Trust · Last reviewed 21 August 2026
How The Timeline protects team data across AI inference, storage, permissions, meetings, analytics, and human access.
Trust / Security / Data privacy
Timeline holds the messages, meetings, files, and decisions teams rarely put in one place. Trust has to come from inspectable boundaries, candid provider choices, and claims the product can actually enforce.
The honest answer is not “nothing leaves our servers.” Relevant content reaches specialist processors. The policy is narrower: Timeline prohibits using Customer Content to train or fine-tune models. Production account and contract evidence is still being completed. Core AI roles require zero-data-retention routes. Non-meeting voice transcription is currently a retained, no-training quality exception, published below instead of hidden behind a blanket claim.
01Data path
People send work to Timeline or an administrator selects a provider resource. Connected does not mean imported by default.
PostgreSQL records, Qdrant vectors, and RustFS objects keep team identity attached. Visibility narrows access again inside the team.
Only the content needed for a feature leaves Timeline. ZDR-required roles stop rather than request a weaker route. Voice-note transcription remains a clearly disclosed retained, no-training quality exception while alternatives are tested.
Derived summaries can change; captured evidence remains attached so a person can check the original source and visibility.
02AI and training
OpenRouter and an upstream endpoint transiently process the necessary input. Text, multimodal, and embedding roles require eligible ZDR endpoints and fail closed instead of using weaker retention. Supported calls send data_collection: deny, zdr: true, and an explicit cache-disable control. OpenRouter may select among eligible ZDR upstreams. The production key must match the code-owned model catalog, with prompt logging, input/output sharing, Broadcast, and persistent response caching disabled.
Voice-note transcription currently uses openai/gpt-4o-transcribe while multilingual ZDR candidates face a strict quality gate. OpenRouter's provider table lists OpenAI as not training on prompts but retaining them; OpenAI API data controls document default API abuse-monitoring retention of inputs and outputs for up to 30 days. This does not weaken ZDR requirements for any other role.
Generation, extraction, summaries, classification, and agent answers
deepseek/deepseek-v4.1-flash-1017 + deepseek/deepseek-v4-pro-0813ZDR required
Quality-first long-context text models; weaker-retention fallback is prohibited.
Images, files, audio, and video text extraction
google/gemini-3.8-flashZDR required
Multimodal extraction with a required ZDR route.
Semantic search vectors
openai/text-embedding-3-smallZDR required
The established retrieval model; changing it would require a measured re-index.
Non-meeting voice-note speech-to-text
openai/gpt-4o-transcribeRetained, no-training quality exception
Quality baseline while multilingual ZDR candidates are evaluated against a non-inferiority gate.
Timeline policy prohibits training or fine-tuning on Customer Content
ZDR-required roles reject endpoints with weaker retention terms; production provider evidence remains a release gate
The voice-transcription retention exception is explicit and quality-gated
03Storage and access
The documented hosted topology places the web app, workers, PostgreSQL, Redis, Qdrant vector search, and RustFS object storage on Railway. The deployment design requires private networking for server traffic and non-public RustFS buckets, with HTTPS exposed only for team-authorized, short-lived signed browser transfers. Current production networking and bucket exposure still require verification. OAuth, API, and MCP secrets use AES-256-GCM authenticated encryption at rest.
Timeline's operating standard prohibits routine browsing of customer workspaces. It permits access only for user-requested support, reliability, security response, legal compliance, or another documented operational need, and requires minimum-necessary, time-bounded access where practical. Production access inventory, enforcement, and audit evidence are still being completed.
Inside a team, records are team-visible, private, or restricted to named users. Administrator status is not a general bypass into a member's private context.
04Processors and retention
Core processors operate the hosted service. Services a customer deliberately connects—such as Slack, Google Drive, GitHub, or a custom MCP server—receive only the data and permissions needed for that connection.
Documented app, worker, PostgreSQL, Redis, private-traffic, and signed-file design; deployed exposure evidence pending
Privacy-filtered routing to eligible AI inference endpoints
Meeting attendance and transcript generation; one-hour media retention requested, with deployed account evidence pending
Ephemeral isolated compute for document extraction
Transactional and inbound email
Personless surface totals, pseudonymous product events, and consent-gated public-browser analytics; provider evidence remains pending
Minimized error diagnostics with targeted credential and URL redaction
Abuse prevention on public forms and password signup
Full categories, lawful bases, disclosures, transfer safeguards, and rights are in the Privacy Policy. Current browser storage and public analytics choices are in the Cookies and similar technologies notice. Contractual service commitments and use rules are in the Terms of Use.
05Assurance status
Current
Public source, team and visibility boundaries, encrypted secrets, meeting consent controls, and a maintained provider register. Browser analytics is confined to reviewed public pages and waits for affirmative consent; protected workspace routes do not load it. Personless surface streams and pseudonymous product events use separate server paths. Provider/account evidence and AI account evidence remain deployment checklist items until captured.
Not claimed today
We do not display badges we have not earned. GDPR is a legal framework, not a certificate. Verified reports and certifications will appear here if and when they are obtained.
Report suspected vulnerabilities privately to [email protected]. Do not put credentials, customer data, or an undisclosed exploit in a public issue.
Trust but verify
The repository is public so teams can inspect architecture, open issues, and propose security patches. If the hosted boundary is not enough, contact us about a dedicated or self-managed deployment and the terms that apply.