/SHA256SUMS.txt and on https://app.tallwright.com/legal. What changed from 1.2: - the operator is David Joebstl, an individual entrepreneur in Austria. - the annex is restructured and shortened. It states each measure and each material limit, and leaves out host names, configuration details and dated readings; the evidence behind each measure is given on request (12.3). - measures added since 1.2: a key ring for the key that seals your credentials; the backoffice's own transport and content policies; rate limits; key-only SSH on every host; push alarms with an all-clear and a weekly test; checks of each host from the others; a dependency-update bot; a nightly encrypted copy of the backoffice database, kept thirty-five days; host access logs kept thirty days, without the one-use Exhibit link. - corrected: an unreachable Witness leaves no signed record of the attempt (7.3), and the backoffice reads your Archives every fifteen minutes, not five (6.4). Annex B to dpa-1.3, the Data Processing Agreement. English is authoritative, as clause 15.3 of the Agreement provides; capitalised terms mean what terms-1.3 and dpa-1.3 say they mean. Operator: David Joebstl, entrepreneur (Einzelunternehmer), Grazer Strasse 1/3, 8120 Peggau, Austria. Security reports: hello@tallwright.com 0. Scope 0.1 This annex states the technical and organisational measures we take for the Service under Article 32 GDPR. A measure not stated here is not claimed. Where a measure is missing or limited in a way that matters to you, the clause says so in the words "Not in place". 0.2 Where. Our servers are rented from Hetzner Online GmbH: the Witness host and the fleet host, which records hosted Archives, are separate machines in Falkenstein, Germany, and the Managed receiver is in Helsinki, Finland. The backoffice runs on Vercel in Frankfurt, with its database at Neon in Amazon Web Services' Frankfurt region. A separate staging host and staging backoffice are used for tests; they hold no customer Archive. Annex C names every provider. 0.3 Per Edition. On Cloud Proof everything here is ours. On Managed the receiver and the Witness host are ours and our reporter runs read-only on your machine; the recording machine and its Archives are yours. On Self-Hosted nothing of ours runs on your machine; the Witness, if you use it, and the backoffice, if you hold an account, are ours. A clause about a host or a component that is not ours in your Edition does not apply to you. 1. Pseudonymisation and data minimisation 1.1 Every Archive the product provisions holds hashes only, and no command changes that: the content of a Record, and the salt that would let anyone test a guess at it, stay with you, and a line carrying a payload or a subject identifier is refused by name. Two older Archives on the fleet host, made before Cloud Proof, hold content - a test of ours and a closed Archive with one test record of a business that piloted the Service - and yours is separated from them as clause 3.3 describes. 1.2 The backoffice stores no IP address and no location in clear. To limit sign-in link requests it keeps, for one day, the SHA-256 of the requested address and a keyed hash of the caller's network address. It keeps a browser's user-agent string, capped in length, on a session and on an acceptance record. It runs no analytics and no tracking. 1.3 Not in place: detection or removal of personal data in what you record. What a Record or a field sent in the clear contains is your decision as controller. Our recorder replaces known credential keys with a redaction marker and, in its model adapter, keeps only an allowlist of message fields; where you declared identifiers opaque, it refuses one that looks like an e-mail address, a telephone number or an IBAN. It checks nothing else. 2. Encryption 2.1 In transit. Every public endpoint of the fleet and receiver hosts and of the backoffice uses TLS and sends Strict Transport Security, and the backoffice loads scripts, styles and fonts only from itself. The Witness host accepts TLS 1.3 only, and the fleet host pins the Witness's public key, so no certificate authority can authorise a substitute Witness. Not in place: a content policy that forbids inline scripts and styles, which the backoffice still allows; and we have not confirmed that its connection to the database verifies the server's certificate. 2.2 At rest. The credentials we hold for your fleet are sealed with AES-256-GCM under a key held in the deployment environment, never in the database; that key is held as a ring, so a new key can seal while the old one still opens. Account keys, export keys and the credentials of the fleet's services are stored only as SHA-256 hashes. Backups are encrypted on the host before they leave it (7.1). Where an Archive holds content, payloads are encrypted with AES-256-GCM under a key per data subject. 2.3 Not in place: a hardware security module or key management service, and disk encryption on our hosts, so a key file sits on the same volume as what it protects; the Witness's signing key never leaves its host. Encryption at rest of the backoffice database is its provider's measure, which we have not verified. The key sealing your credentials has not yet been rotated. 3. Access control 3.1 Sign-in. The backoffice has no passwords: a person signs in with a one-use link sent by e-mail, valid ten minutes, or with a passkey. A session lasts eight hours. Every signed token states its purpose, and it is checked for that purpose before it is believed. Adding a passkey or creating an account key needs a sign-in no older than fifteen minutes. A new address can start an account only after a person of ours has approved it. 3.2 Roles. An account has Owners, Admins, Developers and Auditors. Only an Owner appoints an Owner, nobody changes their own role, and an Auditor is refused every change to Archives and Environments, acceptance of the agreement, provisioning and applications. The role is read from the database on every page, and removing a member revokes their sessions at once. There is no staff login to customer accounts and no impersonation. Not in place: an Auditor seat that is offered for one Archive sees every Archive of the account, and every seat but an Auditor can provision the account's fleet and submit its signed application. 