Titentiten.devInstall
Use case

Similarity is not provenance.

A vector store tells you which text resembles a query. It cannot tell you which source supports the statement, whether it is still true, who was allowed to see it, or what replaced it. Titen answers those four with records rather than with ranking.

Where it stops. Titen holds no certification: no SOC 2, no HIPAA, no ISO, no third-party audit. It ships mechanisms that make an auditable system buildable. It does not make your system compliant, and there is no provisioned enterprise deployment behind these claims.

Install TitenView source
0 certifications
mechanisms, not compliance
metadata-only
audit records actors and ids, never content
before retrieval
authorization filters, it does not post-filter
01

Trace a claim to its sources

Every claim cites the observations it was derived from and keeps its version history. Resolving one is a single authenticated call, and evidence your credential may not read comes back as a count rather than as content.

That is the distinction in the heading. A similarity score explains why a chunk ranked highly. It does not tell a reviewer where the statement came from, and it cannot be replayed six months later to justify a decision.

curl -H "authorization: Bearer $TITEN_API_KEY" \
  http://127.0.0.1:8787/v1/claims/claim_f3963d7b876143f5bfc8230db2757067/evidence

# → evidence.supporting, evidence.contradicting, evidence.qualifying,
#   plus hidden_source_count for evidence this credential may not read
See evidence-backed claims →
02

Conflict before overwrite

When two sources disagree, the cheap answer is to let the later write win. Titen keeps both, and the claim stays disputed until evidence or an authorized resolution exists. No model may silently declare consensus on its own.

The system separates five cases that a single latest-wins rule would flatten into one: conflicting facts from different sources, different observer opinions, decisions that intentionally replace earlier decisions, concurrent checkpoint updates, and duplicate work attempts.

  • Facts stay disputed until resolved; opinions stay observer-scoped.
  • Decisions replace earlier decisions through explicit supersession, and the superseded version is retained.
  • POST /v1/claims/:id/supersede, /revoke and /expire each take an expected version, so a concurrent edit fails loudly instead of quietly winning.
Read the claim lifecycle →
03

Retention, legal hold and approvals

Governance is reachable over the same REST surface as memory, under its own scopes: roles and typed policy, an approval workflow for high-trust memory, retention schedules, and legal hold. These are shipped requirements with contract coverage, not roadmap entries.

  • POST /v1/legal-holds places a hold on an exact claim or observation. A held record survives retention expiry, and purge is blocked while the hold is active.
  • POST /v1/claim-approvals/:id/decide records an approval decision under separation-of-duty policy, against one exact claim version.
  • GET /v1/audit pages metadata-only entries; GET /v1/audit/export streams them as NDJSON behind a titen.audit.header line.
  • Authorization runs before retrieval, so a record the caller may not read contributes no score, no conflict and no count.
Compare the REST surface →
04

Where this stops

This page describes mechanisms, and a mechanism is not a certification. Titen has no SOC 2, HIPAA or ISO attestation and no third-party security audit. The roadmap records governance and enterprise audit as implemented and verified locally, with no claim of a provisioned enterprise deployment.

The audit trail is metadata-only by design. It records which records and which actors were involved, not what they said, so the log cannot become a second uncontrolled copy of the content. That is a deliberate privacy choice, and it means the audit log alone will not satisfy a reviewer who wants full content reconstruction.

  • Titen is pre-1.0. Below 1.0.0 the minor slot is the only breaking-change signal, so pin an exact version and read the changelog before upgrading.
  • Claims are caller-authored: you write the statement and link it to a source observation. Model-assisted derivation ships but is activation-gated, and no candidate model has passed — the best result on record is 65.56% against a 90% threshold.
  • Governance passes the shared local Bun and D1 contract suite. That is local evidence, and it is not a production deployment.
Read the maturity matrix →