Trust

Security and data handling

The most scrutinised part of this business is what leaves your warehouse. The honest answer is: by default, nothing but hashes and booleans. This page states the boundary field by field, and lists what we have not built yet rather than implying it.

Last updated 13 September 2026 · Protocol v9.3

Where we actually are

We are a small, early company. We hold no SOC 2 report and no ISO certification today. Everything below is either implemented or explicitly marked as not yet done. We would rather lose a deal on an honest answer than win one on an implied certification — and we have written that preference into the Independence Charter.

The boundary, in one table

Every engagement runs in a named evidence mode. In the default mode the runner executes inside your VPC or Snowflake account and no human at EvalQA sees a query, a row or a prompt.

ModeWho sees customer dataWhat leaves the perimeter
Mode 1 local-only DefaultNobody at EvalQAHashes, invariant pass/fail, row counts, latency, signed manifest
Mode 2 sanitised escalationNamed EvalQA triage staff, for 14 daysPseudonymised AST snippet and execution plan, encrypted under your tenant key
Mode 3 restricted VDI reviewA qualified reviewer, inside your VDINothing

Data-classification matrix for every evidence field

All customer-derived evidence — invariant failure names, error strings, prompt metadata — is CONFIDENTIAL by default. The matrix below is the one the platform enforces.

Evidence fieldClassificationRedaction protocolRetention
Natural-language promptCONFIDENTIALNamed-entity removalEncrypted; customer opt-out
Generated SQL queryCONFIDENTIALLiteral values masked30 days, or never leaves (Mode 1)
Query result-set dataRESTRICTEDZero transmissionLocal customer boundary only
Invariant pass/failCONFIDENTIALTenant pseudonymisedImmutable object-lock store
Execution latencyCONFIDENTIALAggregate summaryTelemetry database, 90 days
Error stack tracesCONFIDENTIALPath and table pseudonymisation14-day diagnostic TTL

Column redaction is done by stable semantic pseudonymisation with role tagging rather than by masking a column into column_19: the latter destroys the semantic utility that evaluation depends on (audit finding #31).

Key custody, deletion and immutability

There is a real collision between an immutable evidence store and GDPR Article 17 / DPDP deletion rights. We resolve it by separation rather than by hoping nobody asks:

Who can see your data, when a human is involved at all

ControlStatusDetail
Signed confidentiality agreementIn placeEvery reviewer signs before any access, covering all customer content indefinitely.
Identity verificationIn placeGovernment ID verification for every reviewer before their first paid task.
Per-engagement, time-boxed accessIn placeA reviewer sees the items they are assigned, for as long as that adjudication is open. Access is not standing.
Access loggingIn placeEvery item view and rating is attributed and timestamped. You can request the log for your data at any time.
Named roster on requestIn placeHow many reviewers, their qualification tier and their processing location — not merely residency, since residency does not determine where processing happens.
Background screeningRegulated work onlyRun for Tier-1 financial, health or legal schemas. Not by default.
Locked-down VDI for reviewAvailable (Mode 3)Customer-hosted VDI with clipboard, downloads and screen capture disabled. Changes the delivery model and the price; say so early.
Automated PII/PHI redactionIn placeRegex and NER tokenisation (Presidio / GLiNER) before any Mode 2 export.

Platform controls

In place

  • TLS for all traffic; HSTS on the production origin
  • Strict Content-Security-Policy on every page: no inline script, no inline styles, named origins only
  • Authentication on every endpoint that returns stored evaluation data; the notary returns hashes and dates only
  • Allow-listed CORS on data endpoints
  • CSRF protection on state-changing requests; rate limiting on public write and read endpoints
  • Secrets outside the source tree; private data store outside the document root
  • Runner builds signed with Sigstore/Cosign; the runner verifies itself before executing
  • Automated regression tests that fail the build if a private endpoint becomes public or an inline style appears

Not yet

  • SOC 2 Type 2 — scoped; the plan is to start it at customer #4 with a full-time head of security
  • ISO 27001 / ISO 42001
  • Third-party penetration test
  • EAL-3 hardware-attested enclaves (AWS Nitro) — designed, not yet offered
  • SAML single sign-on (SSO through Werify is supported today)
  • Data residency choice for Mode 2 diagnostics
  • E&O insurance at the $5M limit enterprise MSAs ask for — being placed

If one of the items on the right is a hard requirement, tell us at the first conversation. Some are weeks of work and we will say so; some are not on our roadmap at our size and we will say that too.

International transfers

Standard contractual clauses are not a magic solution. Where Mode 2 diagnostics or Mode 3 review involve a transfer out of the UK or EEA we perform a Transfer Impact Assessment, apply sub-processor controls and document the local-law analysis. Reviewer processing location and access authorisation are modelled explicitly, since a reviewer travelling invalidates a residency-based assumption.

Security review pack

We provide a pre-completed CAIQ v4 (CSA STAR Level 1) and standard answers for SIG Lite and bespoke questionnaires. The fastest path through review is almost always Mode 1: nothing leaves, so most of the questionnaire is answered by architecture. Download the DPA and the portability guarantee alongside it.

Reporting a vulnerability

Email [email protected]; also published at /.well-known/security.txt. We acknowledge within 2 business days and give an assessment within 10. Report in good faith — no scanning that degrades service, no accessing data that is not yours, reasonable time before publishing — and we will not pursue legal action. We credit you publicly if you wish. There is no paid bounty programme yet.

If something goes wrong

We notify affected customers within 48 hours of confirming an incident that affects their data, or within 24 hours where your contract or regulator requires it. The notice says what happened, what data was involved, what we have done and what we are asking you to do. Flaws in our method, as opposed to our security, run under the recall protocol.


Need a signed DPA or the CAIQ spreadsheet? Email [email protected].

Evidence modes & the runner adversary model → Data Processing Addendum