3.3 Separation of customers. On the fleet host each customer's Archives belong to a system user of their own, in a directory only that user can open, and every service that serves them runs as that user in a restricted sandbox; once a host serves two customers, access across them is tested by hand in both directions on the running host, and again at each review. In the backoffice database, row-level security is enabled and forced on every table that holds an account's data, and limits a query to that account. Not in place: a second layer behind that row-level security, which the application's own database access can bypass, so a flaw in the application could reach other accounts' rows; and the nightly backup reads every customer's directory into one encrypted repository. 3.4 Hosted Archives are reached through separate services with separate credentials: one records and cannot list, one reads and cannot write, one creates and closes Archives and reads no records, and a one-use link, valid fifteen minutes, or an export key hands over one Exhibit at a time. The read service never returns a subject identifier. On Managed our reporter on your machine only reads, cannot open an Archive and writes nothing. 3.5 Administrative access. One person, the operator named above, holds the keys to our hosts and the accounts at our providers. Our hosts admit SSH by key only and refuse passwords, and, as the operator confirmed, a second factor is on at each provider that offers one; nothing of ours checks that it stays on. Each deploy to the fleet and receiver hosts records who deployed what, and each administrative command on a host is written to that host's admin journal with who, what and why. Not in place: approval of changes by a second person, periodic reviews of access, a scheduled rotation of keys, and signed admin journals; each journal is kept on the host it records. 4. Integrity 4.1 Records are hash-chained, so any later change to a recorded entry is detectable on verification. An Archive remembers the greatest height it has committed, so a shortened log is caught, and a self-check over the Archive runs before anchoring, export, erasure and disclosure, each of which refuses on a finding. 4.2 The head of each Archive is presented to the Witness about every fifteen minutes; that is not a service level. The Witness signs a refusal of a log shorter than one it has countersigned, or of a different root at a size it has countersigned, and it refuses a log that declares our own organisation as its operator. The Witness runs on a machine separate from the one that records; on Cloud Proof, and on Managed where we operate the recording host, we operate both, so an Exhibit from such a log is never evidence of two separate organisations (terms-1.3 clause 5.1). 4.3 The application writes the account audit log and the sign-in log through a database role that can add rows and cannot change them. Not in place: access with the database owner's rights - the operator's, and the application connection's as clause 3.3 says - can change or delete these rows; they are not countersigned by the Witness; and the kernel's append-only file flags are not set on our hosted fleet. 5. Time 5.1 The Witness host's clock follows authenticated network time (NTS) from four sources run by three independent organisations; an unauthenticated source cannot be selected, and a check every fifteen minutes raises an alarm on a stopped or unsynchronised clock or an unauthenticated source. The time in a cosignature is the Witness's own clock. The time-stamp token a recording host adds to an anchor comes from a local test authority on that host, and no measure here covers that host's clock. 5.2 David Joebstl, who operates the Service, provides the Witness as a trust service that is not qualified (Article 3(16) eIDAS). Its time stamps are not qualified electronic time stamps, and no presumption of eIDAS attaches to them (terms-1.3 clause 6). 6. Host security and monitoring 6.1 Hardening. Most of our services run in a systemd sandbox with no capabilities, no new privileges, a read-only system, private temporary files and a system-call filter; the provisioning service and the backup units hold the privileges their task needs, and some auxiliary units set fewer of these restrictions. On the fleet and receiver hosts application ports listen only locally behind a TLS proxy; the Witness listens on its own port. A firewall on each host admits only the ports its service needs, and each host's operating system installs security updates daily; an update that needs a restart waits for a planned one. The Witness host is set up by hand from a written runbook. 6.2 Rate limits. Sign-in link requests are limited per address, per caller and in total, and new early access requests, new books and invitations per day; each fleet service limits the requests per Environment, the receiver those per box, and the Witness the connections one network address may hold. Not in place: protection against guessing beyond those limits, a limit on the backoffice's other routes, and banning an address after failed sign-ins. 6.3 Logging. The backoffice keeps an account audit log of what was done to an account, Archive, seat, passkey or payment, and by whom, and a sign-in log of every attempt, without a network address. A search or query inside an Archive writes a hash-chained access entry before it answers. The proxy in front of the fleet and receiver hosts logs each request with its network address, and a one-use Exhibit link only up to its first segment; those logs are kept thirty days. Each host keeps an admin journal (3.5). Not logged: the Witness writes no request log, a direct read of Records or an export writes no access entry, and access entries are not signed. 6.4 Monitoring. Every service raises an alarm when it fails and an all-clear when it works again, as a push message to the operator's phone; the fleet and receiver hosts also send a test message every week. The hosts check the backoffice, the Witness and each other every ten minutes, and the backoffice checks every account's Archives every fifteen minutes and mails on a change. Not in place: an on-call rotation, since an alarm reaches one person, and monitoring from outside our own estate, so if every host and the platform stop at once nothing reports it. On Managed we detect a silent box, not a false report: the report is neither signed nor bound to the device. 7. Availability and resilience 7.1 Backups. The fleet host backs up every Archive on it and its own configuration, and every night a copy of the backoffice database; the receiver backs up only its roster of boxes; the Witness host backs up nothing. All of it is encrypted on the host before it goes to a storage service of Hetzner in the EU. A daily check raises an alarm when the newest backup is more than 26 hours old, a weekly check reads part of it back, and a drill on 2026-10-26 compared the backups with the live systems. Copies of the backoffice database are kept thirty-five days, without the one-day counters of sign-in link requests. 7.2 Restore. We restore only on your instruction or to recover from a failure, and we re-apply by hand any erasure you had already requested. State that may only move forward - a latched clock, the record of what has already been countersigned, a row recording an act already taken - is never restored as it stood, because a restore would reopen what was closed. Where the law requires it, or a court or an authority orders it, we may make an identified record inaccessible; we tell you what we did and why, and every proof over it stays intact. Not in place: a written procedure for re-applying erasures after a restore. 7.3 Failing safely. If the Witness cannot be reached, no anchor is written, every Record is kept and the head is presented again at the next run, and the attempt leaves no signed record; nothing produces an attestation that only looks valid. No bound holds on how long a Record stays unanchored. 7.4 Not in place: an availability figure, a recovery time or point objective, or a time to restore; a rehearsed restore onto an empty database or a new machine; backups that the host writing them cannot delete; a maximum age for the backups of Archives; a backup of the Witness host, whose signing key is never copied, so losing that machine means a new Witness key, while every Exhibit already made still verifies; and point-in-time recovery of the backoffice database beyond six hours, so a fault found later means restoring the last nightly copy taken before it, and losing every change since. 8. Physical security 8.1 Physical and environmental security is relied on from our infrastructure providers (Annex C). We hold no data-centre controls of our own, have inspected no facility and have not reviewed those providers' certificates or reports. 8.2 The Witness host and the fleet host are separate machines in one data centre, a separation of machines and not of sites; the receiver is in another country. 9. People 9.1 One natural person processes personal data for you: the operator, who does every role and is bound by the duty of data secrecy under Section 6 DSG (dpa-1.3 clause 5). There are no background checks and no security training. Any further person signs a confidentiality undertaking before being given access. 10. Development and supply chain 10.1 A change reaches the main branch only after the full continuous integration has passed: the tests, a language gate, a supply-chain gate, the gates that hold this annex and the data flows to the code, and a check that the contract files served are the bytes whose digests the application and the Witness's policy template pin. Third-party packages install from lock files with hashes. A dependency-update bot proposes updates to the web workspace and the build workflows monthly, and known advisories are checked whenever a lock is regenerated and before a release of the public verifier. 10.2 Not in place: continuous vulnerability scanning of the Python dependencies, static analysis, and secret scanning. 11. Incidents 11.1 An alarm reaches the operator as clause 6.4 describes. A personal data breach is notified to you as clause 12 of dpa-1.3 says. A breach or disruption with a significant impact on the Witness is notified to the supervisory body and to you within 24 hours (Article 19a eIDAS; dpa-1.3 clause 12.5; terms-1.3 clause 6.4). The procedure for both is written in our operations runbook. 12. Review and evidence 12.1 We review this annex at least yearly and whenever the processing or the risk changes materially. Our security reviews are internal; the latest is of 2026-10-18. 12.2 Not in place: a third-party security audit, a penetration test, and a certification against any standard. 12.3 On request (dpa-1.3 clause 13.1), under confidentiality, we give you a completed security questionnaire, the record of the latest restore drill and, for each measure here, the file that implements it and, where one exists, the test that holds it. 13. Changes to this annex 13.1 The Data Processing Agreement governs which version of this annex is in force, and nothing published here varies it. A new version is a new file at its own address, naming the SHA-256 of the version it replaces. We notify the Owners and Admins of every change at least thirty days in advance, clause by clause and in which direction each changed, and you may object within fourteen days on data protection grounds. A change that lowers a measure also follows clause 6.3 of dpa-1.3 and lets you terminate the affected Service without penalty, notice period or reason, with a pro-rata refund of prepaid fees, by writing to hello@tallwright.com, effective on receipt. Corrections are published with their reason, in either direction: understating a measure is as wrong as overstating one.